xref: /linux/drivers/infiniband/hw/mlx5/qp.c (revision 1f2367a39f17bd553a75e179a747f9b257bc9478)
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 <linux/mlx5/fs.h>
38 #include "mlx5_ib.h"
39 #include "ib_rep.h"
40 #include "cmd.h"
41 
42 /* not supported currently */
43 static int wq_signature;
44 
45 enum {
46 	MLX5_IB_ACK_REQ_FREQ	= 8,
47 };
48 
49 enum {
50 	MLX5_IB_DEFAULT_SCHED_QUEUE	= 0x83,
51 	MLX5_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
52 	MLX5_IB_LINK_TYPE_IB		= 0,
53 	MLX5_IB_LINK_TYPE_ETH		= 1
54 };
55 
56 enum {
57 	MLX5_IB_SQ_STRIDE	= 6,
58 	MLX5_IB_SQ_UMR_INLINE_THRESHOLD = 64,
59 };
60 
61 static const u32 mlx5_ib_opcode[] = {
62 	[IB_WR_SEND]				= MLX5_OPCODE_SEND,
63 	[IB_WR_LSO]				= MLX5_OPCODE_LSO,
64 	[IB_WR_SEND_WITH_IMM]			= MLX5_OPCODE_SEND_IMM,
65 	[IB_WR_RDMA_WRITE]			= MLX5_OPCODE_RDMA_WRITE,
66 	[IB_WR_RDMA_WRITE_WITH_IMM]		= MLX5_OPCODE_RDMA_WRITE_IMM,
67 	[IB_WR_RDMA_READ]			= MLX5_OPCODE_RDMA_READ,
68 	[IB_WR_ATOMIC_CMP_AND_SWP]		= MLX5_OPCODE_ATOMIC_CS,
69 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= MLX5_OPCODE_ATOMIC_FA,
70 	[IB_WR_SEND_WITH_INV]			= MLX5_OPCODE_SEND_INVAL,
71 	[IB_WR_LOCAL_INV]			= MLX5_OPCODE_UMR,
72 	[IB_WR_REG_MR]				= MLX5_OPCODE_UMR,
73 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= MLX5_OPCODE_ATOMIC_MASKED_CS,
74 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= MLX5_OPCODE_ATOMIC_MASKED_FA,
75 	[MLX5_IB_WR_UMR]			= MLX5_OPCODE_UMR,
76 };
77 
78 struct mlx5_wqe_eth_pad {
79 	u8 rsvd0[16];
80 };
81 
82 enum raw_qp_set_mask_map {
83 	MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID		= 1UL << 0,
84 	MLX5_RAW_QP_RATE_LIMIT			= 1UL << 1,
85 };
86 
87 struct mlx5_modify_raw_qp_param {
88 	u16 operation;
89 
90 	u32 set_mask; /* raw_qp_set_mask_map */
91 
92 	struct mlx5_rate_limit rl;
93 
94 	u8 rq_q_ctr_id;
95 };
96 
97 static void get_cqs(enum ib_qp_type qp_type,
98 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
99 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq);
100 
101 static int is_qp0(enum ib_qp_type qp_type)
102 {
103 	return qp_type == IB_QPT_SMI;
104 }
105 
106 static int is_sqp(enum ib_qp_type qp_type)
107 {
108 	return is_qp0(qp_type) || is_qp1(qp_type);
109 }
110 
111 /**
112  * mlx5_ib_read_user_wqe_common() - Copy a WQE (or part of) from user WQ
113  * to kernel buffer
114  *
115  * @umem: User space memory where the WQ is
116  * @buffer: buffer to copy to
117  * @buflen: buffer length
118  * @wqe_index: index of WQE to copy from
119  * @wq_offset: offset to start of WQ
120  * @wq_wqe_cnt: number of WQEs in WQ
121  * @wq_wqe_shift: log2 of WQE size
122  * @bcnt: number of bytes to copy
123  * @bytes_copied: number of bytes to copy (return value)
124  *
125  * Copies from start of WQE bcnt or less bytes.
126  * Does not gurantee to copy the entire WQE.
127  *
128  * Return: zero on success, or an error code.
129  */
130 static int mlx5_ib_read_user_wqe_common(struct ib_umem *umem,
131 					void *buffer,
132 					u32 buflen,
133 					int wqe_index,
134 					int wq_offset,
135 					int wq_wqe_cnt,
136 					int wq_wqe_shift,
137 					int bcnt,
138 					size_t *bytes_copied)
139 {
140 	size_t offset = wq_offset + ((wqe_index % wq_wqe_cnt) << wq_wqe_shift);
141 	size_t wq_end = wq_offset + (wq_wqe_cnt << wq_wqe_shift);
142 	size_t copy_length;
143 	int ret;
144 
145 	/* don't copy more than requested, more than buffer length or
146 	 * beyond WQ end
147 	 */
148 	copy_length = min_t(u32, buflen, wq_end - offset);
149 	copy_length = min_t(u32, copy_length, bcnt);
150 
151 	ret = ib_umem_copy_from(buffer, umem, offset, copy_length);
152 	if (ret)
153 		return ret;
154 
155 	if (!ret && bytes_copied)
156 		*bytes_copied = copy_length;
157 
158 	return 0;
159 }
160 
161 int mlx5_ib_read_user_wqe_sq(struct mlx5_ib_qp *qp,
162 			     int wqe_index,
163 			     void *buffer,
164 			     int buflen,
165 			     size_t *bc)
166 {
167 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
168 	struct ib_umem *umem = base->ubuffer.umem;
169 	struct mlx5_ib_wq *wq = &qp->sq;
170 	struct mlx5_wqe_ctrl_seg *ctrl;
171 	size_t bytes_copied;
172 	size_t bytes_copied2;
173 	size_t wqe_length;
174 	int ret;
175 	int ds;
176 
177 	if (buflen < sizeof(*ctrl))
178 		return -EINVAL;
179 
180 	/* at first read as much as possible */
181 	ret = mlx5_ib_read_user_wqe_common(umem,
182 					   buffer,
183 					   buflen,
184 					   wqe_index,
185 					   wq->offset,
186 					   wq->wqe_cnt,
187 					   wq->wqe_shift,
188 					   buflen,
189 					   &bytes_copied);
190 	if (ret)
191 		return ret;
192 
193 	/* we need at least control segment size to proceed */
194 	if (bytes_copied < sizeof(*ctrl))
195 		return -EINVAL;
196 
197 	ctrl = buffer;
198 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
199 	wqe_length = ds * MLX5_WQE_DS_UNITS;
200 
201 	/* if we copied enough then we are done */
202 	if (bytes_copied >= wqe_length) {
203 		*bc = bytes_copied;
204 		return 0;
205 	}
206 
207 	/* otherwise this a wrapped around wqe
208 	 * so read the remaining bytes starting
209 	 * from  wqe_index 0
210 	 */
211 	ret = mlx5_ib_read_user_wqe_common(umem,
212 					   buffer + bytes_copied,
213 					   buflen - bytes_copied,
214 					   0,
215 					   wq->offset,
216 					   wq->wqe_cnt,
217 					   wq->wqe_shift,
218 					   wqe_length - bytes_copied,
219 					   &bytes_copied2);
220 
221 	if (ret)
222 		return ret;
223 	*bc = bytes_copied + bytes_copied2;
224 	return 0;
225 }
226 
227 int mlx5_ib_read_user_wqe_rq(struct mlx5_ib_qp *qp,
228 			     int wqe_index,
229 			     void *buffer,
230 			     int buflen,
231 			     size_t *bc)
232 {
233 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
234 	struct ib_umem *umem = base->ubuffer.umem;
235 	struct mlx5_ib_wq *wq = &qp->rq;
236 	size_t bytes_copied;
237 	int ret;
238 
239 	ret = mlx5_ib_read_user_wqe_common(umem,
240 					   buffer,
241 					   buflen,
242 					   wqe_index,
243 					   wq->offset,
244 					   wq->wqe_cnt,
245 					   wq->wqe_shift,
246 					   buflen,
247 					   &bytes_copied);
248 
249 	if (ret)
250 		return ret;
251 	*bc = bytes_copied;
252 	return 0;
253 }
254 
255 int mlx5_ib_read_user_wqe_srq(struct mlx5_ib_srq *srq,
256 			      int wqe_index,
257 			      void *buffer,
258 			      int buflen,
259 			      size_t *bc)
260 {
261 	struct ib_umem *umem = srq->umem;
262 	size_t bytes_copied;
263 	int ret;
264 
265 	ret = mlx5_ib_read_user_wqe_common(umem,
266 					   buffer,
267 					   buflen,
268 					   wqe_index,
269 					   0,
270 					   srq->msrq.max,
271 					   srq->msrq.wqe_shift,
272 					   buflen,
273 					   &bytes_copied);
274 
275 	if (ret)
276 		return ret;
277 	*bc = bytes_copied;
278 	return 0;
279 }
280 
281 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
282 {
283 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
284 	struct ib_event event;
285 
286 	if (type == MLX5_EVENT_TYPE_PATH_MIG) {
287 		/* This event is only valid for trans_qps */
288 		to_mibqp(qp)->port = to_mibqp(qp)->trans_qp.alt_port;
289 	}
290 
291 	if (ibqp->event_handler) {
292 		event.device     = ibqp->device;
293 		event.element.qp = ibqp;
294 		switch (type) {
295 		case MLX5_EVENT_TYPE_PATH_MIG:
296 			event.event = IB_EVENT_PATH_MIG;
297 			break;
298 		case MLX5_EVENT_TYPE_COMM_EST:
299 			event.event = IB_EVENT_COMM_EST;
300 			break;
301 		case MLX5_EVENT_TYPE_SQ_DRAINED:
302 			event.event = IB_EVENT_SQ_DRAINED;
303 			break;
304 		case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
305 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
306 			break;
307 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
308 			event.event = IB_EVENT_QP_FATAL;
309 			break;
310 		case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
311 			event.event = IB_EVENT_PATH_MIG_ERR;
312 			break;
313 		case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
314 			event.event = IB_EVENT_QP_REQ_ERR;
315 			break;
316 		case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
317 			event.event = IB_EVENT_QP_ACCESS_ERR;
318 			break;
319 		default:
320 			pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
321 			return;
322 		}
323 
324 		ibqp->event_handler(&event, ibqp->qp_context);
325 	}
326 }
327 
328 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
329 		       int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
330 {
331 	int wqe_size;
332 	int wq_size;
333 
334 	/* Sanity check RQ size before proceeding */
335 	if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
336 		return -EINVAL;
337 
338 	if (!has_rq) {
339 		qp->rq.max_gs = 0;
340 		qp->rq.wqe_cnt = 0;
341 		qp->rq.wqe_shift = 0;
342 		cap->max_recv_wr = 0;
343 		cap->max_recv_sge = 0;
344 	} else {
345 		if (ucmd) {
346 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
347 			if (ucmd->rq_wqe_shift > BITS_PER_BYTE * sizeof(ucmd->rq_wqe_shift))
348 				return -EINVAL;
349 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
350 			if ((1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) < qp->wq_sig)
351 				return -EINVAL;
352 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
353 			qp->rq.max_post = qp->rq.wqe_cnt;
354 		} else {
355 			wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
356 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
357 			wqe_size = roundup_pow_of_two(wqe_size);
358 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
359 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
360 			qp->rq.wqe_cnt = wq_size / wqe_size;
361 			if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
362 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
363 					    wqe_size,
364 					    MLX5_CAP_GEN(dev->mdev,
365 							 max_wqe_sz_rq));
366 				return -EINVAL;
367 			}
368 			qp->rq.wqe_shift = ilog2(wqe_size);
369 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
370 			qp->rq.max_post = qp->rq.wqe_cnt;
371 		}
372 	}
373 
374 	return 0;
375 }
376 
377 static int sq_overhead(struct ib_qp_init_attr *attr)
378 {
379 	int size = 0;
380 
381 	switch (attr->qp_type) {
382 	case IB_QPT_XRC_INI:
383 		size += sizeof(struct mlx5_wqe_xrc_seg);
384 		/* fall through */
385 	case IB_QPT_RC:
386 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
387 			max(sizeof(struct mlx5_wqe_atomic_seg) +
388 			    sizeof(struct mlx5_wqe_raddr_seg),
389 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
390 			    sizeof(struct mlx5_mkey_seg) +
391 			    MLX5_IB_SQ_UMR_INLINE_THRESHOLD /
392 			    MLX5_IB_UMR_OCTOWORD);
393 		break;
394 
395 	case IB_QPT_XRC_TGT:
396 		return 0;
397 
398 	case IB_QPT_UC:
399 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
400 			max(sizeof(struct mlx5_wqe_raddr_seg),
401 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
402 			    sizeof(struct mlx5_mkey_seg));
403 		break;
404 
405 	case IB_QPT_UD:
406 		if (attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
407 			size += sizeof(struct mlx5_wqe_eth_pad) +
408 				sizeof(struct mlx5_wqe_eth_seg);
409 		/* fall through */
410 	case IB_QPT_SMI:
411 	case MLX5_IB_QPT_HW_GSI:
412 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
413 			sizeof(struct mlx5_wqe_datagram_seg);
414 		break;
415 
416 	case MLX5_IB_QPT_REG_UMR:
417 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
418 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
419 			sizeof(struct mlx5_mkey_seg);
420 		break;
421 
422 	default:
423 		return -EINVAL;
424 	}
425 
426 	return size;
427 }
428 
429 static int calc_send_wqe(struct ib_qp_init_attr *attr)
430 {
431 	int inl_size = 0;
432 	int size;
433 
434 	size = sq_overhead(attr);
435 	if (size < 0)
436 		return size;
437 
438 	if (attr->cap.max_inline_data) {
439 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
440 			attr->cap.max_inline_data;
441 	}
442 
443 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
444 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN &&
445 	    ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
446 			return MLX5_SIG_WQE_SIZE;
447 	else
448 		return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
449 }
450 
451 static int get_send_sge(struct ib_qp_init_attr *attr, int wqe_size)
452 {
453 	int max_sge;
454 
455 	if (attr->qp_type == IB_QPT_RC)
456 		max_sge = (min_t(int, wqe_size, 512) -
457 			   sizeof(struct mlx5_wqe_ctrl_seg) -
458 			   sizeof(struct mlx5_wqe_raddr_seg)) /
459 			sizeof(struct mlx5_wqe_data_seg);
460 	else if (attr->qp_type == IB_QPT_XRC_INI)
461 		max_sge = (min_t(int, wqe_size, 512) -
462 			   sizeof(struct mlx5_wqe_ctrl_seg) -
463 			   sizeof(struct mlx5_wqe_xrc_seg) -
464 			   sizeof(struct mlx5_wqe_raddr_seg)) /
465 			sizeof(struct mlx5_wqe_data_seg);
466 	else
467 		max_sge = (wqe_size - sq_overhead(attr)) /
468 			sizeof(struct mlx5_wqe_data_seg);
469 
470 	return min_t(int, max_sge, wqe_size - sq_overhead(attr) /
471 		     sizeof(struct mlx5_wqe_data_seg));
472 }
473 
474 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
475 			struct mlx5_ib_qp *qp)
476 {
477 	int wqe_size;
478 	int wq_size;
479 
480 	if (!attr->cap.max_send_wr)
481 		return 0;
482 
483 	wqe_size = calc_send_wqe(attr);
484 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
485 	if (wqe_size < 0)
486 		return wqe_size;
487 
488 	if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
489 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
490 			    wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
491 		return -EINVAL;
492 	}
493 
494 	qp->max_inline_data = wqe_size - sq_overhead(attr) -
495 			      sizeof(struct mlx5_wqe_inline_seg);
496 	attr->cap.max_inline_data = qp->max_inline_data;
497 
498 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN)
499 		qp->signature_en = true;
500 
501 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
502 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
503 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
504 		mlx5_ib_dbg(dev, "send queue size (%d * %d / %d -> %d) exceeds limits(%d)\n",
505 			    attr->cap.max_send_wr, wqe_size, MLX5_SEND_WQE_BB,
506 			    qp->sq.wqe_cnt,
507 			    1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
508 		return -ENOMEM;
509 	}
510 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
511 	qp->sq.max_gs = get_send_sge(attr, wqe_size);
512 	if (qp->sq.max_gs < attr->cap.max_send_sge)
513 		return -ENOMEM;
514 
515 	attr->cap.max_send_sge = qp->sq.max_gs;
516 	qp->sq.max_post = wq_size / wqe_size;
517 	attr->cap.max_send_wr = qp->sq.max_post;
518 
519 	return wq_size;
520 }
521 
522 static int set_user_buf_size(struct mlx5_ib_dev *dev,
523 			    struct mlx5_ib_qp *qp,
524 			    struct mlx5_ib_create_qp *ucmd,
525 			    struct mlx5_ib_qp_base *base,
526 			    struct ib_qp_init_attr *attr)
527 {
528 	int desc_sz = 1 << qp->sq.wqe_shift;
529 
530 	if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
531 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
532 			     desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
533 		return -EINVAL;
534 	}
535 
536 	if (ucmd->sq_wqe_count && !is_power_of_2(ucmd->sq_wqe_count)) {
537 		mlx5_ib_warn(dev, "sq_wqe_count %d is not a power of two\n",
538 			     ucmd->sq_wqe_count);
539 		return -EINVAL;
540 	}
541 
542 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
543 
544 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
545 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
546 			     qp->sq.wqe_cnt,
547 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
548 		return -EINVAL;
549 	}
550 
551 	if (attr->qp_type == IB_QPT_RAW_PACKET ||
552 	    qp->flags & MLX5_IB_QP_UNDERLAY) {
553 		base->ubuffer.buf_size = qp->rq.wqe_cnt << qp->rq.wqe_shift;
554 		qp->raw_packet_qp.sq.ubuffer.buf_size = qp->sq.wqe_cnt << 6;
555 	} else {
556 		base->ubuffer.buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
557 					 (qp->sq.wqe_cnt << 6);
558 	}
559 
560 	return 0;
561 }
562 
563 static int qp_has_rq(struct ib_qp_init_attr *attr)
564 {
565 	if (attr->qp_type == IB_QPT_XRC_INI ||
566 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
567 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
568 	    !attr->cap.max_recv_wr)
569 		return 0;
570 
571 	return 1;
572 }
573 
574 enum {
575 	/* this is the first blue flame register in the array of bfregs assigned
576 	 * to a processes. Since we do not use it for blue flame but rather
577 	 * regular 64 bit doorbells, we do not need a lock for maintaiing
578 	 * "odd/even" order
579 	 */
580 	NUM_NON_BLUE_FLAME_BFREGS = 1,
581 };
582 
583 static int max_bfregs(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi)
584 {
585 	return get_num_static_uars(dev, bfregi) * MLX5_NON_FP_BFREGS_PER_UAR;
586 }
587 
588 static int num_med_bfreg(struct mlx5_ib_dev *dev,
589 			 struct mlx5_bfreg_info *bfregi)
590 {
591 	int n;
592 
593 	n = max_bfregs(dev, bfregi) - bfregi->num_low_latency_bfregs -
594 	    NUM_NON_BLUE_FLAME_BFREGS;
595 
596 	return n >= 0 ? n : 0;
597 }
598 
599 static int first_med_bfreg(struct mlx5_ib_dev *dev,
600 			   struct mlx5_bfreg_info *bfregi)
601 {
602 	return num_med_bfreg(dev, bfregi) ? 1 : -ENOMEM;
603 }
604 
605 static int first_hi_bfreg(struct mlx5_ib_dev *dev,
606 			  struct mlx5_bfreg_info *bfregi)
607 {
608 	int med;
609 
610 	med = num_med_bfreg(dev, bfregi);
611 	return ++med;
612 }
613 
614 static int alloc_high_class_bfreg(struct mlx5_ib_dev *dev,
615 				  struct mlx5_bfreg_info *bfregi)
616 {
617 	int i;
618 
619 	for (i = first_hi_bfreg(dev, bfregi); i < max_bfregs(dev, bfregi); i++) {
620 		if (!bfregi->count[i]) {
621 			bfregi->count[i]++;
622 			return i;
623 		}
624 	}
625 
626 	return -ENOMEM;
627 }
628 
629 static int alloc_med_class_bfreg(struct mlx5_ib_dev *dev,
630 				 struct mlx5_bfreg_info *bfregi)
631 {
632 	int minidx = first_med_bfreg(dev, bfregi);
633 	int i;
634 
635 	if (minidx < 0)
636 		return minidx;
637 
638 	for (i = minidx; i < first_hi_bfreg(dev, bfregi); i++) {
639 		if (bfregi->count[i] < bfregi->count[minidx])
640 			minidx = i;
641 		if (!bfregi->count[minidx])
642 			break;
643 	}
644 
645 	bfregi->count[minidx]++;
646 	return minidx;
647 }
648 
649 static int alloc_bfreg(struct mlx5_ib_dev *dev,
650 		       struct mlx5_bfreg_info *bfregi)
651 {
652 	int bfregn = -ENOMEM;
653 
654 	mutex_lock(&bfregi->lock);
655 	if (bfregi->ver >= 2) {
656 		bfregn = alloc_high_class_bfreg(dev, bfregi);
657 		if (bfregn < 0)
658 			bfregn = alloc_med_class_bfreg(dev, bfregi);
659 	}
660 
661 	if (bfregn < 0) {
662 		BUILD_BUG_ON(NUM_NON_BLUE_FLAME_BFREGS != 1);
663 		bfregn = 0;
664 		bfregi->count[bfregn]++;
665 	}
666 	mutex_unlock(&bfregi->lock);
667 
668 	return bfregn;
669 }
670 
671 void mlx5_ib_free_bfreg(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi, int bfregn)
672 {
673 	mutex_lock(&bfregi->lock);
674 	bfregi->count[bfregn]--;
675 	mutex_unlock(&bfregi->lock);
676 }
677 
678 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
679 {
680 	switch (state) {
681 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
682 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
683 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
684 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
685 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
