xref: /openbmc/linux/drivers/infiniband/hw/mlx5/qp.c (revision 15e3ae36)
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 	if (bfregi->lib_uar_dyn)
701 		return -EINVAL;
702 
703 	mutex_lock(&bfregi->lock);
704 	if (bfregi->ver >= 2) {
705 		bfregn = alloc_high_class_bfreg(dev, bfregi);
706 		if (bfregn < 0)
707 			bfregn = alloc_med_class_bfreg(dev, bfregi);
708 	}
709 
710 	if (bfregn < 0) {
711 		BUILD_BUG_ON(NUM_NON_BLUE_FLAME_BFREGS != 1);
712 		bfregn = 0;
713 		bfregi->count[bfregn]++;
714 	}
715 	mutex_unlock(&bfregi->lock);
716 
717 	return bfregn;
718 }
719 
720 void mlx5_ib_free_bfreg(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi, int bfregn)
721 {
722 	mutex_lock(&bfregi->lock);
723 	bfregi->count[bfregn]--;
724 	mutex_unlock(&bfregi->lock);
725 }
726 
727 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
728 {
729 	switch (state) {
730 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
731 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
732 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
733 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
734 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
735 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
736 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
737 	default:		return -1;
738 	}
739 }
740 
741 static int to_mlx5_st(enum ib_qp_type type)
742 {
743 	switch (type) {
744 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
745 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
746 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
747 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
748 	case IB_QPT_XRC_INI:
749 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
750 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
751 	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
752 	case MLX5_IB_QPT_DCI:		return MLX5_QP_ST_DCI;
753 	case IB_QPT_RAW_IPV6:		return MLX5_QP_ST_RAW_IPV6;
754 	case IB_QPT_RAW_PACKET:
755 	case IB_QPT_RAW_ETHERTYPE:	return MLX5_QP_ST_RAW_ETHERTYPE;
756 	case IB_QPT_MAX:
757 	default:		return -EINVAL;
758 	}
759 }
760 
761 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
762 			     struct mlx5_ib_cq *recv_cq);
763 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
764 			       struct mlx5_ib_cq *recv_cq);
765 
766 int bfregn_to_uar_index(struct mlx5_ib_dev *dev,
767 			struct mlx5_bfreg_info *bfregi, u32 bfregn,
768 			bool dyn_bfreg)
769 {
770 	unsigned int bfregs_per_sys_page;
771 	u32 index_of_sys_page;
772 	u32 offset;
773 
774 	if (bfregi->lib_uar_dyn)
775 		return -EINVAL;
776 
777 	bfregs_per_sys_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k) *
778 				MLX5_NON_FP_BFREGS_PER_UAR;
779 	index_of_sys_page = bfregn / bfregs_per_sys_page;
780 
781 	if (dyn_bfreg) {
782 		index_of_sys_page += bfregi->num_static_sys_pages;
783 
784 		if (index_of_sys_page >= bfregi->num_sys_pages)
785 			return -EINVAL;
786 
787 		if (bfregn > bfregi->num_dyn_bfregs ||
788 		    bfregi->sys_pages[index_of_sys_page] == MLX5_IB_INVALID_UAR_INDEX) {
789 			mlx5_ib_dbg(dev, "Invalid dynamic uar index\n");
790 			return -EINVAL;
791 		}
792 	}
793 
794 	offset = bfregn % bfregs_per_sys_page / MLX5_NON_FP_BFREGS_PER_UAR;
795 	return bfregi->sys_pages[index_of_sys_page] + offset;
796 }
797 
798 static int mlx5_ib_umem_get(struct mlx5_ib_dev *dev, struct ib_udata *udata,
799 			    unsigned long addr, size_t size,
800 			    struct ib_umem **umem, int *npages, int *page_shift,
801 			    int *ncont, u32 *offset)
802 {
803 	int err;
804 
805 	*umem = ib_umem_get(&dev->ib_dev, addr, size, 0);
806 	if (IS_ERR(*umem)) {
807 		mlx5_ib_dbg(dev, "umem_get failed\n");
808 		return PTR_ERR(*umem);
809 	}
810 
811 	mlx5_ib_cont_pages(*umem, addr, 0, npages, page_shift, ncont, NULL);
812 
813 	err = mlx5_ib_get_buf_offset(addr, *page_shift, offset);
814 	if (err) {
815 		mlx5_ib_warn(dev, "bad offset\n");
816 		goto err_umem;
817 	}
818 
819 	mlx5_ib_dbg(dev, "addr 0x%lx, size %zu, npages %d, page_shift %d, ncont %d, offset %d\n",
820 		    addr, size, *npages, *page_shift, *ncont, *offset);
821 
822 	return 0;
823 
824 err_umem:
825 	ib_umem_release(*umem);
826 	*umem = NULL;
827 
828 	return err;
829 }
830 
831 static void destroy_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
832 			    struct mlx5_ib_rwq *rwq, struct ib_udata *udata)
833 {
834 	struct mlx5_ib_ucontext *context =
835 		rdma_udata_to_drv_context(
836 			udata,
837 			struct mlx5_ib_ucontext,
838 			ibucontext);
839 
840 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_DELAY_DROP)
841 		atomic_dec(&dev->delay_drop.rqs_cnt);
842 
843 	mlx5_ib_db_unmap_user(context, &rwq->db);
844 	ib_umem_release(rwq->umem);
845 }
846 
847 static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
848 			  struct ib_udata *udata, struct mlx5_ib_rwq *rwq,
849 			  struct mlx5_ib_create_wq *ucmd)
850 {
851 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
852 		udata, struct mlx5_ib_ucontext, ibucontext);
853 	int page_shift = 0;
854 	int npages;
855 	u32 offset = 0;
856 	int ncont = 0;
857 	int err;
858 
859 	if (!ucmd->buf_addr)
860 		return -EINVAL;
861 
862 	rwq->umem = ib_umem_get(&dev->ib_dev, ucmd->buf_addr, rwq->buf_size, 0);
863 	if (IS_ERR(rwq->umem)) {
864 		mlx5_ib_dbg(dev, "umem_get failed\n");
865 		err = PTR_ERR(rwq->umem);
866 		return err;
867 	}
868 
869 	mlx5_ib_cont_pages(rwq->umem, ucmd->buf_addr, 0, &npages, &page_shift,
870 			   &ncont, NULL);
871 	err = mlx5_ib_get_buf_offset(ucmd->buf_addr, page_shift,
872 				     &rwq->rq_page_offset);
873 	if (err) {
874 		mlx5_ib_warn(dev, "bad offset\n");
875 		goto err_umem;
876 	}
877 
878 	rwq->rq_num_pas = ncont;
879 	rwq->page_shift = page_shift;
880 	rwq->log_page_size =  page_shift - MLX5_ADAPTER_PAGE_SHIFT;
881 	rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
882 
883 	mlx5_ib_dbg(dev, "addr 0x%llx, size %zd, npages %d, page_shift %d, ncont %d, offset %d\n",
884 		    (unsigned long long)ucmd->buf_addr, rwq->buf_size,
885 		    npages, page_shift, ncont, offset);
886 
887 	err = mlx5_ib_db_map_user(ucontext, udata, ucmd->db_addr, &rwq->db);
888 	if (err) {
889 		mlx5_ib_dbg(dev, "map failed\n");
890 		goto err_umem;
891 	}
892 
893 	rwq->create_type = MLX5_WQ_USER;
894 	return 0;
895 
896 err_umem:
897 	ib_umem_release(rwq->umem);
898 	return err;
899 }
900 
901 static int adjust_bfregn(struct mlx5_ib_dev *dev,
902 			 struct mlx5_bfreg_info *bfregi, int bfregn)
903 {
904 	return bfregn / MLX5_NON_FP_BFREGS_PER_UAR * MLX5_BFREGS_PER_UAR +
905 				bfregn % MLX5_NON_FP_BFREGS_PER_UAR;
906 }
907 
908 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
909 			  struct mlx5_ib_qp *qp, struct ib_udata *udata,
910 			  struct ib_qp_init_attr *attr,
911 			  u32 **in,
912 			  struct mlx5_ib_create_qp_resp *resp, int *inlen,
913 			  struct mlx5_ib_qp_base *base)
914 {
915 	struct mlx5_ib_ucontext *context;
916 	struct mlx5_ib_create_qp ucmd;
917 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
918 	int page_shift = 0;
919 	int uar_index = 0;
920 	int npages;
921 	u32 offset = 0;
922 	int bfregn;
923 	int ncont = 0;
924 	__be64 *pas;
925 	void *qpc;
926 	int err;
927 	u16 uid;
928 	u32 uar_flags;
929 
930 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
931 	if (err) {
932 		mlx5_ib_dbg(dev, "copy failed\n");
933 		return err;
934 	}
935 
936 	context = rdma_udata_to_drv_context(udata, struct mlx5_ib_ucontext,
937 					    ibucontext);
938 	uar_flags = ucmd.flags & (MLX5_QP_FLAG_UAR_PAGE_INDEX |
939 				  MLX5_QP_FLAG_BFREG_INDEX);
940 	switch (uar_flags) {
941 	case MLX5_QP_FLAG_UAR_PAGE_INDEX:
942 		uar_index = ucmd.bfreg_index;
943 		bfregn = MLX5_IB_INVALID_BFREG;
944 		break;
945 	case MLX5_QP_FLAG_BFREG_INDEX:
946 		uar_index = bfregn_to_uar_index(dev, &context->bfregi,
947 						ucmd.bfreg_index, true);
948 		if (uar_index < 0)
949 			return uar_index;
950 		bfregn = MLX5_IB_INVALID_BFREG;
951 		break;
952 	case 0:
953 		if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
954 			return -EINVAL;
955 		bfregn = alloc_bfreg(dev, &context->bfregi);
956 		if (bfregn < 0)
957 			return bfregn;
958 		break;
959 	default:
960 		return -EINVAL;
961 	}
962 
963 	mlx5_ib_dbg(dev, "bfregn 0x%x, uar_index 0x%x\n", bfregn, uar_index);
964 	if (bfregn != MLX5_IB_INVALID_BFREG)
965 		uar_index = bfregn_to_uar_index(dev, &context->bfregi, bfregn,
966 						false);
967 
968 	qp->rq.offset = 0;
969 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
970 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
971 
972 	err = set_user_buf_size(dev, qp, &ucmd, base, attr);
973 	if (err)
974 		goto err_bfreg;
975 
976 	if (ucmd.buf_addr && ubuffer->buf_size) {
977 		ubuffer->buf_addr = ucmd.buf_addr;
978 		err = mlx5_ib_umem_get(dev, udata, ubuffer->buf_addr,
979 				       ubuffer->buf_size, &ubuffer->umem,
980 				       &npages, &page_shift, &ncont, &offset);
981 		if (err)
982 			goto err_bfreg;
983 	} else {
984 		ubuffer->umem = NULL;
985 	}
986 
987 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
988 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * ncont;
989 	*in = kvzalloc(*inlen, GFP_KERNEL);
990 	if (!*in) {
991 		err = -ENOMEM;
992 		goto err_umem;
993 	}
994 
995 	uid = (attr->qp_type != IB_QPT_XRC_TGT &&
996 	       attr->qp_type != IB_QPT_XRC_INI) ? to_mpd(pd)->uid : 0;
997 	MLX5_SET(create_qp_in, *in, uid, uid);
998 	pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas);
999 	if (ubuffer->umem)
1000 		mlx5_ib_populate_pas(dev, ubuffer->umem, page_shift, pas, 0);
1001 
1002 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
1003 
1004 	MLX5_SET(qpc, qpc, log_page_size, page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1005 	MLX5_SET(qpc, qpc, page_offset, offset);
1006 
1007 	MLX5_SET(qpc, qpc, uar_page, uar_index);
1008 	if (bfregn != MLX5_IB_INVALID_BFREG)
1009 		resp->bfreg_index = adjust_bfregn(dev, &context->bfregi, bfregn);
1010 	else
1011 		resp->bfreg_index = MLX5_IB_INVALID_BFREG;
1012 	qp->bfregn = bfregn;
1013 
1014 	err = mlx5_ib_db_map_user(context, udata, ucmd.db_addr, &qp->db);
1015 	if (err) {
1016 		mlx5_ib_dbg(dev, "map failed\n");
1017 		goto err_free;
1018 	}
1019 
1020 	err = ib_copy_to_udata(udata, resp, min(udata->outlen, sizeof(*resp)));
1021 	if (err) {
1022 		mlx5_ib_dbg(dev, "copy failed\n");
1023 		goto err_unmap;
1024 	}
1025 	qp->create_type = MLX5_QP_USER;
1026 
1027 	return 0;
1028 
1029 err_unmap:
1030 	mlx5_ib_db_unmap_user(context, &qp->db);
1031 
1032 err_free:
1033 	kvfree(*in);
1034 
1035 err_umem:
1036 	ib_umem_release(ubuffer->umem);
1037 
1038 err_bfreg:
1039 	if (bfregn != MLX5_IB_INVALID_BFREG)
1040 		mlx5_ib_free_bfreg(dev, &context->bfregi, bfregn);
1041 	return err;
1042 }
1043 
1044 static void destroy_qp_user(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1045 			    struct mlx5_ib_qp *qp, struct mlx5_ib_qp_base *base,
1046 			    struct ib_udata *udata)
1047 {
1048 	struct mlx5_ib_ucontext *context =
1049 		rdma_udata_to_drv_context(
1050 			udata,
1051 			struct mlx5_ib_ucontext,
1052 			ibucontext);
1053 
1054 	mlx5_ib_db_unmap_user(context, &qp->db);
1055 	ib_umem_release(base->ubuffer.umem);
1056 
1057 	/*
1058 	 * Free only the BFREGs which are handled by the kernel.
1059 	 * BFREGs of UARs allocated dynamically are handled by user.
1060 	 */
1061 	if (qp->bfregn != MLX5_IB_INVALID_BFREG)
1062 		mlx5_ib_free_bfreg(dev, &context->bfregi, qp->bfregn);
1063 }
1064 
1065 /* get_sq_edge - Get the next nearby edge.
1066  *
1067  * An 'edge' is defined as the first following address after the end
1068  * of the fragment or the SQ. Accordingly, during the WQE construction
1069  * which repetitively increases the pointer to write the next data, it
1070  * simply should check if it gets to an edge.
1071  *
1072  * @sq - SQ buffer.
1073  * @idx - Stride index in the SQ buffer.
1074  *
1075  * Return:
1076  *	The new edge.
1077  */
1078 static void *get_sq_edge(struct mlx5_ib_wq *sq, u32 idx)
1079 {
1080 	void *fragment_end;
1081 
1082 	fragment_end = mlx5_frag_buf_get_wqe
1083 		(&sq->fbc,
1084 		 mlx5_frag_buf_get_idx_last_contig_stride(&sq->fbc, idx));
1085 
1086 	return fragment_end + MLX5_SEND_WQE_BB;
1087 }
1088 
1089 static int create_kernel_qp(struct mlx5_ib_dev *dev,
1090 			    struct ib_qp_init_attr *init_attr,
1091 			    struct mlx5_ib_qp *qp,
1092 			    u32 **in, int *inlen,
1093 			    struct mlx5_ib_qp_base *base)
1094 {
1095 	int uar_index;
1096 	void *qpc;
1097 	int err;
1098 
1099 	if (init_attr->create_flags & ~(IB_QP_CREATE_INTEGRITY_EN |
1100 					IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
1101 					IB_QP_CREATE_IPOIB_UD_LSO |
1102 					IB_QP_CREATE_NETIF_QP |
1103 					MLX5_IB_QP_CREATE_SQPN_QP1 |
1104 					MLX5_IB_QP_CREATE_WC_TEST))
1105 		return -EINVAL;
1106 
1107 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
1108 		qp->bf.bfreg = &dev->fp_bfreg;
1109 	else if (init_attr->create_flags & MLX5_IB_QP_CREATE_WC_TEST)
1110 		qp->bf.bfreg = &dev->wc_bfreg;
1111 	else
1112 		qp->bf.bfreg = &dev->bfreg;
1113 
1114 	/* We need to divide by two since each register is comprised of
1115 	 * two buffers of identical size, namely odd and even
1116 	 */
1117 	qp->bf.buf_size = (1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size)) / 2;
1118 	uar_index = qp->bf.bfreg->index;
1119 
1120 	err = calc_sq_size(dev, init_attr, qp);
1121 	if (err < 0) {
1122 		mlx5_ib_dbg(dev, "err %d\n", err);
1123 		return err;
1124 	}
1125 
1126 	qp->rq.offset = 0;
1127 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
1128 	base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
1129 
1130 	err = mlx5_frag_buf_alloc_node(dev->mdev, base->ubuffer.buf_size,
1131 				       &qp->buf, dev->mdev->priv.numa_node);
1132 	if (err) {
1133 		mlx5_ib_dbg(dev, "err %d\n", err);
1134 		return err;
1135 	}
1136 
1137 	if (qp->rq.wqe_cnt)
1138 		mlx5_init_fbc(qp->buf.frags, qp->rq.wqe_shift,
1139 			      ilog2(qp->rq.wqe_cnt), &qp->rq.fbc);
1140 
1141 	if (qp->sq.wqe_cnt) {
1142 		int sq_strides_offset = (qp->sq.offset  & (PAGE_SIZE - 1)) /
1143 					MLX5_SEND_WQE_BB;
1144 		mlx5_init_fbc_offset(qp->buf.frags +
1145 				     (qp->sq.offset / PAGE_SIZE),
1146 				     ilog2(MLX5_SEND_WQE_BB),
1147 				     ilog2(qp->sq.wqe_cnt),
1148 				     sq_strides_offset, &qp->sq.fbc);
1149 
1150 		qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
1151 	}
1152 
1153 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
1154 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * qp->buf.npages;
1155 	*in = kvzalloc(*inlen, GFP_KERNEL);
1156 	if (!*in) {
1157 		err = -ENOMEM;
1158 		goto err_buf;
1159 	}
1160 
1161 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
1162 	MLX5_SET(qpc, qpc, uar_page, uar_index);
1163 	MLX5_SET(qpc, qpc, log_page_size, qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1164 
1165 	/* Set "fast registration enabled" for all kernel QPs */
1166 	MLX5_SET(qpc, qpc, fre, 1);
1167 	MLX5_SET(qpc, qpc, rlky, 1);
1168 
1169 	if (init_attr->create_flags & MLX5_IB_QP_CREATE_SQPN_QP1) {
1170 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
1171 		qp->flags |= MLX5_IB_QP_SQPN_QP1;
1172 	}
1173 
1174 	mlx5_fill_page_frag_array(&qp->buf,
1175 				  (__be64 *)MLX5_ADDR_OF(create_qp_in,
1176 							 *in, pas));
1177 
1178 	err = mlx5_db_alloc(dev->mdev, &qp->db);
1179 	if (err) {
1180 		mlx5_ib_dbg(dev, "err %d\n", err);
1181 		goto err_free;
1182 	}
1183 
1184 	qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1185 				     sizeof(*qp->sq.wrid), GFP_KERNEL);
1186 	qp->sq.wr_data = kvmalloc_array(qp->sq.wqe_cnt,
1187 					sizeof(*qp->sq.wr_data), GFP_KERNEL);
1188 	qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1189 				     sizeof(*qp->rq.wrid), GFP_KERNEL);
1190 	qp->sq.w_list = kvmalloc_array(qp->sq.wqe_cnt,
1191 				       sizeof(*qp->sq.w_list), GFP_KERNEL);
1192 	qp->sq.wqe_head = kvmalloc_array(qp->sq.wqe_cnt,
1193 					 sizeof(*qp->sq.wqe_head), GFP_KERNEL);
1194 
1195 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
1196 	    !qp->sq.w_list || !qp->sq.wqe_head) {
1197 		err = -ENOMEM;
1198 		goto err_wrid;
1199 	}
1200 	qp->create_type = MLX5_QP_KERNEL;
1201 
1202 	return 0;
1203 
1204 err_wrid:
1205 	kvfree(qp->sq.wqe_head);
1206 	kvfree(qp->sq.w_list);
1207 	kvfree(qp->sq.wrid);
1208 	kvfree(qp->sq.wr_data);
1209 	kvfree(qp->rq.wrid);
1210 	mlx5_db_free(dev->mdev, &qp->db);
1211 
1212 err_free:
1213 	kvfree(*in);
1214 
1215 err_buf:
1216 	mlx5_frag_buf_free(dev->mdev, &qp->buf);
1217 	return err;
1218 }
1219 
1220 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1221 {
1222 	kvfree(qp->sq.wqe_head);
1223 	kvfree(qp->sq.w_list);
1224 	kvfree(qp->sq.wrid);
1225 	kvfree(qp->sq.wr_data);
1226 	kvfree(qp->rq.wrid);
1227 	mlx5_db_free(dev->mdev, &qp->db);
1228 	mlx5_frag_buf_free(dev->mdev, &qp->buf);
1229 }
1230 
1231 static u32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
1232 {
1233 	if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
1234 	    (attr->qp_type == MLX5_IB_QPT_DCI) ||
1235 	    (attr->qp_type == IB_QPT_XRC_INI))
1236 		return MLX5_SRQ_RQ;
1237 	else if (!qp->has_rq)
1238 		return MLX5_ZERO_LEN_RQ;
1239 	else
1240 		return MLX5_NON_ZERO_RQ;
1241 }
1242 
1243 static int is_connected(enum ib_qp_type qp_type)
1244 {
1245 	if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC ||
1246 	    qp_type == MLX5_IB_QPT_DCI)
1247 		return 1;
1248 
1249 	return 0;
1250 }
1251 
1252 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1253 				    struct mlx5_ib_qp *qp,
1254 				    struct mlx5_ib_sq *sq, u32 tdn,
1255 				    struct ib_pd *pd)
1256 {
1257 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {0};
1258 	void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1259 
1260 	MLX5_SET(create_tis_in, in, uid, to_mpd(pd)->uid);
1261 	MLX5_SET(tisc, tisc, transport_domain, tdn);
1262 	if (qp->flags & MLX5_IB_QP_UNDERLAY)
1263 		MLX5_SET(tisc, tisc, underlay_qpn, qp->underlay_qpn);
1264 
1265 	return mlx5_core_create_tis(dev->mdev, in, sizeof(in), &sq->tisn);
1266 }
1267 
1268 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1269 				      struct mlx5_ib_sq *sq, struct ib_pd *pd)
1270 {
1271 	mlx5_cmd_destroy_tis(dev->mdev, sq->tisn, to_mpd(pd)->uid);
1272 }
1273 
1274 static void destroy_flow_rule_vport_sq(struct mlx5_ib_sq *sq)
1275 {
1276 	if (sq->flow_rule)
1277 		mlx5_del_flow_rules(sq->flow_rule);
1278 	sq->flow_rule = NULL;
1279 }
1280 
1281 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1282 				   struct ib_udata *udata,
1283 				   struct mlx5_ib_sq *sq, void *qpin,
1284 				   struct ib_pd *pd)
1285 {
1286 	struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1287 	__be64 *pas;
1288 	void *in;
1289 	void *sqc;
1290 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1291 	void *wq;
1292 	int inlen;
1293 	int err;
1294 	int page_shift = 0;
1295 	int npages;
1296 	int ncont = 0;
1297 	u32 offset = 0;
1298 
1299 	err = mlx5_ib_umem_get(dev, udata, ubuffer->buf_addr, ubuffer->buf_size,
1300 			       &sq->ubuffer.umem, &npages, &page_shift, &ncont,
1301 			       &offset);
1302 	if (err)
1303 		return err;
1304 
1305 	inlen = MLX5_ST_SZ_BYTES(create_sq_in) + sizeof(u64) * ncont;
1306 	in = kvzalloc(inlen, GFP_KERNEL);
1307 	if (!in) {
1308 		err = -ENOMEM;
1309 		goto err_umem;
1310 	}
1311 
1312 	MLX5_SET(create_sq_in, in, uid, to_mpd(pd)->uid);
1313 	sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1314 	MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1315 	if (MLX5_CAP_ETH(dev->mdev, multi_pkt_send_wqe))
1316 		MLX5_SET(sqc, sqc, allow_multi_pkt_send_wqe, 1);
1317 	MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1318 	MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1319 	MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1320 	MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1321 	MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1322 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
1323 	    MLX5_CAP_ETH(dev->mdev, swp))
1324 		MLX5_SET(sqc, sqc, allow_swp, 1);
1325 
1326 	wq = MLX5_ADDR_OF(sqc, sqc, wq);
1327 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1328 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1329 	MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1330 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1331 	MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1332 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1333 	MLX5_SET(wq, wq, log_wq_pg_sz,  page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1334 	MLX5_SET(wq, wq, page_offset, offset);
1335 
1336 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1337 	mlx5_ib_populate_pas(dev, sq->ubuffer.umem, page_shift, pas, 0);
1338 
1339 	err = mlx5_core_create_sq_tracked(dev->mdev, in, inlen, &sq->base.mqp);
1340 
1341 	kvfree(in);
1342 
1343 	if (err)
1344 		goto err_umem;
1345 
1346 	return 0;
1347 
1348 err_umem:
1349 	ib_umem_release(sq->ubuffer.umem);
1350 	sq->ubuffer.umem = NULL;
1351 
1352 	return err;
1353 }
1354 
1355 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1356 				     struct mlx5_ib_sq *sq)
1357 {
1358 	destroy_flow_rule_vport_sq(sq);
1359 	mlx5_core_destroy_sq_tracked(dev->mdev, &sq->base.mqp);
1360 	ib_umem_release(sq->ubuffer.umem);
1361 }
1362 
1363 static size_t get_rq_pas_size(void *qpc)
1364 {
1365 	u32 log_page_size = MLX5_GET(qpc, qpc, log_page_size) + 12;
1366 	u32 log_rq_stride = MLX5_GET(qpc, qpc, log_rq_stride);
1367 	u32 log_rq_size   = MLX5_GET(qpc, qpc, log_rq_size);
1368 	u32 page_offset   = MLX5_GET(qpc, qpc, page_offset);
1369 	u32 po_quanta	  = 1 << (log_page_size - 6);
1370 	u32 rq_sz	  = 1 << (log_rq_size + 4 + log_rq_stride);
1371 	u32 page_size	  = 1 << log_page_size;
1372 	u32 rq_sz_po      = rq_sz + (page_offset * po_quanta);
1373 	u32 rq_num_pas	  = (rq_sz_po + page_size - 1) / page_size;
1374 
1375 	return rq_num_pas * sizeof(u64);
1376 }
1377 
1378 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1379 				   struct mlx5_ib_rq *rq, void *qpin,
1380 				   size_t qpinlen, struct ib_pd *pd)
1381 {
1382 	struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1383 	__be64 *pas;
1384 	__be64 *qp_pas;
1385 	void *in;
1386 	void *rqc;
1387 	void *wq;
1388 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1389 	size_t rq_pas_size = get_rq_pas_size(qpc);
1390 	size_t inlen;
1391 	int err;
1392 
1393 	if (qpinlen < rq_pas_size + MLX5_BYTE_OFF(create_qp_in, pas))
1394 		return -EINVAL;
1395 
1396 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + rq_pas_size;
1397 	in = kvzalloc(inlen, GFP_KERNEL);
1398 	if (!in)
1399 		return -ENOMEM;
1400 
1401 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
1402 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1403 	if (!(rq->flags & MLX5_IB_RQ_CVLAN_STRIPPING))
1404 		MLX5_SET(rqc, rqc, vsd, 1);
1405 	MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
1406 	MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1407 	MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1408 	MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1409 	MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1410 
1411 	if (mqp->flags & MLX5_IB_QP_CAP_SCATTER_FCS)
1412 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
1413 
1414 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
1415 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1416 	if (rq->flags & MLX5_IB_RQ_PCI_WRITE_END_PADDING)
1417 		MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
1418 	MLX5_SET(wq, wq, page_offset, MLX5_GET(qpc, qpc, page_offset));
1419 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1420 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1421 	MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1422 	MLX5_SET(wq, wq, log_wq_pg_sz, MLX5_GET(qpc, qpc, log_page_size));
1423 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1424 
1425 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1426 	qp_pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, qpin, pas);
1427 	memcpy(pas, qp_pas, rq_pas_size);
1428 
1429 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rq->base.mqp);
1430 
1431 	kvfree(in);
1432 
1433 	return err;
1434 }
1435 
1436 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1437 				     struct mlx5_ib_rq *rq)
1438 {
1439 	mlx5_core_destroy_rq_tracked(dev->mdev, &rq->base.mqp);
1440 }
1441 
1442 static bool tunnel_offload_supported(struct mlx5_core_dev *dev)
1443 {
1444 	return  (MLX5_CAP_ETH(dev, tunnel_stateless_vxlan) ||
1445 		 MLX5_CAP_ETH(dev, tunnel_stateless_gre) ||
1446 		 MLX5_CAP_ETH(dev, tunnel_stateless_geneve_rx));
1447 }
1448 
1449 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1450 				      struct mlx5_ib_rq *rq,
1451 				      u32 qp_flags_en,
1452 				      struct ib_pd *pd)
1453 {
1454 	if (qp_flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1455 			   MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1456 		mlx5_ib_disable_lb(dev, false, true);