686 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
687 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
688 	default:		return -1;
689 	}
690 }
691 
692 static int to_mlx5_st(enum ib_qp_type type)
693 {
694 	switch (type) {
695 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
696 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
697 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
698 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
699 	case IB_QPT_XRC_INI:
700 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
701 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
702 	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
703 	case MLX5_IB_QPT_DCI:		return MLX5_QP_ST_DCI;
704 	case IB_QPT_RAW_IPV6:		return MLX5_QP_ST_RAW_IPV6;
705 	case IB_QPT_RAW_PACKET:
706 	case IB_QPT_RAW_ETHERTYPE:	return MLX5_QP_ST_RAW_ETHERTYPE;
707 	case IB_QPT_MAX:
708 	default:		return -EINVAL;
709 	}
710 }
711 
712 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
713 			     struct mlx5_ib_cq *recv_cq);
714 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
715 			       struct mlx5_ib_cq *recv_cq);
716 
717 int bfregn_to_uar_index(struct mlx5_ib_dev *dev,
718 			struct mlx5_bfreg_info *bfregi, u32 bfregn,
719 			bool dyn_bfreg)
720 {
721 	unsigned int bfregs_per_sys_page;
722 	u32 index_of_sys_page;
723 	u32 offset;
724 
725 	bfregs_per_sys_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k) *
726 				MLX5_NON_FP_BFREGS_PER_UAR;
727 	index_of_sys_page = bfregn / bfregs_per_sys_page;
728 
729 	if (dyn_bfreg) {
730 		index_of_sys_page += bfregi->num_static_sys_pages;
731 
732 		if (index_of_sys_page >= bfregi->num_sys_pages)
733 			return -EINVAL;
734 
735 		if (bfregn > bfregi->num_dyn_bfregs ||
736 		    bfregi->sys_pages[index_of_sys_page] == MLX5_IB_INVALID_UAR_INDEX) {
737 			mlx5_ib_dbg(dev, "Invalid dynamic uar index\n");
738 			return -EINVAL;
739 		}
740 	}
741 
742 	offset = bfregn % bfregs_per_sys_page / MLX5_NON_FP_BFREGS_PER_UAR;
743 	return bfregi->sys_pages[index_of_sys_page] + offset;
744 }
745 
746 static int mlx5_ib_umem_get(struct mlx5_ib_dev *dev, struct ib_udata *udata,
747 			    unsigned long addr, size_t size,
748 			    struct ib_umem **umem, int *npages, int *page_shift,
749 			    int *ncont, u32 *offset)
750 {
751 	int err;
752 
753 	*umem = ib_umem_get(udata, addr, size, 0, 0);
754 	if (IS_ERR(*umem)) {
755 		mlx5_ib_dbg(dev, "umem_get failed\n");
756 		return PTR_ERR(*umem);
757 	}
758 
759 	mlx5_ib_cont_pages(*umem, addr, 0, npages, page_shift, ncont, NULL);
760 
761 	err = mlx5_ib_get_buf_offset(addr, *page_shift, offset);
762 	if (err) {
763 		mlx5_ib_warn(dev, "bad offset\n");
764 		goto err_umem;
765 	}
766 
767 	mlx5_ib_dbg(dev, "addr 0x%lx, size %zu, npages %d, page_shift %d, ncont %d, offset %d\n",
768 		    addr, size, *npages, *page_shift, *ncont, *offset);
769 
770 	return 0;
771 
772 err_umem:
773 	ib_umem_release(*umem);
774 	*umem = NULL;
775 
776 	return err;
777 }
778 
779 static void destroy_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
780 			    struct mlx5_ib_rwq *rwq)
781 {
782 	struct mlx5_ib_ucontext *context;
783 
784 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_DELAY_DROP)
785 		atomic_dec(&dev->delay_drop.rqs_cnt);
786 
787 	context = to_mucontext(pd->uobject->context);
788 	mlx5_ib_db_unmap_user(context, &rwq->db);
789 	if (rwq->umem)
790 		ib_umem_release(rwq->umem);
791 }
792 
793 static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
794 			  struct ib_udata *udata, struct mlx5_ib_rwq *rwq,
795 			  struct mlx5_ib_create_wq *ucmd)
796 {
797 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
798 		udata, struct mlx5_ib_ucontext, ibucontext);
799 	int page_shift = 0;
800 	int npages;
801 	u32 offset = 0;
802 	int ncont = 0;
803 	int err;
804 
805 	if (!ucmd->buf_addr)
806 		return -EINVAL;
807 
808 	rwq->umem = ib_umem_get(udata, ucmd->buf_addr, rwq->buf_size, 0, 0);
809 	if (IS_ERR(rwq->umem)) {
810 		mlx5_ib_dbg(dev, "umem_get failed\n");
811 		err = PTR_ERR(rwq->umem);
812 		return err;
813 	}
814 
815 	mlx5_ib_cont_pages(rwq->umem, ucmd->buf_addr, 0, &npages, &page_shift,
816 			   &ncont, NULL);
817 	err = mlx5_ib_get_buf_offset(ucmd->buf_addr, page_shift,
818 				     &rwq->rq_page_offset);
819 	if (err) {
820 		mlx5_ib_warn(dev, "bad offset\n");
821 		goto err_umem;
822 	}
823 
824 	rwq->rq_num_pas = ncont;
825 	rwq->page_shift = page_shift;
826 	rwq->log_page_size =  page_shift - MLX5_ADAPTER_PAGE_SHIFT;
827 	rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
828 
829 	mlx5_ib_dbg(dev, "addr 0x%llx, size %zd, npages %d, page_shift %d, ncont %d, offset %d\n",
830 		    (unsigned long long)ucmd->buf_addr, rwq->buf_size,
831 		    npages, page_shift, ncont, offset);
832 
833 	err = mlx5_ib_db_map_user(ucontext, udata, ucmd->db_addr, &rwq->db);
834 	if (err) {
835 		mlx5_ib_dbg(dev, "map failed\n");
836 		goto err_umem;
837 	}
838 
839 	rwq->create_type = MLX5_WQ_USER;
840 	return 0;
841 
842 err_umem:
843 	ib_umem_release(rwq->umem);
844 	return err;
845 }
846 
847 static int adjust_bfregn(struct mlx5_ib_dev *dev,
848 			 struct mlx5_bfreg_info *bfregi, int bfregn)
849 {
850 	return bfregn / MLX5_NON_FP_BFREGS_PER_UAR * MLX5_BFREGS_PER_UAR +
851 				bfregn % MLX5_NON_FP_BFREGS_PER_UAR;
852 }
853 
854 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
855 			  struct mlx5_ib_qp *qp, struct ib_udata *udata,
856 			  struct ib_qp_init_attr *attr,
857 			  u32 **in,
858 			  struct mlx5_ib_create_qp_resp *resp, int *inlen,
859 			  struct mlx5_ib_qp_base *base)
860 {
861 	struct mlx5_ib_ucontext *context;
862 	struct mlx5_ib_create_qp ucmd;
863 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
864 	int page_shift = 0;
865 	int uar_index = 0;
866 	int npages;
867 	u32 offset = 0;
868 	int bfregn;
869 	int ncont = 0;
870 	__be64 *pas;
871 	void *qpc;
872 	int err;
873 	u16 uid;
874 
875 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
876 	if (err) {
877 		mlx5_ib_dbg(dev, "copy failed\n");
878 		return err;
879 	}
880 
881 	context = rdma_udata_to_drv_context(udata, struct mlx5_ib_ucontext,
882 					    ibucontext);
883 	if (ucmd.flags & MLX5_QP_FLAG_BFREG_INDEX) {
884 		uar_index = bfregn_to_uar_index(dev, &context->bfregi,
885 						ucmd.bfreg_index, true);
886 		if (uar_index < 0)
887 			return uar_index;
888 
889 		bfregn = MLX5_IB_INVALID_BFREG;
890 	} else if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL) {
891 		/*
892 		 * TBD: should come from the verbs when we have the API
893 		 */
894 		/* In CROSS_CHANNEL CQ and QP must use the same UAR */
895 		bfregn = MLX5_CROSS_CHANNEL_BFREG;
896 	}
897 	else {
898 		bfregn = alloc_bfreg(dev, &context->bfregi);
899 		if (bfregn < 0)
900 			return bfregn;
901 	}
902 
903 	mlx5_ib_dbg(dev, "bfregn 0x%x, uar_index 0x%x\n", bfregn, uar_index);
904 	if (bfregn != MLX5_IB_INVALID_BFREG)
905 		uar_index = bfregn_to_uar_index(dev, &context->bfregi, bfregn,
906 						false);
907 
908 	qp->rq.offset = 0;
909 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
910 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
911 
912 	err = set_user_buf_size(dev, qp, &ucmd, base, attr);
913 	if (err)
914 		goto err_bfreg;
915 
916 	if (ucmd.buf_addr && ubuffer->buf_size) {
917 		ubuffer->buf_addr = ucmd.buf_addr;
918 		err = mlx5_ib_umem_get(dev, udata, ubuffer->buf_addr,
919 				       ubuffer->buf_size, &ubuffer->umem,
920 				       &npages, &page_shift, &ncont, &offset);
921 		if (err)
922 			goto err_bfreg;
923 	} else {
924 		ubuffer->umem = NULL;
925 	}
926 
927 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
928 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * ncont;
929 	*in = kvzalloc(*inlen, GFP_KERNEL);
930 	if (!*in) {
931 		err = -ENOMEM;
932 		goto err_umem;
933 	}
934 
935 	uid = (attr->qp_type != IB_QPT_XRC_TGT &&
936 	       attr->qp_type != IB_QPT_XRC_INI) ? to_mpd(pd)->uid : 0;
937 	MLX5_SET(create_qp_in, *in, uid, uid);
938 	pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas);
939 	if (ubuffer->umem)
940 		mlx5_ib_populate_pas(dev, ubuffer->umem, page_shift, pas, 0);
941 
942 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
943 
944 	MLX5_SET(qpc, qpc, log_page_size, page_shift - MLX5_ADAPTER_PAGE_SHIFT);
945 	MLX5_SET(qpc, qpc, page_offset, offset);
946 
947 	MLX5_SET(qpc, qpc, uar_page, uar_index);
948 	if (bfregn != MLX5_IB_INVALID_BFREG)
949 		resp->bfreg_index = adjust_bfregn(dev, &context->bfregi, bfregn);
950 	else
951 		resp->bfreg_index = MLX5_IB_INVALID_BFREG;
952 	qp->bfregn = bfregn;
953 
954 	err = mlx5_ib_db_map_user(context, udata, ucmd.db_addr, &qp->db);
955 	if (err) {
956 		mlx5_ib_dbg(dev, "map failed\n");
957 		goto err_free;
958 	}
959 
960 	err = ib_copy_to_udata(udata, resp, min(udata->outlen, sizeof(*resp)));
961 	if (err) {
962 		mlx5_ib_dbg(dev, "copy failed\n");
963 		goto err_unmap;
964 	}
965 	qp->create_type = MLX5_QP_USER;
966 
967 	return 0;
968 
969 err_unmap:
970 	mlx5_ib_db_unmap_user(context, &qp->db);
971 
972 err_free:
973 	kvfree(*in);
974 
975 err_umem:
976 	if (ubuffer->umem)
977 		ib_umem_release(ubuffer->umem);
978 
979 err_bfreg:
980 	if (bfregn != MLX5_IB_INVALID_BFREG)
981 		mlx5_ib_free_bfreg(dev, &context->bfregi, bfregn);
982 	return err;
983 }
984 
985 static void destroy_qp_user(struct mlx5_ib_dev *dev, struct ib_pd *pd,
986 			    struct mlx5_ib_qp *qp, struct mlx5_ib_qp_base *base)
987 {
988 	struct mlx5_ib_ucontext *context;
989 
990 	context = to_mucontext(pd->uobject->context);
991 	mlx5_ib_db_unmap_user(context, &qp->db);
992 	if (base->ubuffer.umem)
993 		ib_umem_release(base->ubuffer.umem);
994 
995 	/*
996 	 * Free only the BFREGs which are handled by the kernel.
997 	 * BFREGs of UARs allocated dynamically are handled by user.
998 	 */
999 	if (qp->bfregn != MLX5_IB_INVALID_BFREG)
1000 		mlx5_ib_free_bfreg(dev, &context->bfregi, qp->bfregn);
1001 }
1002 
1003 /* get_sq_edge - Get the next nearby edge.
1004  *
1005  * An 'edge' is defined as the first following address after the end
1006  * of the fragment or the SQ. Accordingly, during the WQE construction
1007  * which repetitively increases the pointer to write the next data, it
1008  * simply should check if it gets to an edge.
1009  *
1010  * @sq - SQ buffer.
1011  * @idx - Stride index in the SQ buffer.
1012  *
1013  * Return:
1014  *	The new edge.
1015  */
1016 static void *get_sq_edge(struct mlx5_ib_wq *sq, u32 idx)
1017 {
1018 	void *fragment_end;
1019 
1020 	fragment_end = mlx5_frag_buf_get_wqe
1021 		(&sq->fbc,
1022 		 mlx5_frag_buf_get_idx_last_contig_stride(&sq->fbc, idx));
1023 
1024 	return fragment_end + MLX5_SEND_WQE_BB;
1025 }
1026 
1027 static int create_kernel_qp(struct mlx5_ib_dev *dev,
1028 			    struct ib_qp_init_attr *init_attr,
1029 			    struct mlx5_ib_qp *qp,
1030 			    u32 **in, int *inlen,
1031 			    struct mlx5_ib_qp_base *base)
1032 {
1033 	int uar_index;
1034 	void *qpc;
1035 	int err;
1036 
1037 	if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN |
1038 					IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
1039 					IB_QP_CREATE_IPOIB_UD_LSO |
1040 					IB_QP_CREATE_NETIF_QP |
1041 					mlx5_ib_create_qp_sqpn_qp1()))
1042 		return -EINVAL;
1043 
1044 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
1045 		qp->bf.bfreg = &dev->fp_bfreg;
1046 	else
1047 		qp->bf.bfreg = &dev->bfreg;
1048 
1049 	/* We need to divide by two since each register is comprised of
1050 	 * two buffers of identical size, namely odd and even
1051 	 */
1052 	qp->bf.buf_size = (1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size)) / 2;
1053 	uar_index = qp->bf.bfreg->index;
1054 
1055 	err = calc_sq_size(dev, init_attr, qp);
1056 	if (err < 0) {
1057 		mlx5_ib_dbg(dev, "err %d\n", err);
1058 		return err;
1059 	}
1060 
1061 	qp->rq.offset = 0;
1062 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
1063 	base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
1064 
1065 	err = mlx5_frag_buf_alloc_node(dev->mdev, base->ubuffer.buf_size,
1066 				       &qp->buf, dev->mdev->priv.numa_node);
1067 	if (err) {
1068 		mlx5_ib_dbg(dev, "err %d\n", err);
1069 		return err;
1070 	}
1071 
1072 	if (qp->rq.wqe_cnt)
1073 		mlx5_init_fbc(qp->buf.frags, qp->rq.wqe_shift,
1074 			      ilog2(qp->rq.wqe_cnt), &qp->rq.fbc);
1075 
1076 	if (qp->sq.wqe_cnt) {
1077 		int sq_strides_offset = (qp->sq.offset  & (PAGE_SIZE - 1)) /
1078 					MLX5_SEND_WQE_BB;
1079 		mlx5_init_fbc_offset(qp->buf.frags +
1080 				     (qp->sq.offset / PAGE_SIZE),
1081 				     ilog2(MLX5_SEND_WQE_BB),
1082 				     ilog2(qp->sq.wqe_cnt),
1083 				     sq_strides_offset, &qp->sq.fbc);
1084 
1085 		qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
1086 	}
1087 
1088 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
1089 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * qp->buf.npages;
1090 	*in = kvzalloc(*inlen, GFP_KERNEL);
1091 	if (!*in) {
1092 		err = -ENOMEM;
1093 		goto err_buf;
1094 	}
1095 
1096 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
1097 	MLX5_SET(qpc, qpc, uar_page, uar_index);
1098 	MLX5_SET(qpc, qpc, log_page_size, qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1099 
1100 	/* Set "fast registration enabled" for all kernel QPs */
1101 	MLX5_SET(qpc, qpc, fre, 1);
1102 	MLX5_SET(qpc, qpc, rlky, 1);
1103 
1104 	if (init_attr->create_flags & mlx5_ib_create_qp_sqpn_qp1()) {
1105 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
1106 		qp->flags |= MLX5_IB_QP_SQPN_QP1;
1107 	}
1108 
1109 	mlx5_fill_page_frag_array(&qp->buf,
1110 				  (__be64 *)MLX5_ADDR_OF(create_qp_in,
1111 							 *in, pas));
1112 
1113 	err = mlx5_db_alloc(dev->mdev, &qp->db);
1114 	if (err) {
1115 		mlx5_ib_dbg(dev, "err %d\n", err);
1116 		goto err_free;
1117 	}
1118 
1119 	qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1120 				     sizeof(*qp->sq.wrid), GFP_KERNEL);
1121 	qp->sq.wr_data = kvmalloc_array(qp->sq.wqe_cnt,
1122 					sizeof(*qp->sq.wr_data), GFP_KERNEL);
1123 	qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1124 				     sizeof(*qp->rq.wrid), GFP_KERNEL);
1125 	qp->sq.w_list = kvmalloc_array(qp->sq.wqe_cnt,
1126 				       sizeof(*qp->sq.w_list), GFP_KERNEL);
1127 	qp->sq.wqe_head = kvmalloc_array(qp->sq.wqe_cnt,
1128 					 sizeof(*qp->sq.wqe_head), GFP_KERNEL);
1129 
1130 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
1131 	    !qp->sq.w_list || !qp->sq.wqe_head) {
1132 		err = -ENOMEM;
1133 		goto err_wrid;
1134 	}
1135 	qp->create_type = MLX5_QP_KERNEL;
1136 
1137 	return 0;
1138 
1139 err_wrid:
1140 	kvfree(qp->sq.wqe_head);
1141 	kvfree(qp->sq.w_list);
1142 	kvfree(qp->sq.wrid);
1143 	kvfree(qp->sq.wr_data);
1144 	kvfree(qp->rq.wrid);
1145 	mlx5_db_free(dev->mdev, &qp->db);
1146 
1147 err_free:
1148 	kvfree(*in);
1149 
1150 err_buf:
1151 	mlx5_frag_buf_free(dev->mdev, &qp->buf);
1152 	return err;
1153 }
1154 
1155 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1156 {
1157 	kvfree(qp->sq.wqe_head);
1158 	kvfree(qp->sq.w_list);
1159 	kvfree(qp->sq.wrid);
1160 	kvfree(qp->sq.wr_data);
1161 	kvfree(qp->rq.wrid);
1162 	mlx5_db_free(dev->mdev, &qp->db);
1163 	mlx5_frag_buf_free(dev->mdev, &qp->buf);
1164 }
1165 
1166 static u32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
1167 {
1168 	if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
1169 	    (attr->qp_type == MLX5_IB_QPT_DCI) ||
1170 	    (attr->qp_type == IB_QPT_XRC_INI))
1171 		return MLX5_SRQ_RQ;
1172 	else if (!qp->has_rq)
1173 		return MLX5_ZERO_LEN_RQ;
1174 	else
1175 		return MLX5_NON_ZERO_RQ;
1176 }
1177 
1178 static int is_connected(enum ib_qp_type qp_type)
1179 {
1180 	if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC ||
1181 	    qp_type == MLX5_IB_QPT_DCI)
1182 		return 1;
1183 
1184 	return 0;
1185 }
1186 
1187 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1188 				    struct mlx5_ib_qp *qp,
1189 				    struct mlx5_ib_sq *sq, u32 tdn,
1190 				    struct ib_pd *pd)
1191 {
1192 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {0};
1193 	void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1194 
1195 	MLX5_SET(create_tis_in, in, uid, to_mpd(pd)->uid);
1196 	MLX5_SET(tisc, tisc, transport_domain, tdn);
1197 	if (qp->flags & MLX5_IB_QP_UNDERLAY)
1198 		MLX5_SET(tisc, tisc, underlay_qpn, qp->underlay_qpn);
1199 
1200 	return mlx5_core_create_tis(dev->mdev, in, sizeof(in), &sq->tisn);
1201 }
1202 
1203 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1204 				      struct mlx5_ib_sq *sq, struct ib_pd *pd)
1205 {
1206 	mlx5_cmd_destroy_tis(dev->mdev, sq->tisn, to_mpd(pd)->uid);
1207 }
1208 
1209 static void destroy_flow_rule_vport_sq(struct mlx5_ib_dev *dev,
1210 				       struct mlx5_ib_sq *sq)
1211 {
1212 	if (sq->flow_rule)
1213 		mlx5_del_flow_rules(sq->flow_rule);
1214 }
1215 
1216 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1217 				   struct ib_udata *udata,
1218 				   struct mlx5_ib_sq *sq, void *qpin,
1219 				   struct ib_pd *pd)
1220 {
1221 	struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1222 	__be64 *pas;
1223 	void *in;
1224 	void *sqc;
1225 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1226 	void *wq;
1227 	int inlen;
1228 	int err;
1229 	int page_shift = 0;
1230 	int npages;
1231 	int ncont = 0;
1232 	u32 offset = 0;
1233 
1234 	err = mlx5_ib_umem_get(dev, udata, ubuffer->buf_addr, ubuffer->buf_size,
1235 			       &sq->ubuffer.umem, &npages, &page_shift, &ncont,
1236 			       &offset);
1237 	if (err)
1238 		return err;
1239 
1240 	inlen = MLX5_ST_SZ_BYTES(create_sq_in) + sizeof(u64) * ncont;
1241 	in = kvzalloc(inlen, GFP_KERNEL);
1242 	if (!in) {
1243 		err = -ENOMEM;
1244 		goto err_umem;
1245 	}
1246 
1247 	MLX5_SET(create_sq_in, in, uid, to_mpd(pd)->uid);
1248 	sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1249 	MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1250 	if (MLX5_CAP_ETH(dev->mdev, multi_pkt_send_wqe))
1251 		MLX5_SET(sqc, sqc, allow_multi_pkt_send_wqe, 1);
1252 	MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1253 	MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1254 	MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1255 	MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1256 	MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1257 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
1258 	    MLX5_CAP_ETH(dev->mdev, swp))
1259 		MLX5_SET(sqc, sqc, allow_swp, 1);
1260 
1261 	wq = MLX5_ADDR_OF(sqc, sqc, wq);
1262 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1263 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1264 	MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1265 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1266 	MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1267 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1268 	MLX5_SET(wq, wq, log_wq_pg_sz,  page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1269 	MLX5_SET(wq, wq, page_offset, offset);
1270 
1271 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1272 	mlx5_ib_populate_pas(dev, sq->ubuffer.umem, page_shift, pas, 0);
1273 
1274 	err = mlx5_core_create_sq_tracked(dev->mdev, in, inlen, &sq->base.mqp);
1275 
1276 	kvfree(in);
1277 
1278 	if (err)
1279 		goto err_umem;
1280 