1457 	mlx5_cmd_destroy_tir(dev->mdev, rq->tirn, to_mpd(pd)->uid);
1458 }
1459 
1460 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1461 				    struct mlx5_ib_rq *rq, u32 tdn,
1462 				    u32 *qp_flags_en,
1463 				    struct ib_pd *pd,
1464 				    u32 *out, int outlen)
1465 {
1466 	u8 lb_flag = 0;
1467 	u32 *in;
1468 	void *tirc;
1469 	int inlen;
1470 	int err;
1471 
1472 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1473 	in = kvzalloc(inlen, GFP_KERNEL);
1474 	if (!in)
1475 		return -ENOMEM;
1476 
1477 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1478 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1479 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1480 	MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1481 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1482 	if (*qp_flags_en & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1483 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1484 
1485 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC)
1486 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1487 
1488 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)
1489 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1490 
1491 	if (dev->is_rep) {
1492 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1493 		*qp_flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1494 	}
1495 
1496 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1497 
1498 	err = mlx5_core_create_tir_out(dev->mdev, in, inlen, out, outlen);
1499 
1500 	rq->tirn = MLX5_GET(create_tir_out, out, tirn);
1501 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1502 		err = mlx5_ib_enable_lb(dev, false, true);
1503 
1504 		if (err)
1505 			destroy_raw_packet_qp_tir(dev, rq, 0, pd);
1506 	}
1507 	kvfree(in);
1508 
1509 	return err;
1510 }
1511 
1512 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1513 				u32 *in, size_t inlen,
1514 				struct ib_pd *pd,
1515 				struct ib_udata *udata,
1516 				struct mlx5_ib_create_qp_resp *resp)
1517 {
1518 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1519 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1520 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1521 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1522 		udata, struct mlx5_ib_ucontext, ibucontext);
1523 	int err;
1524 	u32 tdn = mucontext->tdn;
1525 	u16 uid = to_mpd(pd)->uid;
1526 	u32 out[MLX5_ST_SZ_DW(create_tir_out)] = {};
1527 
1528 	if (qp->sq.wqe_cnt) {
1529 		err = create_raw_packet_qp_tis(dev, qp, sq, tdn, pd);
1530 		if (err)
1531 			return err;
1532 
1533 		err = create_raw_packet_qp_sq(dev, udata, sq, in, pd);
1534 		if (err)
1535 			goto err_destroy_tis;
1536 
1537 		if (uid) {
1538 			resp->tisn = sq->tisn;
1539 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TISN;
1540 			resp->sqn = sq->base.mqp.qpn;
1541 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_SQN;
1542 		}
1543 
1544 		sq->base.container_mibqp = qp;
1545 		sq->base.mqp.event = mlx5_ib_qp_event;
1546 	}
1547 
1548 	if (qp->rq.wqe_cnt) {
1549 		rq->base.container_mibqp = qp;
1550 
1551 		if (qp->flags & MLX5_IB_QP_CVLAN_STRIPPING)
1552 			rq->flags |= MLX5_IB_RQ_CVLAN_STRIPPING;
1553 		if (qp->flags & MLX5_IB_QP_PCI_WRITE_END_PADDING)
1554 			rq->flags |= MLX5_IB_RQ_PCI_WRITE_END_PADDING;
1555 		err = create_raw_packet_qp_rq(dev, rq, in, inlen, pd);
1556 		if (err)
1557 			goto err_destroy_sq;
1558 
1559 		err = create_raw_packet_qp_tir(
1560 			dev, rq, tdn, &qp->flags_en, pd, out,
1561 			MLX5_ST_SZ_BYTES(create_tir_out));
1562 		if (err)
1563 			goto err_destroy_rq;
1564 
1565 		if (uid) {
1566 			resp->rqn = rq->base.mqp.qpn;
1567 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_RQN;
1568 			resp->tirn = rq->tirn;
1569 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1570 			if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner)) {
1571 				resp->tir_icm_addr = MLX5_GET(
1572 					create_tir_out, out, icm_address_31_0);
1573 				resp->tir_icm_addr |=
1574 					(u64)MLX5_GET(create_tir_out, out,
1575 						      icm_address_39_32)
1576 					<< 32;
1577 				resp->tir_icm_addr |=
1578 					(u64)MLX5_GET(create_tir_out, out,
1579 						      icm_address_63_40)
1580 					<< 40;
1581 				resp->comp_mask |=
1582 					MLX5_IB_CREATE_QP_RESP_MASK_TIR_ICM_ADDR;
1583 			}
1584 		}
1585 	}
1586 
1587 	qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1588 						     rq->base.mqp.qpn;
1589 	err = ib_copy_to_udata(udata, resp, min(udata->outlen, sizeof(*resp)));
1590 	if (err)
1591 		goto err_destroy_tir;
1592 
1593 	return 0;
1594 
1595 err_destroy_tir:
1596 	destroy_raw_packet_qp_tir(dev, rq, qp->flags_en, pd);
1597 err_destroy_rq:
1598 	destroy_raw_packet_qp_rq(dev, rq);
1599 err_destroy_sq:
1600 	if (!qp->sq.wqe_cnt)
1601 		return err;
1602 	destroy_raw_packet_qp_sq(dev, sq);
1603 err_destroy_tis:
1604 	destroy_raw_packet_qp_tis(dev, sq, pd);
1605 
1606 	return err;
1607 }
1608 
1609 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1610 				  struct mlx5_ib_qp *qp)
1611 {
1612 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1613 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1614 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1615 
1616 	if (qp->rq.wqe_cnt) {
1617 		destroy_raw_packet_qp_tir(dev, rq, qp->flags_en, qp->ibqp.pd);
1618 		destroy_raw_packet_qp_rq(dev, rq);
1619 	}
1620 
1621 	if (qp->sq.wqe_cnt) {
1622 		destroy_raw_packet_qp_sq(dev, sq);
1623 		destroy_raw_packet_qp_tis(dev, sq, qp->ibqp.pd);
1624 	}
1625 }
1626 
1627 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1628 				    struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1629 {
1630 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1631 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1632 
1633 	sq->sq = &qp->sq;
1634 	rq->rq = &qp->rq;
1635 	sq->doorbell = &qp->db;
1636 	rq->doorbell = &qp->db;
1637 }
1638 
1639 static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1640 {
1641 	if (qp->flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1642 			    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1643 		mlx5_ib_disable_lb(dev, false, true);
1644 	mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1645 			     to_mpd(qp->ibqp.pd)->uid);
1646 }
1647 
1648 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1649 				 struct ib_pd *pd,
1650 				 struct ib_qp_init_attr *init_attr,
1651 				 struct ib_udata *udata)
1652 {
1653 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1654 		udata, struct mlx5_ib_ucontext, ibucontext);
1655 	struct mlx5_ib_create_qp_resp resp = {};
1656 	int inlen;
1657 	int outlen;
1658 	int err;
1659 	u32 *in;
1660 	u32 *out;
1661 	void *tirc;
1662 	void *hfso;
1663 	u32 selected_fields = 0;
1664 	u32 outer_l4;
1665 	size_t min_resp_len;
1666 	u32 tdn = mucontext->tdn;
1667 	struct mlx5_ib_create_qp_rss ucmd = {};
1668 	size_t required_cmd_sz;
1669 	u8 lb_flag = 0;
1670 
1671 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1672 		return -EOPNOTSUPP;
1673 
1674 	if (init_attr->create_flags || init_attr->send_cq)
1675 		return -EINVAL;
1676 
1677 	min_resp_len = offsetof(typeof(resp), bfreg_index) + sizeof(resp.bfreg_index);
1678 	if (udata->outlen < min_resp_len)
1679 		return -EINVAL;
1680 
1681 	required_cmd_sz = offsetof(typeof(ucmd), flags) + sizeof(ucmd.flags);
1682 	if (udata->inlen < required_cmd_sz) {
1683 		mlx5_ib_dbg(dev, "invalid inlen\n");
1684 		return -EINVAL;
1685 	}
1686 
1687 	if (udata->inlen > sizeof(ucmd) &&
1688 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
1689 				 udata->inlen - sizeof(ucmd))) {
1690 		mlx5_ib_dbg(dev, "inlen is not supported\n");
1691 		return -EOPNOTSUPP;
1692 	}
1693 
1694 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
1695 		mlx5_ib_dbg(dev, "copy failed\n");
1696 		return -EFAULT;
1697 	}
1698 
1699 	if (ucmd.comp_mask) {
1700 		mlx5_ib_dbg(dev, "invalid comp mask\n");
1701 		return -EOPNOTSUPP;
1702 	}
1703 
1704 	if (ucmd.flags & ~(MLX5_QP_FLAG_TUNNEL_OFFLOADS |
1705 			   MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1706 			   MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)) {
1707 		mlx5_ib_dbg(dev, "invalid flags\n");
1708 		return -EOPNOTSUPP;
1709 	}
1710 
1711 	if (ucmd.flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS &&
1712 	    !tunnel_offload_supported(dev->mdev)) {
1713 		mlx5_ib_dbg(dev, "tunnel offloads isn't supported\n");
1714 		return -EOPNOTSUPP;
1715 	}
1716 
1717 	if (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_INNER &&
1718 	    !(ucmd.flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)) {
1719 		mlx5_ib_dbg(dev, "Tunnel offloads must be set for inner RSS\n");
1720 		return -EOPNOTSUPP;
1721 	}
1722 
1723 	if (ucmd.flags & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC || dev->is_rep) {
1724 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1725 		qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1726 	}
1727 
1728 	if (ucmd.flags & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC) {
1729 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1730 		qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC;
1731 	}
1732 
1733 	err = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp)));
1734 	if (err) {
1735 		mlx5_ib_dbg(dev, "copy failed\n");
1736 		return -EINVAL;
1737 	}
1738 
1739 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1740 	outlen = MLX5_ST_SZ_BYTES(create_tir_out);
1741 	in = kvzalloc(inlen + outlen, GFP_KERNEL);
1742 	if (!in)
1743 		return -ENOMEM;
1744 
1745 	out = in + MLX5_ST_SZ_DW(create_tir_in);
1746 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1747 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1748 	MLX5_SET(tirc, tirc, disp_type,
1749 		 MLX5_TIRC_DISP_TYPE_INDIRECT);
1750 	MLX5_SET(tirc, tirc, indirect_table,
1751 		 init_attr->rwq_ind_tbl->ind_tbl_num);
1752 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1753 
1754 	hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1755 
1756 	if (ucmd.flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1757 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1758 
1759 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1760 
1761 	if (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_INNER)
1762 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_inner);
1763 	else
1764 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1765 
1766 	switch (ucmd.rx_hash_function) {
1767 	case MLX5_RX_HASH_FUNC_TOEPLITZ:
1768 	{
1769 		void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1770 		size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1771 
1772 		if (len != ucmd.rx_key_len) {
1773 			err = -EINVAL;
1774 			goto err;
1775 		}
1776 
1777 		MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_TOEPLITZ);
1778 		memcpy(rss_key, ucmd.rx_hash_key, len);
1779 		break;
1780 	}
1781 	default:
1782 		err = -EOPNOTSUPP;
1783 		goto err;
1784 	}
1785 
1786 	if (!ucmd.rx_hash_fields_mask) {
1787 		/* special case when this TIR serves as steering entry without hashing */
1788 		if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1789 			goto create_tir;
1790 		err = -EINVAL;
1791 		goto err;
1792 	}
1793 
1794 	if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1795 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1796 	     ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1797 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1798 		err = -EINVAL;
1799 		goto err;
1800 	}
1801 
1802 	/* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1803 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1804 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1805 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1806 			 MLX5_L3_PROT_TYPE_IPV4);
1807 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1808 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1809 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1810 			 MLX5_L3_PROT_TYPE_IPV6);
1811 
1812 	outer_l4 = ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1813 		    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP)) << 0 |
1814 		   ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1815 		    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP)) << 1 |
1816 		   (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI) << 2;
1817 
1818 	/* Check that only one l4 protocol is set */
1819 	if (outer_l4 & (outer_l4 - 1)) {
1820 		err = -EINVAL;
1821 		goto err;
1822 	}
1823 
1824 	/* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1825 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1826 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1827 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1828 			 MLX5_L4_PROT_TYPE_TCP);
1829 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1830 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1831 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1832 			 MLX5_L4_PROT_TYPE_UDP);
1833 
1834 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1835 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1836 		selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1837 
1838 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1839 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1840 		selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1841 
1842 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1843 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1844 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1845 
1846 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1847 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1848 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1849 
1850 	if (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI)
1851 		selected_fields |= MLX5_HASH_FIELD_SEL_IPSEC_SPI;
1852 
1853 	MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1854 
1855 create_tir:
1856 	err = mlx5_core_create_tir_out(dev->mdev, in, inlen, out, outlen);
1857 
1858 	qp->rss_qp.tirn = MLX5_GET(create_tir_out, out, tirn);
1859 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1860 		err = mlx5_ib_enable_lb(dev, false, true);
1861 
1862 		if (err)
1863 			mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1864 					     to_mpd(pd)->uid);
1865 	}
1866 
1867 	if (err)
1868 		goto err;
1869 
1870 	if (mucontext->devx_uid) {
1871 		resp.comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1872 		resp.tirn = qp->rss_qp.tirn;
1873 		if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner)) {
1874 			resp.tir_icm_addr =
1875 				MLX5_GET(create_tir_out, out, icm_address_31_0);
1876 			resp.tir_icm_addr |= (u64)MLX5_GET(create_tir_out, out,
1877 							   icm_address_39_32)
1878 					     << 32;
1879 			resp.tir_icm_addr |= (u64)MLX5_GET(create_tir_out, out,
1880 							   icm_address_63_40)
1881 					     << 40;
1882 			resp.comp_mask |=
1883 				MLX5_IB_CREATE_QP_RESP_MASK_TIR_ICM_ADDR;
1884 		}
1885 	}
1886 
1887 	err = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp)));
1888 	if (err)
1889 		goto err_copy;
1890 
1891 	kvfree(in);
1892 	/* qpn is reserved for that QP */
1893 	qp->trans_qp.base.mqp.qpn = 0;
1894 	qp->flags |= MLX5_IB_QP_RSS;
1895 	return 0;
1896 
1897 err_copy:
1898 	mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn, mucontext->devx_uid);
1899 err:
1900 	kvfree(in);
1901 	return err;
1902 }
1903 
1904 static void configure_responder_scat_cqe(struct ib_qp_init_attr *init_attr,
1905 					 void *qpc)
1906 {
1907 	int rcqe_sz;
1908 
1909 	if (init_attr->qp_type == MLX5_IB_QPT_DCI)
1910 		return;
1911 
1912 	rcqe_sz = mlx5_ib_get_cqe_size(init_attr->recv_cq);
1913 
1914 	if (init_attr->qp_type == MLX5_IB_QPT_DCT) {
1915 		if (rcqe_sz == 128)
1916 			MLX5_SET(dctc, qpc, cs_res, MLX5_RES_SCAT_DATA64_CQE);
1917 
1918 		return;
1919 	}
1920 
1921 	MLX5_SET(qpc, qpc, cs_res,
1922 		 rcqe_sz == 128 ? MLX5_RES_SCAT_DATA64_CQE :
1923 				  MLX5_RES_SCAT_DATA32_CQE);
1924 }
1925 
1926 static void configure_requester_scat_cqe(struct mlx5_ib_dev *dev,
1927 					 struct ib_qp_init_attr *init_attr,
1928 					 struct mlx5_ib_create_qp *ucmd,
1929 					 void *qpc)
1930 {
1931 	enum ib_qp_type qpt = init_attr->qp_type;
1932 	int scqe_sz;
1933 	bool allow_scat_cqe = false;
1934 
1935 	if (qpt == IB_QPT_UC || qpt == IB_QPT_UD)
1936 		return;
1937 
1938 	if (ucmd)
1939 		allow_scat_cqe = ucmd->flags & MLX5_QP_FLAG_ALLOW_SCATTER_CQE;
1940 
1941 	if (!allow_scat_cqe && init_attr->sq_sig_type != IB_SIGNAL_ALL_WR)
1942 		return;
1943 
1944 	scqe_sz = mlx5_ib_get_cqe_size(init_attr->send_cq);
1945 	if (scqe_sz == 128) {
1946 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA64_CQE);
1947 		return;
1948 	}
1949 
1950 	if (init_attr->qp_type != MLX5_IB_QPT_DCI ||
1951 	    MLX5_CAP_GEN(dev->mdev, dc_req_scat_data_cqe))
1952 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA32_CQE);
1953 }
1954 
1955 static int atomic_size_to_mode(int size_mask)
1956 {
1957 	/* driver does not support atomic_size > 256B
1958 	 * and does not know how to translate bigger sizes
1959 	 */
1960 	int supported_size_mask = size_mask & 0x1ff;
1961 	int log_max_size;
1962 
1963 	if (!supported_size_mask)
1964 		return -EOPNOTSUPP;
1965 
1966 	log_max_size = __fls(supported_size_mask);
1967 
1968 	if (log_max_size > 3)
1969 		return log_max_size;
1970 
1971 	return MLX5_ATOMIC_MODE_8B;
1972 }
1973 
1974 static int get_atomic_mode(struct mlx5_ib_dev *dev,
1975 			   enum ib_qp_type qp_type)
1976 {
1977 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
1978 	u8 atomic = MLX5_CAP_GEN(dev->mdev, atomic);
1979 	int atomic_mode = -EOPNOTSUPP;
1980 	int atomic_size_mask;
1981 
1982 	if (!atomic)
1983 		return -EOPNOTSUPP;
1984 
1985 	if (qp_type == MLX5_IB_QPT_DCT)
1986 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
1987 	else
1988 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
1989 
1990 	if ((atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_CMP_SWAP) ||
1991 	    (atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_FETCH_ADD))
1992 		atomic_mode = atomic_size_to_mode(atomic_size_mask);
1993 
1994 	if (atomic_mode <= 0 &&
1995 	    (atomic_operations & MLX5_ATOMIC_OPS_CMP_SWAP &&
1996 	     atomic_operations & MLX5_ATOMIC_OPS_FETCH_ADD))
1997 		atomic_mode = MLX5_ATOMIC_MODE_IB_COMP;
1998 
1999 	return atomic_mode;
2000 }
2001 
2002 static inline bool check_flags_mask(uint64_t input, uint64_t supported)
2003 {
2004 	return (input & ~supported) == 0;
2005 }
2006 
2007 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2008 			    struct ib_qp_init_attr *init_attr,
2009 			    struct ib_udata *udata, struct mlx5_ib_qp *qp)
2010 {
2011 	struct mlx5_ib_resources *devr = &dev->devr;
2012 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
2013 	struct mlx5_core_dev *mdev = dev->mdev;
2014 	struct mlx5_ib_create_qp_resp resp = {};
2015 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2016 		udata, struct mlx5_ib_ucontext, ibucontext);
2017 	struct mlx5_ib_cq *send_cq;
2018 	struct mlx5_ib_cq *recv_cq;
2019 	unsigned long flags;
2020 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
2021 	struct mlx5_ib_create_qp ucmd;
2022 	struct mlx5_ib_qp_base *base;
2023 	int mlx5_st;
2024 	void *qpc;
2025 	u32 *in;
2026 	int err;
2027 
2028 	mutex_init(&qp->mutex);
2029 	spin_lock_init(&qp->sq.lock);
2030 	spin_lock_init(&qp->rq.lock);
2031 
2032 	mlx5_st = to_mlx5_st(init_attr->qp_type);
2033 	if (mlx5_st < 0)
2034 		return -EINVAL;
2035 
2036 	if (init_attr->rwq_ind_tbl) {
2037 		if (!udata)
2038 			return -ENOSYS;
2039 
2040 		err = create_rss_raw_qp_tir(dev, qp, pd, init_attr, udata);
2041 		return err;
2042 	}
2043 
2044 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
2045 		if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
2046 			mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
2047 			return -EINVAL;
2048 		} else {
2049 			qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
2050 		}
2051 	}
2052 
2053 	if (init_attr->create_flags &
2054 			(IB_QP_CREATE_CROSS_CHANNEL |
2055 			 IB_QP_CREATE_MANAGED_SEND |
2056 			 IB_QP_CREATE_MANAGED_RECV)) {
2057 		if (!MLX5_CAP_GEN(mdev, cd)) {
2058 			mlx5_ib_dbg(dev, "cross-channel isn't supported\n");
2059 			return -EINVAL;
2060 		}
2061 		if (init_attr->create_flags & IB_QP_CREATE_CROSS_CHANNEL)
2062 			qp->flags |= MLX5_IB_QP_CROSS_CHANNEL;
2063 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_SEND)
2064 			qp->flags |= MLX5_IB_QP_MANAGED_SEND;
2065 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_RECV)
2066 			qp->flags |= MLX5_IB_QP_MANAGED_RECV;
2067 	}
2068 
2069 	if (init_attr->qp_type == IB_QPT_UD &&
2070 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO))
2071 		if (!MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
2072 			mlx5_ib_dbg(dev, "ipoib UD lso qp isn't supported\n");
2073 			return -EOPNOTSUPP;
2074 		}
2075 
2076 	if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
2077 		if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2078 			mlx5_ib_dbg(dev, "Scatter FCS is supported only for Raw Packet QPs");
2079 			return -EOPNOTSUPP;
2080 		}
2081 		if (!MLX5_CAP_GEN(dev->mdev, eth_net_offloads) ||
2082 		    !MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
2083 			mlx5_ib_dbg(dev, "Scatter FCS isn't supported\n");
2084 			return -EOPNOTSUPP;
2085 		}
2086 		qp->flags |= MLX5_IB_QP_CAP_SCATTER_FCS;
2087 	}
2088 
2089 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2090 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2091 
2092 	if (init_attr->create_flags & IB_QP_CREATE_CVLAN_STRIPPING) {
2093 		if (!(MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
2094 		      MLX5_CAP_ETH(dev->mdev, vlan_cap)) ||
2095 		    (init_attr->qp_type != IB_QPT_RAW_PACKET))
2096 			return -EOPNOTSUPP;
2097 		qp->flags |= MLX5_IB_QP_CVLAN_STRIPPING;
2098 	}
2099 
2100 	if (udata) {
2101 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
2102 			mlx5_ib_dbg(dev, "copy failed\n");
2103 			return -EFAULT;
2104 		}
2105 
2106 		if (!check_flags_mask(ucmd.flags,
2107 				      MLX5_QP_FLAG_ALLOW_SCATTER_CQE |
2108 				      MLX5_QP_FLAG_BFREG_INDEX |
2109 				      MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE |
2110 				      MLX5_QP_FLAG_SCATTER_CQE |
2111 				      MLX5_QP_FLAG_SIGNATURE |
2112 				      MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC |
2113 				      MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
2114 				      MLX5_QP_FLAG_TUNNEL_OFFLOADS |
2115 				      MLX5_QP_FLAG_UAR_PAGE_INDEX |
2116 				      MLX5_QP_FLAG_TYPE_DCI |
2117 				      MLX5_QP_FLAG_TYPE_DCT))
2118 			return -EINVAL;
2119 
2120 		err = get_qp_user_index(ucontext, &ucmd, udata->inlen, &uidx);
2121 		if (err)
2122 			return err;
2123 
2124 		qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
2125 		if (MLX5_CAP_GEN(dev->mdev, sctr_data_cqe))
2126 			qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
2127 		if (ucmd.flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS) {
2128 			if (init_attr->qp_type != IB_QPT_RAW_PACKET ||
2129 			    !tunnel_offload_supported(mdev)) {
2130 				mlx5_ib_dbg(dev, "Tunnel offload isn't supported\n");
2131 				return -EOPNOTSUPP;
2132 			}
2133 			qp->flags_en |= MLX5_QP_FLAG_TUNNEL_OFFLOADS;
2134 		}
2135 
2136 		if (ucmd.flags & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC) {
2137 			if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2138 				mlx5_ib_dbg(dev, "Self-LB UC isn't supported\n");
2139 				return -EOPNOTSUPP;
2140 			}
2141 			qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
2142 		}
2143 
2144 		if (ucmd.flags & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC) {
2145 			if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2146 				mlx5_ib_dbg(dev, "Self-LB UM isn't supported\n");
2147 				return -EOPNOTSUPP;
2148 			}
2149 			qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC;
2150 		}
2151 
2152 		if (ucmd.flags & MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE) {
2153 			if (init_attr->qp_type != IB_QPT_RC ||
2154 				!MLX5_CAP_GEN(dev->mdev, qp_packet_based)) {
2155 				mlx5_ib_dbg(dev, "packet based credit mode isn't supported\n");
2156 				return -EOPNOTSUPP;
2157 			}
2158 			qp->flags |= MLX5_IB_QP_PACKET_BASED_CREDIT;
2159 		}
2160 
2161 		if (init_attr->create_flags & IB_QP_CREATE_SOURCE_QPN) {
2162 			if (init_attr->qp_type != IB_QPT_UD ||
2163 			    (MLX5_CAP_GEN(dev->mdev, port_type) !=
2164 			     MLX5_CAP_PORT_TYPE_IB) ||
2165 			    !mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS)) {
2166 				mlx5_ib_dbg(dev, "Source QP option isn't supported\n");
2167 				return -EOPNOTSUPP;
2168 			}
2169 
2170 			qp->flags |= MLX5_IB_QP_UNDERLAY;
2171 			qp->underlay_qpn = init_attr->source_qpn;
2172 		}
2173 	} else {
2174 		qp->wq_sig = !!wq_signature;
2175 	}
2176 
2177 	base = (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2178 		qp->flags & MLX5_IB_QP_UNDERLAY) ?