1281 	err = create_flow_rule_vport_sq(dev, sq);
1282 	if (err)
1283 		goto err_flow;
1284 
1285 	return 0;
1286 
1287 err_flow:
1288 	mlx5_core_destroy_sq_tracked(dev->mdev, &sq->base.mqp);
1289 
1290 err_umem:
1291 	ib_umem_release(sq->ubuffer.umem);
1292 	sq->ubuffer.umem = NULL;
1293 
1294 	return err;
1295 }
1296 
1297 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1298 				     struct mlx5_ib_sq *sq)
1299 {
1300 	destroy_flow_rule_vport_sq(dev, sq);
1301 	mlx5_core_destroy_sq_tracked(dev->mdev, &sq->base.mqp);
1302 	ib_umem_release(sq->ubuffer.umem);
1303 }
1304 
1305 static size_t get_rq_pas_size(void *qpc)
1306 {
1307 	u32 log_page_size = MLX5_GET(qpc, qpc, log_page_size) + 12;
1308 	u32 log_rq_stride = MLX5_GET(qpc, qpc, log_rq_stride);
1309 	u32 log_rq_size   = MLX5_GET(qpc, qpc, log_rq_size);
1310 	u32 page_offset   = MLX5_GET(qpc, qpc, page_offset);
1311 	u32 po_quanta	  = 1 << (log_page_size - 6);
1312 	u32 rq_sz	  = 1 << (log_rq_size + 4 + log_rq_stride);
1313 	u32 page_size	  = 1 << log_page_size;
1314 	u32 rq_sz_po      = rq_sz + (page_offset * po_quanta);
1315 	u32 rq_num_pas	  = (rq_sz_po + page_size - 1) / page_size;
1316 
1317 	return rq_num_pas * sizeof(u64);
1318 }
1319 
1320 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1321 				   struct mlx5_ib_rq *rq, void *qpin,
1322 				   size_t qpinlen, struct ib_pd *pd)
1323 {
1324 	struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1325 	__be64 *pas;
1326 	__be64 *qp_pas;
1327 	void *in;
1328 	void *rqc;
1329 	void *wq;
1330 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1331 	size_t rq_pas_size = get_rq_pas_size(qpc);
1332 	size_t inlen;
1333 	int err;
1334 
1335 	if (qpinlen < rq_pas_size + MLX5_BYTE_OFF(create_qp_in, pas))
1336 		return -EINVAL;
1337 
1338 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + rq_pas_size;
1339 	in = kvzalloc(inlen, GFP_KERNEL);
1340 	if (!in)
1341 		return -ENOMEM;
1342 
1343 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
1344 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1345 	if (!(rq->flags & MLX5_IB_RQ_CVLAN_STRIPPING))
1346 		MLX5_SET(rqc, rqc, vsd, 1);
1347 	MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
1348 	MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1349 	MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1350 	MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1351 	MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1352 
1353 	if (mqp->flags & MLX5_IB_QP_CAP_SCATTER_FCS)
1354 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
1355 
1356 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
1357 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1358 	if (rq->flags & MLX5_IB_RQ_PCI_WRITE_END_PADDING)
1359 		MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
1360 	MLX5_SET(wq, wq, page_offset, MLX5_GET(qpc, qpc, page_offset));
1361 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1362 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1363 	MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1364 	MLX5_SET(wq, wq, log_wq_pg_sz, MLX5_GET(qpc, qpc, log_page_size));
1365 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1366 
1367 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1368 	qp_pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, qpin, pas);
1369 	memcpy(pas, qp_pas, rq_pas_size);
1370 
1371 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rq->base.mqp);
1372 
1373 	kvfree(in);
1374 
1375 	return err;
1376 }
1377 
1378 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1379 				     struct mlx5_ib_rq *rq)
1380 {
1381 	mlx5_core_destroy_rq_tracked(dev->mdev, &rq->base.mqp);
1382 }
1383 
1384 static bool tunnel_offload_supported(struct mlx5_core_dev *dev)
1385 {
1386 	return  (MLX5_CAP_ETH(dev, tunnel_stateless_vxlan) ||
1387 		 MLX5_CAP_ETH(dev, tunnel_stateless_gre) ||
1388 		 MLX5_CAP_ETH(dev, tunnel_stateless_geneve_rx));
1389 }
1390 
1391 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1392 				      struct mlx5_ib_rq *rq,
1393 				      u32 qp_flags_en,
1394 				      struct ib_pd *pd)
1395 {
1396 	if (qp_flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1397 			   MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1398 		mlx5_ib_disable_lb(dev, false, true);
1399 	mlx5_cmd_destroy_tir(dev->mdev, rq->tirn, to_mpd(pd)->uid);
1400 }
1401 
1402 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1403 				    struct mlx5_ib_rq *rq, u32 tdn,
1404 				    u32 *qp_flags_en,
1405 				    struct ib_pd *pd)
1406 {
1407 	u8 lb_flag = 0;
1408 	u32 *in;
1409 	void *tirc;
1410 	int inlen;
1411 	int err;
1412 
1413 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1414 	in = kvzalloc(inlen, GFP_KERNEL);
1415 	if (!in)
1416 		return -ENOMEM;
1417 
1418 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1419 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1420 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1421 	MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1422 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1423 	if (*qp_flags_en & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1424 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1425 
1426 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC)
1427 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1428 
1429 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)
1430 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1431 
1432 	if (dev->rep) {
1433 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1434 		*qp_flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1435 	}
1436 
1437 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1438 
1439 	err = mlx5_core_create_tir(dev->mdev, in, inlen, &rq->tirn);
1440 
1441 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1442 		err = mlx5_ib_enable_lb(dev, false, true);
1443 
1444 		if (err)
1445 			destroy_raw_packet_qp_tir(dev, rq, 0, pd);
1446 	}
1447 	kvfree(in);
1448 
1449 	return err;
1450 }
1451 
1452 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1453 				u32 *in, size_t inlen,
1454 				struct ib_pd *pd,
1455 				struct ib_udata *udata,
1456 				struct mlx5_ib_create_qp_resp *resp)
1457 {
1458 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1459 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1460 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1461 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1462 		udata, struct mlx5_ib_ucontext, ibucontext);
1463 	int err;
1464 	u32 tdn = mucontext->tdn;
1465 	u16 uid = to_mpd(pd)->uid;
1466 
1467 	if (qp->sq.wqe_cnt) {
1468 		err = create_raw_packet_qp_tis(dev, qp, sq, tdn, pd);
1469 		if (err)
1470 			return err;
1471 
1472 		err = create_raw_packet_qp_sq(dev, udata, sq, in, pd);
1473 		if (err)
1474 			goto err_destroy_tis;
1475 
1476 		if (uid) {
1477 			resp->tisn = sq->tisn;
1478 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TISN;
1479 			resp->sqn = sq->base.mqp.qpn;
1480 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_SQN;
1481 		}
1482 
1483 		sq->base.container_mibqp = qp;
1484 		sq->base.mqp.event = mlx5_ib_qp_event;
1485 	}
1486 
1487 	if (qp->rq.wqe_cnt) {
1488 		rq->base.container_mibqp = qp;
1489 
1490 		if (qp->flags & MLX5_IB_QP_CVLAN_STRIPPING)
1491 			rq->flags |= MLX5_IB_RQ_CVLAN_STRIPPING;
1492 		if (qp->flags & MLX5_IB_QP_PCI_WRITE_END_PADDING)
1493 			rq->flags |= MLX5_IB_RQ_PCI_WRITE_END_PADDING;
1494 		err = create_raw_packet_qp_rq(dev, rq, in, inlen, pd);
1495 		if (err)
1496 			goto err_destroy_sq;
1497 
1498 		err = create_raw_packet_qp_tir(dev, rq, tdn, &qp->flags_en, pd);
1499 		if (err)
1500 			goto err_destroy_rq;
1501 
1502 		if (uid) {
1503 			resp->rqn = rq->base.mqp.qpn;
1504 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_RQN;
1505 			resp->tirn = rq->tirn;
1506 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1507 		}
1508 	}
1509 
1510 	qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1511 						     rq->base.mqp.qpn;
1512 	err = ib_copy_to_udata(udata, resp, min(udata->outlen, sizeof(*resp)));
1513 	if (err)
1514 		goto err_destroy_tir;
1515 
1516 	return 0;
1517 
1518 err_destroy_tir:
1519 	destroy_raw_packet_qp_tir(dev, rq, qp->flags_en, pd);
1520 err_destroy_rq:
1521 	destroy_raw_packet_qp_rq(dev, rq);
1522 err_destroy_sq:
1523 	if (!qp->sq.wqe_cnt)
1524 		return err;
1525 	destroy_raw_packet_qp_sq(dev, sq);
1526 err_destroy_tis:
1527 	destroy_raw_packet_qp_tis(dev, sq, pd);
1528 
1529 	return err;
1530 }
1531 
1532 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1533 				  struct mlx5_ib_qp *qp)
1534 {
1535 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1536 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1537 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1538 
1539 	if (qp->rq.wqe_cnt) {
1540 		destroy_raw_packet_qp_tir(dev, rq, qp->flags_en, qp->ibqp.pd);
1541 		destroy_raw_packet_qp_rq(dev, rq);
1542 	}
1543 
1544 	if (qp->sq.wqe_cnt) {
1545 		destroy_raw_packet_qp_sq(dev, sq);
1546 		destroy_raw_packet_qp_tis(dev, sq, qp->ibqp.pd);
1547 	}
1548 }
1549 
1550 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1551 				    struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1552 {
1553 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1554 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1555 
1556 	sq->sq = &qp->sq;
1557 	rq->rq = &qp->rq;
1558 	sq->doorbell = &qp->db;
1559 	rq->doorbell = &qp->db;
1560 }
1561 
1562 static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1563 {
1564 	if (qp->flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1565 			    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1566 		mlx5_ib_disable_lb(dev, false, true);
1567 	mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1568 			     to_mpd(qp->ibqp.pd)->uid);
1569 }
1570 
1571 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1572 				 struct ib_pd *pd,
1573 				 struct ib_qp_init_attr *init_attr,
1574 				 struct ib_udata *udata)
1575 {
1576 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1577 		udata, struct mlx5_ib_ucontext, ibucontext);
1578 	struct mlx5_ib_create_qp_resp resp = {};
1579 	int inlen;
1580 	int err;
1581 	u32 *in;
1582 	void *tirc;
1583 	void *hfso;
1584 	u32 selected_fields = 0;
1585 	u32 outer_l4;
1586 	size_t min_resp_len;
1587 	u32 tdn = mucontext->tdn;
1588 	struct mlx5_ib_create_qp_rss ucmd = {};
1589 	size_t required_cmd_sz;
1590 	u8 lb_flag = 0;
1591 
1592 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1593 		return -EOPNOTSUPP;
1594 
1595 	if (init_attr->create_flags || init_attr->send_cq)
1596 		return -EINVAL;
1597 
1598 	min_resp_len = offsetof(typeof(resp), bfreg_index) + sizeof(resp.bfreg_index);
1599 	if (udata->outlen < min_resp_len)
1600 		return -EINVAL;
1601 
1602 	required_cmd_sz = offsetof(typeof(ucmd), flags) + sizeof(ucmd.flags);
1603 	if (udata->inlen < required_cmd_sz) {
1604 		mlx5_ib_dbg(dev, "invalid inlen\n");
1605 		return -EINVAL;
1606 	}
1607 
1608 	if (udata->inlen > sizeof(ucmd) &&
1609 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
1610 				 udata->inlen - sizeof(ucmd))) {
1611 		mlx5_ib_dbg(dev, "inlen is not supported\n");
1612 		return -EOPNOTSUPP;
1613 	}
1614 
1615 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
1616 		mlx5_ib_dbg(dev, "copy failed\n");
1617 		return -EFAULT;
1618 	}
1619 
1620 	if (ucmd.comp_mask) {
1621 		mlx5_ib_dbg(dev, "invalid comp mask\n");
1622 		return -EOPNOTSUPP;
1623 	}
1624 
1625 	if (ucmd.flags & ~(MLX5_QP_FLAG_TUNNEL_OFFLOADS |
1626 			   MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1627 			   MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)) {
1628 		mlx5_ib_dbg(dev, "invalid flags\n");
1629 		return -EOPNOTSUPP;
1630 	}
1631 
1632 	if (ucmd.flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS &&
1633 	    !tunnel_offload_supported(dev->mdev)) {
1634 		mlx5_ib_dbg(dev, "tunnel offloads isn't supported\n");
1635 		return -EOPNOTSUPP;
1636 	}
1637 
1638 	if (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_INNER &&
1639 	    !(ucmd.flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)) {
1640 		mlx5_ib_dbg(dev, "Tunnel offloads must be set for inner RSS\n");
1641 		return -EOPNOTSUPP;
1642 	}
1643 
1644 	if (ucmd.flags & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC || dev->rep) {
1645 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1646 		qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1647 	}
1648 
1649 	if (ucmd.flags & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC) {
1650 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1651 		qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC;
1652 	}
1653 
1654 	err = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp)));
1655 	if (err) {
1656 		mlx5_ib_dbg(dev, "copy failed\n");
1657 		return -EINVAL;
1658 	}
1659 
1660 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1661 	in = kvzalloc(inlen, GFP_KERNEL);
1662 	if (!in)
1663 		return -ENOMEM;
1664 
1665 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1666 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1667 	MLX5_SET(tirc, tirc, disp_type,
1668 		 MLX5_TIRC_DISP_TYPE_INDIRECT);
1669 	MLX5_SET(tirc, tirc, indirect_table,
1670 		 init_attr->rwq_ind_tbl->ind_tbl_num);
1671 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1672 
1673 	hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1674 
1675 	if (ucmd.flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1676 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1677 
1678 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1679 
1680 	if (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_INNER)
1681 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_inner);
1682 	else
1683 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1684 
1685 	switch (ucmd.rx_hash_function) {
1686 	case MLX5_RX_HASH_FUNC_TOEPLITZ:
1687 	{
1688 		void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1689 		size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1690 
1691 		if (len != ucmd.rx_key_len) {
1692 			err = -EINVAL;
1693 			goto err;
1694 		}
1695 
1696 		MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_TOEPLITZ);
1697 		MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
1698 		memcpy(rss_key, ucmd.rx_hash_key, len);
1699 		break;
1700 	}
1701 	default:
1702 		err = -EOPNOTSUPP;
1703 		goto err;
1704 	}
1705 
1706 	if (!ucmd.rx_hash_fields_mask) {
1707 		/* special case when this TIR serves as steering entry without hashing */
1708 		if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1709 			goto create_tir;
1710 		err = -EINVAL;
1711 		goto err;
1712 	}
1713 
1714 	if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1715 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1716 	     ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1717 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1718 		err = -EINVAL;
1719 		goto err;
1720 	}
1721 
1722 	/* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1723 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1724 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1725 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1726 			 MLX5_L3_PROT_TYPE_IPV4);
1727 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1728 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1729 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1730 			 MLX5_L3_PROT_TYPE_IPV6);
1731 
1732 	outer_l4 = ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1733 		    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP)) << 0 |
1734 		   ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1735 		    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP)) << 1 |
1736 		   (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI) << 2;
1737 
1738 	/* Check that only one l4 protocol is set */
1739 	if (outer_l4 & (outer_l4 - 1)) {
1740 		err = -EINVAL;
1741 		goto err;
1742 	}
1743 
1744 	/* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1745 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1746 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1747 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1748 			 MLX5_L4_PROT_TYPE_TCP);
1749 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1750 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1751 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1752 			 MLX5_L4_PROT_TYPE_UDP);
1753 
1754 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1755 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1756 		selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1757 
1758 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1759 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1760 		selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1761 
1762 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1763 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1764 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1765 
1766 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1767 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1768 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1769 
1770 	if (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI)
1771 		selected_fields |= MLX5_HASH_FIELD_SEL_IPSEC_SPI;
1772 
1773 	MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1774 
1775 create_tir:
1776 	err = mlx5_core_create_tir(dev->mdev, in, inlen, &qp->rss_qp.tirn);
1777 
1778 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1779 		err = mlx5_ib_enable_lb(dev, false, true);
1780 
1781 		if (err)
1782 			mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1783 					     to_mpd(pd)->uid);
1784 	}
1785 
1786 	if (err)
1787 		goto err;
1788 
1789 	if (mucontext->devx_uid) {
1790 		resp.comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1791 		resp.tirn = qp->rss_qp.tirn;
1792 	}
1793 
1794 	err = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp)));
1795 	if (err)
1796 		goto err_copy;
1797 
1798 	kvfree(in);
1799 	/* qpn is reserved for that QP */
1800 	qp->trans_qp.base.mqp.qpn = 0;
1801 	qp->flags |= MLX5_IB_QP_RSS;
1802 	return 0;
1803 
1804 err_copy:
1805 	mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn, mucontext->devx_uid);
1806 err:
1807 	kvfree(in);
1808 	return err;
1809 }
1810 
1811 static void configure_responder_scat_cqe(struct ib_qp_init_attr *init_attr,
1812 					 void *qpc)
1813 {
1814 	int rcqe_sz;
1815 
1816 	if (init_attr->qp_type == MLX5_IB_QPT_DCI)
1817 		return;
1818 
1819 	rcqe_sz = mlx5_ib_get_cqe_size(init_attr->recv_cq);
1820 
1821 	if (rcqe_sz == 128) {
1822 		MLX5_SET(qpc, qpc, cs_res, MLX5_RES_SCAT_DATA64_CQE);
1823 		return;
1824 	}
1825 
1826 	if (init_attr->qp_type != MLX5_IB_QPT_DCT)
1827 		MLX5_SET(qpc, qpc, cs_res, MLX5_RES_SCAT_DATA32_CQE);
1828 }
1829 
1830 static void configure_requester_scat_cqe(struct mlx5_ib_dev *dev,
1831 					 struct ib_qp_init_attr *init_attr,
1832 					 struct mlx5_ib_create_qp *ucmd,
1833 					 void *qpc)
1834 {
1835 	enum ib_qp_type qpt = init_attr->qp_type;
1836 	int scqe_sz;
1837 	bool allow_scat_cqe = 0;
1838 
1839 	if (qpt == IB_QPT_UC || qpt == IB_QPT_UD)
1840 		return;
1841 
1842 	if (ucmd)
1843 		allow_scat_cqe = ucmd->flags & MLX5_QP_FLAG_ALLOW_SCATTER_CQE;
1844 