2179 	       &qp->raw_packet_qp.rq.base :
2180 	       &qp->trans_qp.base;
2181 
2182 	qp->has_rq = qp_has_rq(init_attr);
2183 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
2184 			  qp, udata ? &ucmd : NULL);
2185 	if (err) {
2186 		mlx5_ib_dbg(dev, "err %d\n", err);
2187 		return err;
2188 	}
2189 
2190 	if (pd) {
2191 		if (udata) {
2192 			__u32 max_wqes =
2193 				1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
2194 			mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
2195 			if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
2196 			    ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
2197 				mlx5_ib_dbg(dev, "invalid rq params\n");
2198 				return -EINVAL;
2199 			}
2200 			if (ucmd.sq_wqe_count > max_wqes) {
2201 				mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
2202 					    ucmd.sq_wqe_count, max_wqes);
2203 				return -EINVAL;
2204 			}
2205 			if (init_attr->create_flags &
2206 			    MLX5_IB_QP_CREATE_SQPN_QP1) {
2207 				mlx5_ib_dbg(dev, "user-space is not allowed to create UD QPs spoofing as QP1\n");
2208 				return -EINVAL;
2209 			}
2210 			err = create_user_qp(dev, pd, qp, udata, init_attr, &in,
2211 					     &resp, &inlen, base);
2212 			if (err)
2213 				mlx5_ib_dbg(dev, "err %d\n", err);
2214 		} else {
2215 			err = create_kernel_qp(dev, init_attr, qp, &in, &inlen,
2216 					       base);
2217 			if (err)
2218 				mlx5_ib_dbg(dev, "err %d\n", err);
2219 		}
2220 
2221 		if (err)
2222 			return err;
2223 	} else {
2224 		in = kvzalloc(inlen, GFP_KERNEL);
2225 		if (!in)
2226 			return -ENOMEM;
2227 
2228 		qp->create_type = MLX5_QP_EMPTY;
2229 	}
2230 
2231 	if (is_sqp(init_attr->qp_type))
2232 		qp->port = init_attr->port_num;
2233 
2234 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2235 
2236 	MLX5_SET(qpc, qpc, st, mlx5_st);
2237 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2238 
2239 	if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
2240 		MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
2241 	else
2242 		MLX5_SET(qpc, qpc, latency_sensitive, 1);
2243 
2244 
2245 	if (qp->wq_sig)
2246 		MLX5_SET(qpc, qpc, wq_signature, 1);
2247 
2248 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
2249 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
2250 
2251 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
2252 		MLX5_SET(qpc, qpc, cd_master, 1);
2253 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
2254 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
2255 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
2256 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
2257 	if (qp->flags & MLX5_IB_QP_PACKET_BASED_CREDIT)
2258 		MLX5_SET(qpc, qpc, req_e2e_credit_mode, 1);
2259 	if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
2260 		configure_responder_scat_cqe(init_attr, qpc);
2261 		configure_requester_scat_cqe(dev, init_attr,
2262 					     udata ? &ucmd : NULL,
2263 					     qpc);
2264 	}
2265 
2266 	if (qp->rq.wqe_cnt) {
2267 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2268 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2269 	}
2270 
2271 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
2272 
2273 	if (qp->sq.wqe_cnt) {
2274 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2275 	} else {
2276 		MLX5_SET(qpc, qpc, no_sq, 1);
2277 		if (init_attr->srq &&
2278 		    init_attr->srq->srq_type == IB_SRQT_TM)
2279 			MLX5_SET(qpc, qpc, offload_type,
2280 				 MLX5_QPC_OFFLOAD_TYPE_RNDV);
2281 	}
2282 
2283 	/* Set default resources */
2284 	switch (init_attr->qp_type) {
2285 	case IB_QPT_XRC_TGT:
2286 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
2287 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(devr->c0)->mcq.cqn);
2288 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
2289 		MLX5_SET(qpc, qpc, xrcd, to_mxrcd(init_attr->xrcd)->xrcdn);
2290 		break;
2291 	case IB_QPT_XRC_INI:
2292 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
2293 		MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
2294 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
2295 		break;
2296 	default:
2297 		if (init_attr->srq) {
2298 			MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x0)->xrcdn);
2299 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(init_attr->srq)->msrq.srqn);
2300 		} else {
2301 			MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
2302 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s1)->msrq.srqn);
2303 		}
2304 	}
2305 
2306 	if (init_attr->send_cq)
2307 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(init_attr->send_cq)->mcq.cqn);
2308 
2309 	if (init_attr->recv_cq)
2310 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(init_attr->recv_cq)->mcq.cqn);
2311 
2312 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2313 
2314 	/* 0xffffff means we ask to work with cqe version 0 */
2315 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2316 		MLX5_SET(qpc, qpc, user_index, uidx);
2317 
2318 	/* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
2319 	if (init_attr->qp_type == IB_QPT_UD &&
2320 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)) {
2321 		MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
2322 		qp->flags |= MLX5_IB_QP_LSO;
2323 	}
2324 
2325 	if (init_attr->create_flags & IB_QP_CREATE_PCI_WRITE_END_PADDING) {
2326 		if (!MLX5_CAP_GEN(dev->mdev, end_pad)) {
2327 			mlx5_ib_dbg(dev, "scatter end padding is not supported\n");
2328 			err = -EOPNOTSUPP;
2329 			goto err;
2330 		} else if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2331 			MLX5_SET(qpc, qpc, end_padding_mode,
2332 				 MLX5_WQ_END_PAD_MODE_ALIGN);
2333 		} else {
2334 			qp->flags |= MLX5_IB_QP_PCI_WRITE_END_PADDING;
2335 		}
2336 	}
2337 
2338 	if (inlen < 0) {
2339 		err = -EINVAL;
2340 		goto err;
2341 	}
2342 
2343 	if (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2344 	    qp->flags & MLX5_IB_QP_UNDERLAY) {
2345 		qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd.sq_buf_addr;
2346 		raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
2347 		err = create_raw_packet_qp(dev, qp, in, inlen, pd, udata,
2348 					   &resp);
2349 	} else {
2350 		err = mlx5_core_create_qp(dev->mdev, &base->mqp, in, inlen);
2351 	}
2352 
2353 	if (err) {
2354 		mlx5_ib_dbg(dev, "create qp failed\n");
2355 		goto err_create;
2356 	}
2357 
2358 	kvfree(in);
2359 
2360 	base->container_mibqp = qp;
2361 	base->mqp.event = mlx5_ib_qp_event;
2362 
2363 	get_cqs(init_attr->qp_type, init_attr->send_cq, init_attr->recv_cq,
2364 		&send_cq, &recv_cq);
2365 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2366 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2367 	/* Maintain device to QPs access, needed for further handling via reset
2368 	 * flow
2369 	 */
2370 	list_add_tail(&qp->qps_list, &dev->qp_list);
2371 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2372 	 */
2373 	if (send_cq)
2374 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2375 	if (recv_cq)
2376 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2377 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2378 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2379 
2380 	return 0;
2381 
2382 err_create:
2383 	if (qp->create_type == MLX5_QP_USER)
2384 		destroy_qp_user(dev, pd, qp, base, udata);
2385 	else if (qp->create_type == MLX5_QP_KERNEL)
2386 		destroy_qp_kernel(dev, qp);
2387 
2388 err:
2389 	kvfree(in);
2390 	return err;
2391 }
2392 
2393 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2394 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
2395 {
2396 	if (send_cq) {
2397 		if (recv_cq) {
2398 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2399 				spin_lock(&send_cq->lock);
2400 				spin_lock_nested(&recv_cq->lock,
2401 						 SINGLE_DEPTH_NESTING);
2402 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2403 				spin_lock(&send_cq->lock);
2404 				__acquire(&recv_cq->lock);
2405 			} else {
2406 				spin_lock(&recv_cq->lock);
2407 				spin_lock_nested(&send_cq->lock,
2408 						 SINGLE_DEPTH_NESTING);
2409 			}
2410 		} else {
2411 			spin_lock(&send_cq->lock);
2412 			__acquire(&recv_cq->lock);
2413 		}
2414 	} else if (recv_cq) {
2415 		spin_lock(&recv_cq->lock);
2416 		__acquire(&send_cq->lock);
2417 	} else {
2418 		__acquire(&send_cq->lock);
2419 		__acquire(&recv_cq->lock);
2420 	}
2421 }
2422 
2423 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2424 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
2425 {
2426 	if (send_cq) {
2427 		if (recv_cq) {
2428 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2429 				spin_unlock(&recv_cq->lock);
2430 				spin_unlock(&send_cq->lock);
2431 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2432 				__release(&recv_cq->lock);
2433 				spin_unlock(&send_cq->lock);
2434 			} else {
2435 				spin_unlock(&send_cq->lock);
2436 				spin_unlock(&recv_cq->lock);
2437 			}
2438 		} else {
2439 			__release(&recv_cq->lock);
2440 			spin_unlock(&send_cq->lock);
2441 		}
2442 	} else if (recv_cq) {
2443 		__release(&send_cq->lock);
2444 		spin_unlock(&recv_cq->lock);
2445 	} else {
2446 		__release(&recv_cq->lock);
2447 		__release(&send_cq->lock);
2448 	}
2449 }
2450 
2451 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
2452 {
2453 	return to_mpd(qp->ibqp.pd);
2454 }
2455 
2456 static void get_cqs(enum ib_qp_type qp_type,
2457 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
2458 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
2459 {
2460 	switch (qp_type) {
2461 	case IB_QPT_XRC_TGT:
2462 		*send_cq = NULL;
2463 		*recv_cq = NULL;
2464 		break;
2465 	case MLX5_IB_QPT_REG_UMR:
2466 	case IB_QPT_XRC_INI:
2467 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2468 		*recv_cq = NULL;
2469 		break;
2470 
2471 	case IB_QPT_SMI:
2472 	case MLX5_IB_QPT_HW_GSI:
2473 	case IB_QPT_RC:
2474 	case IB_QPT_UC:
2475 	case IB_QPT_UD:
2476 	case IB_QPT_RAW_IPV6:
2477 	case IB_QPT_RAW_ETHERTYPE:
2478 	case IB_QPT_RAW_PACKET:
2479 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2480 		*recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
2481 		break;
2482 
2483 	case IB_QPT_MAX:
2484 	default:
2485 		*send_cq = NULL;
2486 		*recv_cq = NULL;
2487 		break;
2488 	}
2489 }
2490 
2491 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2492 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
2493 				u8 lag_tx_affinity);
2494 
2495 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2496 			      struct ib_udata *udata)
2497 {
2498 	struct mlx5_ib_cq *send_cq, *recv_cq;
2499 	struct mlx5_ib_qp_base *base;
2500 	unsigned long flags;
2501 	int err;
2502 
2503 	if (qp->ibqp.rwq_ind_tbl) {
2504 		destroy_rss_raw_qp_tir(dev, qp);
2505 		return;
2506 	}
2507 
2508 	base = (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
2509 		qp->flags & MLX5_IB_QP_UNDERLAY) ?
2510 	       &qp->raw_packet_qp.rq.base :
2511 	       &qp->trans_qp.base;
2512 
2513 	if (qp->state != IB_QPS_RESET) {
2514 		if (qp->ibqp.qp_type != IB_QPT_RAW_PACKET &&
2515 		    !(qp->flags & MLX5_IB_QP_UNDERLAY)) {
2516 			err = mlx5_core_qp_modify(dev->mdev,
2517 						  MLX5_CMD_OP_2RST_QP, 0,
2518 						  NULL, &base->mqp);
2519 		} else {
2520 			struct mlx5_modify_raw_qp_param raw_qp_param = {
2521 				.operation = MLX5_CMD_OP_2RST_QP
2522 			};
2523 
2524 			err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
2525 		}
2526 		if (err)
2527 			mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
2528 				     base->mqp.qpn);
2529 	}
2530 
2531 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
2532 		&send_cq, &recv_cq);
2533 
2534 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2535 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2536 	/* del from lists under both locks above to protect reset flow paths */
2537 	list_del(&qp->qps_list);
2538 	if (send_cq)
2539 		list_del(&qp->cq_send_list);
2540 
2541 	if (recv_cq)
2542 		list_del(&qp->cq_recv_list);
2543 
2544 	if (qp->create_type == MLX5_QP_KERNEL) {
2545 		__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2546 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
2547 		if (send_cq != recv_cq)
2548 			__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
2549 					   NULL);
2550 	}
2551 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2552 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2553 
2554 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
2555 	    qp->flags & MLX5_IB_QP_UNDERLAY) {
2556 		destroy_raw_packet_qp(dev, qp);
2557 	} else {
2558 		err = mlx5_core_destroy_qp(dev->mdev, &base->mqp);
2559 		if (err)
2560 			mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
2561 				     base->mqp.qpn);
2562 	}
2563 
2564 	if (qp->create_type == MLX5_QP_KERNEL)
2565 		destroy_qp_kernel(dev, qp);
2566 	else if (qp->create_type == MLX5_QP_USER)
2567 		destroy_qp_user(dev, &get_pd(qp)->ibpd, qp, base, udata);
2568 }
2569 
2570 static const char *ib_qp_type_str(enum ib_qp_type type)
2571 {
2572 	switch (type) {
2573 	case IB_QPT_SMI:
2574 		return "IB_QPT_SMI";
2575 	case IB_QPT_GSI:
2576 		return "IB_QPT_GSI";
2577 	case IB_QPT_RC:
2578 		return "IB_QPT_RC";
2579 	case IB_QPT_UC:
2580 		return "IB_QPT_UC";
2581 	case IB_QPT_UD:
2582 		return "IB_QPT_UD";
2583 	case IB_QPT_RAW_IPV6:
2584 		return "IB_QPT_RAW_IPV6";
2585 	case IB_QPT_RAW_ETHERTYPE:
2586 		return "IB_QPT_RAW_ETHERTYPE";
2587 	case IB_QPT_XRC_INI:
2588 		return "IB_QPT_XRC_INI";
2589 	case IB_QPT_XRC_TGT:
2590 		return "IB_QPT_XRC_TGT";
2591 	case IB_QPT_RAW_PACKET:
2592 		return "IB_QPT_RAW_PACKET";
2593 	case MLX5_IB_QPT_REG_UMR:
2594 		return "MLX5_IB_QPT_REG_UMR";
2595 	case IB_QPT_DRIVER:
2596 		return "IB_QPT_DRIVER";
2597 	case IB_QPT_MAX:
2598 	default:
2599 		return "Invalid QP type";
2600 	}
2601 }
2602 
2603 static struct ib_qp *mlx5_ib_create_dct(struct ib_pd *pd,
2604 					struct ib_qp_init_attr *attr,
2605 					struct mlx5_ib_create_qp *ucmd,
2606 					struct ib_udata *udata)
2607 {
2608 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2609 		udata, struct mlx5_ib_ucontext, ibucontext);
2610 	struct mlx5_ib_qp *qp;
2611 	int err = 0;
2612 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
2613 	void *dctc;
2614 
2615 	if (!attr->srq || !attr->recv_cq)
2616 		return ERR_PTR(-EINVAL);
2617 
2618 	err = get_qp_user_index(ucontext, ucmd, sizeof(*ucmd), &uidx);
2619 	if (err)
2620 		return ERR_PTR(err);
2621 
2622 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2623 	if (!qp)
2624 		return ERR_PTR(-ENOMEM);
2625 
2626 	qp->dct.in = kzalloc(MLX5_ST_SZ_BYTES(create_dct_in), GFP_KERNEL);
2627 	if (!qp->dct.in) {
2628 		err = -ENOMEM;
2629 		goto err_free;
2630 	}
2631 
2632 	MLX5_SET(create_dct_in, qp->dct.in, uid, to_mpd(pd)->uid);
2633 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
2634 	qp->qp_sub_type = MLX5_IB_QPT_DCT;
2635 	MLX5_SET(dctc, dctc, pd, to_mpd(pd)->pdn);
2636 	MLX5_SET(dctc, dctc, srqn_xrqn, to_msrq(attr->srq)->msrq.srqn);
2637 	MLX5_SET(dctc, dctc, cqn, to_mcq(attr->recv_cq)->mcq.cqn);
2638 	MLX5_SET64(dctc, dctc, dc_access_key, ucmd->access_key);
2639 	MLX5_SET(dctc, dctc, user_index, uidx);
2640 
2641 	if (ucmd->flags & MLX5_QP_FLAG_SCATTER_CQE)
2642 		configure_responder_scat_cqe(attr, dctc);
2643 
2644 	qp->state = IB_QPS_RESET;
2645 
2646 	return &qp->ibqp;
2647 err_free:
2648 	kfree(qp);
2649 	return ERR_PTR(err);
2650 }
2651 
2652 static int set_mlx_qp_type(struct mlx5_ib_dev *dev,
2653 			   struct ib_qp_init_attr *init_attr,
2654 			   struct mlx5_ib_create_qp *ucmd,
2655 			   struct ib_udata *udata)
2656 {
2657 	enum { MLX_QP_FLAGS = MLX5_QP_FLAG_TYPE_DCT | MLX5_QP_FLAG_TYPE_DCI };
2658 	int err;
2659 
2660 	if (!udata)
2661 		return -EINVAL;
2662 
2663 	if (udata->inlen < sizeof(*ucmd)) {
2664 		mlx5_ib_dbg(dev, "create_qp user command is smaller than expected\n");
2665 		return -EINVAL;
2666 	}
2667 	err = ib_copy_from_udata(ucmd, udata, sizeof(*ucmd));
2668 	if (err)
2669 		return err;
2670 
2671 	if ((ucmd->flags & MLX_QP_FLAGS) == MLX5_QP_FLAG_TYPE_DCI) {
2672 		init_attr->qp_type = MLX5_IB_QPT_DCI;
2673 	} else {
2674 		if ((ucmd->flags & MLX_QP_FLAGS) == MLX5_QP_FLAG_TYPE_DCT) {
2675 			init_attr->qp_type = MLX5_IB_QPT_DCT;
2676 		} else {
2677 			mlx5_ib_dbg(dev, "Invalid QP flags\n");
2678 			return -EINVAL;
2679 		}
2680 	}
2681 
2682 	if (!MLX5_CAP_GEN(dev->mdev, dct)) {
2683 		mlx5_ib_dbg(dev, "DC transport is not supported\n");
2684 		return -EOPNOTSUPP;
2685 	}
2686 
2687 	return 0;
2688 }
2689 
2690 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
2691 				struct ib_qp_init_attr *verbs_init_attr,
2692 				struct ib_udata *udata)
2693 {
2694 	struct mlx5_ib_dev *dev;
2695 	struct mlx5_ib_qp *qp;
2696 	u16 xrcdn = 0;
2697 	int err;
2698 	struct ib_qp_init_attr mlx_init_attr;
2699 	struct ib_qp_init_attr *init_attr = verbs_init_attr;
2700 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2701 		udata, struct mlx5_ib_ucontext, ibucontext);
2702 
2703 	if (pd) {
2704 		dev = to_mdev(pd->device);
2705 
2706 		if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
2707 			if (!ucontext) {
2708 				mlx5_ib_dbg(dev, "Raw Packet QP is not supported for kernel consumers\n");
2709 				return ERR_PTR(-EINVAL);
2710 			} else if (!ucontext->cqe_version) {
2711 				mlx5_ib_dbg(dev, "Raw Packet QP is only supported for CQE version > 0\n");
2712 				return ERR_PTR(-EINVAL);
2713 			}
2714 		}
2715 	} else {
2716 		/* being cautious here */
2717 		if (init_attr->qp_type != IB_QPT_XRC_TGT &&
2718 		    init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
2719 			pr_warn("%s: no PD for transport %s\n", __func__,
2720 				ib_qp_type_str(init_attr->qp_type));
2721 			return ERR_PTR(-EINVAL);
2722 		}
2723 		dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
2724 	}
2725 
2726 	if (init_attr->qp_type == IB_QPT_DRIVER) {
2727 		struct mlx5_ib_create_qp ucmd;
2728 
2729 		init_attr = &mlx_init_attr;
2730 		memcpy(init_attr, verbs_init_attr, sizeof(*verbs_init_attr));
2731 		err = set_mlx_qp_type(dev, init_attr, &ucmd, udata);
2732 		if (err)
2733 			return ERR_PTR(err);
2734 
2735 		if (init_attr->qp_type == MLX5_IB_QPT_DCI) {
2736 			if (init_attr->cap.max_recv_wr ||
2737 			    init_attr->cap.max_recv_sge) {
2738 				mlx5_ib_dbg(dev, "DCI QP requires zero size receive queue\n");
2739 				return ERR_PTR(-EINVAL);
2740 			}
2741 		} else {
2742 			return mlx5_ib_create_dct(pd, init_attr, &ucmd, udata);
2743 		}
2744 	}
2745 
2746 	switch (init_attr->qp_type) {
2747 	case IB_QPT_XRC_TGT:
2748 	case IB_QPT_XRC_INI:
2749 		if (!MLX5_CAP_GEN(dev->mdev, xrc)) {
2750 			mlx5_ib_dbg(dev, "XRC not supported\n");
2751 			return ERR_PTR(-ENOSYS);
2752 		}
2753 		init_attr->recv_cq = NULL;
2754 		if (init_attr->qp_type == IB_QPT_XRC_TGT) {
2755 			xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
2756 			init_attr->send_cq = NULL;
2757 		}
2758 
2759 		/* fall through */
2760 	case IB_QPT_RAW_PACKET:
2761 	case IB_QPT_RC:
2762 	case IB_QPT_UC:
2763 	case IB_QPT_UD:
2764 	case IB_QPT_SMI:
2765 	case MLX5_IB_QPT_HW_GSI:
2766 	case MLX5_IB_QPT_REG_UMR:
2767 	case MLX5_IB_QPT_DCI:
2768 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2769 		if (!qp)
2770 			return ERR_PTR(-ENOMEM);
2771 
2772 		err = create_qp_common(dev, pd, init_attr, udata, qp);
2773 		if (err) {
2774 			mlx5_ib_dbg(dev, "create_qp_common failed\n");
2775 			kfree(qp);
2776 			return ERR_PTR(err);
2777 		}
2778 
2779 		if (is_qp0(init_attr->qp_type))
2780 			qp->ibqp.qp_num = 0;
2781 		else if (is_qp1(init_attr->qp_type))
2782 			qp->ibqp.qp_num = 1;
2783 		else
2784 			qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
2785 
2786 		mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
2787 			    qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
2788 			    init_attr->recv_cq ? to_mcq(init_attr->recv_cq)->mcq.cqn : -1,
2789 			    init_attr->send_cq ? to_mcq(init_attr->send_cq)->mcq.cqn : -1);
2790 
2791 		qp->trans_qp.xrcdn = xrcdn;
2792 
2793 		break;
2794 
2795 	case IB_QPT_GSI:
2796 		return mlx5_ib_gsi_create_qp(pd, init_attr);
2797 
2798 	case IB_QPT_RAW_IPV6:
2799 	case IB_QPT_RAW_ETHERTYPE:
2800 	case IB_QPT_MAX:
2801 	default:
2802 		mlx5_ib_dbg(dev, "unsupported qp type %d\n",
2803 			    init_attr->qp_type);
2804 		/* Don't support raw QPs */
2805 		return ERR_PTR(-EOPNOTSUPP);
2806 	}
2807 
2808 	if (verbs_init_attr->qp_type == IB_QPT_DRIVER)
2809 		qp->qp_sub_type = init_attr->qp_type;
2810 
2811 	return &qp->ibqp;
2812 }
2813 
2814 static int mlx5_ib_destroy_dct(struct mlx5_ib_qp *mqp)
2815 {
2816 	struct mlx5_ib_dev *dev = to_mdev(mqp->ibqp.device);
2817 
2818 	if (mqp->state == IB_QPS_RTR) {
2819 		int err;
2820 
2821 		err = mlx5_core_destroy_dct(dev->mdev, &mqp->dct.mdct);
2822 		if (err) {
2823 			mlx5_ib_warn(dev, "failed to destroy DCT %d\n", err);
2824 			return err;
2825 		}
2826 	}
2827 
2828 	kfree(mqp->dct.in);
2829 	kfree(mqp);
2830 	return 0;
2831 }
2832 
2833 int mlx5_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
2834 {
2835 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2836 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2837 
2838 	if (unlikely(qp->qp_type == IB_QPT_GSI))
2839 		return mlx5_ib_gsi_destroy_qp(qp);
2840 
2841 	if (mqp->qp_sub_type == MLX5_IB_QPT_DCT)
2842 		return mlx5_ib_destroy_dct(mqp);
2843 
2844 	destroy_qp_common(dev, mqp, udata);
2845 
2846 	kfree(mqp);
2847 
2848 	return 0;
2849 }
2850 
2851 static int to_mlx5_access_flags(struct mlx5_ib_qp *qp,
2852 				const struct ib_qp_attr *attr,
2853 				int attr_mask, __be32 *hw_access_flags_be)
2854 {
2855 	u8 dest_rd_atomic;
2856 	u32 access_flags, hw_access_flags = 0;
2857 
2858 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
2859 
2860 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2861 		dest_rd_atomic = attr->max_dest_rd_atomic;
2862 	else
2863 		dest_rd_atomic = qp->trans_qp.resp_depth;
2864 
2865 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2866 		access_flags = attr->qp_access_flags;
2867 	else
2868 		access_flags = qp->trans_qp.atomic_rd_en;
2869 
2870 	if (!dest_rd_atomic)
2871 		access_flags &= IB_ACCESS_REMOTE_WRITE;
2872 
2873 	if (access_flags & IB_ACCESS_REMOTE_READ)
2874 		hw_access_flags |= MLX5_QP_BIT_RRE;
2875 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC) {
2876 		int atomic_mode;
2877 
2878 		atomic_mode = get_atomic_mode(dev, qp->ibqp.qp_type);
2879 		if (atomic_mode < 0)
2880 			return -EOPNOTSUPP;
2881 
2882 		hw_access_flags |= MLX5_QP_BIT_RAE;
2883 		hw_access_flags |= atomic_mode << MLX5_ATOMIC_MODE_OFFSET;
2884 	}
2885 
2886 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
2887 		hw_access_flags |= MLX5_QP_BIT_RWE;
2888 
2889 	*hw_access_flags_be = cpu_to_be32(hw_access_flags);
2890 
2891 	return 0;
2892 }
2893 
2894 enum {
2895 	MLX5_PATH_FLAG_FL	= 1 << 0,
2896 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
2897 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
2898 };
2899 
2900 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
2901 {
2902 	if (rate == IB_RATE_PORT_CURRENT)
2903 		return 0;
2904 
2905 	if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_600_GBPS)
2906 		return -EINVAL;
2907 
2908 	while (rate != IB_RATE_PORT_CURRENT &&
2909 	       !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
2910 		 MLX5_CAP_GEN(dev->mdev, stat_rate_support)))
2911 		--rate;
2912 
2913 	return rate ? rate + MLX5_STAT_RATE_OFFSET : rate;
2914 }
2915 
2916 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
2917 				      struct mlx5_ib_sq *sq, u8 sl,
2918 				      struct ib_pd *pd)
2919 {
2920 	void *in;
2921 	void *tisc;
2922 	int inlen;
2923 	int err;
2924 
2925 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2926 	in = kvzalloc(inlen, GFP_KERNEL);
2927 	if (!in)
2928 		return -ENOMEM;
2929 
2930 	MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
2931 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
2932 
2933 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2934 	MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
2935 
2936 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2937 
2938 	kvfree(in);
2939 
2940 	return err;
2941 }
2942 
2943 static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,
2944 					 struct mlx5_ib_sq *sq, u8 tx_affinity,
2945 					 struct ib_pd *pd)
2946 {
2947 	void *in;
2948 	void *tisc;
2949 	int inlen;
2950 	int err;
2951 
2952 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2953 	in = kvzalloc(inlen, GFP_KERNEL);
2954 	if (!in)
2955 		return -ENOMEM;
2956 
2957 	MLX5_SET(modify_tis_in, in, bitmask.lag_tx_port_affinity, 1);
2958 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
2959 
2960 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2961 	MLX5_SET(tisc, tisc, lag_tx_port_affinity, tx_affinity);
2962 
2963 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2964 
2965 	kvfree(in);
2966 
2967 	return err;
2968 }
2969 
2970 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2971 			 const struct rdma_ah_attr *ah,
2972 			 struct mlx5_qp_path *path, u8 port, int attr_mask,
2973 			 u32 path_flags, const struct ib_qp_attr *attr,
2974 			 bool alt)
2975 {
2976 	const struct ib_global_route *grh = rdma_ah_read_grh(ah);
2977 	int err;
2978 	enum ib_gid_type gid_type;
2979 	u8 ah_flags = rdma_ah_get_ah_flags(ah);
2980 	u8 sl = rdma_ah_get_sl(ah);
2981 
2982 	if (attr_mask & IB_QP_PKEY_INDEX)
2983 		path->pkey_index = cpu_to_be16(alt ? attr->alt_pkey_index :
2984 						     attr->pkey_index);
2985 
2986 	if (ah_flags & IB_AH_GRH) {
2987 		if (grh->sgid_index >=
2988 		    dev->mdev->port_caps[port - 1].gid_table_len) {
2989 			pr_err("sgid_index (%u) too large. max is %d\n",
2990 			       grh->sgid_index,
2991 			       dev->mdev->port_caps[port - 1].gid_table_len);
2992 			return -EINVAL;
2993 		}
2994 	}
2995 
2996 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
2997 		if (!(ah_flags & IB_AH_GRH))
2998 			return -EINVAL;
2999 
3000 		memcpy(path->rmac, ah->roce.dmac, sizeof(ah->roce.dmac));
3001 		if (qp->ibqp.qp_type == IB_QPT_RC ||
3002 		    qp->ibqp.qp_type == IB_QPT_UC ||
3003 		    qp->ibqp.qp_type == IB_QPT_XRC_INI ||
3004 		    qp->ibqp.qp_type == IB_QPT_XRC_TGT)
3005 			path->udp_sport =
3006 				mlx5_get_roce_udp_sport(dev, ah->grh.sgid_attr);
3007 		path->dci_cfi_prio_sl = (sl & 0x7) << 4;
3008 		gid_type = ah->grh.sgid_attr->gid_type;
3009 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
3010 			path->ecn_dscp = (grh->traffic_class >> 2) & 0x3f;
3011 	} else {
3012 		path->fl_free_ar = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
3013 		path->fl_free_ar |=
3014 			(path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x40 : 0;
3015 		path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
3016 		path->grh_mlid = rdma_ah_get_path_bits(ah) & 0x7f;
3017 		if (ah_flags & IB_AH_GRH)
3018 			path->grh_mlid	|= 1 << 7;
3019 		path->dci_cfi_prio_sl = sl & 0xf;
3020 	}
3021 
3022 	if (ah_flags & IB_AH_GRH) {
3023 		path->mgid_index = grh->sgid_index;
3024 		path->hop_limit  = grh->hop_limit;
3025 		path->tclass_flowlabel =
3026 			cpu_to_be32((grh->traffic_class << 20) |
3027 				    (grh->flow_label));
3028 		memcpy(path->rgid, grh->dgid.raw, 16);
3029 	}
3030 
3031 	err = ib_rate_to_mlx5(dev, rdma_ah_get_static_rate(ah));
3032 	if (err < 0)
3033 		return err;
3034 	path->static_rate = err;
3035 	path->port = port;
3036 
3037 	if (attr_mask & IB_QP_TIMEOUT)
3038 		path->ackto_lt = (alt ? attr->alt_timeout : attr->timeout) << 3;
3039 
3040 	if ((qp->ibqp.qp_type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
3041 		return modify_raw_packet_eth_prio(dev->mdev,
3042 						  &qp->raw_packet_qp.sq,
3043 						  sl & 0xf, qp->ibqp.pd);
3044 
3045 	return 0;
3046 }
3047 
3048 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
3049 	[MLX5_QP_STATE_INIT] = {
3050 		[MLX5_QP_STATE_INIT] = {
3051 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3052 					  MLX5_QP_OPTPAR_RAE		|
3053 					  MLX5_QP_OPTPAR_RWE		|
3054 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3055 					  MLX5_QP_OPTPAR_PRI_PORT,
3056 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3057 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3058 					  MLX5_QP_OPTPAR_PRI_PORT,
3059 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
3060 					  MLX5_QP_OPTPAR_Q_KEY		|
3061 					  MLX5_QP_OPTPAR_PRI_PORT,
3062 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_RRE		|
3063 					  MLX5_QP_OPTPAR_RAE		|
3064 					  MLX5_QP_OPTPAR_RWE		|
3065 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3066 					  MLX5_QP_OPTPAR_PRI_PORT,
3067 		},
3068 		[MLX5_QP_STATE_RTR] = {
3069 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
3070 					  MLX5_QP_OPTPAR_RRE            |
3071 					  MLX5_QP_OPTPAR_RAE            |
3072 					  MLX5_QP_OPTPAR_RWE            |
3073 					  MLX5_QP_OPTPAR_PKEY_INDEX,
3074 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
3075 					  MLX5_QP_OPTPAR_RWE            |
3076 					  MLX5_QP_OPTPAR_PKEY_INDEX,
3077 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
3078 					  MLX5_QP_OPTPAR_Q_KEY,
3079 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
3080 					   MLX5_QP_OPTPAR_Q_KEY,
3081 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
3082 					  MLX5_QP_OPTPAR_RRE            |
3083 					  MLX5_QP_OPTPAR_RAE            |
3084 					  MLX5_QP_OPTPAR_RWE            |
3085 					  MLX5_QP_OPTPAR_PKEY_INDEX,
3086 		},
3087 	},
3088 	[MLX5_QP_STATE_RTR] = {
3089 		[MLX5_QP_STATE_RTS] = {
3090 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3091 					  MLX5_QP_OPTPAR_RRE		|
3092 					  MLX5_QP_OPTPAR_RAE		|
3093 					  MLX5_QP_OPTPAR_RWE		|
3094 					  MLX5_QP_OPTPAR_PM_STATE	|
3095 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
3096 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3097 					  MLX5_QP_OPTPAR_RWE		|
3098 					  MLX5_QP_OPTPAR_PM_STATE,
3099 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
3100 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3101 					  MLX5_QP_OPTPAR_RRE		|
3102 					  MLX5_QP_OPTPAR_RAE		|
3103 					  MLX5_QP_OPTPAR_RWE		|
3104 					  MLX5_QP_OPTPAR_PM_STATE	|
3105 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
3106 		},
3107 	},
3108 	[MLX5_QP_STATE_RTS] = {
3109 		[MLX5_QP_STATE_RTS] = {
3110 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3111 					  MLX5_QP_OPTPAR_RAE		|
3112 					  MLX5_QP_OPTPAR_RWE		|
3113 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3114 					  MLX5_QP_OPTPAR_PM_STATE	|
3115 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3116 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3117 					  MLX5_QP_OPTPAR_PM_STATE	|
3118 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3119 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
3120 					  MLX5_QP_OPTPAR_SRQN		|
3121 					  MLX5_QP_OPTPAR_CQN_RCV,
3122 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_RRE		|
3123 					  MLX5_QP_OPTPAR_RAE		|
3124 					  MLX5_QP_OPTPAR_RWE		|
3125 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3126 					  MLX5_QP_OPTPAR_PM_STATE	|
3127 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3128 		},
3129 	},
3130 	[MLX5_QP_STATE_SQER] = {
3131 		[MLX5_QP_STATE_RTS] = {
3132 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
3133 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
3134 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
3135 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3136 					   MLX5_QP_OPTPAR_RWE		|
3137 					   MLX5_QP_OPTPAR_RAE		|
3138 					   MLX5_QP_OPTPAR_RRE,
3139 			[MLX5_QP_ST_XRC]  = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3140 					   MLX5_QP_OPTPAR_RWE		|
3141 					   MLX5_QP_OPTPAR_RAE		|
3142 					   MLX5_QP_OPTPAR_RRE,
3143 		},
3144 	},
3145 };
3146 
3147 static int ib_nr_to_mlx5_nr(int ib_mask)
3148 {
3149 	switch (ib_mask) {
3150 	case IB_QP_STATE:
3151 		return 0;
3152 	case IB_QP_CUR_STATE:
3153 		return 0;
3154 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
3155 		return 0;
3156 	case IB_QP_ACCESS_FLAGS:
3157 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
3158 			MLX5_QP_OPTPAR_RAE;
3159 	case IB_QP_PKEY_INDEX:
3160 		return MLX5_QP_OPTPAR_PKEY_INDEX;
3161 	case IB_QP_PORT:
3162 		return MLX5_QP_OPTPAR_PRI_PORT;
3163 	case IB_QP_QKEY:
3164 		return MLX5_QP_OPTPAR_Q_KEY;
3165 	case IB_QP_AV:
3166 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
3167 			MLX5_QP_OPTPAR_PRI_PORT;
3168 	case IB_QP_PATH_MTU:
3169 		return 0;
3170 	case IB_QP_TIMEOUT:
3171 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
3172 	case IB_QP_RETRY_CNT:
3173 		return MLX5_QP_OPTPAR_RETRY_COUNT;
3174 	case IB_QP_RNR_RETRY:
3175 		return MLX5_QP_OPTPAR_RNR_RETRY;
3176 	case IB_QP_RQ_PSN:
3177 		return 0;
3178 	case IB_QP_MAX_QP_RD_ATOMIC:
3179 		return MLX5_QP_OPTPAR_SRA_MAX;
3180 	case IB_QP_ALT_PATH:
3181 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
3182 	case IB_QP_MIN_RNR_TIMER:
3183 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
3184 	case IB_QP_SQ_PSN:
3185 		return 0;
3186 	case IB_QP_MAX_DEST_RD_ATOMIC:
3187 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
3188 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
3189 	case IB_QP_PATH_MIG_STATE:
3190 		return MLX5_QP_OPTPAR_PM_STATE;
3191 	case IB_QP_CAP:
3192 		return 0;
3193 	case IB_QP_DEST_QPN:
3194 		return 0;
3195 	}
3196 	return 0;
3197 }
3198 
3199 static int ib_mask_to_mlx5_opt(int ib_mask)
3200 {
3201 	int result = 0;
3202 	int i;
3203 
3204 	for (i = 0; i < 8 * sizeof(int); i++) {
3205 		if ((1 << i) & ib_mask)
3206 			result |= ib_nr_to_mlx5_nr(1 << i);
3207 	}
3208 
3209 	return result;
3210 }
3211 
3212 static int modify_raw_packet_qp_rq(
3213 	struct mlx5_ib_dev *dev, struct mlx5_ib_rq *rq, int new_state,
3214 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3215 {
3216 	void *in;
3217 	void *rqc;
3218 	int inlen;
3219 	int err;
3220 
3221 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
3222 	in = kvzalloc(inlen, GFP_KERNEL);
3223 	if (!in)
3224 		return -ENOMEM;
3225 
3226 	MLX5_SET(modify_rq_in, in, rq_state, rq->state);
3227 	MLX5_SET(modify_rq_in, in, uid, to_mpd(pd)->uid);
3228 
3229 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
3230 	MLX5_SET(rqc, rqc, state, new_state);
3231 
3232 	if (raw_qp_param->set_mask & MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID) {
3233 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
3234 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
3235 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
3236 			MLX5_SET(rqc, rqc, counter_set_id, raw_qp_param->rq_q_ctr_id);
3237 		} else
3238 			dev_info_once(
3239 				&dev->ib_dev.dev,
3240 				"RAW PACKET QP counters are not supported on current FW\n");
3241 	}
3242 
3243 	err = mlx5_core_modify_rq(dev->mdev, rq->base.mqp.qpn, in, inlen);
3244 	if (err)
3245 		goto out;
3246 
3247 	rq->state = new_state;
3248 
3249 out:
3250 	kvfree(in);
3251 	return err;
3252 }
3253 
3254 static int modify_raw_packet_qp_sq(
3255 	struct mlx5_core_dev *dev, struct mlx5_ib_sq *sq, int new_state,
3256 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3257 {
3258 	struct mlx5_ib_qp *ibqp = sq->base.container_mibqp;
3259 	struct mlx5_rate_limit old_rl = ibqp->rl;
3260 	struct mlx5_rate_limit new_rl = old_rl;
3261 	bool new_rate_added = false;
3262 	u16 rl_index = 0;
3263 	void *in;
3264 	void *sqc;
3265 	int inlen;
3266 	int err;
3267 
3268 	inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
3269 	in = kvzalloc(inlen, GFP_KERNEL);
3270 	if (!in)
3271 		return -ENOMEM;
3272 
3273 	MLX5_SET(modify_sq_in, in, uid, to_mpd(pd)->uid);
3274 	MLX5_SET(modify_sq_in, in, sq_state, sq->state);
3275 
3276 	sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
3277 	MLX5_SET(sqc, sqc, state, new_state);
3278 
3279 	if (raw_qp_param->set_mask & MLX5_RAW_QP_RATE_LIMIT) {
3280 		if (new_state != MLX5_SQC_STATE_RDY)
3281 			pr_warn("%s: Rate limit can only be changed when SQ is moving to RDY\n",
3282 				__func__);
3283 		else
3284 			new_rl = raw_qp_param->rl;
3285 	}
3286 
3287 	if (!mlx5_rl_are_equal(&old_rl, &new_rl)) {
3288 		if (new_rl.rate) {
3289 			err = mlx5_rl_add_rate(dev, &rl_index, &new_rl);
3290 			if (err) {
3291 				pr_err("Failed configuring rate limit(err %d): \
3292 				       rate %u, max_burst_sz %u, typical_pkt_sz %u\n",
3293 				       err, new_rl.rate, new_rl.max_burst_sz,
3294 				       new_rl.typical_pkt_sz);
3295 
3296 				goto out;
3297 			}
3298 			new_rate_added = true;
3299 		}
3300 
3301 		MLX5_SET64(modify_sq_in, in, modify_bitmask, 1);
3302 		/* index 0 means no limit */
3303 		MLX5_SET(sqc, sqc, packet_pacing_rate_limit_index, rl_index);
3304 	}
3305 
3306 	err = mlx5_core_modify_sq(dev, sq->base.mqp.qpn, in, inlen);
3307 	if (err) {
3308 		/* Remove new rate from table if failed */
3309 		if (new_rate_added)
3310 			mlx5_rl_remove_rate(dev, &new_rl);
3311 		goto out;
3312 	}
3313 
3314 	/* Only remove the old rate after new rate was set */
3315 	if ((old_rl.rate && !mlx5_rl_are_equal(&old_rl, &new_rl)) ||
3316 	    (new_state != MLX5_SQC_STATE_RDY)) {
3317 		mlx5_rl_remove_rate(dev, &old_rl);
3318 		if (new_state != MLX5_SQC_STATE_RDY)
3319 			memset(&new_rl, 0, sizeof(new_rl));
3320 	}
3321 
3322 	ibqp->rl = new_rl;
3323 	sq->state = new_state;
3324 
3325 out:
3326 	kvfree(in);
3327 	return err;
3328 }
3329 
3330 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3331 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
3332 				u8 tx_affinity)
3333 {
3334 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
3335 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
3336 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
3337 	int modify_rq = !!qp->rq.wqe_cnt;
3338 	int modify_sq = !!qp->sq.wqe_cnt;
3339 	int rq_state;
3340 	int sq_state;
3341 	int err;
3342 
3343 	switch (raw_qp_param->operation) {
3344 	case MLX5_CMD_OP_RST2INIT_QP:
3345 		rq_state = MLX5_RQC_STATE_RDY;
3346 		sq_state = MLX5_SQC_STATE_RDY;
3347 		break;
3348 	case MLX5_CMD_OP_2ERR_QP:
3349 		rq_state = MLX5_RQC_STATE_ERR;
3350 		sq_state = MLX5_SQC_STATE_ERR;
3351 		break;
3352 	case MLX5_CMD_OP_2RST_QP:
3353 		rq_state = MLX5_RQC_STATE_RST;
3354 		sq_state = MLX5_SQC_STATE_RST;
3355 		break;
3356 	case MLX5_CMD_OP_RTR2RTS_QP:
3357 	case MLX5_CMD_OP_RTS2RTS_QP:
3358 		if (raw_qp_param->set_mask ==
3359 		    MLX5_RAW_QP_RATE_LIMIT) {
3360 			modify_rq = 0;
3361 			sq_state = sq->state;
3362 		} else {
3363 			return raw_qp_param->set_mask ? -EINVAL : 0;
3364 		}
3365 		break;
3366 	case MLX5_CMD_OP_INIT2INIT_QP:
3367 	case MLX5_CMD_OP_INIT2RTR_QP:
3368 		if (raw_qp_param->set_mask)
3369 			return -EINVAL;
3370 		else
3371 			return 0;
3372 	default:
3373 		WARN_ON(1);
3374 		return -EINVAL;
3375 	}
3376 
3377 	if (modify_rq) {
3378 		err =  modify_raw_packet_qp_rq(dev, rq, rq_state, raw_qp_param,
3379 					       qp->ibqp.pd);
3380 		if (err)
3381 			return err;
3382 	}
3383 
3384 	if (modify_sq) {
3385 		struct mlx5_flow_handle *flow_rule;
3386 
3387 		if (tx_affinity) {
3388 			err = modify_raw_packet_tx_affinity(dev->mdev, sq,
3389 							    tx_affinity,
3390 							    qp->ibqp.pd);
3391 			if (err)
3392 				return err;
3393 		}
3394 
3395 		flow_rule = create_flow_rule_vport_sq(dev, sq,
3396 						      raw_qp_param->port);
3397 		if (IS_ERR(flow_rule))
3398 			return PTR_ERR(flow_rule);
3399 
3400 		err = modify_raw_packet_qp_sq(dev->mdev, sq, sq_state,
3401 					      raw_qp_param, qp->ibqp.pd);
3402 		if (err) {
3403 			if (flow_rule)
3404 				mlx5_del_flow_rules(flow_rule);
3405 			return err;
3406 		}
3407 
3408 		if (flow_rule) {
3409 			destroy_flow_rule_vport_sq(sq);
3410 			sq->flow_rule = flow_rule;
3411 		}
3412 
3413 		return err;
3414 	}
3415 
3416 	return 0;
3417 }
3418 
3419 static unsigned int get_tx_affinity(struct mlx5_ib_dev *dev,
3420 				    struct mlx5_ib_pd *pd,
3421 				    struct mlx5_ib_qp_base *qp_base,
3422 				    u8 port_num, struct ib_udata *udata)
3423 {
3424 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3425 		udata, struct mlx5_ib_ucontext, ibucontext);
3426 	unsigned int tx_port_affinity;
3427 
3428 	if (ucontext) {
3429 		tx_port_affinity = (unsigned int)atomic_add_return(
3430 					   1, &ucontext->tx_port_affinity) %
3431 					   MLX5_MAX_PORTS +
3432 				   1;
3433 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x ucontext %p\n",
3434 				tx_port_affinity, qp_base->mqp.qpn, ucontext);
3435 	} else {
3436 		tx_port_affinity =
3437 			(unsigned int)atomic_add_return(
3438 				1, &dev->port[port_num].roce.tx_port_affinity) %
3439 				MLX5_MAX_PORTS +
3440 			1;
3441 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x\n",
3442 				tx_port_affinity, qp_base->mqp.qpn);
3443 	}
3444 
3445 	return tx_port_affinity;
3446 }
3447 
3448 static int __mlx5_ib_qp_set_counter(struct ib_qp *qp,
3449 				    struct rdma_counter *counter)
3450 {
3451 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3452 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3453 	struct mlx5_qp_context context = {};
3454 	struct mlx5_ib_qp_base *base;
3455 	u32 set_id;
3456 
3457 	if (counter)
3458 		set_id = counter->id;
3459 	else
3460 		set_id = mlx5_ib_get_counters_id(dev, mqp->port - 1);
3461 
3462 	base = &mqp->trans_qp.base;
3463 	context.qp_counter_set_usr_page &= cpu_to_be32(0xffffff);
3464 	context.qp_counter_set_usr_page |= cpu_to_be32(set_id << 24);
3465 	return mlx5_core_qp_modify(dev->mdev,
3466 				   MLX5_CMD_OP_RTS2RTS_QP,
3467 				   MLX5_QP_OPTPAR_COUNTER_SET_ID,
3468 				   &context, &base->mqp);
3469 }
3470 
3471 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
3472 			       const struct ib_qp_attr *attr, int attr_mask,
3473 			       enum ib_qp_state cur_state,
3474 			       enum ib_qp_state new_state,
3475 			       const struct mlx5_ib_modify_qp *ucmd,
3476 			       struct ib_udata *udata)
3477 {
3478 	static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
3479 		[MLX5_QP_STATE_RST] = {
3480 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3481 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3482 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_RST2INIT_QP,
3483 		},
3484 		[MLX5_QP_STATE_INIT]  = {
3485 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3486 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3487 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_INIT2INIT_QP,
3488 			[MLX5_QP_STATE_RTR]	= MLX5_CMD_OP_INIT2RTR_QP,
3489 		},
3490 		[MLX5_QP_STATE_RTR]   = {
3491 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3492 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3493 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTR2RTS_QP,
3494 		},
3495 		[MLX5_QP_STATE_RTS]   = {
3496 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3497 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3498 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTS2RTS_QP,
3499 		},
3500 		[MLX5_QP_STATE_SQD] = {
3501 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3502 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3503 		},
3504 		[MLX5_QP_STATE_SQER] = {
3505 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3506 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3507 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQERR2RTS_QP,
3508 		},
3509 		[MLX5_QP_STATE_ERR] = {
3510 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3511 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3512 		}
3513 	};
3514 
3515 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3516 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3517 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
3518 	struct mlx5_ib_cq *send_cq, *recv_cq;
3519 	struct mlx5_qp_context *context;
3520 	struct mlx5_ib_pd *pd;
3521 	enum mlx5_qp_state mlx5_cur, mlx5_new;
3522 	enum mlx5_qp_optpar optpar;
3523 	u32 set_id = 0;
3524 	int mlx5_st;
3525 	int err;
3526 	u16 op;
3527 	u8 tx_affinity = 0;
3528 
3529 	mlx5_st = to_mlx5_st(ibqp->qp_type == IB_QPT_DRIVER ?
3530 			     qp->qp_sub_type : ibqp->qp_type);
3531 	if (mlx5_st < 0)
3532 		return -EINVAL;
3533 
3534 	context = kzalloc(sizeof(*context), GFP_KERNEL);
3535 	if (!context)
3536 		return -ENOMEM;
3537 
3538 	pd = get_pd(qp);
3539 	context->flags = cpu_to_be32(mlx5_st << 16);
3540 
3541 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
3542 		context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
3543 	} else {
3544 		switch (attr->path_mig_state) {
3545 		case IB_MIG_MIGRATED:
3546 			context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
3547 			break;
3548 		case IB_MIG_REARM:
3549 			context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
3550 			break;
3551 		case IB_MIG_ARMED:
3552 			context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
3553 			break;
3554 		}
3555 	}
3556 
3557 	if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
3558 		if ((ibqp->qp_type == IB_QPT_RC) ||
3559 		    (ibqp->qp_type == IB_QPT_UD &&
3560 		     !(qp->flags & MLX5_IB_QP_SQPN_QP1)) ||
3561 		    (ibqp->qp_type == IB_QPT_UC) ||
3562 		    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
3563 		    (ibqp->qp_type == IB_QPT_XRC_INI) ||
3564 		    (ibqp->qp_type == IB_QPT_XRC_TGT)) {
3565 			if (dev->lag_active) {
3566 				u8 p = mlx5_core_native_port_num(dev->mdev) - 1;
3567 				tx_affinity = get_tx_affinity(dev, pd, base, p,
3568 							      udata);
3569 				context->flags |= cpu_to_be32(tx_affinity << 24);
3570 			}
3571 		}
3572 	}
3573 
3574 	if (is_sqp(ibqp->qp_type)) {
3575 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
3576 	} else if ((ibqp->qp_type == IB_QPT_UD &&
3577 		    !(qp->flags & MLX5_IB_QP_UNDERLAY)) ||
3578 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
3579 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
3580 	} else if (attr_mask & IB_QP_PATH_MTU) {
3581 		if (attr->path_mtu < IB_MTU_256 ||
3582 		    attr->path_mtu > IB_MTU_4096) {
3583 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
3584 			err = -EINVAL;
3585 			goto out;
3586 		}
3587 		context->mtu_msgmax = (attr->path_mtu << 5) |
3588 				      (u8)MLX5_CAP_GEN(dev->mdev, log_max_msg);
3589 	}
3590 
3591 	if (attr_mask & IB_QP_DEST_QPN)
3592 		context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
3593 
3594 	if (attr_mask & IB_QP_PKEY_INDEX)
3595 		context->pri_path.pkey_index = cpu_to_be16(attr->pkey_index);
3596 
3597 	/* todo implement counter_index functionality */
3598 
3599 	if (is_sqp(ibqp->qp_type))
3600 		context->pri_path.port = qp->port;
3601 
3602 	if (attr_mask & IB_QP_PORT)
3603 		context->pri_path.port = attr->port_num;
3604 
3605 	if (attr_mask & IB_QP_AV) {
3606 		err = mlx5_set_path(dev, qp, &attr->ah_attr, &context->pri_path,
3607 				    attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
3608 				    attr_mask, 0, attr, false);
3609 		if (err)
3610 			goto out;
3611 	}
3612 
3613 	if (attr_mask & IB_QP_TIMEOUT)
3614 		context->pri_path.ackto_lt |= attr->timeout << 3;
3615 
3616 	if (attr_mask & IB_QP_ALT_PATH) {
3617 		err = mlx5_set_path(dev, qp, &attr->alt_ah_attr,
3618 				    &context->alt_path,
3619 				    attr->alt_port_num,
3620 				    attr_mask | IB_QP_PKEY_INDEX | IB_QP_TIMEOUT,
3621 				    0, attr, true);
3622 		if (err)
3623 			goto out;
3624 	}
3625 
3626 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
3627 		&send_cq, &recv_cq);
3628 
3629 	context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
3630 	context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
3631 	context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
3632 	context->params1  = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
3633 
3634 	if (attr_mask & IB_QP_RNR_RETRY)
3635 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
3636 
3637 	if (attr_mask & IB_QP_RETRY_CNT)
3638 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
3639 
3640 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
3641 		if (attr->max_rd_atomic)
3642 			context->params1 |=
3643 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
3644 	}
3645 
3646 	if (attr_mask & IB_QP_SQ_PSN)
3647 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
3648 
3649 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
3650 		if (attr->max_dest_rd_atomic)
3651 			context->params2 |=
3652 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
3653 	}
3654 
3655 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
3656 		__be32 access_flags;
3657 
3658 		err = to_mlx5_access_flags(qp, attr, attr_mask, &access_flags);
3659 		if (err)
3660 			goto out;
3661 
3662 		context->params2 |= access_flags;
3663 	}
3664 
3665 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
3666 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
3667 
3668 	if (attr_mask & IB_QP_RQ_PSN)
3669 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
3670 
3671 	if (attr_mask & IB_QP_QKEY)
3672 		context->qkey = cpu_to_be32(attr->qkey);
3673 
3674 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
3675 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
3676 
3677 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
3678 		u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
3679 			       qp->port) - 1;
3680 
3681 		/* Underlay port should be used - index 0 function per port */
3682 		if (qp->flags & MLX5_IB_QP_UNDERLAY)
3683 			port_num = 0;
3684 
3685 		if (ibqp->counter)
3686 			set_id = ibqp->counter->id;
3687 		else
3688 			set_id = mlx5_ib_get_counters_id(dev, port_num);
3689 		context->qp_counter_set_usr_page |=
3690 			cpu_to_be32(set_id << 24);
3691 	}
3692 
3693 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
3694 		context->sq_crq_size |= cpu_to_be16(1 << 4);
3695 
3696 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
3697 		context->deth_sqpn = cpu_to_be32(1);
3698 
3699 	mlx5_cur = to_mlx5_state(cur_state);
3700 	mlx5_new = to_mlx5_state(new_state);
3701 
3702 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
3703 	    !optab[mlx5_cur][mlx5_new]) {
3704 		err = -EINVAL;
3705 		goto out;
3706 	}
3707 
3708 	op = optab[mlx5_cur][mlx5_new];
3709 	optpar = ib_mask_to_mlx5_opt(attr_mask);
3710 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
3711 
3712 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
3713 	    qp->flags & MLX5_IB_QP_UNDERLAY) {
3714 		struct mlx5_modify_raw_qp_param raw_qp_param = {};
3715 
3716 		raw_qp_param.operation = op;
3717 		if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
3718 			raw_qp_param.rq_q_ctr_id = set_id;
3719 			raw_qp_param.set_mask |= MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID;
3720 		}
3721 
3722 		if (attr_mask & IB_QP_PORT)
3723 			raw_qp_param.port = attr->port_num;
3724 
3725 		if (attr_mask & IB_QP_RATE_LIMIT) {
3726 			raw_qp_param.rl.rate = attr->rate_limit;
3727 
3728 			if (ucmd->burst_info.max_burst_sz) {
3729 				if (attr->rate_limit &&
3730 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_burst_bound)) {
3731 					raw_qp_param.rl.max_burst_sz =
3732 						ucmd->burst_info.max_burst_sz;
3733 				} else {
3734 					err = -EINVAL;
3735 					goto out;
3736 				}
3737 			}
3738 
3739 			if (ucmd->burst_info.typical_pkt_sz) {
3740 				if (attr->rate_limit &&
3741 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_typical_size)) {
3742 					raw_qp_param.rl.typical_pkt_sz =
3743 						ucmd->burst_info.typical_pkt_sz;
3744 				} else {
3745 					err = -EINVAL;
3746 					goto out;
3747 				}
3748 			}
3749 
3750 			raw_qp_param.set_mask |= MLX5_RAW_QP_RATE_LIMIT;
3751 		}
3752 
3753 		err = modify_raw_packet_qp(dev, qp, &raw_qp_param, tx_affinity);
3754 	} else {
3755 		err = mlx5_core_qp_modify(dev->mdev, op, optpar, context,
3756 					  &base->mqp);
3757 	}
3758 
3759 	if (err)
3760 		goto out;
3761 
3762 	qp->state = new_state;
3763 
3764 	if (attr_mask & IB_QP_ACCESS_FLAGS)
3765 		qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
3766 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
3767 		qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
3768 	if (attr_mask & IB_QP_PORT)
3769 		qp->port = attr->port_num;
3770 	if (attr_mask & IB_QP_ALT_PATH)
3771 		qp->trans_qp.alt_port = attr->alt_port_num;
3772 
3773 	/*
3774 	 * If we moved a kernel QP to RESET, clean up all old CQ
3775 	 * entries and reinitialize the QP.