1845 	if (!allow_scat_cqe && init_attr->sq_sig_type != IB_SIGNAL_ALL_WR)
1846 		return;
1847 
1848 	scqe_sz = mlx5_ib_get_cqe_size(init_attr->send_cq);
1849 	if (scqe_sz == 128) {
1850 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA64_CQE);
1851 		return;
1852 	}
1853 
1854 	if (init_attr->qp_type != MLX5_IB_QPT_DCI ||
1855 	    MLX5_CAP_GEN(dev->mdev, dc_req_scat_data_cqe))
1856 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA32_CQE);
1857 }
1858 
1859 static int atomic_size_to_mode(int size_mask)
1860 {
1861 	/* driver does not support atomic_size > 256B
1862 	 * and does not know how to translate bigger sizes
1863 	 */
1864 	int supported_size_mask = size_mask & 0x1ff;
1865 	int log_max_size;
1866 
1867 	if (!supported_size_mask)
1868 		return -EOPNOTSUPP;
1869 
1870 	log_max_size = __fls(supported_size_mask);
1871 
1872 	if (log_max_size > 3)
1873 		return log_max_size;
1874 
1875 	return MLX5_ATOMIC_MODE_8B;
1876 }
1877 
1878 static int get_atomic_mode(struct mlx5_ib_dev *dev,
1879 			   enum ib_qp_type qp_type)
1880 {
1881 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
1882 	u8 atomic = MLX5_CAP_GEN(dev->mdev, atomic);
1883 	int atomic_mode = -EOPNOTSUPP;
1884 	int atomic_size_mask;
1885 
1886 	if (!atomic)
1887 		return -EOPNOTSUPP;
1888 
1889 	if (qp_type == MLX5_IB_QPT_DCT)
1890 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
1891 	else
1892 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
1893 
1894 	if ((atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_CMP_SWAP) ||
1895 	    (atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_FETCH_ADD))
1896 		atomic_mode = atomic_size_to_mode(atomic_size_mask);
1897 
1898 	if (atomic_mode <= 0 &&
1899 	    (atomic_operations & MLX5_ATOMIC_OPS_CMP_SWAP &&
1900 	     atomic_operations & MLX5_ATOMIC_OPS_FETCH_ADD))
1901 		atomic_mode = MLX5_ATOMIC_MODE_IB_COMP;
1902 
1903 	return atomic_mode;
1904 }
1905 
1906 static inline bool check_flags_mask(uint64_t input, uint64_t supported)
1907 {
1908 	return (input & ~supported) == 0;
1909 }
1910 
1911 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1912 			    struct ib_qp_init_attr *init_attr,
1913 			    struct ib_udata *udata, struct mlx5_ib_qp *qp)
1914 {
1915 	struct mlx5_ib_resources *devr = &dev->devr;
1916 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1917 	struct mlx5_core_dev *mdev = dev->mdev;
1918 	struct mlx5_ib_create_qp_resp resp = {};
1919 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
1920 		udata, struct mlx5_ib_ucontext, ibucontext);
1921 	struct mlx5_ib_cq *send_cq;
1922 	struct mlx5_ib_cq *recv_cq;
1923 	unsigned long flags;
1924 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
1925 	struct mlx5_ib_create_qp ucmd;
1926 	struct mlx5_ib_qp_base *base;
1927 	int mlx5_st;
1928 	void *qpc;
1929 	u32 *in;
1930 	int err;
1931 
1932 	mutex_init(&qp->mutex);
1933 	spin_lock_init(&qp->sq.lock);
1934 	spin_lock_init(&qp->rq.lock);
1935 
1936 	mlx5_st = to_mlx5_st(init_attr->qp_type);
1937 	if (mlx5_st < 0)
1938 		return -EINVAL;
1939 
1940 	if (init_attr->rwq_ind_tbl) {
1941 		if (!udata)
1942 			return -ENOSYS;
1943 
1944 		err = create_rss_raw_qp_tir(dev, qp, pd, init_attr, udata);
1945 		return err;
1946 	}
1947 
1948 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
1949 		if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
1950 			mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
1951 			return -EINVAL;
1952 		} else {
1953 			qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1954 		}
1955 	}
1956 
1957 	if (init_attr->create_flags &
1958 			(IB_QP_CREATE_CROSS_CHANNEL |
1959 			 IB_QP_CREATE_MANAGED_SEND |
1960 			 IB_QP_CREATE_MANAGED_RECV)) {
1961 		if (!MLX5_CAP_GEN(mdev, cd)) {
1962 			mlx5_ib_dbg(dev, "cross-channel isn't supported\n");
1963 			return -EINVAL;
1964 		}
1965 		if (init_attr->create_flags & IB_QP_CREATE_CROSS_CHANNEL)
1966 			qp->flags |= MLX5_IB_QP_CROSS_CHANNEL;
1967 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_SEND)
1968 			qp->flags |= MLX5_IB_QP_MANAGED_SEND;
1969 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_RECV)
1970 			qp->flags |= MLX5_IB_QP_MANAGED_RECV;
1971 	}
1972 
1973 	if (init_attr->qp_type == IB_QPT_UD &&
1974 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO))
1975 		if (!MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
1976 			mlx5_ib_dbg(dev, "ipoib UD lso qp isn't supported\n");
1977 			return -EOPNOTSUPP;
1978 		}
1979 
1980 	if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1981 		if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
1982 			mlx5_ib_dbg(dev, "Scatter FCS is supported only for Raw Packet QPs");
1983 			return -EOPNOTSUPP;
1984 		}
1985 		if (!MLX5_CAP_GEN(dev->mdev, eth_net_offloads) ||
1986 		    !MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
1987 			mlx5_ib_dbg(dev, "Scatter FCS isn't supported\n");
1988 			return -EOPNOTSUPP;
1989 		}
1990 		qp->flags |= MLX5_IB_QP_CAP_SCATTER_FCS;
1991 	}
1992 
1993 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1994 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
1995 
1996 	if (init_attr->create_flags & IB_QP_CREATE_CVLAN_STRIPPING) {
1997 		if (!(MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
1998 		      MLX5_CAP_ETH(dev->mdev, vlan_cap)) ||
1999 		    (init_attr->qp_type != IB_QPT_RAW_PACKET))
2000 			return -EOPNOTSUPP;
2001 		qp->flags |= MLX5_IB_QP_CVLAN_STRIPPING;
2002 	}
2003 
2004 	if (udata) {
2005 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
2006 			mlx5_ib_dbg(dev, "copy failed\n");
2007 			return -EFAULT;
2008 		}
2009 
2010 		if (!check_flags_mask(ucmd.flags,
2011 				      MLX5_QP_FLAG_ALLOW_SCATTER_CQE |
2012 				      MLX5_QP_FLAG_BFREG_INDEX |
2013 				      MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE |
2014 				      MLX5_QP_FLAG_SCATTER_CQE |
2015 				      MLX5_QP_FLAG_SIGNATURE |
2016 				      MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC |
2017 				      MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
2018 				      MLX5_QP_FLAG_TUNNEL_OFFLOADS |
2019 				      MLX5_QP_FLAG_TYPE_DCI |
2020 				      MLX5_QP_FLAG_TYPE_DCT))
2021 			return -EINVAL;
2022 
2023 		err = get_qp_user_index(ucontext, &ucmd, udata->inlen, &uidx);
2024 		if (err)
2025 			return err;
2026 
2027 		qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
2028 		if (MLX5_CAP_GEN(dev->mdev, sctr_data_cqe))
2029 			qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
2030 		if (ucmd.flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS) {
2031 			if (init_attr->qp_type != IB_QPT_RAW_PACKET ||
2032 			    !tunnel_offload_supported(mdev)) {
2033 				mlx5_ib_dbg(dev, "Tunnel offload isn't supported\n");
2034 				return -EOPNOTSUPP;
2035 			}
2036 			qp->flags_en |= MLX5_QP_FLAG_TUNNEL_OFFLOADS;
2037 		}
2038 
2039 		if (ucmd.flags & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC) {
2040 			if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2041 				mlx5_ib_dbg(dev, "Self-LB UC isn't supported\n");
2042 				return -EOPNOTSUPP;
2043 			}
2044 			qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
2045 		}
2046 
2047 		if (ucmd.flags & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC) {
2048 			if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2049 				mlx5_ib_dbg(dev, "Self-LB UM isn't supported\n");
2050 				return -EOPNOTSUPP;
2051 			}
2052 			qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC;
2053 		}
2054 
2055 		if (ucmd.flags & MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE) {
2056 			if (init_attr->qp_type != IB_QPT_RC ||
2057 				!MLX5_CAP_GEN(dev->mdev, qp_packet_based)) {
2058 				mlx5_ib_dbg(dev, "packet based credit mode isn't supported\n");
2059 				return -EOPNOTSUPP;
2060 			}
2061 			qp->flags |= MLX5_IB_QP_PACKET_BASED_CREDIT;
2062 		}
2063 
2064 		if (init_attr->create_flags & IB_QP_CREATE_SOURCE_QPN) {
2065 			if (init_attr->qp_type != IB_QPT_UD ||
2066 			    (MLX5_CAP_GEN(dev->mdev, port_type) !=
2067 			     MLX5_CAP_PORT_TYPE_IB) ||
2068 			    !mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS)) {
2069 				mlx5_ib_dbg(dev, "Source QP option isn't supported\n");
2070 				return -EOPNOTSUPP;
2071 			}
2072 
2073 			qp->flags |= MLX5_IB_QP_UNDERLAY;
2074 			qp->underlay_qpn = init_attr->source_qpn;
2075 		}
2076 	} else {
2077 		qp->wq_sig = !!wq_signature;
2078 	}
2079 
2080 	base = (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2081 		qp->flags & MLX5_IB_QP_UNDERLAY) ?
2082 	       &qp->raw_packet_qp.rq.base :
2083 	       &qp->trans_qp.base;
2084 
2085 	qp->has_rq = qp_has_rq(init_attr);
2086 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
2087 			  qp, udata ? &ucmd : NULL);
2088 	if (err) {
2089 		mlx5_ib_dbg(dev, "err %d\n", err);
2090 		return err;
2091 	}
2092 
2093 	if (pd) {
2094 		if (udata) {
2095 			__u32 max_wqes =
2096 				1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
2097 			mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
2098 			if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
2099 			    ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
2100 				mlx5_ib_dbg(dev, "invalid rq params\n");
2101 				return -EINVAL;
2102 			}
2103 			if (ucmd.sq_wqe_count > max_wqes) {
2104 				mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
2105 					    ucmd.sq_wqe_count, max_wqes);
2106 				return -EINVAL;
2107 			}
2108 			if (init_attr->create_flags &
2109 			    mlx5_ib_create_qp_sqpn_qp1()) {
2110 				mlx5_ib_dbg(dev, "user-space is not allowed to create UD QPs spoofing as QP1\n");
2111 				return -EINVAL;
2112 			}
2113 			err = create_user_qp(dev, pd, qp, udata, init_attr, &in,
2114 					     &resp, &inlen, base);
2115 			if (err)
2116 				mlx5_ib_dbg(dev, "err %d\n", err);
2117 		} else {
2118 			err = create_kernel_qp(dev, init_attr, qp, &in, &inlen,
2119 					       base);
2120 			if (err)
2121 				mlx5_ib_dbg(dev, "err %d\n", err);
2122 		}
2123 
2124 		if (err)
2125 			return err;
2126 	} else {
2127 		in = kvzalloc(inlen, GFP_KERNEL);
2128 		if (!in)
2129 			return -ENOMEM;
2130 
2131 		qp->create_type = MLX5_QP_EMPTY;
2132 	}
2133 
2134 	if (is_sqp(init_attr->qp_type))
2135 		qp->port = init_attr->port_num;
2136 
2137 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2138 
2139 	MLX5_SET(qpc, qpc, st, mlx5_st);
2140 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2141 
2142 	if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
2143 		MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
2144 	else
2145 		MLX5_SET(qpc, qpc, latency_sensitive, 1);
2146 
2147 
2148 	if (qp->wq_sig)
2149 		MLX5_SET(qpc, qpc, wq_signature, 1);
2150 
2151 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
2152 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
2153 
2154 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
2155 		MLX5_SET(qpc, qpc, cd_master, 1);
2156 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
2157 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
2158 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
2159 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
2160 	if (qp->flags & MLX5_IB_QP_PACKET_BASED_CREDIT)
2161 		MLX5_SET(qpc, qpc, req_e2e_credit_mode, 1);
2162 	if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
2163 		configure_responder_scat_cqe(init_attr, qpc);
2164 		configure_requester_scat_cqe(dev, init_attr,
2165 					     udata ? &ucmd : NULL,
2166 					     qpc);
2167 	}
2168 
2169 	if (qp->rq.wqe_cnt) {
2170 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2171 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2172 	}
2173 
2174 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
2175 
2176 	if (qp->sq.wqe_cnt) {
2177 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2178 	} else {
2179 		MLX5_SET(qpc, qpc, no_sq, 1);
2180 		if (init_attr->srq &&
2181 		    init_attr->srq->srq_type == IB_SRQT_TM)
2182 			MLX5_SET(qpc, qpc, offload_type,
2183 				 MLX5_QPC_OFFLOAD_TYPE_RNDV);
2184 	}
2185 
2186 	/* Set default resources */
2187 	switch (init_attr->qp_type) {
2188 	case IB_QPT_XRC_TGT:
2189 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
2190 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(devr->c0)->mcq.cqn);
2191 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
2192 		MLX5_SET(qpc, qpc, xrcd, to_mxrcd(init_attr->xrcd)->xrcdn);
2193 		break;
2194 	case IB_QPT_XRC_INI:
2195 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
2196 		MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
2197 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
2198 		break;
2199 	default:
2200 		if (init_attr->srq) {
2201 			MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x0)->xrcdn);
2202 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(init_attr->srq)->msrq.srqn);
2203 		} else {
2204 			MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
2205 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s1)->msrq.srqn);
2206 		}
2207 	}
2208 
2209 	if (init_attr->send_cq)
2210 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(init_attr->send_cq)->mcq.cqn);
2211 
2212 	if (init_attr->recv_cq)
2213 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(init_attr->recv_cq)->mcq.cqn);
2214 
2215 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2216 
2217 	/* 0xffffff means we ask to work with cqe version 0 */
2218 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2219 		MLX5_SET(qpc, qpc, user_index, uidx);
2220 
2221 	/* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
2222 	if (init_attr->qp_type == IB_QPT_UD &&
2223 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)) {
2224 		MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
2225 		qp->flags |= MLX5_IB_QP_LSO;
2226 	}
2227 
2228 	if (init_attr->create_flags & IB_QP_CREATE_PCI_WRITE_END_PADDING) {
2229 		if (!MLX5_CAP_GEN(dev->mdev, end_pad)) {
2230 			mlx5_ib_dbg(dev, "scatter end padding is not supported\n");
2231 			err = -EOPNOTSUPP;
2232 			goto err;
2233 		} else if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2234 			MLX5_SET(qpc, qpc, end_padding_mode,
2235 				 MLX5_WQ_END_PAD_MODE_ALIGN);
2236 		} else {
2237 			qp->flags |= MLX5_IB_QP_PCI_WRITE_END_PADDING;
2238 		}
2239 	}
2240 
2241 	if (inlen < 0) {
2242 		err = -EINVAL;
2243 		goto err;
2244 	}
2245 
2246 	if (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2247 	    qp->flags & MLX5_IB_QP_UNDERLAY) {
2248 		qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd.sq_buf_addr;
2249 		raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
2250 		err = create_raw_packet_qp(dev, qp, in, inlen, pd, udata,
2251 					   &resp);
2252 	} else {
2253 		err = mlx5_core_create_qp(dev->mdev, &base->mqp, in, inlen);
2254 	}
2255 
2256 	if (err) {
2257 		mlx5_ib_dbg(dev, "create qp failed\n");
2258 		goto err_create;
2259 	}
2260 
2261 	kvfree(in);
2262 
2263 	base->container_mibqp = qp;
2264 	base->mqp.event = mlx5_ib_qp_event;
2265 
2266 	get_cqs(init_attr->qp_type, init_attr->send_cq, init_attr->recv_cq,
2267 		&send_cq, &recv_cq);
2268 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2269 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2270 	/* Maintain device to QPs access, needed for further handling via reset
2271 	 * flow
2272 	 */
2273 	list_add_tail(&qp->qps_list, &dev->qp_list);
2274 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2275 	 */
2276 	if (send_cq)
2277 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2278 	if (recv_cq)
2279 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2280 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2281 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2282 
2283 	return 0;
2284 
2285 err_create:
2286 	if (qp->create_type == MLX5_QP_USER)
2287 		destroy_qp_user(dev, pd, qp, base);
2288 	else if (qp->create_type == MLX5_QP_KERNEL)
2289 		destroy_qp_kernel(dev, qp);
2290 
2291 err:
2292 	kvfree(in);
2293 	return err;
2294 }
2295 
2296 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2297 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
2298 {
2299 	if (send_cq) {
2300 		if (recv_cq) {
2301 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2302 				spin_lock(&send_cq->lock);
2303 				spin_lock_nested(&recv_cq->lock,
2304 						 SINGLE_DEPTH_NESTING);
2305 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2306 				spin_lock(&send_cq->lock);
2307 				__acquire(&recv_cq->lock);
2308 			} else {
2309 				spin_lock(&recv_cq->lock);
2310 				spin_lock_nested(&send_cq->lock,
2311 						 SINGLE_DEPTH_NESTING);
2312 			}
2313 		} else {
2314 			spin_lock(&send_cq->lock);
2315 			__acquire(&recv_cq->lock);
2316 		}
2317 	} else if (recv_cq) {
2318 		spin_lock(&recv_cq->lock);
2319 		__acquire(&send_cq->lock);
2320 	} else {
2321 		__acquire(&send_cq->lock);
2322 		__acquire(&recv_cq->lock);
2323 	}
2324 }
2325 
2326 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2327 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
2328 {
2329 	if (send_cq) {
2330 		if (recv_cq) {
2331 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2332 				spin_unlock(&recv_cq->lock);
2333 				spin_unlock(&send_cq->lock);
2334 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2335 				__release(&recv_cq->lock);
2336 				spin_unlock(&send_cq->lock);
2337 			} else {
2338 				spin_unlock(&send_cq->lock);
2339 				spin_unlock(&recv_cq->lock);
2340 			}
2341 		} else {
2342 			__release(&recv_cq->lock);
2343 			spin_unlock(&send_cq->lock);
2344 		}
2345 	} else if (recv_cq) {
2346 		__release(&send_cq->lock);
2347 		spin_unlock(&recv_cq->lock);
2348 	} else {
2349 		__release(&recv_cq->lock);
2350 		__release(&send_cq->lock);
2351 	}
2352 }
2353 
2354 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
2355 {
2356 	return to_mpd(qp->ibqp.pd);
2357 }
2358 
2359 static void get_cqs(enum ib_qp_type qp_type,
2360 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
2361 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
2362 {
2363 	switch (qp_type) {
2364 	case IB_QPT_XRC_TGT:
2365 		*send_cq = NULL;
2366 		*recv_cq = NULL;
2367 		break;
2368 	case MLX5_IB_QPT_REG_UMR:
2369 	case IB_QPT_XRC_INI:
2370 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2371 		*recv_cq = NULL;
2372 		break;
2373 
2374 	case IB_QPT_SMI:
2375 	case MLX5_IB_QPT_HW_GSI:
2376 	case IB_QPT_RC:
2377 	case IB_QPT_UC:
2378 	case IB_QPT_UD:
2379 	case IB_QPT_RAW_IPV6:
2380 	case IB_QPT_RAW_ETHERTYPE:
2381 	case IB_QPT_RAW_PACKET:
2382 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2383 		*recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
2384 		break;
2385 
2386 	case IB_QPT_MAX:
2387 	default:
2388 		*send_cq = NULL;
2389 		*recv_cq = NULL;
2390 		break;
2391 	}
2392 }
2393 
2394 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2395 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
2396 				u8 lag_tx_affinity);
2397 
2398 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
2399 {
2400 	struct mlx5_ib_cq *send_cq, *recv_cq;
2401 	struct mlx5_ib_qp_base *base;
2402 	unsigned long flags;
2403 	int err;
2404 
2405 	if (qp->ibqp.rwq_ind_tbl) {
2406 		destroy_rss_raw_qp_tir(dev, qp);
2407 		return;
2408 	}
2409 
2410 	base = (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
2411 		qp->flags & MLX5_IB_QP_UNDERLAY) ?