3776 	 */
3777 	if (new_state == IB_QPS_RESET &&
3778 	    !ibqp->uobject && ibqp->qp_type != IB_QPT_XRC_TGT) {
3779 		mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
3780 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
3781 		if (send_cq != recv_cq)
3782 			mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
3783 
3784 		qp->rq.head = 0;
3785 		qp->rq.tail = 0;
3786 		qp->sq.head = 0;
3787 		qp->sq.tail = 0;
3788 		qp->sq.cur_post = 0;
3789 		if (qp->sq.wqe_cnt)
3790 			qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
3791 		qp->sq.last_poll = 0;
3792 		qp->db.db[MLX5_RCV_DBR] = 0;
3793 		qp->db.db[MLX5_SND_DBR] = 0;
3794 	}
3795 
3796 	if ((new_state == IB_QPS_RTS) && qp->counter_pending) {
3797 		err = __mlx5_ib_qp_set_counter(ibqp, ibqp->counter);
3798 		if (!err)
3799 			qp->counter_pending = 0;
3800 	}
3801 
3802 out:
3803 	kfree(context);
3804 	return err;
3805 }
3806 
3807 static inline bool is_valid_mask(int mask, int req, int opt)
3808 {
3809 	if ((mask & req) != req)
3810 		return false;
3811 
3812 	if (mask & ~(req | opt))
3813 		return false;
3814 
3815 	return true;
3816 }
3817 
3818 /* check valid transition for driver QP types
3819  * for now the only QP type that this function supports is DCI
3820  */
3821 static bool modify_dci_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state new_state,
3822 				enum ib_qp_attr_mask attr_mask)
3823 {
3824 	int req = IB_QP_STATE;
3825 	int opt = 0;
3826 
3827 	if (new_state == IB_QPS_RESET) {
3828 		return is_valid_mask(attr_mask, req, opt);
3829 	} else if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
3830 		req |= IB_QP_PKEY_INDEX | IB_QP_PORT;
3831 		return is_valid_mask(attr_mask, req, opt);
3832 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_INIT) {
3833 		opt = IB_QP_PKEY_INDEX | IB_QP_PORT;
3834 		return is_valid_mask(attr_mask, req, opt);
3835 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
3836 		req |= IB_QP_PATH_MTU;
3837 		opt = IB_QP_PKEY_INDEX | IB_QP_AV;
3838 		return is_valid_mask(attr_mask, req, opt);
3839 	} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
3840 		req |= IB_QP_TIMEOUT | IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
3841 		       IB_QP_MAX_QP_RD_ATOMIC | IB_QP_SQ_PSN;
3842 		opt = IB_QP_MIN_RNR_TIMER;
3843 		return is_valid_mask(attr_mask, req, opt);
3844 	} else if (cur_state == IB_QPS_RTS && new_state == IB_QPS_RTS) {
3845 		opt = IB_QP_MIN_RNR_TIMER;
3846 		return is_valid_mask(attr_mask, req, opt);
3847 	} else if (cur_state != IB_QPS_RESET && new_state == IB_QPS_ERR) {
3848 		return is_valid_mask(attr_mask, req, opt);
3849 	}
3850 	return false;
3851 }
3852 
3853 /* mlx5_ib_modify_dct: modify a DCT QP
3854  * valid transitions are:
3855  * RESET to INIT: must set access_flags, pkey_index and port
3856  * INIT  to RTR : must set min_rnr_timer, tclass, flow_label,
3857  *			   mtu, gid_index and hop_limit
3858  * Other transitions and attributes are illegal
3859  */
3860 static int mlx5_ib_modify_dct(struct ib_qp *ibqp, struct ib_qp_attr *attr,
3861 			      int attr_mask, struct ib_udata *udata)
3862 {
3863 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3864 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3865 	enum ib_qp_state cur_state, new_state;
3866 	int err = 0;
3867 	int required = IB_QP_STATE;
3868 	void *dctc;
3869 
3870 	if (!(attr_mask & IB_QP_STATE))
3871 		return -EINVAL;
3872 
3873 	cur_state = qp->state;
3874 	new_state = attr->qp_state;
3875 
3876 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
3877 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
3878 		u16 set_id;
3879 
3880 		required |= IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
3881 		if (!is_valid_mask(attr_mask, required, 0))
3882 			return -EINVAL;
3883 
3884 		if (attr->port_num == 0 ||
3885 		    attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports)) {
3886 			mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
3887 				    attr->port_num, dev->num_ports);
3888 			return -EINVAL;
3889 		}
3890 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
3891 			MLX5_SET(dctc, dctc, rre, 1);
3892 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
3893 			MLX5_SET(dctc, dctc, rwe, 1);
3894 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC) {
3895 			int atomic_mode;
3896 
3897 			atomic_mode = get_atomic_mode(dev, MLX5_IB_QPT_DCT);
3898 			if (atomic_mode < 0)
3899 				return -EOPNOTSUPP;
3900 
3901 			MLX5_SET(dctc, dctc, atomic_mode, atomic_mode);
3902 			MLX5_SET(dctc, dctc, rae, 1);
3903 		}
3904 		MLX5_SET(dctc, dctc, pkey_index, attr->pkey_index);
3905 		MLX5_SET(dctc, dctc, port, attr->port_num);
3906 
3907 		set_id = mlx5_ib_get_counters_id(dev, attr->port_num - 1);
3908 		MLX5_SET(dctc, dctc, counter_set_id, set_id);
3909 
3910 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
3911 		struct mlx5_ib_modify_qp_resp resp = {};
3912 		u32 out[MLX5_ST_SZ_DW(create_dct_out)] = {0};
3913 		u32 min_resp_len = offsetof(typeof(resp), dctn) +
3914 				   sizeof(resp.dctn);
3915 
3916 		if (udata->outlen < min_resp_len)
3917 			return -EINVAL;
3918 		resp.response_length = min_resp_len;
3919 
3920 		required |= IB_QP_MIN_RNR_TIMER | IB_QP_AV | IB_QP_PATH_MTU;
3921 		if (!is_valid_mask(attr_mask, required, 0))
3922 			return -EINVAL;
3923 		MLX5_SET(dctc, dctc, min_rnr_nak, attr->min_rnr_timer);
3924 		MLX5_SET(dctc, dctc, tclass, attr->ah_attr.grh.traffic_class);
3925 		MLX5_SET(dctc, dctc, flow_label, attr->ah_attr.grh.flow_label);
3926 		MLX5_SET(dctc, dctc, mtu, attr->path_mtu);
3927 		MLX5_SET(dctc, dctc, my_addr_index, attr->ah_attr.grh.sgid_index);
3928 		MLX5_SET(dctc, dctc, hop_limit, attr->ah_attr.grh.hop_limit);
3929 
3930 		err = mlx5_core_create_dct(dev->mdev, &qp->dct.mdct, qp->dct.in,
3931 					   MLX5_ST_SZ_BYTES(create_dct_in), out,
3932 					   sizeof(out));
3933 		if (err)
3934 			return err;
3935 		resp.dctn = qp->dct.mdct.mqp.qpn;
3936 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
3937 		if (err) {
3938 			mlx5_core_destroy_dct(dev->mdev, &qp->dct.mdct);
3939 			return err;
3940 		}
3941 	} else {
3942 		mlx5_ib_warn(dev, "Modify DCT: Invalid transition from %d to %d\n", cur_state, new_state);
3943 		return -EINVAL;
3944 	}
3945 	if (err)
3946 		qp->state = IB_QPS_ERR;
3947 	else
3948 		qp->state = new_state;
3949 	return err;
3950 }
3951 
3952 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
3953 		      int attr_mask, struct ib_udata *udata)
3954 {
3955 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3956 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3957 	struct mlx5_ib_modify_qp ucmd = {};
3958 	enum ib_qp_type qp_type;
3959 	enum ib_qp_state cur_state, new_state;
3960 	size_t required_cmd_sz;
3961 	int err = -EINVAL;
3962 	int port;
3963 
3964 	if (ibqp->rwq_ind_tbl)
3965 		return -ENOSYS;
3966 
3967 	if (udata && udata->inlen) {
3968 		required_cmd_sz = offsetof(typeof(ucmd), reserved) +
3969 			sizeof(ucmd.reserved);
3970 		if (udata->inlen < required_cmd_sz)
3971 			return -EINVAL;
3972 
3973 		if (udata->inlen > sizeof(ucmd) &&
3974 		    !ib_is_udata_cleared(udata, sizeof(ucmd),
3975 					 udata->inlen - sizeof(ucmd)))
3976 			return -EOPNOTSUPP;
3977 
3978 		if (ib_copy_from_udata(&ucmd, udata,
3979 				       min(udata->inlen, sizeof(ucmd))))
3980 			return -EFAULT;
3981 
3982 		if (ucmd.comp_mask ||
3983 		    memchr_inv(&ucmd.reserved, 0, sizeof(ucmd.reserved)) ||
3984 		    memchr_inv(&ucmd.burst_info.reserved, 0,
3985 			       sizeof(ucmd.burst_info.reserved)))
3986 			return -EOPNOTSUPP;
3987 	}
3988 
3989 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
3990 		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
3991 
3992 	if (ibqp->qp_type == IB_QPT_DRIVER)
3993 		qp_type = qp->qp_sub_type;
3994 	else
3995 		qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ?
3996 			IB_QPT_GSI : ibqp->qp_type;
3997 
3998 	if (qp_type == MLX5_IB_QPT_DCT)
3999 		return mlx5_ib_modify_dct(ibqp, attr, attr_mask, udata);
4000 
4001 	mutex_lock(&qp->mutex);
4002 
4003 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
4004 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
4005 
4006 	if (!(cur_state == new_state && cur_state == IB_QPS_RESET)) {
4007 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
4008 	}
4009 
4010 	if (qp->flags & MLX5_IB_QP_UNDERLAY) {
4011 		if (attr_mask & ~(IB_QP_STATE | IB_QP_CUR_STATE)) {
4012 			mlx5_ib_dbg(dev, "invalid attr_mask 0x%x when underlay QP is used\n",
4013 				    attr_mask);
4014 			goto out;
4015 		}
4016 	} else if (qp_type != MLX5_IB_QPT_REG_UMR &&
4017 		   qp_type != MLX5_IB_QPT_DCI &&
4018 		   !ib_modify_qp_is_ok(cur_state, new_state, qp_type,
4019 				       attr_mask)) {
4020 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
4021 			    cur_state, new_state, ibqp->qp_type, attr_mask);
4022 		goto out;
4023 	} else if (qp_type == MLX5_IB_QPT_DCI &&
4024 		   !modify_dci_qp_is_ok(cur_state, new_state, attr_mask)) {
4025 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
4026 			    cur_state, new_state, qp_type, attr_mask);
4027 		goto out;
4028 	}
4029 
4030 	if ((attr_mask & IB_QP_PORT) &&
4031 	    (attr->port_num == 0 ||
4032 	     attr->port_num > dev->num_ports)) {
4033 		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
4034 			    attr->port_num, dev->num_ports);
4035 		goto out;
4036 	}
4037 
4038 	if (attr_mask & IB_QP_PKEY_INDEX) {
4039 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
4040 		if (attr->pkey_index >=
4041 		    dev->mdev->port_caps[port - 1].pkey_table_len) {
4042 			mlx5_ib_dbg(dev, "invalid pkey index %d\n",
4043 				    attr->pkey_index);
4044 			goto out;
4045 		}
4046 	}
4047 
4048 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
4049 	    attr->max_rd_atomic >
4050 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
4051 		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
4052 			    attr->max_rd_atomic);
4053 		goto out;
4054 	}
4055 
4056 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
4057 	    attr->max_dest_rd_atomic >
4058 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
4059 		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
4060 			    attr->max_dest_rd_atomic);
4061 		goto out;
4062 	}
4063 
4064 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
4065 		err = 0;
4066 		goto out;
4067 	}
4068 
4069 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state,
4070 				  new_state, &ucmd, udata);
4071 
4072 out:
4073 	mutex_unlock(&qp->mutex);
4074 	return err;
4075 }
4076 
4077 static void _handle_post_send_edge(struct mlx5_ib_wq *sq, void **seg,
4078 				   u32 wqe_sz, void **cur_edge)
4079 {
4080 	u32 idx;
4081 
4082 	idx = (sq->cur_post + (wqe_sz >> 2)) & (sq->wqe_cnt - 1);
4083 	*cur_edge = get_sq_edge(sq, idx);
4084 
4085 	*seg = mlx5_frag_buf_get_wqe(&sq->fbc, idx);
4086 }
4087 
4088 /* handle_post_send_edge - Check if we get to SQ edge. If yes, update to the
4089  * next nearby edge and get new address translation for current WQE position.
4090  * @sq - SQ buffer.
4091  * @seg: Current WQE position (16B aligned).
4092  * @wqe_sz: Total current WQE size [16B].
4093  * @cur_edge: Updated current edge.
4094  */
4095 static inline void handle_post_send_edge(struct mlx5_ib_wq *sq, void **seg,
4096 					 u32 wqe_sz, void **cur_edge)
4097 {
4098 	if (likely(*seg != *cur_edge))
4099 		return;
4100 
4101 	_handle_post_send_edge(sq, seg, wqe_sz, cur_edge);
4102 }
4103 
4104 /* memcpy_send_wqe - copy data from src to WQE and update the relevant WQ's
4105  * pointers. At the end @seg is aligned to 16B regardless the copied size.
4106  * @sq - SQ buffer.
4107  * @cur_edge: Updated current edge.
4108  * @seg: Current WQE position (16B aligned).
4109  * @wqe_sz: Total current WQE size [16B].
4110  * @src: Pointer to copy from.
4111  * @n: Number of bytes to copy.
4112  */
4113 static inline void memcpy_send_wqe(struct mlx5_ib_wq *sq, void **cur_edge,
4114 				   void **seg, u32 *wqe_sz, const void *src,
4115 				   size_t n)
4116 {
4117 	while (likely(n)) {
4118 		size_t leftlen = *cur_edge - *seg;
4119 		size_t copysz = min_t(size_t, leftlen, n);
4120 		size_t stride;
4121 
4122 		memcpy(*seg, src, copysz);
4123 
4124 		n -= copysz;
4125 		src += copysz;
4126 		stride = !n ? ALIGN(copysz, 16) : copysz;
4127 		*seg += stride;
4128 		*wqe_sz += stride >> 4;
4129 		handle_post_send_edge(sq, seg, *wqe_sz, cur_edge);
4130 	}
4131 }
4132 
4133 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
4134 {
4135 	struct mlx5_ib_cq *cq;
4136 	unsigned cur;
4137 
4138 	cur = wq->head - wq->tail;
4139 	if (likely(cur + nreq < wq->max_post))
4140 		return 0;
4141 
4142 	cq = to_mcq(ib_cq);
4143 	spin_lock(&cq->lock);
4144 	cur = wq->head - wq->tail;
4145 	spin_unlock(&cq->lock);
4146 
4147 	return cur + nreq >= wq->max_post;
4148 }
4149 
4150 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
4151 					  u64 remote_addr, u32 rkey)
4152 {
4153 	rseg->raddr    = cpu_to_be64(remote_addr);
4154 	rseg->rkey     = cpu_to_be32(rkey);
4155 	rseg->reserved = 0;
4156 }
4157 
4158 static void set_eth_seg(const struct ib_send_wr *wr, struct mlx5_ib_qp *qp,
4159 			void **seg, int *size, void **cur_edge)
4160 {
4161 	struct mlx5_wqe_eth_seg *eseg = *seg;
4162 
4163 	memset(eseg, 0, sizeof(struct mlx5_wqe_eth_seg));
4164 
4165 	if (wr->send_flags & IB_SEND_IP_CSUM)
4166 		eseg->cs_flags = MLX5_ETH_WQE_L3_CSUM |
4167 				 MLX5_ETH_WQE_L4_CSUM;
4168 
4169 	if (wr->opcode == IB_WR_LSO) {
4170 		struct ib_ud_wr *ud_wr = container_of(wr, struct ib_ud_wr, wr);
4171 		size_t left, copysz;
4172 		void *pdata = ud_wr->header;
4173 		size_t stride;
4174 
4175 		left = ud_wr->hlen;
4176 		eseg->mss = cpu_to_be16(ud_wr->mss);
4177 		eseg->inline_hdr.sz = cpu_to_be16(left);
4178 
4179 		/* memcpy_send_wqe should get a 16B align address. Hence, we
4180 		 * first copy up to the current edge and then, if needed,
4181 		 * fall-through to memcpy_send_wqe.