2412 	       &qp->raw_packet_qp.rq.base :
2413 	       &qp->trans_qp.base;
2414 
2415 	if (qp->state != IB_QPS_RESET) {
2416 		if (qp->ibqp.qp_type != IB_QPT_RAW_PACKET &&
2417 		    !(qp->flags & MLX5_IB_QP_UNDERLAY)) {
2418 			err = mlx5_core_qp_modify(dev->mdev,
2419 						  MLX5_CMD_OP_2RST_QP, 0,
2420 						  NULL, &base->mqp);
2421 		} else {
2422 			struct mlx5_modify_raw_qp_param raw_qp_param = {
2423 				.operation = MLX5_CMD_OP_2RST_QP
2424 			};
2425 
2426 			err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
2427 		}
2428 		if (err)
2429 			mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
2430 				     base->mqp.qpn);
2431 	}
2432 
2433 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
2434 		&send_cq, &recv_cq);
2435 
2436 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2437 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2438 	/* del from lists under both locks above to protect reset flow paths */
2439 	list_del(&qp->qps_list);
2440 	if (send_cq)
2441 		list_del(&qp->cq_send_list);
2442 
2443 	if (recv_cq)
2444 		list_del(&qp->cq_recv_list);
2445 
2446 	if (qp->create_type == MLX5_QP_KERNEL) {
2447 		__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2448 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
2449 		if (send_cq != recv_cq)
2450 			__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
2451 					   NULL);
2452 	}
2453 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2454 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2455 
2456 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
2457 	    qp->flags & MLX5_IB_QP_UNDERLAY) {
2458 		destroy_raw_packet_qp(dev, qp);
2459 	} else {
2460 		err = mlx5_core_destroy_qp(dev->mdev, &base->mqp);
2461 		if (err)
2462 			mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
2463 				     base->mqp.qpn);
2464 	}
2465 
2466 	if (qp->create_type == MLX5_QP_KERNEL)
2467 		destroy_qp_kernel(dev, qp);
2468 	else if (qp->create_type == MLX5_QP_USER)
2469 		destroy_qp_user(dev, &get_pd(qp)->ibpd, qp, base);
2470 }
2471 
2472 static const char *ib_qp_type_str(enum ib_qp_type type)
2473 {
2474 	switch (type) {
2475 	case IB_QPT_SMI:
2476 		return "IB_QPT_SMI";
2477 	case IB_QPT_GSI:
2478 		return "IB_QPT_GSI";
2479 	case IB_QPT_RC:
2480 		return "IB_QPT_RC";
2481 	case IB_QPT_UC:
2482 		return "IB_QPT_UC";
2483 	case IB_QPT_UD:
2484 		return "IB_QPT_UD";
2485 	case IB_QPT_RAW_IPV6:
2486 		return "IB_QPT_RAW_IPV6";
2487 	case IB_QPT_RAW_ETHERTYPE:
2488 		return "IB_QPT_RAW_ETHERTYPE";
2489 	case IB_QPT_XRC_INI:
2490 		return "IB_QPT_XRC_INI";
2491 	case IB_QPT_XRC_TGT:
2492 		return "IB_QPT_XRC_TGT";
2493 	case IB_QPT_RAW_PACKET:
2494 		return "IB_QPT_RAW_PACKET";
2495 	case MLX5_IB_QPT_REG_UMR:
2496 		return "MLX5_IB_QPT_REG_UMR";
2497 	case IB_QPT_DRIVER:
2498 		return "IB_QPT_DRIVER";
2499 	case IB_QPT_MAX:
2500 	default:
2501 		return "Invalid QP type";
2502 	}
2503 }
2504 
2505 static struct ib_qp *mlx5_ib_create_dct(struct ib_pd *pd,
2506 					struct ib_qp_init_attr *attr,
2507 					struct mlx5_ib_create_qp *ucmd,
2508 					struct ib_udata *udata)
2509 {
2510 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2511 		udata, struct mlx5_ib_ucontext, ibucontext);
2512 	struct mlx5_ib_qp *qp;
2513 	int err = 0;
2514 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
2515 	void *dctc;
2516 
2517 	if (!attr->srq || !attr->recv_cq)
2518 		return ERR_PTR(-EINVAL);
2519 
2520 	err = get_qp_user_index(ucontext, ucmd, sizeof(*ucmd), &uidx);
2521 	if (err)
2522 		return ERR_PTR(err);
2523 
2524 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2525 	if (!qp)
2526 		return ERR_PTR(-ENOMEM);
2527 
2528 	qp->dct.in = kzalloc(MLX5_ST_SZ_BYTES(create_dct_in), GFP_KERNEL);
2529 	if (!qp->dct.in) {
2530 		err = -ENOMEM;
2531 		goto err_free;
2532 	}
2533 
2534 	MLX5_SET(create_dct_in, qp->dct.in, uid, to_mpd(pd)->uid);
2535 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
2536 	qp->qp_sub_type = MLX5_IB_QPT_DCT;
2537 	MLX5_SET(dctc, dctc, pd, to_mpd(pd)->pdn);
2538 	MLX5_SET(dctc, dctc, srqn_xrqn, to_msrq(attr->srq)->msrq.srqn);
2539 	MLX5_SET(dctc, dctc, cqn, to_mcq(attr->recv_cq)->mcq.cqn);
2540 	MLX5_SET64(dctc, dctc, dc_access_key, ucmd->access_key);
2541 	MLX5_SET(dctc, dctc, user_index, uidx);
2542 
2543 	if (ucmd->flags & MLX5_QP_FLAG_SCATTER_CQE)
2544 		configure_responder_scat_cqe(attr, dctc);
2545 
2546 	qp->state = IB_QPS_RESET;
2547 
2548 	return &qp->ibqp;
2549 err_free:
2550 	kfree(qp);
2551 	return ERR_PTR(err);
2552 }
2553 
2554 static int set_mlx_qp_type(struct mlx5_ib_dev *dev,
2555 			   struct ib_qp_init_attr *init_attr,
2556 			   struct mlx5_ib_create_qp *ucmd,
2557 			   struct ib_udata *udata)
2558 {
2559 	enum { MLX_QP_FLAGS = MLX5_QP_FLAG_TYPE_DCT | MLX5_QP_FLAG_TYPE_DCI };
2560 	int err;
2561 
2562 	if (!udata)
2563 		return -EINVAL;
2564 
2565 	if (udata->inlen < sizeof(*ucmd)) {
2566 		mlx5_ib_dbg(dev, "create_qp user command is smaller than expected\n");
2567 		return -EINVAL;
2568 	}
2569 	err = ib_copy_from_udata(ucmd, udata, sizeof(*ucmd));
2570 	if (err)
2571 		return err;
2572 
2573 	if ((ucmd->flags & MLX_QP_FLAGS) == MLX5_QP_FLAG_TYPE_DCI) {
2574 		init_attr->qp_type = MLX5_IB_QPT_DCI;
2575 	} else {
2576 		if ((ucmd->flags & MLX_QP_FLAGS) == MLX5_QP_FLAG_TYPE_DCT) {
2577 			init_attr->qp_type = MLX5_IB_QPT_DCT;
2578 		} else {
2579 			mlx5_ib_dbg(dev, "Invalid QP flags\n");
2580 			return -EINVAL;
2581 		}
2582 	}
2583 
2584 	if (!MLX5_CAP_GEN(dev->mdev, dct)) {
2585 		mlx5_ib_dbg(dev, "DC transport is not supported\n");
2586 		return -EOPNOTSUPP;
2587 	}
2588 
2589 	return 0;
2590 }
2591 
2592 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
2593 				struct ib_qp_init_attr *verbs_init_attr,
2594 				struct ib_udata *udata)
2595 {
2596 	struct mlx5_ib_dev *dev;
2597 	struct mlx5_ib_qp *qp;
2598 	u16 xrcdn = 0;
2599 	int err;
2600 	struct ib_qp_init_attr mlx_init_attr;
2601 	struct ib_qp_init_attr *init_attr = verbs_init_attr;
2602 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2603 		udata, struct mlx5_ib_ucontext, ibucontext);
2604 
2605 	if (pd) {
2606 		dev = to_mdev(pd->device);
2607 
2608 		if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
2609 			if (!ucontext) {
2610 				mlx5_ib_dbg(dev, "Raw Packet QP is not supported for kernel consumers\n");
2611 				return ERR_PTR(-EINVAL);
2612 			} else if (!ucontext->cqe_version) {
2613 				mlx5_ib_dbg(dev, "Raw Packet QP is only supported for CQE version > 0\n");
2614 				return ERR_PTR(-EINVAL);
2615 			}
2616 		}
2617 	} else {
2618 		/* being cautious here */
2619 		if (init_attr->qp_type != IB_QPT_XRC_TGT &&
2620 		    init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
2621 			pr_warn("%s: no PD for transport %s\n", __func__,
2622 				ib_qp_type_str(init_attr->qp_type));
2623 			return ERR_PTR(-EINVAL);
2624 		}
2625 		dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
2626 	}
2627 
2628 	if (init_attr->qp_type == IB_QPT_DRIVER) {
2629 		struct mlx5_ib_create_qp ucmd;
2630 
2631 		init_attr = &mlx_init_attr;
2632 		memcpy(init_attr, verbs_init_attr, sizeof(*verbs_init_attr));
2633 		err = set_mlx_qp_type(dev, init_attr, &ucmd, udata);
2634 		if (err)
2635 			return ERR_PTR(err);
2636 
2637 		if (init_attr->qp_type == MLX5_IB_QPT_DCI) {
2638 			if (init_attr->cap.max_recv_wr ||
2639 			    init_attr->cap.max_recv_sge) {
2640 				mlx5_ib_dbg(dev, "DCI QP requires zero size receive queue\n");
2641 				return ERR_PTR(-EINVAL);
2642 			}
2643 		} else {
2644 			return mlx5_ib_create_dct(pd, init_attr, &ucmd, udata);
2645 		}
2646 	}
2647 
2648 	switch (init_attr->qp_type) {
2649 	case IB_QPT_XRC_TGT:
2650 	case IB_QPT_XRC_INI:
2651 		if (!MLX5_CAP_GEN(dev->mdev, xrc)) {
2652 			mlx5_ib_dbg(dev, "XRC not supported\n");
2653 			return ERR_PTR(-ENOSYS);
2654 		}
2655 		init_attr->recv_cq = NULL;
2656 		if (init_attr->qp_type == IB_QPT_XRC_TGT) {
2657 			xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
2658 			init_attr->send_cq = NULL;
2659 		}
2660 
2661 		/* fall through */
2662 	case IB_QPT_RAW_PACKET:
2663 	case IB_QPT_RC:
2664 	case IB_QPT_UC:
2665 	case IB_QPT_UD:
2666 	case IB_QPT_SMI:
2667 	case MLX5_IB_QPT_HW_GSI:
2668 	case MLX5_IB_QPT_REG_UMR:
2669 	case MLX5_IB_QPT_DCI:
2670 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2671 		if (!qp)
2672 			return ERR_PTR(-ENOMEM);
2673 
2674 		err = create_qp_common(dev, pd, init_attr, udata, qp);
2675 		if (err) {
2676 			mlx5_ib_dbg(dev, "create_qp_common failed\n");
2677 			kfree(qp);
2678 			return ERR_PTR(err);
2679 		}
2680 
2681 		if (is_qp0(init_attr->qp_type))
2682 			qp->ibqp.qp_num = 0;
2683 		else if (is_qp1(init_attr->qp_type))
2684 			qp->ibqp.qp_num = 1;
2685 		else
2686 			qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
2687 
2688 		mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
2689 			    qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
2690 			    init_attr->recv_cq ? to_mcq(init_attr->recv_cq)->mcq.cqn : -1,
2691 			    init_attr->send_cq ? to_mcq(init_attr->send_cq)->mcq.cqn : -1);
2692 
2693 		qp->trans_qp.xrcdn = xrcdn;
2694 
2695 		break;
2696 
2697 	case IB_QPT_GSI:
2698 		return mlx5_ib_gsi_create_qp(pd, init_attr);
2699 
2700 	case IB_QPT_RAW_IPV6:
2701 	case IB_QPT_RAW_ETHERTYPE:
2702 	case IB_QPT_MAX:
2703 	default:
2704 		mlx5_ib_dbg(dev, "unsupported qp type %d\n",
2705 			    init_attr->qp_type);
2706 		/* Don't support raw QPs */
2707 		return ERR_PTR(-EINVAL);
2708 	}
2709 
2710 	if (verbs_init_attr->qp_type == IB_QPT_DRIVER)
2711 		qp->qp_sub_type = init_attr->qp_type;
2712 
2713 	return &qp->ibqp;
2714 }
2715 
2716 static int mlx5_ib_destroy_dct(struct mlx5_ib_qp *mqp)
2717 {
2718 	struct mlx5_ib_dev *dev = to_mdev(mqp->ibqp.device);
2719 
2720 	if (mqp->state == IB_QPS_RTR) {
2721 		int err;
2722 
2723 		err = mlx5_core_destroy_dct(dev->mdev, &mqp->dct.mdct);
2724 		if (err) {
2725 			mlx5_ib_warn(dev, "failed to destroy DCT %d\n", err);
2726 			return err;
2727 		}
2728 	}
2729 
2730 	kfree(mqp->dct.in);
2731 	kfree(mqp);
2732 	return 0;
2733 }
2734 
2735 int mlx5_ib_destroy_qp(struct ib_qp *qp)
2736 {
2737 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2738 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2739 
2740 	if (unlikely(qp->qp_type == IB_QPT_GSI))
2741 		return mlx5_ib_gsi_destroy_qp(qp);
2742 
2743 	if (mqp->qp_sub_type == MLX5_IB_QPT_DCT)
2744 		return mlx5_ib_destroy_dct(mqp);
2745 
2746 	destroy_qp_common(dev, mqp);
2747 
2748 	kfree(mqp);
2749 
2750 	return 0;
2751 }
2752 
2753 static int to_mlx5_access_flags(struct mlx5_ib_qp *qp,
2754 				const struct ib_qp_attr *attr,
2755 				int attr_mask, __be32 *hw_access_flags_be)
2756 {
2757 	u8 dest_rd_atomic;
2758 	u32 access_flags, hw_access_flags = 0;
2759 
2760 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
2761 
2762 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2763 		dest_rd_atomic = attr->max_dest_rd_atomic;
2764 	else
2765 		dest_rd_atomic = qp->trans_qp.resp_depth;
2766 
2767 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2768 		access_flags = attr->qp_access_flags;
2769 	else
2770 		access_flags = qp->trans_qp.atomic_rd_en;
2771 
2772 	if (!dest_rd_atomic)
2773 		access_flags &= IB_ACCESS_REMOTE_WRITE;
2774 
2775 	if (access_flags & IB_ACCESS_REMOTE_READ)
2776 		hw_access_flags |= MLX5_QP_BIT_RRE;
2777 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC) {
2778 		int atomic_mode;
2779 
2780 		atomic_mode = get_atomic_mode(dev, qp->ibqp.qp_type);
2781 		if (atomic_mode < 0)
2782 			return -EOPNOTSUPP;
2783 
2784 		hw_access_flags |= MLX5_QP_BIT_RAE;
2785 		hw_access_flags |= atomic_mode << MLX5_ATOMIC_MODE_OFFSET;
2786 	}
2787 
2788 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
2789 		hw_access_flags |= MLX5_QP_BIT_RWE;
2790 
2791 	*hw_access_flags_be = cpu_to_be32(hw_access_flags);
2792 
2793 	return 0;
2794 }
2795 
2796 enum {
2797 	MLX5_PATH_FLAG_FL	= 1 << 0,
2798 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
2799 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
2800 };
2801 
2802 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
2803 {
2804 	if (rate == IB_RATE_PORT_CURRENT)
2805 		return 0;
2806 
2807 	if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_600_GBPS)
2808 		return -EINVAL;
2809 
2810 	while (rate != IB_RATE_PORT_CURRENT &&
2811 	       !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
2812 		 MLX5_CAP_GEN(dev->mdev, stat_rate_support)))
2813 		--rate;
2814 
2815 	return rate ? rate + MLX5_STAT_RATE_OFFSET : rate;
2816 }
2817 
2818 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
2819 				      struct mlx5_ib_sq *sq, u8 sl,
2820 				      struct ib_pd *pd)
2821 {
2822 	void *in;
2823 	void *tisc;
2824 	int inlen;
2825 	int err;
2826 
2827 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2828 	in = kvzalloc(inlen, GFP_KERNEL);
2829 	if (!in)
2830 		return -ENOMEM;
2831 
2832 	MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
2833 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
2834 
2835 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2836 	MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
2837 
2838 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2839 
2840 	kvfree(in);
2841 
2842 	return err;
2843 }
2844 
2845 static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,
2846 					 struct mlx5_ib_sq *sq, u8 tx_affinity,
2847 					 struct ib_pd *pd)
2848 {
2849 	void *in;
2850 	void *tisc;
2851 	int inlen;
2852 	int err;
2853 
2854 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2855 	in = kvzalloc(inlen, GFP_KERNEL);
2856 	if (!in)
2857 		return -ENOMEM;
2858 
2859 	MLX5_SET(modify_tis_in, in, bitmask.lag_tx_port_affinity, 1);
2860 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
2861 
2862 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2863 	MLX5_SET(tisc, tisc, lag_tx_port_affinity, tx_affinity);
2864 
2865 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2866 
2867 	kvfree(in);
2868 
2869 	return err;
2870 }
2871 
2872 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2873 			 const struct rdma_ah_attr *ah,
2874 			 struct mlx5_qp_path *path, u8 port, int attr_mask,
2875 			 u32 path_flags, const struct ib_qp_attr *attr,
2876 			 bool alt)
2877 {
2878 	const struct ib_global_route *grh = rdma_ah_read_grh(ah);
2879 	int err;
2880 	enum ib_gid_type gid_type;
2881 	u8 ah_flags = rdma_ah_get_ah_flags(ah);
2882 	u8 sl = rdma_ah_get_sl(ah);
2883 
2884 	if (attr_mask & IB_QP_PKEY_INDEX)
2885 		path->pkey_index = cpu_to_be16(alt ? attr->alt_pkey_index :
2886 						     attr->pkey_index);
2887 
2888 	if (ah_flags & IB_AH_GRH) {
2889 		if (grh->sgid_index >=
2890 		    dev->mdev->port_caps[port - 1].gid_table_len) {
2891 			pr_err("sgid_index (%u) too large. max is %d\n",
2892 			       grh->sgid_index,
2893 			       dev->mdev->port_caps[port - 1].gid_table_len);
2894 			return -EINVAL;
2895 		}
2896 	}
2897 
2898 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
2899 		if (!(ah_flags & IB_AH_GRH))
2900 			return -EINVAL;
2901 
2902 		memcpy(path->rmac, ah->roce.dmac, sizeof(ah->roce.dmac));
2903 		if (qp->ibqp.qp_type == IB_QPT_RC ||
2904 		    qp->ibqp.qp_type == IB_QPT_UC ||
2905 		    qp->ibqp.qp_type == IB_QPT_XRC_INI ||
2906 		    qp->ibqp.qp_type == IB_QPT_XRC_TGT)
2907 			path->udp_sport =
2908 				mlx5_get_roce_udp_sport(dev, ah->grh.sgid_attr);
2909 		path->dci_cfi_prio_sl = (sl & 0x7) << 4;
2910 		gid_type = ah->grh.sgid_attr->gid_type;
2911 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
2912 			path->ecn_dscp = (grh->traffic_class >> 2) & 0x3f;
2913 	} else {
2914 		path->fl_free_ar = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
2915 		path->fl_free_ar |=
2916 			(path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x40 : 0;
2917 		path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
2918 		path->grh_mlid = rdma_ah_get_path_bits(ah) & 0x7f;
2919 		if (ah_flags & IB_AH_GRH)
2920 			path->grh_mlid	|= 1 << 7;
2921 		path->dci_cfi_prio_sl = sl & 0xf;
2922 	}
2923 
2924 	if (ah_flags & IB_AH_GRH) {
2925 		path->mgid_index = grh->sgid_index;
2926 		path->hop_limit  = grh->hop_limit;
2927 		path->tclass_flowlabel =
2928 			cpu_to_be32((grh->traffic_class << 20) |
2929 				    (grh->flow_label));
2930 		memcpy(path->rgid, grh->dgid.raw, 16);
2931 	}
2932 
2933 	err = ib_rate_to_mlx5(dev, rdma_ah_get_static_rate(ah));
2934 	if (err < 0)
2935 		return err;
2936 	path->static_rate = err;
2937 	path->port = port;
2938 
2939 	if (attr_mask & IB_QP_TIMEOUT)
2940 		path->ackto_lt = (alt ? attr->alt_timeout : attr->timeout) << 3;
2941 
2942 	if ((qp->ibqp.qp_type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
2943 		return modify_raw_packet_eth_prio(dev->mdev,
2944 						  &qp->raw_packet_qp.sq,
2945 						  sl & 0xf, qp->ibqp.pd);
2946 
2947 	return 0;
2948 }
2949 
2950 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
2951 	[MLX5_QP_STATE_INIT] = {
2952 		[MLX5_QP_STATE_INIT] = {
2953 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
2954 					  MLX5_QP_OPTPAR_RAE		|
2955 					  MLX5_QP_OPTPAR_RWE		|
2956 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
2957 					  MLX5_QP_OPTPAR_PRI_PORT,
2958 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
2959 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
2960 					  MLX5_QP_OPTPAR_PRI_PORT,
2961 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
2962 					  MLX5_QP_OPTPAR_Q_KEY		|
2963 					  MLX5_QP_OPTPAR_PRI_PORT,
2964 		},
2965 		[MLX5_QP_STATE_RTR] = {
2966 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
2967 					  MLX5_QP_OPTPAR_RRE            |
2968 					  MLX5_QP_OPTPAR_RAE            |
2969 					  MLX5_QP_OPTPAR_RWE            |
2970 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2971 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
2972 					  MLX5_QP_OPTPAR_RWE            |
2973 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2974 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
2975 					  MLX5_QP_OPTPAR_Q_KEY,
2976 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
2977 					   MLX5_QP_OPTPAR_Q_KEY,
2978 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2979 					  MLX5_QP_OPTPAR_RRE            |
2980 					  MLX5_QP_OPTPAR_RAE            |
2981 					  MLX5_QP_OPTPAR_RWE            |
2982 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2983 		},
2984 	},
2985 	[MLX5_QP_STATE_RTR] = {
2986 		[MLX5_QP_STATE_RTS] = {
2987 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
2988 					  MLX5_QP_OPTPAR_RRE		|
2989 					  MLX5_QP_OPTPAR_RAE		|
2990 					  MLX5_QP_OPTPAR_RWE		|
2991 					  MLX5_QP_OPTPAR_PM_STATE	|
2992 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
2993 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
2994 					  MLX5_QP_OPTPAR_RWE		|
2995 					  MLX5_QP_OPTPAR_PM_STATE,
2996 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
2997 		},
2998 	},
2999 	[MLX5_QP_STATE_RTS] = {
3000 		[MLX5_QP_STATE_RTS] = {
3001 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3002 					  MLX5_QP_OPTPAR_RAE		|
3003 					  MLX5_QP_OPTPAR_RWE		|
3004 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3005 					  MLX5_QP_OPTPAR_PM_STATE	|
3006 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3007 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3008 					  MLX5_QP_OPTPAR_PM_STATE	|
3009 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3010 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
3011 					  MLX5_QP_OPTPAR_SRQN		|