4182 		 */
4183 		copysz = min_t(u64, *cur_edge - (void *)eseg->inline_hdr.start,
4184 			       left);
4185 		memcpy(eseg->inline_hdr.start, pdata, copysz);
4186 		stride = ALIGN(sizeof(struct mlx5_wqe_eth_seg) -
4187 			       sizeof(eseg->inline_hdr.start) + copysz, 16);
4188 		*size += stride / 16;
4189 		*seg += stride;
4190 
4191 		if (copysz < left) {
4192 			handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4193 			left -= copysz;
4194 			pdata += copysz;
4195 			memcpy_send_wqe(&qp->sq, cur_edge, seg, size, pdata,
4196 					left);
4197 		}
4198 
4199 		return;
4200 	}
4201 
4202 	*seg += sizeof(struct mlx5_wqe_eth_seg);
4203 	*size += sizeof(struct mlx5_wqe_eth_seg) / 16;
4204 }
4205 
4206 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
4207 			     const struct ib_send_wr *wr)
4208 {
4209 	memcpy(&dseg->av, &to_mah(ud_wr(wr)->ah)->av, sizeof(struct mlx5_av));
4210 	dseg->av.dqp_dct = cpu_to_be32(ud_wr(wr)->remote_qpn | MLX5_EXTENDED_UD_AV);
4211 	dseg->av.key.qkey.qkey = cpu_to_be32(ud_wr(wr)->remote_qkey);
4212 }
4213 
4214 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
4215 {
4216 	dseg->byte_count = cpu_to_be32(sg->length);
4217 	dseg->lkey       = cpu_to_be32(sg->lkey);
4218 	dseg->addr       = cpu_to_be64(sg->addr);
4219 }
4220 
4221 static u64 get_xlt_octo(u64 bytes)
4222 {
4223 	return ALIGN(bytes, MLX5_IB_UMR_XLT_ALIGNMENT) /
4224 	       MLX5_IB_UMR_OCTOWORD;
4225 }
4226 
4227 static __be64 frwr_mkey_mask(bool atomic)
4228 {
4229 	u64 result;
4230 
4231 	result = MLX5_MKEY_MASK_LEN		|
4232 		MLX5_MKEY_MASK_PAGE_SIZE	|
4233 		MLX5_MKEY_MASK_START_ADDR	|
4234 		MLX5_MKEY_MASK_EN_RINVAL	|
4235 		MLX5_MKEY_MASK_KEY		|
4236 		MLX5_MKEY_MASK_LR		|
4237 		MLX5_MKEY_MASK_LW		|
4238 		MLX5_MKEY_MASK_RR		|
4239 		MLX5_MKEY_MASK_RW		|
4240 		MLX5_MKEY_MASK_SMALL_FENCE	|
4241 		MLX5_MKEY_MASK_FREE;
4242 
4243 	if (atomic)
4244 		result |= MLX5_MKEY_MASK_A;
4245 
4246 	return cpu_to_be64(result);
4247 }
4248 
4249 static __be64 sig_mkey_mask(void)
4250 {
4251 	u64 result;
4252 
4253 	result = MLX5_MKEY_MASK_LEN		|
4254 		MLX5_MKEY_MASK_PAGE_SIZE	|
4255 		MLX5_MKEY_MASK_START_ADDR	|
4256 		MLX5_MKEY_MASK_EN_SIGERR	|
4257 		MLX5_MKEY_MASK_EN_RINVAL	|
4258 		MLX5_MKEY_MASK_KEY		|
4259 		MLX5_MKEY_MASK_LR		|
4260 		MLX5_MKEY_MASK_LW		|
4261 		MLX5_MKEY_MASK_RR		|
4262 		MLX5_MKEY_MASK_RW		|
4263 		MLX5_MKEY_MASK_SMALL_FENCE	|
4264 		MLX5_MKEY_MASK_FREE		|
4265 		MLX5_MKEY_MASK_BSF_EN;
4266 
4267 	return cpu_to_be64(result);
4268 }
4269 
4270 static void set_reg_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr,
4271 			    struct mlx5_ib_mr *mr, u8 flags, bool atomic)
4272 {
4273 	int size = (mr->ndescs + mr->meta_ndescs) * mr->desc_size;
4274 
4275 	memset(umr, 0, sizeof(*umr));
4276 
4277 	umr->flags = flags;
4278 	umr->xlt_octowords = cpu_to_be16(get_xlt_octo(size));
4279 	umr->mkey_mask = frwr_mkey_mask(atomic);
4280 }
4281 
4282 static void set_linv_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr)
4283 {
4284 	memset(umr, 0, sizeof(*umr));
4285 	umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
4286 	umr->flags = MLX5_UMR_INLINE;
4287 }
4288 
4289 static __be64 get_umr_enable_mr_mask(void)
4290 {
4291 	u64 result;
4292 
4293 	result = MLX5_MKEY_MASK_KEY |
4294 		 MLX5_MKEY_MASK_FREE;
4295 
4296 	return cpu_to_be64(result);
4297 }
4298 
4299 static __be64 get_umr_disable_mr_mask(void)
4300 {
4301 	u64 result;
4302 
4303 	result = MLX5_MKEY_MASK_FREE;
4304 
4305 	return cpu_to_be64(result);
4306 }
4307 
4308 static __be64 get_umr_update_translation_mask(void)
4309 {
4310 	u64 result;
4311 
4312 	result = MLX5_MKEY_MASK_LEN |
4313 		 MLX5_MKEY_MASK_PAGE_SIZE |
4314 		 MLX5_MKEY_MASK_START_ADDR;
4315 
4316 	return cpu_to_be64(result);
4317 }
4318 
4319 static __be64 get_umr_update_access_mask(int atomic)
4320 {
4321 	u64 result;
4322 
4323 	result = MLX5_MKEY_MASK_LR |
4324 		 MLX5_MKEY_MASK_LW |
4325 		 MLX5_MKEY_MASK_RR |
4326 		 MLX5_MKEY_MASK_RW;
4327 
4328 	if (atomic)
4329 		result |= MLX5_MKEY_MASK_A;
4330 
4331 	return cpu_to_be64(result);
4332 }
4333 
4334 static __be64 get_umr_update_pd_mask(void)
4335 {
4336 	u64 result;
4337 
4338 	result = MLX5_MKEY_MASK_PD;
4339 
4340 	return cpu_to_be64(result);
4341 }
4342 
4343 static int umr_check_mkey_mask(struct mlx5_ib_dev *dev, u64 mask)
4344 {
4345 	if ((mask & MLX5_MKEY_MASK_PAGE_SIZE &&
4346 	     MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled)) ||
4347 	    (mask & MLX5_MKEY_MASK_A &&
4348 	     MLX5_CAP_GEN(dev->mdev, umr_modify_atomic_disabled)))
4349 		return -EPERM;
4350 	return 0;
4351 }
4352 
4353 static int set_reg_umr_segment(struct mlx5_ib_dev *dev,
4354 			       struct mlx5_wqe_umr_ctrl_seg *umr,
4355 			       const struct ib_send_wr *wr, int atomic)
4356 {
4357 	const struct mlx5_umr_wr *umrwr = umr_wr(wr);
4358 
4359 	memset(umr, 0, sizeof(*umr));
4360 
4361 	if (!umrwr->ignore_free_state) {
4362 		if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
4363 			 /* fail if free */
4364 			umr->flags = MLX5_UMR_CHECK_FREE;
4365 		else
4366 			/* fail if not free */
4367 			umr->flags = MLX5_UMR_CHECK_NOT_FREE;
4368 	}
4369 
4370 	umr->xlt_octowords = cpu_to_be16(get_xlt_octo(umrwr->xlt_size));
4371 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_XLT) {
4372 		u64 offset = get_xlt_octo(umrwr->offset);
4373 
4374 		umr->xlt_offset = cpu_to_be16(offset & 0xffff);
4375 		umr->xlt_offset_47_16 = cpu_to_be32(offset >> 16);
4376 		umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
4377 	}
4378 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION)
4379 		umr->mkey_mask |= get_umr_update_translation_mask();
4380 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS) {
4381 		umr->mkey_mask |= get_umr_update_access_mask(atomic);
4382 		umr->mkey_mask |= get_umr_update_pd_mask();
4383 	}
4384 	if (wr->send_flags & MLX5_IB_SEND_UMR_ENABLE_MR)
4385 		umr->mkey_mask |= get_umr_enable_mr_mask();
4386 	if (wr->send_flags & MLX5_IB_SEND_UMR_DISABLE_MR)
4387 		umr->mkey_mask |= get_umr_disable_mr_mask();
4388 
4389 	if (!wr->num_sge)
4390 		umr->flags |= MLX5_UMR_INLINE;
4391 
4392 	return umr_check_mkey_mask(dev, be64_to_cpu(umr->mkey_mask));
4393 }
4394 
4395 static u8 get_umr_flags(int acc)
4396 {
4397 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
4398 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
4399 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
4400 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
4401 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
4402 }
4403 
4404 static void set_reg_mkey_seg(struct mlx5_mkey_seg *seg,
4405 			     struct mlx5_ib_mr *mr,
4406 			     u32 key, int access)
4407 {
4408 	int ndescs = ALIGN(mr->ndescs + mr->meta_ndescs, 8) >> 1;
4409 
4410 	memset(seg, 0, sizeof(*seg));
4411 
4412 	if (mr->access_mode == MLX5_MKC_ACCESS_MODE_MTT)
4413 		seg->log2_page_size = ilog2(mr->ibmr.page_size);
4414 	else if (mr->access_mode == MLX5_MKC_ACCESS_MODE_KLMS)
4415 		/* KLMs take twice the size of MTTs */
4416 		ndescs *= 2;
4417 
4418 	seg->flags = get_umr_flags(access) | mr->access_mode;
4419 	seg->qpn_mkey7_0 = cpu_to_be32((key & 0xff) | 0xffffff00);
4420 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
4421 	seg->start_addr = cpu_to_be64(mr->ibmr.iova);
4422 	seg->len = cpu_to_be64(mr->ibmr.length);
4423 	seg->xlt_oct_size = cpu_to_be32(ndescs);
4424 }
4425 
4426 static void set_linv_mkey_seg(struct mlx5_mkey_seg *seg)
4427 {
4428 	memset(seg, 0, sizeof(*seg));
4429 	seg->status = MLX5_MKEY_STATUS_FREE;
4430 }
4431 
4432 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg,
4433 				 const struct ib_send_wr *wr)
4434 {
4435 	const struct mlx5_umr_wr *umrwr = umr_wr(wr);
4436 
4437 	memset(seg, 0, sizeof(*seg));
4438 	if (wr->send_flags & MLX5_IB_SEND_UMR_DISABLE_MR)
4439 		seg->status = MLX5_MKEY_STATUS_FREE;
4440 
4441 	seg->flags = convert_access(umrwr->access_flags);
4442 	if (umrwr->pd)
4443 		seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
4444 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION &&
4445 	    !umrwr->length)
4446 		seg->flags_pd |= cpu_to_be32(MLX5_MKEY_LEN64);
4447 
4448 	seg->start_addr = cpu_to_be64(umrwr->virt_addr);
4449 	seg->len = cpu_to_be64(umrwr->length);
4450 	seg->log2_page_size = umrwr->page_shift;
4451 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
4452 				       mlx5_mkey_variant(umrwr->mkey));
4453 }
4454 
4455 static void set_reg_data_seg(struct mlx5_wqe_data_seg *dseg,
4456 			     struct mlx5_ib_mr *mr,
4457 			     struct mlx5_ib_pd *pd)
4458 {
4459 	int bcount = mr->desc_size * (mr->ndescs + mr->meta_ndescs);
4460 
4461 	dseg->addr = cpu_to_be64(mr->desc_map);
4462 	dseg->byte_count = cpu_to_be32(ALIGN(bcount, 64));
4463 	dseg->lkey = cpu_to_be32(pd->ibpd.local_dma_lkey);
4464 }
4465 
4466 static __be32 send_ieth(const struct ib_send_wr *wr)
4467 {
4468 	switch (wr->opcode) {
4469 	case IB_WR_SEND_WITH_IMM:
4470 	case IB_WR_RDMA_WRITE_WITH_IMM:
4471 		return wr->ex.imm_data;
4472 
4473 	case IB_WR_SEND_WITH_INV:
4474 		return cpu_to_be32(wr->ex.invalidate_rkey);
4475 
4476 	default:
4477 		return 0;
4478 	}
4479 }
4480 
4481 static u8 calc_sig(void *wqe, int size)
4482 {
4483 	u8 *p = wqe;
4484 	u8 res = 0;
4485 	int i;
4486 
4487 	for (i = 0; i < size; i++)
4488 		res ^= p[i];
4489 
4490 	return ~res;
4491 }
4492 
4493 static u8 wq_sig(void *wqe)
4494 {
4495 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
4496 }
4497 
4498 static int set_data_inl_seg(struct mlx5_ib_qp *qp, const struct ib_send_wr *wr,
4499 			    void **wqe, int *wqe_sz, void **cur_edge)
4500 {
4501 	struct mlx5_wqe_inline_seg *seg;
4502 	size_t offset;
4503 	int inl = 0;
4504 	int i;
4505 
4506 	seg = *wqe;
4507 	*wqe += sizeof(*seg);
4508 	offset = sizeof(*seg);
4509 
4510 	for (i = 0; i < wr->num_sge; i++) {
4511 		size_t len  = wr->sg_list[i].length;
4512 		void *addr = (void *)(unsigned long)(wr->sg_list[i].addr);
4513 
4514 		inl += len;
4515 
4516 		if (unlikely(inl > qp->max_inline_data))
4517 			return -ENOMEM;
4518 
4519 		while (likely(len)) {
4520 			size_t leftlen;
4521 			size_t copysz;
4522 
4523 			handle_post_send_edge(&qp->sq, wqe,
4524 					      *wqe_sz + (offset >> 4),
4525 					      cur_edge);
4526 
4527 			leftlen = *cur_edge - *wqe;
4528 			copysz = min_t(size_t, leftlen, len);
4529 
4530 			memcpy(*wqe, addr, copysz);
4531 			len -= copysz;
4532 			addr += copysz;
4533 			*wqe += copysz;
4534 			offset += copysz;
4535 		}
4536 	}
4537 
4538 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
4539 
4540 	*wqe_sz +=  ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
4541 
4542 	return 0;
4543 }
4544 
4545 static u16 prot_field_size(enum ib_signature_type type)
4546 {
4547 	switch (type) {
4548 	case IB_SIG_TYPE_T10_DIF:
4549 		return MLX5_DIF_SIZE;
4550 	default:
4551 		return 0;
4552 	}
4553 }
4554 
4555 static u8 bs_selector(int block_size)
4556 {
4557 	switch (block_size) {
4558 	case 512:	    return 0x1;
4559 	case 520:	    return 0x2;
4560 	case 4096:	    return 0x3;
4561 	case 4160:	    return 0x4;
4562 	case 1073741824:    return 0x5;
4563 	default:	    return 0;
4564 	}
4565 }
4566 
4567 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
4568 			      struct mlx5_bsf_inl *inl)
4569 {
4570 	/* Valid inline section and allow BSF refresh */
4571 	inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
4572 				       MLX5_BSF_REFRESH_DIF);
4573 	inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
4574 	inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
4575 	/* repeating block */
4576 	inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
4577 	inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
4578 			MLX5_DIF_CRC : MLX5_DIF_IPCS;
4579 
4580 	if (domain->sig.dif.ref_remap)
4581 		inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
4582 
4583 	if (domain->sig.dif.app_escape) {
4584 		if (domain->sig.dif.ref_escape)
4585 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
4586 		else
4587 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
4588 	}
4589 
4590 	inl->dif_app_bitmask_check =
4591 		cpu_to_be16(domain->sig.dif.apptag_check_mask);
4592 }
4593 
4594 static int mlx5_set_bsf(struct ib_mr *sig_mr,
4595 			struct ib_sig_attrs *sig_attrs,
4596 			struct mlx5_bsf *bsf, u32 data_size)
4597 {
4598 	struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
4599 	struct mlx5_bsf_basic *basic = &bsf->basic;
4600 	struct ib_sig_domain *mem = &sig_attrs->mem;
4601 	struct ib_sig_domain *wire = &sig_attrs->wire;
4602 
4603 	memset(bsf, 0, sizeof(*bsf));
4604 
4605 	/* Basic + Extended + Inline */
4606 	basic->bsf_size_sbs = 1 << 7;
4607 	/* Input domain check byte mask */
4608 	basic->check_byte_mask = sig_attrs->check_mask;
4609 	basic->raw_data_size = cpu_to_be32(data_size);
4610 
4611 	/* Memory domain */
4612 	switch (sig_attrs->mem.sig_type) {
4613 	case IB_SIG_TYPE_NONE:
4614 		break;
4615 	case IB_SIG_TYPE_T10_DIF:
4616 		basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
4617 		basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
4618 		mlx5_fill_inl_bsf(mem, &bsf->m_inl);
4619 		break;
4620 	default:
4621 		return -EINVAL;
4622 	}
4623 
4624 	/* Wire domain */
4625 	switch (sig_attrs->wire.sig_type) {
4626 	case IB_SIG_TYPE_NONE:
4627 		break;
4628 	case IB_SIG_TYPE_T10_DIF:
4629 		if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
4630 		    mem->sig_type == wire->sig_type) {
4631 			/* Same block structure */
4632 			basic->bsf_size_sbs |= 1 << 4;
4633 			if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
4634 				basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
4635 			if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
4636 				basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
4637 			if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
4638 				basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
4639 		} else
4640 			basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
4641 
4642 		basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
4643 		mlx5_fill_inl_bsf(wire, &bsf->w_inl);
4644 		break;
4645 	default:
4646 		return -EINVAL;
4647 	}
4648 
4649 	return 0;
4650 }
4651 
4652 static int set_sig_data_segment(const struct ib_send_wr *send_wr,
4653 				struct ib_mr *sig_mr,
4654 				struct ib_sig_attrs *sig_attrs,
4655 				struct mlx5_ib_qp *qp, void **seg, int *size,
4656 				void **cur_edge)
4657 {
4658 	struct mlx5_bsf *bsf;
4659 	u32 data_len;
4660 	u32 data_key;
4661 	u64 data_va;
4662 	u32 prot_len = 0;
4663 	u32 prot_key = 0;
4664 	u64 prot_va = 0;
4665 	bool prot = false;
4666 	int ret;
4667 	int wqe_size;
4668 	struct mlx5_ib_mr *mr = to_mmr(sig_mr);
4669 	struct mlx5_ib_mr *pi_mr = mr->pi_mr;
4670 
4671 	data_len = pi_mr->data_length;
4672 	data_key = pi_mr->ibmr.lkey;
4673 	data_va = pi_mr->data_iova;
4674 	if (pi_mr->meta_ndescs) {
4675 		prot_len = pi_mr->meta_length;
4676 		prot_key = pi_mr->ibmr.lkey;
4677 		prot_va = pi_mr->pi_iova;
4678 		prot = true;
4679 	}
4680 
4681 	if (!prot || (data_key == prot_key && data_va == prot_va &&
4682 		      data_len == prot_len)) {
4683 		/**
4684 		 * Source domain doesn't contain signature information
4685 		 * or data and protection are interleaved in memory.
4686 		 * So need construct:
4687 		 *                  ------------------
4688 		 *                 |     data_klm     |
4689 		 *                  ------------------
4690 		 *                 |       BSF        |
4691 		 *                  ------------------
4692 		 **/
4693 		struct mlx5_klm *data_klm = *seg;
4694 
4695 		data_klm->bcount = cpu_to_be32(data_len);
4696 		data_klm->key = cpu_to_be32(data_key);
4697 		data_klm->va = cpu_to_be64(data_va);
4698 		wqe_size = ALIGN(sizeof(*data_klm), 64);
4699 	} else {
4700 		/**
4701 		 * Source domain contains signature information
4702 		 * So need construct a strided block format:
4703 		 *               ---------------------------
4704 		 *              |     stride_block_ctrl     |
4705 		 *               ---------------------------
4706 		 *              |          data_klm         |
4707 		 *               ---------------------------
4708 		 *              |          prot_klm         |
4709 		 *               ---------------------------
4710 		 *              |             BSF           |
4711 		 *               ---------------------------
4712 		 **/
4713 		struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
4714 		struct mlx5_stride_block_entry *data_sentry;
4715 		struct mlx5_stride_block_entry *prot_sentry;
4716 		u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
4717 		int prot_size;
4718 
4719 		sblock_ctrl = *seg;
4720 		data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
4721 		prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
4722 
4723 		prot_size = prot_field_size(sig_attrs->mem.sig_type);
4724 		if (!prot_size) {
4725 			pr_err("Bad block size given: %u\n", block_size);
4726 			return -EINVAL;
4727 		}
4728 		sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
4729 							    prot_size);
4730 		sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
4731 		sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
4732 		sblock_ctrl->num_entries = cpu_to_be16(2);
4733 
4734 		data_sentry->bcount = cpu_to_be16(block_size);
4735 		data_sentry->key = cpu_to_be32(data_key);
4736 		data_sentry->va = cpu_to_be64(data_va);
4737 		data_sentry->stride = cpu_to_be16(block_size);
4738 
4739 		prot_sentry->bcount = cpu_to_be16(prot_size);
4740 		prot_sentry->key = cpu_to_be32(prot_key);
4741 		prot_sentry->va = cpu_to_be64(prot_va);
4742 		prot_sentry->stride = cpu_to_be16(prot_size);
4743 
4744 		wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
4745 				 sizeof(*prot_sentry), 64);
4746 	}
4747 
4748 	*seg += wqe_size;
4749 	*size += wqe_size / 16;
4750 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4751 
4752 	bsf = *seg;
4753 	ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
4754 	if (ret)
4755 		return -EINVAL;
4756 
4757 	*seg += sizeof(*bsf);
4758 	*size += sizeof(*bsf) / 16;
4759 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4760 
4761 	return 0;
4762 }
4763 
4764 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
4765 				 struct ib_mr *sig_mr, int access_flags,
4766 				 u32 size, u32 length, u32 pdn)
4767 {
4768 	u32 sig_key = sig_mr->rkey;
4769 	u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
4770 
4771 	memset(seg, 0, sizeof(*seg));
4772 
4773 	seg->flags = get_umr_flags(access_flags) | MLX5_MKC_ACCESS_MODE_KLMS;
4774 	seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
4775 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
4776 				    MLX5_MKEY_BSF_EN | pdn);
4777 	seg->len = cpu_to_be64(length);
4778 	seg->xlt_oct_size = cpu_to_be32(get_xlt_octo(size));
4779 	seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
4780 }
4781 
4782 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
4783 				u32 size)
4784 {
4785 	memset(umr, 0, sizeof(*umr));
4786 
4787 	umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
4788 	umr->xlt_octowords = cpu_to_be16(get_xlt_octo(size));
4789 	umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
4790 	umr->mkey_mask = sig_mkey_mask();
4791 }
4792 
4793 static int set_pi_umr_wr(const struct ib_send_wr *send_wr,
4794 			 struct mlx5_ib_qp *qp, void **seg, int *size,
4795 			 void **cur_edge)
4796 {
4797 	const struct ib_reg_wr *wr = reg_wr(send_wr);
4798 	struct mlx5_ib_mr *sig_mr = to_mmr(wr->mr);
4799 	struct mlx5_ib_mr *pi_mr = sig_mr->pi_mr;
4800 	struct ib_sig_attrs *sig_attrs = sig_mr->ibmr.sig_attrs;
4801 	u32 pdn = get_pd(qp)->pdn;
4802 	u32 xlt_size;
4803 	int region_len, ret;
4804 
4805 	if (unlikely(send_wr->num_sge != 0) ||
4806 	    unlikely(wr->access & IB_ACCESS_REMOTE_ATOMIC) ||
4807 	    unlikely(!sig_mr->sig) || unlikely(!qp->ibqp.integrity_en) ||
4808 	    unlikely(!sig_mr->sig->sig_status_checked))
4809 		return -EINVAL;
4810 
4811 	/* length of the protected region, data + protection */
4812 	region_len = pi_mr->ibmr.length;
4813 
4814 	/**
4815 	 * KLM octoword size - if protection was provided
4816 	 * then we use strided block format (3 octowords),
4817 	 * else we use single KLM (1 octoword)
4818 	 **/
4819 	if (sig_attrs->mem.sig_type != IB_SIG_TYPE_NONE)
4820 		xlt_size = 0x30;
4821 	else
4822 		xlt_size = sizeof(struct mlx5_klm);
4823 
4824 	set_sig_umr_segment(*seg, xlt_size);
4825 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4826 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4827 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4828 
4829 	set_sig_mkey_segment(*seg, wr->mr, wr->access, xlt_size, region_len,
4830 			     pdn);
4831 	*seg += sizeof(struct mlx5_mkey_seg);
4832 	*size += sizeof(struct mlx5_mkey_seg) / 16;
4833 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4834 
4835 	ret = set_sig_data_segment(send_wr, wr->mr, sig_attrs, qp, seg, size,
4836 				   cur_edge);
4837 	if (ret)
4838 		return ret;
4839 
4840 	sig_mr->sig->sig_status_checked = false;
4841 	return 0;
4842 }
4843 
4844 static int set_psv_wr(struct ib_sig_domain *domain,
4845 		      u32 psv_idx, void **seg, int *size)
4846 {
4847 	struct mlx5_seg_set_psv *psv_seg = *seg;
4848 
4849 	memset(psv_seg, 0, sizeof(*psv_seg));
4850 	psv_seg->psv_num = cpu_to_be32(psv_idx);
4851 	switch (domain->sig_type) {
4852 	case IB_SIG_TYPE_NONE:
4853 		break;
4854 	case IB_SIG_TYPE_T10_DIF:
4855 		psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
4856 						     domain->sig.dif.app_tag);
4857 		psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
4858 		break;
4859 	default:
4860 		pr_err("Bad signature type (%d) is given.\n",
4861 		       domain->sig_type);
4862 		return -EINVAL;
4863 	}
4864 
4865 	*seg += sizeof(*psv_seg);
4866 	*size += sizeof(*psv_seg) / 16;
4867 
4868 	return 0;
4869 }
4870 
4871 static int set_reg_wr(struct mlx5_ib_qp *qp,
4872 		      const struct ib_reg_wr *wr,
4873 		      void **seg, int *size, void **cur_edge,
4874 		      bool check_not_free)
4875 {
4876 	struct mlx5_ib_mr *mr = to_mmr(wr->mr);
4877 	struct mlx5_ib_pd *pd = to_mpd(qp->ibqp.pd);
4878 	struct mlx5_ib_dev *dev = to_mdev(pd->ibpd.device);
4879 	int mr_list_size = (mr->ndescs + mr->meta_ndescs) * mr->desc_size;
4880 	bool umr_inline = mr_list_size <= MLX5_IB_SQ_UMR_INLINE_THRESHOLD;
4881 	bool atomic = wr->access & IB_ACCESS_REMOTE_ATOMIC;
4882 	u8 flags = 0;
4883 
4884 	if (!mlx5_ib_can_use_umr(dev, atomic, wr->access)) {
4885 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
4886 			     "Fast update of %s for MR is disabled\n",
4887 			     (MLX5_CAP_GEN(dev->mdev,
4888 					   umr_modify_entity_size_disabled)) ?
4889 				     "entity size" :
4890 				     "atomic access");
4891 		return -EINVAL;
4892 	}
4893 
4894 	if (unlikely(wr->wr.send_flags & IB_SEND_INLINE)) {
4895 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
4896 			     "Invalid IB_SEND_INLINE send flag\n");
4897 		return -EINVAL;
4898 	}
4899 
4900 	if (check_not_free)
4901 		flags |= MLX5_UMR_CHECK_NOT_FREE;
4902 	if (umr_inline)
4903 		flags |= MLX5_UMR_INLINE;
4904 
4905 	set_reg_umr_seg(*seg, mr, flags, atomic);
4906 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4907 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4908 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4909 
4910 	set_reg_mkey_seg(*seg, mr, wr->key, wr->access);
4911 	*seg += sizeof(struct mlx5_mkey_seg);
4912 	*size += sizeof(struct mlx5_mkey_seg) / 16;
4913 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4914 
4915 	if (umr_inline) {
4916 		memcpy_send_wqe(&qp->sq, cur_edge, seg, size, mr->descs,
4917 				mr_list_size);
4918 		*size = ALIGN(*size, MLX5_SEND_WQE_BB >> 4);
4919 	} else {
4920 		set_reg_data_seg(*seg, mr, pd);
4921 		*seg += sizeof(struct mlx5_wqe_data_seg);
4922 		*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
4923 	}
4924 	return 0;
4925 }
4926 
4927 static void set_linv_wr(struct mlx5_ib_qp *qp, void **seg, int *size,
4928 			void **cur_edge)
4929 {
4930 	set_linv_umr_seg(*seg);
4931 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4932 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4933 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4934 	set_linv_mkey_seg(*seg);
4935 	*seg += sizeof(struct mlx5_mkey_seg);
4936 	*size += sizeof(struct mlx5_mkey_seg) / 16;
4937 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4938 }
4939 
4940 static void dump_wqe(struct mlx5_ib_qp *qp, u32 idx, int size_16)
4941 {
4942 	__be32 *p = NULL;
4943 	int i, j;
4944 
4945 	pr_debug("dump WQE index %u:\n", idx);
4946 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
4947 		if ((i & 0xf) == 0) {
4948 			p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, idx);
4949 			pr_debug("WQBB at %p:\n", (void *)p);
4950 			j = 0;
4951 			idx = (idx + 1) & (qp->sq.wqe_cnt - 1);
4952 		}
4953 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
4954 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
4955 			 be32_to_cpu(p[j + 3]));
4956 	}
4957 }
4958 
4959 static int __begin_wqe(struct mlx5_ib_qp *qp, void **seg,
4960 		       struct mlx5_wqe_ctrl_seg **ctrl,
4961 		       const struct ib_send_wr *wr, unsigned int *idx,
4962 		       int *size, void **cur_edge, int nreq,
4963 		       bool send_signaled, bool solicited)
4964 {
4965 	if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)))
4966 		return -ENOMEM;
4967 
4968 	*idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
4969 	*seg = mlx5_frag_buf_get_wqe(&qp->sq.fbc, *idx);
4970 	*ctrl = *seg;
4971 	*(uint32_t *)(*seg + 8) = 0;
4972 	(*ctrl)->imm = send_ieth(wr);
4973 	(*ctrl)->fm_ce_se = qp->sq_signal_bits |
4974 		(send_signaled ? MLX5_WQE_CTRL_CQ_UPDATE : 0) |
4975 		(solicited ? MLX5_WQE_CTRL_SOLICITED : 0);
4976 
4977 	*seg += sizeof(**ctrl);
4978 	*size = sizeof(**ctrl) / 16;
4979 	*cur_edge = qp->sq.cur_edge;
4980 
4981 	return 0;
4982 }
4983 
4984 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
4985 		     struct mlx5_wqe_ctrl_seg **ctrl,
4986 		     const struct ib_send_wr *wr, unsigned *idx,
4987 		     int *size, void **cur_edge, int nreq)
4988 {
4989 	return __begin_wqe(qp, seg, ctrl, wr, idx, size, cur_edge, nreq,
4990 			   wr->send_flags & IB_SEND_SIGNALED,
4991 			   wr->send_flags & IB_SEND_SOLICITED);
4992 }
4993 
4994 static void finish_wqe(struct mlx5_ib_qp *qp,
4995 		       struct mlx5_wqe_ctrl_seg *ctrl,
4996 		       void *seg, u8 size, void *cur_edge,
4997 		       unsigned int idx, u64 wr_id, int nreq, u8 fence,
4998 		       u32 mlx5_opcode)
4999 {
5000 	u8 opmod = 0;
5001 
5002 	ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
5003 					     mlx5_opcode | ((u32)opmod << 24));
5004 	ctrl->qpn_ds = cpu_to_be32(size | (qp->trans_qp.base.mqp.qpn << 8));
5005 	ctrl->fm_ce_se |= fence;
5006 	if (unlikely(qp->wq_sig))
5007 		ctrl->signature = wq_sig(ctrl);
5008 
5009 	qp->sq.wrid[idx] = wr_id;
5010 	qp->sq.w_list[idx].opcode = mlx5_opcode;
5011 	qp->sq.wqe_head[idx] = qp->sq.head + nreq;
5012 	qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
5013 	qp->sq.w_list[idx].next = qp->sq.cur_post;
5014 
5015 	/* We save the edge which was possibly updated during the WQE
5016 	 * construction, into SQ's cache.
5017 	 */
5018 	seg = PTR_ALIGN(seg, MLX5_SEND_WQE_BB);
5019 	qp->sq.cur_edge = (unlikely(seg == cur_edge)) ?
5020 			  get_sq_edge(&qp->sq, qp->sq.cur_post &
5021 				      (qp->sq.wqe_cnt - 1)) :
5022 			  cur_edge;
5023 }
5024 
5025 static int _mlx5_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
5026 			      const struct ib_send_wr **bad_wr, bool drain)
5027 {
5028 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
5029 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
5030 	struct mlx5_core_dev *mdev = dev->mdev;
5031 	struct ib_reg_wr reg_pi_wr;
5032 	struct mlx5_ib_qp *qp;
5033 	struct mlx5_ib_mr *mr;
5034 	struct mlx5_ib_mr *pi_mr;
5035 	struct mlx5_ib_mr pa_pi_mr;
5036 	struct ib_sig_attrs *sig_attrs;
5037 	struct mlx5_wqe_xrc_seg *xrc;
5038 	struct mlx5_bf *bf;
5039 	void *cur_edge;
5040 	int uninitialized_var(size);
5041 	unsigned long flags;
5042 	unsigned idx;
5043 	int err = 0;
5044 	int num_sge;
5045 	void *seg;
5046 	int nreq;
5047 	int i;
5048 	u8 next_fence = 0;
5049 	u8 fence;
5050 
5051 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR &&
5052 		     !drain)) {
5053 		*bad_wr = wr;
5054 		return -EIO;
5055 	}
5056 
5057 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
5058 		return mlx5_ib_gsi_post_send(ibqp, wr, bad_wr);
5059 
5060 	qp = to_mqp(ibqp);
5061 	bf = &qp->bf;
5062 
5063 	spin_lock_irqsave(&qp->sq.lock, flags);
5064 
5065 	for (nreq = 0; wr; nreq++, wr = wr->next) {
5066 		if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
5067 			mlx5_ib_warn(dev, "\n");
5068 			err = -EINVAL;
5069 			*bad_wr = wr;
5070 			goto out;
5071 		}
5072 
5073 		num_sge = wr->num_sge;
5074 		if (unlikely(num_sge > qp->sq.max_gs)) {
5075 			mlx5_ib_warn(dev, "\n");
5076 			err = -EINVAL;
5077 			*bad_wr = wr;
5078 			goto out;
5079 		}
5080 
5081 		err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, &cur_edge,
5082 				nreq);
5083 		if (err) {
5084 			mlx5_ib_warn(dev, "\n");
5085 			err = -ENOMEM;
5086 			*bad_wr = wr;
5087 			goto out;
5088 		}
5089 
5090 		if (wr->opcode == IB_WR_REG_MR ||
5091 		    wr->opcode == IB_WR_REG_MR_INTEGRITY) {
5092 			fence = dev->umr_fence;
5093 			next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
5094 		} else  {
5095 			if (wr->send_flags & IB_SEND_FENCE) {
5096 				if (qp->next_fence)
5097 					fence = MLX5_FENCE_MODE_SMALL_AND_FENCE;
5098 				else
5099 					fence = MLX5_FENCE_MODE_FENCE;
5100 			} else {
5101 				fence = qp->next_fence;
5102 			}
5103 		}
5104 
5105 		switch (ibqp->qp_type) {
5106 		case IB_QPT_XRC_INI:
5107 			xrc = seg;
5108 			seg += sizeof(*xrc);
5109 			size += sizeof(*xrc) / 16;
5110 			/* fall through */
5111 		case IB_QPT_RC:
5112 			switch (wr->opcode) {
5113 			case IB_WR_RDMA_READ:
5114 			case IB_WR_RDMA_WRITE:
5115 			case IB_WR_RDMA_WRITE_WITH_IMM:
5116 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
5117 					      rdma_wr(wr)->rkey);
5118 				seg += sizeof(struct mlx5_wqe_raddr_seg);
5119 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
5120 				break;
5121 
5122 			case IB_WR_ATOMIC_CMP_AND_SWP:
5123 			case IB_WR_ATOMIC_FETCH_AND_ADD:
5124 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
5125 				mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
5126 				err = -ENOSYS;
5127 				*bad_wr = wr;
5128 				goto out;
5129 
5130 			case IB_WR_LOCAL_INV:
5131 				qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
5132 				ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
5133 				set_linv_wr(qp, &seg, &size, &cur_edge);
5134 				num_sge = 0;
5135 				break;
5136 
5137 			case IB_WR_REG_MR:
5138 				qp->sq.wr_data[idx] = IB_WR_REG_MR;
5139 				ctrl->imm = cpu_to_be32(reg_wr(wr)->key);
5140 				err = set_reg_wr(qp, reg_wr(wr), &seg, &size,
5141 						 &cur_edge, true);
5142 				if (err) {
5143 					*bad_wr = wr;
5144 					goto out;
5145 				}
5146 				num_sge = 0;
5147 				break;
5148 
5149 			case IB_WR_REG_MR_INTEGRITY:
5150 				qp->sq.wr_data[idx] = IB_WR_REG_MR_INTEGRITY;
5151 
5152 				mr = to_mmr(reg_wr(wr)->mr);
5153 				pi_mr = mr->pi_mr;
5154 
5155 				if (pi_mr) {
5156 					memset(&reg_pi_wr, 0,
5157 					       sizeof(struct ib_reg_wr));
5158 
5159 					reg_pi_wr.mr = &pi_mr->ibmr;
5160 					reg_pi_wr.access = reg_wr(wr)->access;
5161 					reg_pi_wr.key = pi_mr->ibmr.rkey;
5162 
5163 					ctrl->imm = cpu_to_be32(reg_pi_wr.key);
5164 					/* UMR for data + prot registration */
5165 					err = set_reg_wr(qp, &reg_pi_wr, &seg,
5166 							 &size, &cur_edge,
5167 							 false);
5168 					if (err) {
5169 						*bad_wr = wr;
5170 						goto out;
5171 					}
5172 					finish_wqe(qp, ctrl, seg, size,
5173 						   cur_edge, idx, wr->wr_id,
5174 						   nreq, fence,
5175 						   MLX5_OPCODE_UMR);
5176 
5177 					err = begin_wqe(qp, &seg, &ctrl, wr,
5178 							&idx, &size, &cur_edge,
5179 							nreq);
5180 					if (err) {
5181 						mlx5_ib_warn(dev, "\n");
5182 						err = -ENOMEM;
5183 						*bad_wr = wr;
5184 						goto out;
5185 					}
5186 				} else {
5187 					memset(&pa_pi_mr, 0,
5188 					       sizeof(struct mlx5_ib_mr));
5189 					/* No UMR, use local_dma_lkey */
5190 					pa_pi_mr.ibmr.lkey =
5191 						mr->ibmr.pd->local_dma_lkey;
5192 
5193 					pa_pi_mr.ndescs = mr->ndescs;
5194 					pa_pi_mr.data_length = mr->data_length;
5195 					pa_pi_mr.data_iova = mr->data_iova;
5196 					if (mr->meta_ndescs) {
5197 						pa_pi_mr.meta_ndescs =
5198 							mr->meta_ndescs;
5199 						pa_pi_mr.meta_length =
5200 							mr->meta_length;
5201 						pa_pi_mr.pi_iova = mr->pi_iova;
5202 					}
5203 
5204 					pa_pi_mr.ibmr.length = mr->ibmr.length;
5205 					mr->pi_mr = &pa_pi_mr;
5206 				}
5207 				ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
5208 				/* UMR for sig MR */
5209 				err = set_pi_umr_wr(wr, qp, &seg, &size,
5210 						    &cur_edge);
5211 				if (err) {
5212 					mlx5_ib_warn(dev, "\n");
5213 					*bad_wr = wr;
5214 					goto out;
5215 				}
5216 				finish_wqe(qp, ctrl, seg, size, cur_edge, idx,
5217 					   wr->wr_id, nreq, fence,
5218 					   MLX5_OPCODE_UMR);
5219 
5220 				/*
5221 				 * SET_PSV WQEs are not signaled and solicited
5222 				 * on error
5223 				 */
5224 				sig_attrs = mr->ibmr.sig_attrs;
5225 				err = __begin_wqe(qp, &seg, &ctrl, wr, &idx,
5226 						  &size, &cur_edge, nreq, false,
5227 						  true);
5228 				if (err) {
5229 					mlx5_ib_warn(dev, "\n");
5230 					err = -ENOMEM;
5231 					*bad_wr = wr;
5232 					goto out;
5233 				}
5234 				err = set_psv_wr(&sig_attrs->mem,
5235 						 mr->sig->psv_memory.psv_idx,
5236 						 &seg, &size);
5237 				if (err) {
5238 					mlx5_ib_warn(dev, "\n");
5239 					*bad_wr = wr;
5240 					goto out;
5241 				}
5242 				finish_wqe(qp, ctrl, seg, size, cur_edge, idx,
5243 					   wr->wr_id, nreq, next_fence,
5244 					   MLX5_OPCODE_SET_PSV);
5245 
5246 				err = __begin_wqe(qp, &seg, &ctrl, wr, &idx,
5247 						  &size, &cur_edge, nreq, false,
5248 						  true);
5249 				if (err) {
5250 					mlx5_ib_warn(dev, "\n");
5251 					err = -ENOMEM;
5252 					*bad_wr = wr;
5253 					goto out;
5254 				}
5255 				err = set_psv_wr(&sig_attrs->wire,
5256 						 mr->sig->psv_wire.psv_idx,
5257 						 &seg, &size);
5258 				if (err) {
5259 					mlx5_ib_warn(dev, "\n");
5260 					*bad_wr = wr;
5261 					goto out;
5262 				}
5263 				finish_wqe(qp, ctrl, seg, size, cur_edge, idx,
5264 					   wr->wr_id, nreq, next_fence,
5265 					   MLX5_OPCODE_SET_PSV);
5266 
5267 				qp->next_fence =
5268 					MLX5_FENCE_MODE_INITIATOR_SMALL;
5269 				num_sge = 0;
5270 				goto skip_psv;
5271 
5272 			default:
5273 				break;
5274 			}
5275 			break;
5276 
5277 		case IB_QPT_UC:
5278 			switch (wr->opcode) {
5279 			case IB_WR_RDMA_WRITE:
5280 			case IB_WR_RDMA_WRITE_WITH_IMM:
5281 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
5282 					      rdma_wr(wr)->rkey);
5283 				seg  += sizeof(struct mlx5_wqe_raddr_seg);
5284 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
5285 				break;
5286 
5287 			default:
5288 				break;
5289 			}
5290 			break;
5291 
5292 		case IB_QPT_SMI:
5293 			if (unlikely(!mdev->port_caps[qp->port - 1].has_smi)) {
5294 				mlx5_ib_warn(dev, "Send SMP MADs is not allowed\n");
5295 				err = -EPERM;
5296 				*bad_wr = wr;
5297 				goto out;
5298 			}
5299 			/* fall through */
5300 		case MLX5_IB_QPT_HW_GSI:
5301 			set_datagram_seg(seg, wr);
5302 			seg += sizeof(struct mlx5_wqe_datagram_seg);
5303 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
5304 			handle_post_send_edge(&qp->sq, &seg, size, &cur_edge);
5305 
5306 			break;
5307 		case IB_QPT_UD:
5308 			set_datagram_seg(seg, wr);
5309 			seg += sizeof(struct mlx5_wqe_datagram_seg);
5310 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
5311 			handle_post_send_edge(&qp->sq, &seg, size, &cur_edge);
5312 
5313 			/* handle qp that supports ud offload */
5314 			if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO) {
5315 				struct mlx5_wqe_eth_pad *pad;
5316 
5317 				pad = seg;
5318 				memset(pad, 0, sizeof(struct mlx5_wqe_eth_pad));
5319 				seg += sizeof(struct mlx5_wqe_eth_pad);
5320 				size += sizeof(struct mlx5_wqe_eth_pad) / 16;
5321 				set_eth_seg(wr, qp, &seg, &size, &cur_edge);
5322 				handle_post_send_edge(&qp->sq, &seg, size,
5323 						      &cur_edge);
5324 			}
5325 			break;
5326 		case MLX5_IB_QPT_REG_UMR:
5327 			if (wr->opcode != MLX5_IB_WR_UMR) {
5328 				err = -EINVAL;
5329 				mlx5_ib_warn(dev, "bad opcode\n");
5330 				goto out;
5331 			}
5332 			qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
5333 			ctrl->imm = cpu_to_be32(umr_wr(wr)->mkey);
5334 			err = set_reg_umr_segment(dev, seg, wr, !!(MLX5_CAP_GEN(mdev, atomic)));
5335 			if (unlikely(err))
5336 				goto out;
5337 			seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
5338 			size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
5339 			handle_post_send_edge(&qp->sq, &seg, size, &cur_edge);
5340 			set_reg_mkey_segment(seg, wr);
5341 			seg += sizeof(struct mlx5_mkey_seg);
5342 			size += sizeof(struct mlx5_mkey_seg) / 16;
5343 			handle_post_send_edge(&qp->sq, &seg, size, &cur_edge);
5344 			break;
5345 
5346 		default:
5347 			break;
5348 		}
5349 
5350 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
5351 			err = set_data_inl_seg(qp, wr, &seg, &size, &cur_edge);
5352 			if (unlikely(err)) {
5353 				mlx5_ib_warn(dev, "\n");
5354 				*bad_wr = wr;
5355 				goto out;
5356 			}
5357 		} else {
5358 			for (i = 0; i < num_sge; i++) {
5359 				handle_post_send_edge(&qp->sq, &seg, size,
5360 						      &cur_edge);
5361 				if (likely(wr->sg_list[i].length)) {
5362 					set_data_ptr_seg
5363 					((struct mlx5_wqe_data_seg *)seg,
5364 					 wr->sg_list + i);
5365 					size += sizeof(struct mlx5_wqe_data_seg) / 16;
5366 					seg += sizeof(struct mlx5_wqe_data_seg);
5367 				}
5368 			}
5369 		}
5370 
5371 		qp->next_fence = next_fence;
5372 		finish_wqe(qp, ctrl, seg, size, cur_edge, idx, wr->wr_id, nreq,
5373 			   fence, mlx5_ib_opcode[wr->opcode]);
5374 skip_psv:
5375 		if (0)
5376 			dump_wqe(qp, idx, size);
5377 	}
5378 
5379 out:
5380 	if (likely(nreq)) {
5381 		qp->sq.head += nreq;
5382 
5383 		/* Make sure that descriptors are written before
5384 		 * updating doorbell record and ringing the doorbell
5385 		 */
5386 		wmb();
5387 
5388 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
5389 
5390 		/* Make sure doorbell record is visible to the HCA before
5391 		 * we hit doorbell */
5392 		wmb();
5393 
5394 		mlx5_write64((__be32 *)ctrl, bf->bfreg->map + bf->offset);
5395 		/* Make sure doorbells don't leak out of SQ spinlock
5396 		 * and reach the HCA out of order.
5397 		 */
5398 		bf->offset ^= bf->buf_size;
5399 	}
5400 
5401 	spin_unlock_irqrestore(&qp->sq.lock, flags);
5402 
5403 	return err;
5404 }
5405 
5406 int mlx5_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
5407 		      const struct ib_send_wr **bad_wr)
5408 {
5409 	return _mlx5_ib_post_send(ibqp, wr, bad_wr, false);
5410 }
5411 
5412 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
5413 {
5414 	sig->signature = calc_sig(sig, size);
5415 }
5416 
5417 static int _mlx5_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
5418 		      const struct ib_recv_wr **bad_wr, bool drain)
5419 {
5420 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
5421 	struct mlx5_wqe_data_seg *scat;
5422 	struct mlx5_rwqe_sig *sig;
5423 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
5424 	struct mlx5_core_dev *mdev = dev->mdev;
5425 	unsigned long flags;
5426 	int err = 0;
5427 	int nreq;
5428 	int ind;
5429 	int i;
5430 
5431 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR &&
5432 		     !drain)) {
5433 		*bad_wr = wr;
5434 		return -EIO;
5435 	}
5436 
5437 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
5438 		return mlx5_ib_gsi_post_recv(ibqp, wr, bad_wr);
5439 
5440 	spin_lock_irqsave(&qp->rq.lock, flags);
5441 
5442 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
5443 
5444 	for (nreq = 0; wr; nreq++, wr = wr->next) {
5445 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
5446 			err = -ENOMEM;
5447 			*bad_wr = wr;
5448 			goto out;
5449 		}
5450 
5451 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
5452 			err = -EINVAL;
5453 			*bad_wr = wr;
5454 			goto out;
5455 		}
5456 
5457 		scat = mlx5_frag_buf_get_wqe(&qp->rq.fbc, ind);
5458 		if (qp->wq_sig)
5459 			scat++;
5460 
5461 		for (i = 0; i < wr->num_sge; i++)
5462 			set_data_ptr_seg(scat + i, wr->sg_list + i);
5463 
5464 		if (i < qp->rq.max_gs) {
5465 			scat[i].byte_count = 0;
5466 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
5467 			scat[i].addr       = 0;
5468 		}
5469 
5470 		if (qp->wq_sig) {
5471 			sig = (struct mlx5_rwqe_sig *)scat;
5472 			set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
5473 		}
5474 
5475 		qp->rq.wrid[ind] = wr->wr_id;
5476 
5477 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
5478 	}
5479 
5480 out:
5481 	if (likely(nreq)) {
5482 		qp->rq.head += nreq;
5483 
5484 		/* Make sure that descriptors are written before
5485 		 * doorbell record.
5486 		 */
5487 		wmb();
5488 
5489 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
5490 	}
5491 
5492 	spin_unlock_irqrestore(&qp->rq.lock, flags);
5493 
5494 	return err;
5495 }
5496 
5497 int mlx5_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
5498 		      const struct ib_recv_wr **bad_wr)
5499 {
5500 	return _mlx5_ib_post_recv(ibqp, wr, bad_wr, false);
5501 }
5502 
5503 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
5504 {
5505 	switch (mlx5_state) {
5506 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
5507 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
5508 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
5509 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
5510 	case MLX5_QP_STATE_SQ_DRAINING:
5511 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
5512 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
5513 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
5514 	default:		     return -1;
5515 	}
5516 }
5517 
5518 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
5519 {
5520 	switch (mlx5_mig_state) {
5521 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
5522 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
5523 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
5524 	default: return -1;
5525 	}
5526 }
5527 
5528 static int to_ib_qp_access_flags(int mlx5_flags)
5529 {
5530 	int ib_flags = 0;
5531 
5532 	if (mlx5_flags & MLX5_QP_BIT_RRE)
5533 		ib_flags |= IB_ACCESS_REMOTE_READ;
5534 	if (mlx5_flags & MLX5_QP_BIT_RWE)
5535 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
5536 	if (mlx5_flags & MLX5_QP_BIT_RAE)
5537 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
5538 
5539 	return ib_flags;
5540 }
5541 
5542 static void to_rdma_ah_attr(struct mlx5_ib_dev *ibdev,
5543 			    struct rdma_ah_attr *ah_attr,
5544 			    struct mlx5_qp_path *path)
5545 {
5546 
5547 	memset(ah_attr, 0, sizeof(*ah_attr));
5548 
5549 	if (!path->port || path->port > ibdev->num_ports)
5550 		return;
5551 
5552 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, path->port);
5553 
5554 	rdma_ah_set_port_num(ah_attr, path->port);
5555 	rdma_ah_set_sl(ah_attr, path->dci_cfi_prio_sl & 0xf);
5556 
5557 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
5558 	rdma_ah_set_path_bits(ah_attr, path->grh_mlid & 0x7f);
5559 	rdma_ah_set_static_rate(ah_attr,
5560 				path->static_rate ? path->static_rate - 5 : 0);
5561 	if (path->grh_mlid & (1 << 7)) {
5562 		u32 tc_fl = be32_to_cpu(path->tclass_flowlabel);
5563 
5564 		rdma_ah_set_grh(ah_attr, NULL,
5565 				tc_fl & 0xfffff,
5566 				path->mgid_index,
5567 				path->hop_limit,
5568 				(tc_fl >> 20) & 0xff);
5569 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
5570 	}
5571 }
5572 
5573 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
5574 					struct mlx5_ib_sq *sq,
5575 					u8 *sq_state)
5576 {
5577 	int err;
5578 
5579 	err = mlx5_core_query_sq_state(dev->mdev, sq->base.mqp.qpn, sq_state);
5580 	if (err)
5581 		goto out;
5582 	sq->state = *sq_state;
5583 
5584 out:
5585 	return err;
5586 }
5587 
5588 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
5589 					struct mlx5_ib_rq *rq,
5590 					u8 *rq_state)
5591 {
5592 	void *out;
5593 	void *rqc;
5594 	int inlen;
5595 	int err;
5596 
5597 	inlen = MLX5_ST_SZ_BYTES(query_rq_out);
5598 	out = kvzalloc(inlen, GFP_KERNEL);
5599 	if (!out)
5600 		return -ENOMEM;
5601 
5602 	err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
5603 	if (err)
5604 		goto out;
5605 
5606 	rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
5607 	*rq_state = MLX5_GET(rqc, rqc, state);
5608 	rq->state = *rq_state;
5609 
5610 out:
5611 	kvfree(out);
5612 	return err;
5613 }
5614 
5615 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
5616 				  struct mlx5_ib_qp *qp, u8 *qp_state)
5617 {
5618 	static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
5619 		[MLX5_RQC_STATE_RST] = {
5620 			[MLX5_SQC_STATE_RST]	= IB_QPS_RESET,
5621 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
5622 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE_BAD,
5623 			[MLX5_SQ_STATE_NA]	= IB_QPS_RESET,
5624 		},
5625 		[MLX5_RQC_STATE_RDY] = {
5626 			[MLX5_SQC_STATE_RST]	= MLX5_QP_STATE_BAD,
5627 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
5628 			[MLX5_SQC_STATE_ERR]	= IB_QPS_SQE,
5629 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE,
5630 		},
5631 		[MLX5_RQC_STATE_ERR] = {
5632 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE_BAD,
5633 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
5634 			[MLX5_SQC_STATE_ERR]	= IB_QPS_ERR,
5635 			[MLX5_SQ_STATE_NA]	= IB_QPS_ERR,
5636 		},
5637 		[MLX5_RQ_STATE_NA] = {
5638 			[MLX5_SQC_STATE_RST]    = IB_QPS_RESET,
5639 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
5640 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE,
5641 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE_BAD,
5642 		},
5643 	};
5644 
5645 	*qp_state = sqrq_trans[rq_state][sq_state];
5646 
5647 	if (*qp_state == MLX5_QP_STATE_BAD) {
5648 		WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
5649 		     qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
5650 		     qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
5651 		return -EINVAL;
5652 	}
5653 
5654 	if (*qp_state == MLX5_QP_STATE)
5655 		*qp_state = qp->state;
5656 
5657 	return 0;
5658 }
5659 
5660 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
5661 				     struct mlx5_ib_qp *qp,
5662 				     u8 *raw_packet_qp_state)
5663 {
5664 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
5665 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
5666 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
5667 	int err;
5668 	u8 sq_state = MLX5_SQ_STATE_NA;
5669 	u8 rq_state = MLX5_RQ_STATE_NA;
5670 
5671 	if (qp->sq.wqe_cnt) {
5672 		err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
5673 		if (err)
5674 			return err;
5675 	}
5676 
5677 	if (qp->rq.wqe_cnt) {
5678 		err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
5679 		if (err)
5680 			return err;
5681 	}
5682 
5683 	return sqrq_state_to_qp_state(sq_state, rq_state, qp,
5684 				      raw_packet_qp_state);
5685 }
5686 
5687 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
5688 			 struct ib_qp_attr *qp_attr)
5689 {
5690 	int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
5691 	struct mlx5_qp_context *context;
5692 	int mlx5_state;
5693 	u32 *outb;
5694 	int err = 0;
5695 
5696 	outb = kzalloc(outlen, GFP_KERNEL);
5697 	if (!outb)
5698 		return -ENOMEM;
5699 
5700 	err = mlx5_core_qp_query(dev->mdev, &qp->trans_qp.base.mqp, outb,
5701 				 outlen);
5702 	if (err)
5703 		goto out;
5704 
5705 	/* FIXME: use MLX5_GET rather than mlx5_qp_context manual struct */
5706 	context = (struct mlx5_qp_context *)MLX5_ADDR_OF(query_qp_out, outb, qpc);
5707 
5708 	mlx5_state = be32_to_cpu(context->flags) >> 28;
5709 
5710 	qp->state		     = to_ib_qp_state(mlx5_state);
5711 	qp_attr->path_mtu	     = context->mtu_msgmax >> 5;
5712 	qp_attr->path_mig_state	     =
5713 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
5714 	qp_attr->qkey		     = be32_to_cpu(context->qkey);
5715 	qp_attr->rq_psn		     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
5716 	qp_attr->sq_psn		     = be32_to_cpu(context->next_send_psn) & 0xffffff;
5717 	qp_attr->dest_qp_num	     = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
5718 	qp_attr->qp_access_flags     =
5719 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
5720 
5721 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
5722 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
5723 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
5724 		qp_attr->alt_pkey_index =
5725 			be16_to_cpu(context->alt_path.pkey_index);
5726 		qp_attr->alt_port_num	=
5727 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
5728 	}
5729 
5730 	qp_attr->pkey_index = be16_to_cpu(context->pri_path.pkey_index);
5731 	qp_attr->port_num = context->pri_path.port;
5732 
5733 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
5734 	qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
5735 
5736 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
5737 
5738 	qp_attr->max_dest_rd_atomic =
5739 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
5740 	qp_attr->min_rnr_timer	    =
5741 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
5742 	qp_attr->timeout	    = context->pri_path.ackto_lt >> 3;
5743 	qp_attr->retry_cnt	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
5744 	qp_attr->rnr_retry	    = (be32_to_cpu(context->params1) >> 13) & 0x7;
5745 	qp_attr->alt_timeout	    = context->alt_path.ackto_lt >> 3;
5746 
5747 out:
5748 	kfree(outb);
5749 	return err;
5750 }
5751 
5752 static int mlx5_ib_dct_query_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *mqp,
5753 				struct ib_qp_attr *qp_attr, int qp_attr_mask,
5754 				struct ib_qp_init_attr *qp_init_attr)
5755 {
5756 	struct mlx5_core_dct	*dct = &mqp->dct.mdct;
5757 	u32 *out;
5758 	u32 access_flags = 0;
5759 	int outlen = MLX5_ST_SZ_BYTES(query_dct_out);
5760 	void *dctc;
5761 	int err;
5762 	int supported_mask = IB_QP_STATE |
5763 			     IB_QP_ACCESS_FLAGS |
5764 			     IB_QP_PORT |
5765 			     IB_QP_MIN_RNR_TIMER |
5766 			     IB_QP_AV |
5767 			     IB_QP_PATH_MTU |
5768 			     IB_QP_PKEY_INDEX;
5769 
5770 	if (qp_attr_mask & ~supported_mask)
5771 		return -EINVAL;
5772 	if (mqp->state != IB_QPS_RTR)
5773 		return -EINVAL;
5774 
5775 	out = kzalloc(outlen, GFP_KERNEL);
5776 	if (!out)
5777 		return -ENOMEM;
5778 
5779 	err = mlx5_core_dct_query(dev->mdev, dct, out, outlen);
5780 	if (err)
5781 		goto out;
5782 
5783 	dctc = MLX5_ADDR_OF(query_dct_out, out, dct_context_entry);
5784 
5785 	if (qp_attr_mask & IB_QP_STATE)
5786 		qp_attr->qp_state = IB_QPS_RTR;
5787 
5788 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
5789 		if (MLX5_GET(dctc, dctc, rre))
5790 			access_flags |= IB_ACCESS_REMOTE_READ;
5791 		if (MLX5_GET(dctc, dctc, rwe))
5792 			access_flags |= IB_ACCESS_REMOTE_WRITE;
5793 		if (MLX5_GET(dctc, dctc, rae))
5794 			access_flags |= IB_ACCESS_REMOTE_ATOMIC;
5795 		qp_attr->qp_access_flags = access_flags;
5796 	}
5797 
5798 	if (qp_attr_mask & IB_QP_PORT)
5799 		qp_attr->port_num = MLX5_GET(dctc, dctc, port);
5800 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER)
5801 		qp_attr->min_rnr_timer = MLX5_GET(dctc, dctc, min_rnr_nak);
5802 	if (qp_attr_mask & IB_QP_AV) {
5803 		qp_attr->ah_attr.grh.traffic_class = MLX5_GET(dctc, dctc, tclass);
5804 		qp_attr->ah_attr.grh.flow_label = MLX5_GET(dctc, dctc, flow_label);
5805 		qp_attr->ah_attr.grh.sgid_index = MLX5_GET(dctc, dctc, my_addr_index);
5806 		qp_attr->ah_attr.grh.hop_limit = MLX5_GET(dctc, dctc, hop_limit);
5807 	}
5808 	if (qp_attr_mask & IB_QP_PATH_MTU)
5809 		qp_attr->path_mtu = MLX5_GET(dctc, dctc, mtu);
5810 	if (qp_attr_mask & IB_QP_PKEY_INDEX)
5811 		qp_attr->pkey_index = MLX5_GET(dctc, dctc, pkey_index);
5812 out:
5813 	kfree(out);
5814 	return err;
5815 }
5816 
5817 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
5818 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
5819 {
5820 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
5821 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
5822 	int err = 0;
5823 	u8 raw_packet_qp_state;
5824 
5825 	if (ibqp->rwq_ind_tbl)
5826 		return -ENOSYS;
5827 
5828 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
5829 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
5830 					    qp_init_attr);
5831 
5832 	/* Not all of output fields are applicable, make sure to zero them */
5833 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
5834 	memset(qp_attr, 0, sizeof(*qp_attr));
5835 
5836 	if (unlikely(qp->qp_sub_type == MLX5_IB_QPT_DCT))
5837 		return mlx5_ib_dct_query_qp(dev, qp, qp_attr,
5838 					    qp_attr_mask, qp_init_attr);
5839 
5840 	mutex_lock(&qp->mutex);
5841 
5842 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
5843 	    qp->flags & MLX5_IB_QP_UNDERLAY) {
5844 		err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
5845 		if (err)
5846 			goto out;
5847 		qp->state = raw_packet_qp_state;
5848 		qp_attr->port_num = 1;
5849 	} else {
5850 		err = query_qp_attr(dev, qp, qp_attr);
5851 		if (err)
5852 			goto out;
5853 	}
5854 
5855 	qp_attr->qp_state	     = qp->state;
5856 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
5857 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
5858 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
5859 
5860 	if (!ibqp->uobject) {
5861 		qp_attr->cap.max_send_wr  = qp->sq.max_post;
5862 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
5863 		qp_init_attr->qp_context = ibqp->qp_context;
5864 	} else {
5865 		qp_attr->cap.max_send_wr  = 0;
5866 		qp_attr->cap.max_send_sge = 0;
5867 	}
5868 
5869 	qp_init_attr->qp_type = ibqp->qp_type;
5870 	qp_init_attr->recv_cq = ibqp->recv_cq;
5871 	qp_init_attr->send_cq = ibqp->send_cq;
5872 	qp_init_attr->srq = ibqp->srq;
5873 	qp_attr->cap.max_inline_data = qp->max_inline_data;
5874 
5875 	qp_init_attr->cap	     = qp_attr->cap;
5876 
5877 	qp_init_attr->create_flags = 0;
5878 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
5879 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
5880 
5881 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
5882 		qp_init_attr->create_flags |= IB_QP_CREATE_CROSS_CHANNEL;
5883 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
5884 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_SEND;
5885 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
5886 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_RECV;
5887 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
5888 		qp_init_attr->create_flags |= MLX5_IB_QP_CREATE_SQPN_QP1;
5889 
5890 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
5891 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
5892 
5893 out:
5894 	mutex_unlock(&qp->mutex);
5895 	return err;
5896 }
5897 
5898 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
5899 				   struct ib_udata *udata)
5900 {
5901 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5902 	struct mlx5_ib_xrcd *xrcd;
5903 	int err;
5904 
5905 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
5906 		return ERR_PTR(-ENOSYS);
5907 
5908 	xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
5909 	if (!xrcd)
5910 		return ERR_PTR(-ENOMEM);
5911 
5912 	err = mlx5_cmd_xrcd_alloc(dev->mdev, &xrcd->xrcdn, 0);
5913 	if (err) {
5914 		kfree(xrcd);
5915 		return ERR_PTR(-ENOMEM);
5916 	}
5917 
5918 	return &xrcd->ibxrcd;
5919 }
5920 
5921 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata)
5922 {
5923 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
5924 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
5925 	int err;
5926 
5927 	err = mlx5_cmd_xrcd_dealloc(dev->mdev, xrcdn, 0);
5928 	if (err)
5929 		mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
5930 
5931 	kfree(xrcd);
5932 	return 0;
5933 }
5934 
5935 static void mlx5_ib_wq_event(struct mlx5_core_qp *core_qp, int type)
5936 {
5937 	struct mlx5_ib_rwq *rwq = to_mibrwq(core_qp);
5938 	struct mlx5_ib_dev *dev = to_mdev(rwq->ibwq.device);
5939 	struct ib_event event;
5940 
5941 	if (rwq->ibwq.event_handler) {
5942 		event.device     = rwq->ibwq.device;
5943 		event.element.wq = &rwq->ibwq;
5944 		switch (type) {
5945 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
5946 			event.event = IB_EVENT_WQ_FATAL;
5947 			break;
5948 		default:
5949 			mlx5_ib_warn(dev, "Unexpected event type %d on WQ %06x\n", type, core_qp->qpn);
5950 			return;
5951 		}
5952 
5953 		rwq->ibwq.event_handler(&event, rwq->ibwq.wq_context);
5954 	}
5955 }
5956 
5957 static int set_delay_drop(struct mlx5_ib_dev *dev)
5958 {
5959 	int err = 0;
5960 
5961 	mutex_lock(&dev->delay_drop.lock);
5962 	if (dev->delay_drop.activate)
5963 		goto out;
5964 
5965 	err = mlx5_core_set_delay_drop(dev->mdev, dev->delay_drop.timeout);
5966 	if (err)
5967 		goto out;
5968 
5969 	dev->delay_drop.activate = true;
5970 out:
5971 	mutex_unlock(&dev->delay_drop.lock);
5972 
5973 	if (!err)
5974 		atomic_inc(&dev->delay_drop.rqs_cnt);
5975 	return err;
5976 }
5977 
5978 static int  create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
5979 		      struct ib_wq_init_attr *init_attr)
5980 {
5981 	struct mlx5_ib_dev *dev;
5982 	int has_net_offloads;
5983 	__be64 *rq_pas0;
5984 	void *in;
5985 	void *rqc;
5986 	void *wq;
5987 	int inlen;
5988 	int err;
5989 
5990 	dev = to_mdev(pd->device);
5991 
5992 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
5993 	in = kvzalloc(inlen, GFP_KERNEL);
5994 	if (!in)
5995 		return -ENOMEM;
5996 
5997 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
5998 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
5999 	MLX5_SET(rqc,  rqc, mem_rq_type,
6000 		 MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
6001 	MLX5_SET(rqc, rqc, user_index, rwq->user_index);
6002 	MLX5_SET(rqc,  rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
6003 	MLX5_SET(rqc,  rqc, state, MLX5_RQC_STATE_RST);
6004 	MLX5_SET(rqc,  rqc, flush_in_error_en, 1);
6005 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
6006 	MLX5_SET(wq, wq, wq_type,
6007 		 rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ ?