3012 					  MLX5_QP_OPTPAR_CQN_RCV,
3013 		},
3014 	},
3015 	[MLX5_QP_STATE_SQER] = {
3016 		[MLX5_QP_STATE_RTS] = {
3017 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
3018 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
3019 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
3020 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3021 					   MLX5_QP_OPTPAR_RWE		|
3022 					   MLX5_QP_OPTPAR_RAE		|
3023 					   MLX5_QP_OPTPAR_RRE,
3024 		},
3025 	},
3026 };
3027 
3028 static int ib_nr_to_mlx5_nr(int ib_mask)
3029 {
3030 	switch (ib_mask) {
3031 	case IB_QP_STATE:
3032 		return 0;
3033 	case IB_QP_CUR_STATE:
3034 		return 0;
3035 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
3036 		return 0;
3037 	case IB_QP_ACCESS_FLAGS:
3038 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
3039 			MLX5_QP_OPTPAR_RAE;
3040 	case IB_QP_PKEY_INDEX:
3041 		return MLX5_QP_OPTPAR_PKEY_INDEX;
3042 	case IB_QP_PORT:
3043 		return MLX5_QP_OPTPAR_PRI_PORT;
3044 	case IB_QP_QKEY:
3045 		return MLX5_QP_OPTPAR_Q_KEY;
3046 	case IB_QP_AV:
3047 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
3048 			MLX5_QP_OPTPAR_PRI_PORT;
3049 	case IB_QP_PATH_MTU:
3050 		return 0;
3051 	case IB_QP_TIMEOUT:
3052 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
3053 	case IB_QP_RETRY_CNT:
3054 		return MLX5_QP_OPTPAR_RETRY_COUNT;
3055 	case IB_QP_RNR_RETRY:
3056 		return MLX5_QP_OPTPAR_RNR_RETRY;
3057 	case IB_QP_RQ_PSN:
3058 		return 0;
3059 	case IB_QP_MAX_QP_RD_ATOMIC:
3060 		return MLX5_QP_OPTPAR_SRA_MAX;
3061 	case IB_QP_ALT_PATH:
3062 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
3063 	case IB_QP_MIN_RNR_TIMER:
3064 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
3065 	case IB_QP_SQ_PSN:
3066 		return 0;
3067 	case IB_QP_MAX_DEST_RD_ATOMIC:
3068 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
3069 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
3070 	case IB_QP_PATH_MIG_STATE:
3071 		return MLX5_QP_OPTPAR_PM_STATE;
3072 	case IB_QP_CAP:
3073 		return 0;
3074 	case IB_QP_DEST_QPN:
3075 		return 0;
3076 	}
3077 	return 0;
3078 }
3079 
3080 static int ib_mask_to_mlx5_opt(int ib_mask)
3081 {
3082 	int result = 0;
3083 	int i;
3084 
3085 	for (i = 0; i < 8 * sizeof(int); i++) {
3086 		if ((1 << i) & ib_mask)
3087 			result |= ib_nr_to_mlx5_nr(1 << i);
3088 	}
3089 
3090 	return result;
3091 }
3092 
3093 static int modify_raw_packet_qp_rq(
3094 	struct mlx5_ib_dev *dev, struct mlx5_ib_rq *rq, int new_state,
3095 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3096 {
3097 	void *in;
3098 	void *rqc;
3099 	int inlen;
3100 	int err;
3101 
3102 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
3103 	in = kvzalloc(inlen, GFP_KERNEL);
3104 	if (!in)
3105 		return -ENOMEM;
3106 
3107 	MLX5_SET(modify_rq_in, in, rq_state, rq->state);
3108 	MLX5_SET(modify_rq_in, in, uid, to_mpd(pd)->uid);
3109 
3110 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
3111 	MLX5_SET(rqc, rqc, state, new_state);
3112 
3113 	if (raw_qp_param->set_mask & MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID) {
3114 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
3115 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
3116 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
3117 			MLX5_SET(rqc, rqc, counter_set_id, raw_qp_param->rq_q_ctr_id);
3118 		} else
3119 			dev_info_once(
3120 				&dev->ib_dev.dev,
3121 				"RAW PACKET QP counters are not supported on current FW\n");
3122 	}
3123 
3124 	err = mlx5_core_modify_rq(dev->mdev, rq->base.mqp.qpn, in, inlen);
3125 	if (err)
3126 		goto out;
3127 
3128 	rq->state = new_state;
3129 
3130 out:
3131 	kvfree(in);
3132 	return err;
3133 }
3134 
3135 static int modify_raw_packet_qp_sq(
3136 	struct mlx5_core_dev *dev, struct mlx5_ib_sq *sq, int new_state,
3137 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3138 {
3139 	struct mlx5_ib_qp *ibqp = sq->base.container_mibqp;
3140 	struct mlx5_rate_limit old_rl = ibqp->rl;
3141 	struct mlx5_rate_limit new_rl = old_rl;
3142 	bool new_rate_added = false;
3143 	u16 rl_index = 0;
3144 	void *in;
3145 	void *sqc;
3146 	int inlen;
3147 	int err;
3148 
3149 	inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
3150 	in = kvzalloc(inlen, GFP_KERNEL);
3151 	if (!in)
3152 		return -ENOMEM;
3153 
3154 	MLX5_SET(modify_sq_in, in, uid, to_mpd(pd)->uid);
3155 	MLX5_SET(modify_sq_in, in, sq_state, sq->state);
3156 
3157 	sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
3158 	MLX5_SET(sqc, sqc, state, new_state);
3159 
3160 	if (raw_qp_param->set_mask & MLX5_RAW_QP_RATE_LIMIT) {
3161 		if (new_state != MLX5_SQC_STATE_RDY)
3162 			pr_warn("%s: Rate limit can only be changed when SQ is moving to RDY\n",
3163 				__func__);
3164 		else
3165 			new_rl = raw_qp_param->rl;
3166 	}
3167 
3168 	if (!mlx5_rl_are_equal(&old_rl, &new_rl)) {
3169 		if (new_rl.rate) {
3170 			err = mlx5_rl_add_rate(dev, &rl_index, &new_rl);
3171 			if (err) {
3172 				pr_err("Failed configuring rate limit(err %d): \
3173 				       rate %u, max_burst_sz %u, typical_pkt_sz %u\n",
3174 				       err, new_rl.rate, new_rl.max_burst_sz,
3175 				       new_rl.typical_pkt_sz);
3176 
3177 				goto out;
3178 			}
3179 			new_rate_added = true;
3180 		}
3181 
3182 		MLX5_SET64(modify_sq_in, in, modify_bitmask, 1);
3183 		/* index 0 means no limit */
3184 		MLX5_SET(sqc, sqc, packet_pacing_rate_limit_index, rl_index);
3185 	}
3186 
3187 	err = mlx5_core_modify_sq(dev, sq->base.mqp.qpn, in, inlen);
3188 	if (err) {
3189 		/* Remove new rate from table if failed */
3190 		if (new_rate_added)
3191 			mlx5_rl_remove_rate(dev, &new_rl);
3192 		goto out;
3193 	}
3194 
3195 	/* Only remove the old rate after new rate was set */
3196 	if ((old_rl.rate &&
3197 	     !mlx5_rl_are_equal(&old_rl, &new_rl)) ||
3198 	    (new_state != MLX5_SQC_STATE_RDY))
3199 		mlx5_rl_remove_rate(dev, &old_rl);
3200 
3201 	ibqp->rl = new_rl;
3202 	sq->state = new_state;
3203 
3204 out:
3205 	kvfree(in);
3206 	return err;
3207 }
3208 
3209 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3210 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
3211 				u8 tx_affinity)
3212 {
3213 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
3214 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
3215 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
3216 	int modify_rq = !!qp->rq.wqe_cnt;
3217 	int modify_sq = !!qp->sq.wqe_cnt;
3218 	int rq_state;
3219 	int sq_state;
3220 	int err;
3221 
3222 	switch (raw_qp_param->operation) {
3223 	case MLX5_CMD_OP_RST2INIT_QP:
3224 		rq_state = MLX5_RQC_STATE_RDY;
3225 		sq_state = MLX5_SQC_STATE_RDY;
3226 		break;
3227 	case MLX5_CMD_OP_2ERR_QP:
3228 		rq_state = MLX5_RQC_STATE_ERR;
3229 		sq_state = MLX5_SQC_STATE_ERR;
3230 		break;
3231 	case MLX5_CMD_OP_2RST_QP:
3232 		rq_state = MLX5_RQC_STATE_RST;
3233 		sq_state = MLX5_SQC_STATE_RST;
3234 		break;
3235 	case MLX5_CMD_OP_RTR2RTS_QP:
3236 	case MLX5_CMD_OP_RTS2RTS_QP:
3237 		if (raw_qp_param->set_mask ==
3238 		    MLX5_RAW_QP_RATE_LIMIT) {
3239 			modify_rq = 0;
3240 			sq_state = sq->state;
3241 		} else {
3242 			return raw_qp_param->set_mask ? -EINVAL : 0;
3243 		}
3244 		break;
3245 	case MLX5_CMD_OP_INIT2INIT_QP:
3246 	case MLX5_CMD_OP_INIT2RTR_QP:
3247 		if (raw_qp_param->set_mask)
3248 			return -EINVAL;
3249 		else
3250 			return 0;
3251 	default:
3252 		WARN_ON(1);
3253 		return -EINVAL;
3254 	}
3255 
3256 	if (modify_rq) {
3257 		err =  modify_raw_packet_qp_rq(dev, rq, rq_state, raw_qp_param,
3258 					       qp->ibqp.pd);
3259 		if (err)
3260 			return err;
3261 	}
3262 
3263 	if (modify_sq) {
3264 		if (tx_affinity) {
3265 			err = modify_raw_packet_tx_affinity(dev->mdev, sq,
3266 							    tx_affinity,
3267 							    qp->ibqp.pd);
3268 			if (err)
3269 				return err;
3270 		}
3271 
3272 		return modify_raw_packet_qp_sq(dev->mdev, sq, sq_state,
3273 					       raw_qp_param, qp->ibqp.pd);
3274 	}
3275 
3276 	return 0;
3277 }
3278 
3279 static unsigned int get_tx_affinity(struct mlx5_ib_dev *dev,
3280 				    struct mlx5_ib_pd *pd,
3281 				    struct mlx5_ib_qp_base *qp_base,
3282 				    u8 port_num, struct ib_udata *udata)
3283 {
3284 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3285 		udata, struct mlx5_ib_ucontext, ibucontext);
3286 	unsigned int tx_port_affinity;
3287 
3288 	if (ucontext) {
3289 		tx_port_affinity = (unsigned int)atomic_add_return(
3290 					   1, &ucontext->tx_port_affinity) %
3291 					   MLX5_MAX_PORTS +
3292 				   1;
3293 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x ucontext %p\n",
3294 				tx_port_affinity, qp_base->mqp.qpn, ucontext);
3295 	} else {
3296 		tx_port_affinity =
3297 			(unsigned int)atomic_add_return(
3298 				1, &dev->roce[port_num].tx_port_affinity) %
3299 				MLX5_MAX_PORTS +
3300 			1;
3301 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x\n",
3302 				tx_port_affinity, qp_base->mqp.qpn);
3303 	}
3304 
3305 	return tx_port_affinity;
3306 }
3307 
3308 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
3309 			       const struct ib_qp_attr *attr, int attr_mask,
3310 			       enum ib_qp_state cur_state,
3311 			       enum ib_qp_state new_state,
3312 			       const struct mlx5_ib_modify_qp *ucmd,
3313 			       struct ib_udata *udata)
3314 {
3315 	static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
3316 		[MLX5_QP_STATE_RST] = {
3317 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3318 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3319 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_RST2INIT_QP,
3320 		},
3321 		[MLX5_QP_STATE_INIT]  = {
3322 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3323 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3324 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_INIT2INIT_QP,
3325 			[MLX5_QP_STATE_RTR]	= MLX5_CMD_OP_INIT2RTR_QP,
3326 		},
3327 		[MLX5_QP_STATE_RTR]   = {
3328 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3329 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3330 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTR2RTS_QP,
3331 		},
3332 		[MLX5_QP_STATE_RTS]   = {
3333 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3334 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3335 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTS2RTS_QP,
3336 		},
3337 		[MLX5_QP_STATE_SQD] = {
3338 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3339 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3340 		},
3341 		[MLX5_QP_STATE_SQER] = {
3342 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3343 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3344 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQERR2RTS_QP,
3345 		},
3346 		[MLX5_QP_STATE_ERR] = {
3347 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3348 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3349 		}
3350 	};
3351 
3352 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3353 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3354 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
3355 	struct mlx5_ib_cq *send_cq, *recv_cq;
3356 	struct mlx5_qp_context *context;
3357 	struct mlx5_ib_pd *pd;
3358 	struct mlx5_ib_port *mibport = NULL;
3359 	enum mlx5_qp_state mlx5_cur, mlx5_new;
3360 	enum mlx5_qp_optpar optpar;
3361 	int mlx5_st;
3362 	int err;
3363 	u16 op;
3364 	u8 tx_affinity = 0;
3365 
3366 	mlx5_st = to_mlx5_st(ibqp->qp_type == IB_QPT_DRIVER ?
3367 			     qp->qp_sub_type : ibqp->qp_type);
3368 	if (mlx5_st < 0)
3369 		return -EINVAL;
3370 
3371 	context = kzalloc(sizeof(*context), GFP_KERNEL);
3372 	if (!context)
3373 		return -ENOMEM;
3374 
3375 	pd = get_pd(qp);
3376 	context->flags = cpu_to_be32(mlx5_st << 16);
3377 
3378 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
3379 		context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
3380 	} else {
3381 		switch (attr->path_mig_state) {
3382 		case IB_MIG_MIGRATED:
3383 			context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
3384 			break;
3385 		case IB_MIG_REARM:
3386 			context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
3387 			break;
3388 		case IB_MIG_ARMED:
3389 			context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
3390 			break;
3391 		}
3392 	}
3393 
3394 	if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
3395 		if ((ibqp->qp_type == IB_QPT_RC) ||
3396 		    (ibqp->qp_type == IB_QPT_UD &&
3397 		     !(qp->flags & MLX5_IB_QP_SQPN_QP1)) ||
3398 		    (ibqp->qp_type == IB_QPT_UC) ||
3399 		    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
3400 		    (ibqp->qp_type == IB_QPT_XRC_INI) ||
3401 		    (ibqp->qp_type == IB_QPT_XRC_TGT)) {
3402 			if (dev->lag_active) {
3403 				u8 p = mlx5_core_native_port_num(dev->mdev);
3404 				tx_affinity = get_tx_affinity(dev, pd, base, p,
3405 							      udata);
3406 				context->flags |= cpu_to_be32(tx_affinity << 24);
3407 			}
3408 		}
3409 	}
3410 
3411 	if (is_sqp(ibqp->qp_type)) {
3412 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
3413 	} else if ((ibqp->qp_type == IB_QPT_UD &&
3414 		    !(qp->flags & MLX5_IB_QP_UNDERLAY)) ||
3415 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
3416 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
3417 	} else if (attr_mask & IB_QP_PATH_MTU) {
3418 		if (attr->path_mtu < IB_MTU_256 ||
3419 		    attr->path_mtu > IB_MTU_4096) {
3420 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
3421 			err = -EINVAL;
3422 			goto out;
3423 		}
3424 		context->mtu_msgmax = (attr->path_mtu << 5) |
3425 				      (u8)MLX5_CAP_GEN(dev->mdev, log_max_msg);
3426 	}
3427 
3428 	if (attr_mask & IB_QP_DEST_QPN)
3429 		context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
3430 
3431 	if (attr_mask & IB_QP_PKEY_INDEX)
3432 		context->pri_path.pkey_index = cpu_to_be16(attr->pkey_index);
3433 
3434 	/* todo implement counter_index functionality */
3435 
3436 	if (is_sqp(ibqp->qp_type))
3437 		context->pri_path.port = qp->port;
3438 
3439 	if (attr_mask & IB_QP_PORT)
3440 		context->pri_path.port = attr->port_num;
3441 
3442 	if (attr_mask & IB_QP_AV) {
3443 		err = mlx5_set_path(dev, qp, &attr->ah_attr, &context->pri_path,
3444 				    attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
3445 				    attr_mask, 0, attr, false);
3446 		if (err)
3447 			goto out;
3448 	}
3449 
3450 	if (attr_mask & IB_QP_TIMEOUT)
3451 		context->pri_path.ackto_lt |= attr->timeout << 3;
3452 
3453 	if (attr_mask & IB_QP_ALT_PATH) {
3454 		err = mlx5_set_path(dev, qp, &attr->alt_ah_attr,
3455 				    &context->alt_path,
3456 				    attr->alt_port_num,
3457 				    attr_mask | IB_QP_PKEY_INDEX | IB_QP_TIMEOUT,
3458 				    0, attr, true);
3459 		if (err)
3460 			goto out;
3461 	}
3462 
3463 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
3464 		&send_cq, &recv_cq);
3465 
3466 	context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
3467 	context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
3468 	context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
3469 	context->params1  = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
3470 
3471 	if (attr_mask & IB_QP_RNR_RETRY)
3472 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
3473 
3474 	if (attr_mask & IB_QP_RETRY_CNT)
3475 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
3476 
3477 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
3478 		if (attr->max_rd_atomic)
3479 			context->params1 |=
3480 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
3481 	}
3482 
3483 	if (attr_mask & IB_QP_SQ_PSN)
3484 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
3485 
3486 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
3487 		if (attr->max_dest_rd_atomic)
3488 			context->params2 |=
3489 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
3490 	}
3491 
3492 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
3493 		__be32 access_flags;
3494 
3495 		err = to_mlx5_access_flags(qp, attr, attr_mask, &access_flags);
3496 		if (err)
3497 			goto out;
3498 
3499 		context->params2 |= access_flags;
3500 	}
3501 
3502 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
3503 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
3504 
3505 	if (attr_mask & IB_QP_RQ_PSN)
3506 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
3507 
3508 	if (attr_mask & IB_QP_QKEY)
3509 		context->qkey = cpu_to_be32(attr->qkey);
3510 
3511 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
3512 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
3513 
3514 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
3515 		u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
3516 			       qp->port) - 1;
3517 
3518 		/* Underlay port should be used - index 0 function per port */
3519 		if (qp->flags & MLX5_IB_QP_UNDERLAY)
3520 			port_num = 0;
3521 
3522 		mibport = &dev->port[port_num];
3523 		context->qp_counter_set_usr_page |=
3524 			cpu_to_be32((u32)(mibport->cnts.set_id) << 24);
3525 	}
3526 
3527 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
3528 		context->sq_crq_size |= cpu_to_be16(1 << 4);
3529 
3530 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
3531 		context->deth_sqpn = cpu_to_be32(1);
3532 
3533 	mlx5_cur = to_mlx5_state(cur_state);
3534 	mlx5_new = to_mlx5_state(new_state);
3535 
3536 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
3537 	    !optab[mlx5_cur][mlx5_new]) {
3538 		err = -EINVAL;
3539 		goto out;
3540 	}
3541 
3542 	op = optab[mlx5_cur][mlx5_new];
3543 	optpar = ib_mask_to_mlx5_opt(attr_mask);
3544 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
3545 
3546 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
3547 	    qp->flags & MLX5_IB_QP_UNDERLAY) {
3548 		struct mlx5_modify_raw_qp_param raw_qp_param = {};
3549 
3550 		raw_qp_param.operation = op;
3551 		if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
3552 			raw_qp_param.rq_q_ctr_id = mibport->cnts.set_id;
3553 			raw_qp_param.set_mask |= MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID;
3554 		}
3555 
3556 		if (attr_mask & IB_QP_RATE_LIMIT) {
3557 			raw_qp_param.rl.rate = attr->rate_limit;
3558 
3559 			if (ucmd->burst_info.max_burst_sz) {
3560 				if (attr->rate_limit &&
3561 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_burst_bound)) {
3562 					raw_qp_param.rl.max_burst_sz =
3563 						ucmd->burst_info.max_burst_sz;
3564 				} else {
3565 					err = -EINVAL;
3566 					goto out;
3567 				}
3568 			}
3569 
3570 			if (ucmd->burst_info.typical_pkt_sz) {
3571 				if (attr->rate_limit &&
3572 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_typical_size)) {
3573 					raw_qp_param.rl.typical_pkt_sz =
3574 						ucmd->burst_info.typical_pkt_sz;
3575 				} else {
3576 					err = -EINVAL;
3577 					goto out;
3578 				}
3579 			}
3580 
3581 			raw_qp_param.set_mask |= MLX5_RAW_QP_RATE_LIMIT;
3582 		}
3583 
3584 		err = modify_raw_packet_qp(dev, qp, &raw_qp_param, tx_affinity);
3585 	} else {
3586 		err = mlx5_core_qp_modify(dev->mdev, op, optpar, context,
3587 					  &base->mqp);
3588 	}
3589 
3590 	if (err)
3591 		goto out;
3592 
3593 	qp->state = new_state;
3594 
3595 	if (attr_mask & IB_QP_ACCESS_FLAGS)
3596 		qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
3597 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
3598 		qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
3599 	if (attr_mask & IB_QP_PORT)
3600 		qp->port = attr->port_num;
3601 	if (attr_mask & IB_QP_ALT_PATH)
3602 		qp->trans_qp.alt_port = attr->alt_port_num;
3603 
3604 	/*
3605 	 * If we moved a kernel QP to RESET, clean up all old CQ
3606 	 * entries and reinitialize the QP.
3607 	 */
3608 	if (new_state == IB_QPS_RESET &&
3609 	    !ibqp->uobject && ibqp->qp_type != IB_QPT_XRC_TGT) {
3610 		mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
3611 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
3612 		if (send_cq != recv_cq)
3613 			mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
3614 
3615 		qp->rq.head = 0;
3616 		qp->rq.tail = 0;
3617 		qp->sq.head = 0;
3618 		qp->sq.tail = 0;
3619 		qp->sq.cur_post = 0;
3620 		if (qp->sq.wqe_cnt)
3621 			qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
3622 		qp->db.db[MLX5_RCV_DBR] = 0;
3623 		qp->db.db[MLX5_SND_DBR] = 0;
3624 	}
3625 
3626 out:
3627 	kfree(context);
3628 	return err;
3629 }
3630 
3631 static inline bool is_valid_mask(int mask, int req, int opt)
3632 {
3633 	if ((mask & req) != req)
3634 		return false;
3635 
3636 	if (mask & ~(req | opt))
3637 		return false;
3638 
3639 	return true;
3640 }
3641 
3642 /* check valid transition for driver QP types
3643  * for now the only QP type that this function supports is DCI
3644  */
3645 static bool modify_dci_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state new_state,
3646 				enum ib_qp_attr_mask attr_mask)
3647 {
3648 	int req = IB_QP_STATE;
3649 	int opt = 0;
3650 
3651 	if (new_state == IB_QPS_RESET) {
3652 		return is_valid_mask(attr_mask, req, opt);
3653 	} else if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
3654 		req |= IB_QP_PKEY_INDEX | IB_QP_PORT;
3655 		return is_valid_mask(attr_mask, req, opt);
3656 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_INIT) {
3657 		opt = IB_QP_PKEY_INDEX | IB_QP_PORT;
3658 		return is_valid_mask(attr_mask, req, opt);
3659 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
3660 		req |= IB_QP_PATH_MTU;
3661 		opt = IB_QP_PKEY_INDEX | IB_QP_AV;
3662 		return is_valid_mask(attr_mask, req, opt);
3663 	} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
3664 		req |= IB_QP_TIMEOUT | IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
3665 		       IB_QP_MAX_QP_RD_ATOMIC | IB_QP_SQ_PSN;
3666 		opt = IB_QP_MIN_RNR_TIMER;
3667 		return is_valid_mask(attr_mask, req, opt);
3668 	} else if (cur_state == IB_QPS_RTS && new_state == IB_QPS_RTS) {
3669 		opt = IB_QP_MIN_RNR_TIMER;
3670 		return is_valid_mask(attr_mask, req, opt);
3671 	} else if (cur_state != IB_QPS_RESET && new_state == IB_QPS_ERR) {
3672 		return is_valid_mask(attr_mask, req, opt);
3673 	}
3674 	return false;
3675 }
3676 
3677 /* mlx5_ib_modify_dct: modify a DCT QP
3678  * valid transitions are:
3679  * RESET to INIT: must set access_flags, pkey_index and port
3680  * INIT  to RTR : must set min_rnr_timer, tclass, flow_label,
3681  *			   mtu, gid_index and hop_limit
3682  * Other transitions and attributes are illegal
3683  */
3684 static int mlx5_ib_modify_dct(struct ib_qp *ibqp, struct ib_qp_attr *attr,
3685 			      int attr_mask, struct ib_udata *udata)
3686 {
3687 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3688 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3689 	enum ib_qp_state cur_state, new_state;
3690 	int err = 0;
3691 	int required = IB_QP_STATE;
3692 	void *dctc;
3693 
3694 	if (!(attr_mask & IB_QP_STATE))
3695 		return -EINVAL;
3696 
3697 	cur_state = qp->state;
3698 	new_state = attr->qp_state;
3699 
3700 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
3701 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
3702 		required |= IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
3703 		if (!is_valid_mask(attr_mask, required, 0))
3704 			return -EINVAL;
3705 
3706 		if (attr->port_num == 0 ||
3707 		    attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports)) {
3708 			mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
3709 				    attr->port_num, dev->num_ports);
3710 			return -EINVAL;
3711 		}
3712 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
3713 			MLX5_SET(dctc, dctc, rre, 1);
3714 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
3715 			MLX5_SET(dctc, dctc, rwe, 1);
3716 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC) {
3717 			int atomic_mode;
3718 
3719 			atomic_mode = get_atomic_mode(dev, MLX5_IB_QPT_DCT);
3720 			if (atomic_mode < 0)
3721 				return -EOPNOTSUPP;
3722 
3723 			MLX5_SET(dctc, dctc, atomic_mode, atomic_mode);
3724 			MLX5_SET(dctc, dctc, rae, 1);
3725 		}
3726 		MLX5_SET(dctc, dctc, pkey_index, attr->pkey_index);
3727 		MLX5_SET(dctc, dctc, port, attr->port_num);
3728 		MLX5_SET(dctc, dctc, counter_set_id, dev->port[attr->port_num - 1].cnts.set_id);
3729 
3730 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
3731 		struct mlx5_ib_modify_qp_resp resp = {};
3732 		u32 min_resp_len = offsetof(typeof(resp), dctn) +
3733 				   sizeof(resp.dctn);
3734 
3735 		if (udata->outlen < min_resp_len)
3736 			return -EINVAL;
3737 		resp.response_length = min_resp_len;
3738 
3739 		required |= IB_QP_MIN_RNR_TIMER | IB_QP_AV | IB_QP_PATH_MTU;
3740 		if (!is_valid_mask(attr_mask, required, 0))
3741 			return -EINVAL;
3742 		MLX5_SET(dctc, dctc, min_rnr_nak, attr->min_rnr_timer);
3743 		MLX5_SET(dctc, dctc, tclass, attr->ah_attr.grh.traffic_class);
3744 		MLX5_SET(dctc, dctc, flow_label, attr->ah_attr.grh.flow_label);
3745 		MLX5_SET(dctc, dctc, mtu, attr->path_mtu);
3746 		MLX5_SET(dctc, dctc, my_addr_index, attr->ah_attr.grh.sgid_index);
3747 		MLX5_SET(dctc, dctc, hop_limit, attr->ah_attr.grh.hop_limit);
3748 
3749 		err = mlx5_core_create_dct(dev->mdev, &qp->dct.mdct, qp->dct.in,
3750 					   MLX5_ST_SZ_BYTES(create_dct_in));
3751 		if (err)
3752 			return err;
3753 		resp.dctn = qp->dct.mdct.mqp.qpn;
3754 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
3755 		if (err) {
3756 			mlx5_core_destroy_dct(dev->mdev, &qp->dct.mdct);
3757 			return err;
3758 		}
3759 	} else {
3760 		mlx5_ib_warn(dev, "Modify DCT: Invalid transition from %d to %d\n", cur_state, new_state);
3761 		return -EINVAL;
3762 	}
3763 	if (err)
3764 		qp->state = IB_QPS_ERR;
3765 	else
3766 		qp->state = new_state;
3767 	return err;
3768 }
3769 
3770 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
3771 		      int attr_mask, struct ib_udata *udata)
3772 {
3773 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3774 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3775 	struct mlx5_ib_modify_qp ucmd = {};
3776 	enum ib_qp_type qp_type;
3777 	enum ib_qp_state cur_state, new_state;
3778 	size_t required_cmd_sz;
3779 	int err = -EINVAL;
3780 	int port;
3781 
3782 	if (ibqp->rwq_ind_tbl)
3783 		return -ENOSYS;
3784 
3785 	if (udata && udata->inlen) {
3786 		required_cmd_sz = offsetof(typeof(ucmd), reserved) +
3787 			sizeof(ucmd.reserved);
3788 		if (udata->inlen < required_cmd_sz)
3789 			return -EINVAL;
3790 
3791 		if (udata->inlen > sizeof(ucmd) &&
3792 		    !ib_is_udata_cleared(udata, sizeof(ucmd),
3793 					 udata->inlen - sizeof(ucmd)))
3794 			return -EOPNOTSUPP;
3795 
3796 		if (ib_copy_from_udata(&ucmd, udata,
3797 				       min(udata->inlen, sizeof(ucmd))))
3798 			return -EFAULT;
3799 
3800 		if (ucmd.comp_mask ||
3801 		    memchr_inv(&ucmd.reserved, 0, sizeof(ucmd.reserved)) ||
3802 		    memchr_inv(&ucmd.burst_info.reserved, 0,
3803 			       sizeof(ucmd.burst_info.reserved)))
3804 			return -EOPNOTSUPP;
3805 	}
3806 
3807 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
3808 		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
3809 
3810 	if (ibqp->qp_type == IB_QPT_DRIVER)
3811 		qp_type = qp->qp_sub_type;
3812 	else
3813 		qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ?
3814 			IB_QPT_GSI : ibqp->qp_type;
3815 
3816 	if (qp_type == MLX5_IB_QPT_DCT)
3817 		return mlx5_ib_modify_dct(ibqp, attr, attr_mask, udata);
3818 
3819 	mutex_lock(&qp->mutex);
3820 
3821 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
3822 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
3823 
3824 	if (!(cur_state == new_state && cur_state == IB_QPS_RESET)) {
3825 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
3826 	}
3827 
3828 	if (qp->flags & MLX5_IB_QP_UNDERLAY) {
3829 		if (attr_mask & ~(IB_QP_STATE | IB_QP_CUR_STATE)) {
3830 			mlx5_ib_dbg(dev, "invalid attr_mask 0x%x when underlay QP is used\n",
3831 				    attr_mask);
3832 			goto out;
3833 		}
3834 	} else if (qp_type != MLX5_IB_QPT_REG_UMR &&
3835 		   qp_type != MLX5_IB_QPT_DCI &&
3836 		   !ib_modify_qp_is_ok(cur_state, new_state, qp_type,
3837 				       attr_mask)) {
3838 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
3839 			    cur_state, new_state, ibqp->qp_type, attr_mask);
3840 		goto out;
3841 	} else if (qp_type == MLX5_IB_QPT_DCI &&
3842 		   !modify_dci_qp_is_ok(cur_state, new_state, attr_mask)) {
3843 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
3844 			    cur_state, new_state, qp_type, attr_mask);
3845 		goto out;
3846 	}
3847 
3848 	if ((attr_mask & IB_QP_PORT) &&
3849 	    (attr->port_num == 0 ||
3850 	     attr->port_num > dev->num_ports)) {
3851 		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
3852 			    attr->port_num, dev->num_ports);
3853 		goto out;
3854 	}
3855 
3856 	if (attr_mask & IB_QP_PKEY_INDEX) {
3857 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
3858 		if (attr->pkey_index >=
3859 		    dev->mdev->port_caps[port - 1].pkey_table_len) {
3860 			mlx5_ib_dbg(dev, "invalid pkey index %d\n",
3861 				    attr->pkey_index);
3862 			goto out;
3863 		}
3864 	}
3865 
3866 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
3867 	    attr->max_rd_atomic >
3868 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
3869 		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
3870 			    attr->max_rd_atomic);
3871 		goto out;
3872 	}
3873 
3874 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
3875 	    attr->max_dest_rd_atomic >
3876 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
3877 		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
3878 			    attr->max_dest_rd_atomic);
3879 		goto out;
3880 	}
3881 
3882 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
3883 		err = 0;
3884 		goto out;
3885 	}
3886 
3887 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state,
3888 				  new_state, &ucmd, udata);
3889 
3890 out:
3891 	mutex_unlock(&qp->mutex);
3892 	return err;
3893 }
3894 
3895 static void _handle_post_send_edge(struct mlx5_ib_wq *sq, void **seg,
3896 				   u32 wqe_sz, void **cur_edge)
3897 {
3898 	u32 idx;
3899 
3900 	idx = (sq->cur_post + (wqe_sz >> 2)) & (sq->wqe_cnt - 1);
3901 	*cur_edge = get_sq_edge(sq, idx);
3902 
3903 	*seg = mlx5_frag_buf_get_wqe(&sq->fbc, idx);
3904 }
3905 
3906 /* handle_post_send_edge - Check if we get to SQ edge. If yes, update to the
3907  * next nearby edge and get new address translation for current WQE position.
3908  * @sq - SQ buffer.
3909  * @seg: Current WQE position (16B aligned).
3910  * @wqe_sz: Total current WQE size [16B].
3911  * @cur_edge: Updated current edge.
3912  */
3913 static inline void handle_post_send_edge(struct mlx5_ib_wq *sq, void **seg,
3914 					 u32 wqe_sz, void **cur_edge)
3915 {
3916 	if (likely(*seg != *cur_edge))
3917 		return;
3918 
3919 	_handle_post_send_edge(sq, seg, wqe_sz, cur_edge);
3920 }
3921 
3922 /* memcpy_send_wqe - copy data from src to WQE and update the relevant WQ's
3923  * pointers. At the end @seg is aligned to 16B regardless the copied size.
3924  * @sq - SQ buffer.
3925  * @cur_edge: Updated current edge.
3926  * @seg: Current WQE position (16B aligned).
3927  * @wqe_sz: Total current WQE size [16B].
3928  * @src: Pointer to copy from.
3929  * @n: Number of bytes to copy.
3930  */
3931 static inline void memcpy_send_wqe(struct mlx5_ib_wq *sq, void **cur_edge,
3932 				   void **seg, u32 *wqe_sz, const void *src,
3933 				   size_t n)
3934 {
3935 	while (likely(n)) {
3936 		size_t leftlen = *cur_edge - *seg;
3937 		size_t copysz = min_t(size_t, leftlen, n);
3938 		size_t stride;
3939 
3940 		memcpy(*seg, src, copysz);
3941 
3942 		n -= copysz;
3943 		src += copysz;
3944 		stride = !n ? ALIGN(copysz, 16) : copysz;
3945 		*seg += stride;
3946 		*wqe_sz += stride >> 4;
3947 		handle_post_send_edge(sq, seg, *wqe_sz, cur_edge);
3948 	}
3949 }
3950 
3951 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3952 {
3953 	struct mlx5_ib_cq *cq;
3954 	unsigned cur;
3955 
3956 	cur = wq->head - wq->tail;
3957 	if (likely(cur + nreq < wq->max_post))
3958 		return 0;
3959 
3960 	cq = to_mcq(ib_cq);
3961 	spin_lock(&cq->lock);
3962 	cur = wq->head - wq->tail;
3963 	spin_unlock(&cq->lock);
3964 
3965 	return cur + nreq >= wq->max_post;
3966 }
3967 
3968 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
3969 					  u64 remote_addr, u32 rkey)
3970 {
3971 	rseg->raddr    = cpu_to_be64(remote_addr);
3972 	rseg->rkey     = cpu_to_be32(rkey);
3973 	rseg->reserved = 0;
3974 }
3975 
3976 static void set_eth_seg(const struct ib_send_wr *wr, struct mlx5_ib_qp *qp,
3977 			void **seg, int *size, void **cur_edge)
3978 {
3979 	struct mlx5_wqe_eth_seg *eseg = *seg;
3980 
3981 	memset(eseg, 0, sizeof(struct mlx5_wqe_eth_seg));
3982 
3983 	if (wr->send_flags & IB_SEND_IP_CSUM)
3984 		eseg->cs_flags = MLX5_ETH_WQE_L3_CSUM |
3985 				 MLX5_ETH_WQE_L4_CSUM;
3986 
3987 	if (wr->opcode == IB_WR_LSO) {
3988 		struct ib_ud_wr *ud_wr = container_of(wr, struct ib_ud_wr, wr);
3989 		size_t left, copysz;
3990 		void *pdata = ud_wr->header;
3991 		size_t stride;
3992 
3993 		left = ud_wr->hlen;
3994 		eseg->mss = cpu_to_be16(ud_wr->mss);
3995 		eseg->inline_hdr.sz = cpu_to_be16(left);
3996 
3997 		/* memcpy_send_wqe should get a 16B align address. Hence, we
3998 		 * first copy up to the current edge and then, if needed,
3999 		 * fall-through to memcpy_send_wqe.