6008 		 MLX5_WQ_TYPE_CYCLIC_STRIDING_RQ : MLX5_WQ_TYPE_CYCLIC);
6009 	if (init_attr->create_flags & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
6010 		if (!MLX5_CAP_GEN(dev->mdev, end_pad)) {
6011 			mlx5_ib_dbg(dev, "Scatter end padding is not supported\n");
6012 			err = -EOPNOTSUPP;
6013 			goto out;
6014 		} else {
6015 			MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
6016 		}
6017 	}
6018 	MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
6019 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ) {
6020 		/*
6021 		 * In Firmware number of strides in each WQE is:
6022 		 *   "512 * 2^single_wqe_log_num_of_strides"
6023 		 * Values 3 to 8 are accepted as 10 to 15, 9 to 18 are
6024 		 * accepted as 0 to 9
6025 		 */
6026 		static const u8 fw_map[] = { 10, 11, 12, 13, 14, 15, 0, 1,
6027 					     2,  3,  4,  5,  6,  7,  8, 9 };
6028 		MLX5_SET(wq, wq, two_byte_shift_en, rwq->two_byte_shift_en);
6029 		MLX5_SET(wq, wq, log_wqe_stride_size,
6030 			 rwq->single_stride_log_num_of_bytes -
6031 			 MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES);
6032 		MLX5_SET(wq, wq, log_wqe_num_of_strides,
6033 			 fw_map[rwq->log_num_strides -
6034 				MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES]);
6035 	}
6036 	MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
6037 	MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
6038 	MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
6039 	MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
6040 	MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
6041 	MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
6042 	has_net_offloads = MLX5_CAP_GEN(dev->mdev, eth_net_offloads);
6043 	if (init_attr->create_flags & IB_WQ_FLAGS_CVLAN_STRIPPING) {
6044 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
6045 			mlx5_ib_dbg(dev, "VLAN offloads are not supported\n");
6046 			err = -EOPNOTSUPP;
6047 			goto out;
6048 		}
6049 	} else {
6050 		MLX5_SET(rqc, rqc, vsd, 1);
6051 	}
6052 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS) {
6053 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, scatter_fcs))) {
6054 			mlx5_ib_dbg(dev, "Scatter FCS is not supported\n");
6055 			err = -EOPNOTSUPP;
6056 			goto out;
6057 		}
6058 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
6059 	}
6060 	if (init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
6061 		if (!(dev->ib_dev.attrs.raw_packet_caps &
6062 		      IB_RAW_PACKET_CAP_DELAY_DROP)) {
6063 			mlx5_ib_dbg(dev, "Delay drop is not supported\n");
6064 			err = -EOPNOTSUPP;
6065 			goto out;
6066 		}
6067 		MLX5_SET(rqc, rqc, delay_drop_en, 1);
6068 	}
6069 	rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
6070 	mlx5_ib_populate_pas(dev, rwq->umem, rwq->page_shift, rq_pas0, 0);
6071 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rwq->core_qp);
6072 	if (!err && init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
6073 		err = set_delay_drop(dev);
6074 		if (err) {
6075 			mlx5_ib_warn(dev, "Failed to enable delay drop err=%d\n",
6076 				     err);
6077 			mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
6078 		} else {
6079 			rwq->create_flags |= MLX5_IB_WQ_FLAGS_DELAY_DROP;
6080 		}
6081 	}
6082 out:
6083 	kvfree(in);
6084 	return err;
6085 }
6086 
6087 static int set_user_rq_size(struct mlx5_ib_dev *dev,
6088 			    struct ib_wq_init_attr *wq_init_attr,
6089 			    struct mlx5_ib_create_wq *ucmd,
6090 			    struct mlx5_ib_rwq *rwq)
6091 {
6092 	/* Sanity check RQ size before proceeding */
6093 	if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
6094 		return -EINVAL;
6095 
6096 	if (!ucmd->rq_wqe_count)
6097 		return -EINVAL;
6098 
6099 	rwq->wqe_count = ucmd->rq_wqe_count;
6100 	rwq->wqe_shift = ucmd->rq_wqe_shift;
6101 	if (check_shl_overflow(rwq->wqe_count, rwq->wqe_shift, &rwq->buf_size))
6102 		return -EINVAL;
6103 
6104 	rwq->log_rq_stride = rwq->wqe_shift;
6105 	rwq->log_rq_size = ilog2(rwq->wqe_count);
6106 	return 0;
6107 }
6108 
6109 static bool log_of_strides_valid(struct mlx5_ib_dev *dev, u32 log_num_strides)
6110 {
6111 	if ((log_num_strides > MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES) ||
6112 	    (log_num_strides < MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES))
6113 		return false;
6114 
6115 	if (!MLX5_CAP_GEN(dev->mdev, ext_stride_num_range) &&
6116 	    (log_num_strides < MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES))
6117 		return false;
6118 
6119 	return true;
6120 }
6121 
6122 static int prepare_user_rq(struct ib_pd *pd,
6123 			   struct ib_wq_init_attr *init_attr,
6124 			   struct ib_udata *udata,
6125 			   struct mlx5_ib_rwq *rwq)
6126 {
6127 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
6128 	struct mlx5_ib_create_wq ucmd = {};
6129 	int err;
6130 	size_t required_cmd_sz;
6131 
6132 	required_cmd_sz = offsetof(typeof(ucmd), single_stride_log_num_of_bytes)
6133 		+ sizeof(ucmd.single_stride_log_num_of_bytes);
6134 	if (udata->inlen < required_cmd_sz) {
6135 		mlx5_ib_dbg(dev, "invalid inlen\n");
6136 		return -EINVAL;
6137 	}
6138 
6139 	if (udata->inlen > sizeof(ucmd) &&
6140 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
6141 				 udata->inlen - sizeof(ucmd))) {
6142 		mlx5_ib_dbg(dev, "inlen is not supported\n");
6143 		return -EOPNOTSUPP;
6144 	}
6145 
6146 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
6147 		mlx5_ib_dbg(dev, "copy failed\n");
6148 		return -EFAULT;
6149 	}
6150 
6151 	if (ucmd.comp_mask & (~MLX5_IB_CREATE_WQ_STRIDING_RQ)) {
6152 		mlx5_ib_dbg(dev, "invalid comp mask\n");
6153 		return -EOPNOTSUPP;
6154 	} else if (ucmd.comp_mask & MLX5_IB_CREATE_WQ_STRIDING_RQ) {
6155 		if (!MLX5_CAP_GEN(dev->mdev, striding_rq)) {
6156 			mlx5_ib_dbg(dev, "Striding RQ is not supported\n");
6157 			return -EOPNOTSUPP;
6158 		}
6159 		if ((ucmd.single_stride_log_num_of_bytes <
6160 		    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES) ||
6161 		    (ucmd.single_stride_log_num_of_bytes >
6162 		     MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES)) {
6163 			mlx5_ib_dbg(dev, "Invalid log stride size (%u. Range is %u - %u)\n",
6164 				    ucmd.single_stride_log_num_of_bytes,
6165 				    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES,
6166 				    MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES);
6167 			return -EINVAL;
6168 		}
6169 		if (!log_of_strides_valid(dev,
6170 					  ucmd.single_wqe_log_num_of_strides)) {
6171 			mlx5_ib_dbg(
6172 				dev,
6173 				"Invalid log num strides (%u. Range is %u - %u)\n",
6174 				ucmd.single_wqe_log_num_of_strides,
6175 				MLX5_CAP_GEN(dev->mdev, ext_stride_num_range) ?
6176 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES :
6177 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES,
6178 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES);
6179 			return -EINVAL;
6180 		}
6181 		rwq->single_stride_log_num_of_bytes =
6182 			ucmd.single_stride_log_num_of_bytes;
6183 		rwq->log_num_strides = ucmd.single_wqe_log_num_of_strides;
6184 		rwq->two_byte_shift_en = !!ucmd.two_byte_shift_en;
6185 		rwq->create_flags |= MLX5_IB_WQ_FLAGS_STRIDING_RQ;
6186 	}
6187 
6188 	err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
6189 	if (err) {
6190 		mlx5_ib_dbg(dev, "err %d\n", err);
6191 		return err;
6192 	}
6193 
6194 	err = create_user_rq(dev, pd, udata, rwq, &ucmd);
6195 	if (err) {
6196 		mlx5_ib_dbg(dev, "err %d\n", err);
6197 		return err;
6198 	}
6199 
6200 	rwq->user_index = ucmd.user_index;
6201 	return 0;
6202 }
6203 
6204 struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
6205 				struct ib_wq_init_attr *init_attr,
6206 				struct ib_udata *udata)
6207 {
6208 	struct mlx5_ib_dev *dev;
6209 	struct mlx5_ib_rwq *rwq;
6210 	struct mlx5_ib_create_wq_resp resp = {};
6211 	size_t min_resp_len;
6212 	int err;
6213 
6214 	if (!udata)
6215 		return ERR_PTR(-ENOSYS);
6216 
6217 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
6218 	if (udata->outlen && udata->outlen < min_resp_len)
6219 		return ERR_PTR(-EINVAL);
6220 
6221 	if (!capable(CAP_SYS_RAWIO) &&
6222 	    init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP)
6223 		return ERR_PTR(-EPERM);
6224 
6225 	dev = to_mdev(pd->device);
6226 	switch (init_attr->wq_type) {
6227 	case IB_WQT_RQ:
6228 		rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
6229 		if (!rwq)
6230 			return ERR_PTR(-ENOMEM);
6231 		err = prepare_user_rq(pd, init_attr, udata, rwq);
6232 		if (err)
6233 			goto err;
6234 		err = create_rq(rwq, pd, init_attr);
6235 		if (err)
6236 			goto err_user_rq;
6237 		break;
6238 	default:
6239 		mlx5_ib_dbg(dev, "unsupported wq type %d\n",
6240 			    init_attr->wq_type);
6241 		return ERR_PTR(-EINVAL);
6242 	}
6243 
6244 	rwq->ibwq.wq_num = rwq->core_qp.qpn;
6245 	rwq->ibwq.state = IB_WQS_RESET;
6246 	if (udata->outlen) {
6247 		resp.response_length = offsetof(typeof(resp), response_length) +
6248 				sizeof(resp.response_length);
6249 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
6250 		if (err)
6251 			goto err_copy;
6252 	}
6253 
6254 	rwq->core_qp.event = mlx5_ib_wq_event;
6255 	rwq->ibwq.event_handler = init_attr->event_handler;
6256 	return &rwq->ibwq;
6257 
6258 err_copy:
6259 	mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
6260 err_user_rq:
6261 	destroy_user_rq(dev, pd, rwq, udata);
6262 err:
6263 	kfree(rwq);
6264 	return ERR_PTR(err);
6265 }
6266 
6267 void mlx5_ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata)
6268 {
6269 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
6270 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
6271 
6272 	mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
6273 	destroy_user_rq(dev, wq->pd, rwq, udata);
6274 	kfree(rwq);
6275 }
6276 
6277 struct ib_rwq_ind_table *mlx5_ib_create_rwq_ind_table(struct ib_device *device,
6278 						      struct ib_rwq_ind_table_init_attr *init_attr,
6279 						      struct ib_udata *udata)
6280 {
6281 	struct mlx5_ib_dev *dev = to_mdev(device);
6282 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl;
6283 	int sz = 1 << init_attr->log_ind_tbl_size;
6284 	struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
6285 	size_t min_resp_len;
6286 	int inlen;
6287 	int err;
6288 	int i;
6289 	u32 *in;
6290 	void *rqtc;
6291 
6292 	if (udata->inlen > 0 &&
6293 	    !ib_is_udata_cleared(udata, 0,
6294 				 udata->inlen))
6295 		return ERR_PTR(-EOPNOTSUPP);
6296 
6297 	if (init_attr->log_ind_tbl_size >
6298 	    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size)) {
6299 		mlx5_ib_dbg(dev, "log_ind_tbl_size = %d is bigger than supported = %d\n",
6300 			    init_attr->log_ind_tbl_size,
6301 			    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size));
6302 		return ERR_PTR(-EINVAL);
6303 	}
6304 
6305 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
6306 	if (udata->outlen && udata->outlen < min_resp_len)
6307 		return ERR_PTR(-EINVAL);
6308 
6309 	rwq_ind_tbl = kzalloc(sizeof(*rwq_ind_tbl), GFP_KERNEL);
6310 	if (!rwq_ind_tbl)
6311 		return ERR_PTR(-ENOMEM);
6312 
6313 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
6314 	in = kvzalloc(inlen, GFP_KERNEL);
6315 	if (!in) {
6316 		err = -ENOMEM;
6317 		goto err;
6318 	}
6319 
6320 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
6321 
6322 	MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
6323 	MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
6324 
6325 	for (i = 0; i < sz; i++)
6326 		MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
6327 
6328 	rwq_ind_tbl->uid = to_mpd(init_attr->ind_tbl[0]->pd)->uid;
6329 	MLX5_SET(create_rqt_in, in, uid, rwq_ind_tbl->uid);
6330 
6331 	err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
6332 	kvfree(in);
6333 
6334 	if (err)
6335 		goto err;
6336 
6337 	rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
6338 	if (udata->outlen) {
6339 		resp.response_length = offsetof(typeof(resp), response_length) +
6340 					sizeof(resp.response_length);
6341 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
6342 		if (err)
6343 			goto err_copy;
6344 	}
6345 
6346 	return &rwq_ind_tbl->ib_rwq_ind_tbl;
6347 
6348 err_copy:
6349 	mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
6350 err:
6351 	kfree(rwq_ind_tbl);
6352 	return ERR_PTR(err);
6353 }
6354 
6355 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
6356 {
6357 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
6358 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
6359 
6360 	mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
6361 
6362 	kfree(rwq_ind_tbl);
6363 	return 0;
6364 }
6365 
6366 int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
6367 		      u32 wq_attr_mask, struct ib_udata *udata)
6368 {
6369 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
6370 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
6371 	struct mlx5_ib_modify_wq ucmd = {};
6372 	size_t required_cmd_sz;
6373 	int curr_wq_state;
6374 	int wq_state;
6375 	int inlen;
6376 	int err;
6377 	void *rqc;
6378 	void *in;
6379 
6380 	required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
6381 	if (udata->inlen < required_cmd_sz)
6382 		return -EINVAL;
6383 
6384 	if (udata->inlen > sizeof(ucmd) &&
6385 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
6386 				 udata->inlen - sizeof(ucmd)))
6387 		return -EOPNOTSUPP;
6388 
6389 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
6390 		return -EFAULT;
6391 
6392 	if (ucmd.comp_mask || ucmd.reserved)
6393 		return -EOPNOTSUPP;
6394 
6395 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
6396 	in = kvzalloc(inlen, GFP_KERNEL);
6397 	if (!in)
6398 		return -ENOMEM;
6399 
6400 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
6401 
6402 	curr_wq_state = (wq_attr_mask & IB_WQ_CUR_STATE) ?
6403 		wq_attr->curr_wq_state : wq->state;
6404 	wq_state = (wq_attr_mask & IB_WQ_STATE) ?
6405 		wq_attr->wq_state : curr_wq_state;
6406 	if (curr_wq_state == IB_WQS_ERR)
6407 		curr_wq_state = MLX5_RQC_STATE_ERR;
6408 	if (wq_state == IB_WQS_ERR)
6409 		wq_state = MLX5_RQC_STATE_ERR;
6410 	MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
6411 	MLX5_SET(modify_rq_in, in, uid, to_mpd(wq->pd)->uid);
6412 	MLX5_SET(rqc, rqc, state, wq_state);
6413 
6414 	if (wq_attr_mask & IB_WQ_FLAGS) {
6415 		if (wq_attr->flags_mask & IB_WQ_FLAGS_CVLAN_STRIPPING) {
6416 			if (!(MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
6417 			      MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
6418 				mlx5_ib_dbg(dev, "VLAN offloads are not "
6419 					    "supported\n");
6420 				err = -EOPNOTSUPP;
6421 				goto out;
6422 			}
6423 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
6424 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD);
6425 			MLX5_SET(rqc, rqc, vsd,
6426 				 (wq_attr->flags & IB_WQ_FLAGS_CVLAN_STRIPPING) ? 0 : 1);
6427 		}
6428 
6429 		if (wq_attr->flags_mask & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
6430 			mlx5_ib_dbg(dev, "Modifying scatter end padding is not supported\n");
6431 			err = -EOPNOTSUPP;
6432 			goto out;
6433 		}
6434 	}
6435 
6436 	if (curr_wq_state == IB_WQS_RESET && wq_state == IB_WQS_RDY) {
6437 		u16 set_id;
6438 
6439 		set_id = mlx5_ib_get_counters_id(dev, 0);
6440 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
6441 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
6442 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
6443 			MLX5_SET(rqc, rqc, counter_set_id, set_id);
6444 		} else
6445 			dev_info_once(
6446 				&dev->ib_dev.dev,
6447 				"Receive WQ counters are not supported on current FW\n");
6448 	}
6449 
6450 	err = mlx5_core_modify_rq(dev->mdev, rwq->core_qp.qpn, in, inlen);
6451 	if (!err)
6452 		rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;
6453 
6454 out:
6455 	kvfree(in);
6456 	return err;
6457 }
6458 
6459 struct mlx5_ib_drain_cqe {
6460 	struct ib_cqe cqe;
6461 	struct completion done;
6462 };
6463 
6464 static void mlx5_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
6465 {
6466 	struct mlx5_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
6467 						     struct mlx5_ib_drain_cqe,
6468 						     cqe);
6469 
6470 	complete(&cqe->done);
6471 }
6472 
6473 /* This function returns only once the drained WR was completed */
6474 static void handle_drain_completion(struct ib_cq *cq,
6475 				    struct mlx5_ib_drain_cqe *sdrain,
6476 				    struct mlx5_ib_dev *dev)
6477 {
6478 	struct mlx5_core_dev *mdev = dev->mdev;
6479 
6480 	if (cq->poll_ctx == IB_POLL_DIRECT) {
6481 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
6482 			ib_process_cq_direct(cq, -1);
6483 		return;
6484 	}
6485 
6486 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
6487 		struct mlx5_ib_cq *mcq = to_mcq(cq);
6488 		bool triggered = false;
6489 		unsigned long flags;
6490 
6491 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
6492 		/* Make sure that the CQ handler won't run if wasn't run yet */
6493 		if (!mcq->mcq.reset_notify_added)
6494 			mcq->mcq.reset_notify_added = 1;
6495 		else
6496 			triggered = true;
6497 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
6498 
6499 		if (triggered) {
6500 			/* Wait for any scheduled/running task to be ended */
6501 			switch (cq->poll_ctx) {
6502 			case IB_POLL_SOFTIRQ:
6503 				irq_poll_disable(&cq->iop);
6504 				irq_poll_enable(&cq->iop);
6505 				break;
6506 			case IB_POLL_WORKQUEUE:
6507 				cancel_work_sync(&cq->work);
6508 				break;
6509 			default:
6510 				WARN_ON_ONCE(1);
6511 			}
6512 		}
6513 
6514 		/* Run the CQ handler - this makes sure that the drain WR will
6515 		 * be processed if wasn't processed yet.
6516 		 */
6517 		mcq->mcq.comp(&mcq->mcq, NULL);
6518 	}
6519 
6520 	wait_for_completion(&sdrain->done);
6521 }
6522 
6523 void mlx5_ib_drain_sq(struct ib_qp *qp)
6524 {
6525 	struct ib_cq *cq = qp->send_cq;
6526 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
6527 	struct mlx5_ib_drain_cqe sdrain;
6528 	const struct ib_send_wr *bad_swr;
6529 	struct ib_rdma_wr swr = {
6530 		.wr = {
6531 			.next = NULL,
6532 			{ .wr_cqe	= &sdrain.cqe, },
6533 			.opcode	= IB_WR_RDMA_WRITE,
6534 		},
6535 	};
6536 	int ret;
6537 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
6538 	struct mlx5_core_dev *mdev = dev->mdev;
6539 
6540 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
6541 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
6542 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
6543 		return;
6544 	}
6545 
6546 	sdrain.cqe.done = mlx5_ib_drain_qp_done;
6547 	init_completion(&sdrain.done);
6548 
6549 	ret = _mlx5_ib_post_send(qp, &swr.wr, &bad_swr, true);
6550 	if (ret) {
6551 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
6552 		return;
6553 	}
6554 
6555 	handle_drain_completion(cq, &sdrain, dev);
6556 }
6557 
6558 void mlx5_ib_drain_rq(struct ib_qp *qp)
6559 {
6560 	struct ib_cq *cq = qp->recv_cq;
6561 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
6562 	struct mlx5_ib_drain_cqe rdrain;
6563 	struct ib_recv_wr rwr = {};
6564 	const struct ib_recv_wr *bad_rwr;
6565 	int ret;
6566 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
6567 	struct mlx5_core_dev *mdev = dev->mdev;
6568 
6569 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
6570 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
6571 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
6572 		return;
6573 	}
6574 
6575 	rwr.wr_cqe = &rdrain.cqe;
6576 	rdrain.cqe.done = mlx5_ib_drain_qp_done;
6577 	init_completion(&rdrain.done);
6578 
6579 	ret = _mlx5_ib_post_recv(qp, &rwr, &bad_rwr, true);
6580 	if (ret) {
6581 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
6582 		return;
6583 	}
6584 
6585 	handle_drain_completion(cq, &rdrain, dev);
6586 }
6587 
6588 /**
6589  * Bind a qp to a counter. If @counter is NULL then bind the qp to
6590  * the default counter
6591  */
6592 int mlx5_ib_qp_set_counter(struct ib_qp *qp, struct rdma_counter *counter)
6593 {
6594 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
6595 	struct mlx5_ib_qp *mqp = to_mqp(qp);
6596 	int err = 0;
6597 
6598 	mutex_lock(&mqp->mutex);
6599 	if (mqp->state == IB_QPS_RESET) {
6600 		qp->counter = counter;
6601 		goto out;
6602 	}
6603 
6604 	if (!MLX5_CAP_GEN(dev->mdev, rts2rts_qp_counters_set_id)) {
6605 		err = -EOPNOTSUPP;
6606 		goto out;
6607 	}
6608 
6609 	if (mqp->state == IB_QPS_RTS) {
6610 		err = __mlx5_ib_qp_set_counter(qp, counter);
6611 		if (!err)
6612 			qp->counter = counter;
6613 
6614 		goto out;
6615 	}
6616 
6617 	mqp->counter_pending = 1;
6618 	qp->counter = counter;
6619 
6620 out:
6621 	mutex_unlock(&mqp->mutex);
6622 	return err;
6623 }
6624