4000 		 */
4001 		copysz = min_t(u64, *cur_edge - (void *)eseg->inline_hdr.start,
4002 			       left);
4003 		memcpy(eseg->inline_hdr.start, pdata, copysz);
4004 		stride = ALIGN(sizeof(struct mlx5_wqe_eth_seg) -
4005 			       sizeof(eseg->inline_hdr.start) + copysz, 16);
4006 		*size += stride / 16;
4007 		*seg += stride;
4008 
4009 		if (copysz < left) {
4010 			handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4011 			left -= copysz;
4012 			pdata += copysz;
4013 			memcpy_send_wqe(&qp->sq, cur_edge, seg, size, pdata,
4014 					left);
4015 		}
4016 
4017 		return;
4018 	}
4019 
4020 	*seg += sizeof(struct mlx5_wqe_eth_seg);
4021 	*size += sizeof(struct mlx5_wqe_eth_seg) / 16;
4022 }
4023 
4024 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
4025 			     const struct ib_send_wr *wr)
4026 {
4027 	memcpy(&dseg->av, &to_mah(ud_wr(wr)->ah)->av, sizeof(struct mlx5_av));
4028 	dseg->av.dqp_dct = cpu_to_be32(ud_wr(wr)->remote_qpn | MLX5_EXTENDED_UD_AV);
4029 	dseg->av.key.qkey.qkey = cpu_to_be32(ud_wr(wr)->remote_qkey);
4030 }
4031 
4032 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
4033 {
4034 	dseg->byte_count = cpu_to_be32(sg->length);
4035 	dseg->lkey       = cpu_to_be32(sg->lkey);
4036 	dseg->addr       = cpu_to_be64(sg->addr);
4037 }
4038 
4039 static u64 get_xlt_octo(u64 bytes)
4040 {
4041 	return ALIGN(bytes, MLX5_IB_UMR_XLT_ALIGNMENT) /
4042 	       MLX5_IB_UMR_OCTOWORD;
4043 }
4044 
4045 static __be64 frwr_mkey_mask(void)
4046 {
4047 	u64 result;
4048 
4049 	result = MLX5_MKEY_MASK_LEN		|
4050 		MLX5_MKEY_MASK_PAGE_SIZE	|
4051 		MLX5_MKEY_MASK_START_ADDR	|
4052 		MLX5_MKEY_MASK_EN_RINVAL	|
4053 		MLX5_MKEY_MASK_KEY		|
4054 		MLX5_MKEY_MASK_LR		|
4055 		MLX5_MKEY_MASK_LW		|
4056 		MLX5_MKEY_MASK_RR		|
4057 		MLX5_MKEY_MASK_RW		|
4058 		MLX5_MKEY_MASK_A		|
4059 		MLX5_MKEY_MASK_SMALL_FENCE	|
4060 		MLX5_MKEY_MASK_FREE;
4061 
4062 	return cpu_to_be64(result);
4063 }
4064 
4065 static __be64 sig_mkey_mask(void)
4066 {
4067 	u64 result;
4068 
4069 	result = MLX5_MKEY_MASK_LEN		|
4070 		MLX5_MKEY_MASK_PAGE_SIZE	|
4071 		MLX5_MKEY_MASK_START_ADDR	|
4072 		MLX5_MKEY_MASK_EN_SIGERR	|
4073 		MLX5_MKEY_MASK_EN_RINVAL	|
4074 		MLX5_MKEY_MASK_KEY		|
4075 		MLX5_MKEY_MASK_LR		|
4076 		MLX5_MKEY_MASK_LW		|
4077 		MLX5_MKEY_MASK_RR		|
4078 		MLX5_MKEY_MASK_RW		|
4079 		MLX5_MKEY_MASK_SMALL_FENCE	|
4080 		MLX5_MKEY_MASK_FREE		|
4081 		MLX5_MKEY_MASK_BSF_EN;
4082 
4083 	return cpu_to_be64(result);
4084 }
4085 
4086 static void set_reg_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr,
4087 			    struct mlx5_ib_mr *mr, bool umr_inline)
4088 {
4089 	int size = mr->ndescs * mr->desc_size;
4090 
4091 	memset(umr, 0, sizeof(*umr));
4092 
4093 	umr->flags = MLX5_UMR_CHECK_NOT_FREE;
4094 	if (umr_inline)
4095 		umr->flags |= MLX5_UMR_INLINE;
4096 	umr->xlt_octowords = cpu_to_be16(get_xlt_octo(size));
4097 	umr->mkey_mask = frwr_mkey_mask();
4098 }
4099 
4100 static void set_linv_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr)
4101 {
4102 	memset(umr, 0, sizeof(*umr));
4103 	umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
4104 	umr->flags = MLX5_UMR_INLINE;
4105 }
4106 
4107 static __be64 get_umr_enable_mr_mask(void)
4108 {
4109 	u64 result;
4110 
4111 	result = MLX5_MKEY_MASK_KEY |
4112 		 MLX5_MKEY_MASK_FREE;
4113 
4114 	return cpu_to_be64(result);
4115 }
4116 
4117 static __be64 get_umr_disable_mr_mask(void)
4118 {
4119 	u64 result;
4120 
4121 	result = MLX5_MKEY_MASK_FREE;
4122 
4123 	return cpu_to_be64(result);
4124 }
4125 
4126 static __be64 get_umr_update_translation_mask(void)
4127 {
4128 	u64 result;
4129 
4130 	result = MLX5_MKEY_MASK_LEN |
4131 		 MLX5_MKEY_MASK_PAGE_SIZE |
4132 		 MLX5_MKEY_MASK_START_ADDR;
4133 
4134 	return cpu_to_be64(result);
4135 }
4136 
4137 static __be64 get_umr_update_access_mask(int atomic)
4138 {
4139 	u64 result;
4140 
4141 	result = MLX5_MKEY_MASK_LR |
4142 		 MLX5_MKEY_MASK_LW |
4143 		 MLX5_MKEY_MASK_RR |
4144 		 MLX5_MKEY_MASK_RW;
4145 
4146 	if (atomic)
4147 		result |= MLX5_MKEY_MASK_A;
4148 
4149 	return cpu_to_be64(result);
4150 }
4151 
4152 static __be64 get_umr_update_pd_mask(void)
4153 {
4154 	u64 result;
4155 
4156 	result = MLX5_MKEY_MASK_PD;
4157 
4158 	return cpu_to_be64(result);
4159 }
4160 
4161 static int umr_check_mkey_mask(struct mlx5_ib_dev *dev, u64 mask)
4162 {
4163 	if ((mask & MLX5_MKEY_MASK_PAGE_SIZE &&
4164 	     MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled)) ||
4165 	    (mask & MLX5_MKEY_MASK_A &&
4166 	     MLX5_CAP_GEN(dev->mdev, umr_modify_atomic_disabled)))
4167 		return -EPERM;
4168 	return 0;
4169 }
4170 
4171 static int set_reg_umr_segment(struct mlx5_ib_dev *dev,
4172 			       struct mlx5_wqe_umr_ctrl_seg *umr,
4173 			       const struct ib_send_wr *wr, int atomic)
4174 {
4175 	const struct mlx5_umr_wr *umrwr = umr_wr(wr);
4176 
4177 	memset(umr, 0, sizeof(*umr));
4178 
4179 	if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
4180 		umr->flags = MLX5_UMR_CHECK_FREE; /* fail if free */
4181 	else
4182 		umr->flags = MLX5_UMR_CHECK_NOT_FREE; /* fail if not free */
4183 
4184 	umr->xlt_octowords = cpu_to_be16(get_xlt_octo(umrwr->xlt_size));
4185 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_XLT) {
4186 		u64 offset = get_xlt_octo(umrwr->offset);
4187 
4188 		umr->xlt_offset = cpu_to_be16(offset & 0xffff);
4189 		umr->xlt_offset_47_16 = cpu_to_be32(offset >> 16);
4190 		umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
4191 	}
4192 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION)
4193 		umr->mkey_mask |= get_umr_update_translation_mask();
4194 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS) {
4195 		umr->mkey_mask |= get_umr_update_access_mask(atomic);
4196 		umr->mkey_mask |= get_umr_update_pd_mask();
4197 	}
4198 	if (wr->send_flags & MLX5_IB_SEND_UMR_ENABLE_MR)
4199 		umr->mkey_mask |= get_umr_enable_mr_mask();
4200 	if (wr->send_flags & MLX5_IB_SEND_UMR_DISABLE_MR)
4201 		umr->mkey_mask |= get_umr_disable_mr_mask();
4202 
4203 	if (!wr->num_sge)
4204 		umr->flags |= MLX5_UMR_INLINE;
4205 
4206 	return umr_check_mkey_mask(dev, be64_to_cpu(umr->mkey_mask));
4207 }
4208 
4209 static u8 get_umr_flags(int acc)
4210 {
4211 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
4212 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
4213 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
4214 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
4215 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
4216 }
4217 
4218 static void set_reg_mkey_seg(struct mlx5_mkey_seg *seg,
4219 			     struct mlx5_ib_mr *mr,
4220 			     u32 key, int access)
4221 {
4222 	int ndescs = ALIGN(mr->ndescs, 8) >> 1;
4223 
4224 	memset(seg, 0, sizeof(*seg));
4225 
4226 	if (mr->access_mode == MLX5_MKC_ACCESS_MODE_MTT)
4227 		seg->log2_page_size = ilog2(mr->ibmr.page_size);
4228 	else if (mr->access_mode == MLX5_MKC_ACCESS_MODE_KLMS)
4229 		/* KLMs take twice the size of MTTs */
4230 		ndescs *= 2;
4231 
4232 	seg->flags = get_umr_flags(access) | mr->access_mode;
4233 	seg->qpn_mkey7_0 = cpu_to_be32((key & 0xff) | 0xffffff00);
4234 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
4235 	seg->start_addr = cpu_to_be64(mr->ibmr.iova);
4236 	seg->len = cpu_to_be64(mr->ibmr.length);
4237 	seg->xlt_oct_size = cpu_to_be32(ndescs);
4238 }
4239 
4240 static void set_linv_mkey_seg(struct mlx5_mkey_seg *seg)
4241 {
4242 	memset(seg, 0, sizeof(*seg));
4243 	seg->status = MLX5_MKEY_STATUS_FREE;
4244 }
4245 
4246 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg,
4247 				 const struct ib_send_wr *wr)
4248 {
4249 	const struct mlx5_umr_wr *umrwr = umr_wr(wr);
4250 
4251 	memset(seg, 0, sizeof(*seg));
4252 	if (wr->send_flags & MLX5_IB_SEND_UMR_DISABLE_MR)
4253 		seg->status = MLX5_MKEY_STATUS_FREE;
4254 
4255 	seg->flags = convert_access(umrwr->access_flags);
4256 	if (umrwr->pd)
4257 		seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
4258 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION &&
4259 	    !umrwr->length)
4260 		seg->flags_pd |= cpu_to_be32(MLX5_MKEY_LEN64);
4261 
4262 	seg->start_addr = cpu_to_be64(umrwr->virt_addr);
4263 	seg->len = cpu_to_be64(umrwr->length);
4264 	seg->log2_page_size = umrwr->page_shift;
4265 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
4266 				       mlx5_mkey_variant(umrwr->mkey));
4267 }
4268 
4269 static void set_reg_data_seg(struct mlx5_wqe_data_seg *dseg,
4270 			     struct mlx5_ib_mr *mr,
4271 			     struct mlx5_ib_pd *pd)
4272 {
4273 	int bcount = mr->desc_size * mr->ndescs;
4274 
4275 	dseg->addr = cpu_to_be64(mr->desc_map);
4276 	dseg->byte_count = cpu_to_be32(ALIGN(bcount, 64));
4277 	dseg->lkey = cpu_to_be32(pd->ibpd.local_dma_lkey);
4278 }
4279 
4280 static __be32 send_ieth(const struct ib_send_wr *wr)
4281 {
4282 	switch (wr->opcode) {
4283 	case IB_WR_SEND_WITH_IMM:
4284 	case IB_WR_RDMA_WRITE_WITH_IMM:
4285 		return wr->ex.imm_data;
4286 
4287 	case IB_WR_SEND_WITH_INV:
4288 		return cpu_to_be32(wr->ex.invalidate_rkey);
4289 
4290 	default:
4291 		return 0;
4292 	}
4293 }
4294 
4295 static u8 calc_sig(void *wqe, int size)
4296 {
4297 	u8 *p = wqe;
4298 	u8 res = 0;
4299 	int i;
4300 
4301 	for (i = 0; i < size; i++)
4302 		res ^= p[i];
4303 
4304 	return ~res;
4305 }
4306 
4307 static u8 wq_sig(void *wqe)
4308 {
4309 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
4310 }
4311 
4312 static int set_data_inl_seg(struct mlx5_ib_qp *qp, const struct ib_send_wr *wr,
4313 			    void **wqe, int *wqe_sz, void **cur_edge)
4314 {
4315 	struct mlx5_wqe_inline_seg *seg;
4316 	size_t offset;
4317 	int inl = 0;
4318 	int i;
4319 
4320 	seg = *wqe;
4321 	*wqe += sizeof(*seg);
4322 	offset = sizeof(*seg);
4323 
4324 	for (i = 0; i < wr->num_sge; i++) {
4325 		size_t len  = wr->sg_list[i].length;
4326 		void *addr = (void *)(unsigned long)(wr->sg_list[i].addr);
4327 
4328 		inl += len;
4329 
4330 		if (unlikely(inl > qp->max_inline_data))
4331 			return -ENOMEM;
4332 
4333 		while (likely(len)) {
4334 			size_t leftlen;
4335 			size_t copysz;
4336 
4337 			handle_post_send_edge(&qp->sq, wqe,
4338 					      *wqe_sz + (offset >> 4),
4339 					      cur_edge);
4340 
4341 			leftlen = *cur_edge - *wqe;
4342 			copysz = min_t(size_t, leftlen, len);
4343 
4344 			memcpy(*wqe, addr, copysz);
4345 			len -= copysz;
4346 			addr += copysz;
4347 			*wqe += copysz;
4348 			offset += copysz;
4349 		}
4350 	}
4351 
4352 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
4353 
4354 	*wqe_sz +=  ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
4355 
4356 	return 0;
4357 }
4358 
4359 static u16 prot_field_size(enum ib_signature_type type)
4360 {
4361 	switch (type) {
4362 	case IB_SIG_TYPE_T10_DIF:
4363 		return MLX5_DIF_SIZE;
4364 	default:
4365 		return 0;
4366 	}
4367 }
4368 
4369 static u8 bs_selector(int block_size)
4370 {
4371 	switch (block_size) {
4372 	case 512:	    return 0x1;
4373 	case 520:	    return 0x2;
4374 	case 4096:	    return 0x3;
4375 	case 4160:	    return 0x4;
4376 	case 1073741824:    return 0x5;
4377 	default:	    return 0;
4378 	}
4379 }
4380 
4381 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
4382 			      struct mlx5_bsf_inl *inl)
4383 {
4384 	/* Valid inline section and allow BSF refresh */
4385 	inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
4386 				       MLX5_BSF_REFRESH_DIF);
4387 	inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
4388 	inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
4389 	/* repeating block */
4390 	inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
4391 	inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
4392 			MLX5_DIF_CRC : MLX5_DIF_IPCS;
4393 
4394 	if (domain->sig.dif.ref_remap)
4395 		inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
4396 
4397 	if (domain->sig.dif.app_escape) {
4398 		if (domain->sig.dif.ref_escape)
4399 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
4400 		else
4401 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
4402 	}
4403 
4404 	inl->dif_app_bitmask_check =
4405 		cpu_to_be16(domain->sig.dif.apptag_check_mask);
4406 }
4407 
4408 static int mlx5_set_bsf(struct ib_mr *sig_mr,
4409 			struct ib_sig_attrs *sig_attrs,
4410 			struct mlx5_bsf *bsf, u32 data_size)
4411 {
4412 	struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
4413 	struct mlx5_bsf_basic *basic = &bsf->basic;
4414 	struct ib_sig_domain *mem = &sig_attrs->mem;
4415 	struct ib_sig_domain *wire = &sig_attrs->wire;
4416 
4417 	memset(bsf, 0, sizeof(*bsf));
4418 
4419 	/* Basic + Extended + Inline */
4420 	basic->bsf_size_sbs = 1 << 7;
4421 	/* Input domain check byte mask */
4422 	basic->check_byte_mask = sig_attrs->check_mask;
4423 	basic->raw_data_size = cpu_to_be32(data_size);
4424 
4425 	/* Memory domain */
4426 	switch (sig_attrs->mem.sig_type) {
4427 	case IB_SIG_TYPE_NONE:
4428 		break;
4429 	case IB_SIG_TYPE_T10_DIF:
4430 		basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
4431 		basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
4432 		mlx5_fill_inl_bsf(mem, &bsf->m_inl);
4433 		break;
4434 	default:
4435 		return -EINVAL;
4436 	}
4437 
4438 	/* Wire domain */
4439 	switch (sig_attrs->wire.sig_type) {
4440 	case IB_SIG_TYPE_NONE:
4441 		break;
4442 	case IB_SIG_TYPE_T10_DIF:
4443 		if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
4444 		    mem->sig_type == wire->sig_type) {
4445 			/* Same block structure */
4446 			basic->bsf_size_sbs |= 1 << 4;
4447 			if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
4448 				basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
4449 			if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
4450 				basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
4451 			if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
4452 				basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
4453 		} else
4454 			basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
4455 
4456 		basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
4457 		mlx5_fill_inl_bsf(wire, &bsf->w_inl);
4458 		break;
4459 	default:
4460 		return -EINVAL;
4461 	}
4462 
4463 	return 0;
4464 }
4465 
4466 static int set_sig_data_segment(const struct ib_sig_handover_wr *wr,
4467 				struct mlx5_ib_qp *qp, void **seg,
4468 				int *size, void **cur_edge)
4469 {
4470 	struct ib_sig_attrs *sig_attrs = wr->sig_attrs;
4471 	struct ib_mr *sig_mr = wr->sig_mr;
4472 	struct mlx5_bsf *bsf;
4473 	u32 data_len = wr->wr.sg_list->length;
4474 	u32 data_key = wr->wr.sg_list->lkey;
4475 	u64 data_va = wr->wr.sg_list->addr;
4476 	int ret;
4477 	int wqe_size;
4478 
4479 	if (!wr->prot ||
4480 	    (data_key == wr->prot->lkey &&
4481 	     data_va == wr->prot->addr &&
4482 	     data_len == wr->prot->length)) {
4483 		/**
4484 		 * Source domain doesn't contain signature information
4485 		 * or data and protection are interleaved in memory.
4486 		 * So need construct:
4487 		 *                  ------------------
4488 		 *                 |     data_klm     |
4489 		 *                  ------------------
4490 		 *                 |       BSF        |
4491 		 *                  ------------------
4492 		 **/
4493 		struct mlx5_klm *data_klm = *seg;
4494 
4495 		data_klm->bcount = cpu_to_be32(data_len);
4496 		data_klm->key = cpu_to_be32(data_key);
4497 		data_klm->va = cpu_to_be64(data_va);
4498 		wqe_size = ALIGN(sizeof(*data_klm), 64);
4499 	} else {
4500 		/**
4501 		 * Source domain contains signature information
4502 		 * So need construct a strided block format:
4503 		 *               ---------------------------
4504 		 *              |     stride_block_ctrl     |
4505 		 *               ---------------------------
4506 		 *              |          data_klm         |
4507 		 *               ---------------------------
4508 		 *              |          prot_klm         |
4509 		 *               ---------------------------
4510 		 *              |             BSF           |
4511 		 *               ---------------------------
4512 		 **/
4513 		struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
4514 		struct mlx5_stride_block_entry *data_sentry;
4515 		struct mlx5_stride_block_entry *prot_sentry;
4516 		u32 prot_key = wr->prot->lkey;
4517 		u64 prot_va = wr->prot->addr;
4518 		u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
4519 		int prot_size;
4520 
4521 		sblock_ctrl = *seg;
4522 		data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
4523 		prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
4524 
4525 		prot_size = prot_field_size(sig_attrs->mem.sig_type);
4526 		if (!prot_size) {
4527 			pr_err("Bad block size given: %u\n", block_size);
4528 			return -EINVAL;
4529 		}
4530 		sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
4531 							    prot_size);
4532 		sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
4533 		sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
4534 		sblock_ctrl->num_entries = cpu_to_be16(2);
4535 
4536 		data_sentry->bcount = cpu_to_be16(block_size);
4537 		data_sentry->key = cpu_to_be32(data_key);
4538 		data_sentry->va = cpu_to_be64(data_va);
4539 		data_sentry->stride = cpu_to_be16(block_size);
4540 
4541 		prot_sentry->bcount = cpu_to_be16(prot_size);
4542 		prot_sentry->key = cpu_to_be32(prot_key);
4543 		prot_sentry->va = cpu_to_be64(prot_va);
4544 		prot_sentry->stride = cpu_to_be16(prot_size);
4545 
4546 		wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
4547 				 sizeof(*prot_sentry), 64);
4548 	}
4549 
4550 	*seg += wqe_size;
4551 	*size += wqe_size / 16;
4552 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4553 
4554 	bsf = *seg;
4555 	ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
4556 	if (ret)
4557 		return -EINVAL;
4558 
4559 	*seg += sizeof(*bsf);
4560 	*size += sizeof(*bsf) / 16;
4561 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4562 
4563 	return 0;
4564 }
4565 
4566 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
4567 				 const struct ib_sig_handover_wr *wr, u32 size,
4568 				 u32 length, u32 pdn)
4569 {
4570 	struct ib_mr *sig_mr = wr->sig_mr;
4571 	u32 sig_key = sig_mr->rkey;
4572 	u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
4573 
4574 	memset(seg, 0, sizeof(*seg));
4575 
4576 	seg->flags = get_umr_flags(wr->access_flags) |
4577 				   MLX5_MKC_ACCESS_MODE_KLMS;
4578 	seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
4579 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
4580 				    MLX5_MKEY_BSF_EN | pdn);
4581 	seg->len = cpu_to_be64(length);
4582 	seg->xlt_oct_size = cpu_to_be32(get_xlt_octo(size));
4583 	seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
4584 }
4585 
4586 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
4587 				u32 size)
4588 {
4589 	memset(umr, 0, sizeof(*umr));
4590 
4591 	umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
4592 	umr->xlt_octowords = cpu_to_be16(get_xlt_octo(size));
4593 	umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
4594 	umr->mkey_mask = sig_mkey_mask();
4595 }
4596 
4597 
4598 static int set_sig_umr_wr(const struct ib_send_wr *send_wr,
4599 			  struct mlx5_ib_qp *qp, void **seg, int *size,
4600 			  void **cur_edge)
4601 {
4602 	const struct ib_sig_handover_wr *wr = sig_handover_wr(send_wr);
4603 	struct mlx5_ib_mr *sig_mr = to_mmr(wr->sig_mr);
4604 	u32 pdn = get_pd(qp)->pdn;
4605 	u32 xlt_size;
4606 	int region_len, ret;
4607 
4608 	if (unlikely(wr->wr.num_sge != 1) ||
4609 	    unlikely(wr->access_flags & IB_ACCESS_REMOTE_ATOMIC) ||
4610 	    unlikely(!sig_mr->sig) || unlikely(!qp->signature_en) ||
4611 	    unlikely(!sig_mr->sig->sig_status_checked))
4612 		return -EINVAL;
4613 
4614 	/* length of the protected region, data + protection */
4615 	region_len = wr->wr.sg_list->length;
4616 	if (wr->prot &&
4617 	    (wr->prot->lkey != wr->wr.sg_list->lkey  ||
4618 	     wr->prot->addr != wr->wr.sg_list->addr  ||
4619 	     wr->prot->length != wr->wr.sg_list->length))
4620 		region_len += wr->prot->length;
4621 
4622 	/**
4623 	 * KLM octoword size - if protection was provided
4624 	 * then we use strided block format (3 octowords),
4625 	 * else we use single KLM (1 octoword)
4626 	 **/
4627 	xlt_size = wr->prot ? 0x30 : sizeof(struct mlx5_klm);
4628 
4629 	set_sig_umr_segment(*seg, xlt_size);
4630 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4631 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4632 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4633 
4634 	set_sig_mkey_segment(*seg, wr, xlt_size, region_len, pdn);
4635 	*seg += sizeof(struct mlx5_mkey_seg);
4636 	*size += sizeof(struct mlx5_mkey_seg) / 16;
4637 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4638 
4639 	ret = set_sig_data_segment(wr, qp, seg, size, cur_edge);
4640 	if (ret)
4641 		return ret;
4642 
4643 	sig_mr->sig->sig_status_checked = false;
4644 	return 0;
4645 }
4646 
4647 static int set_psv_wr(struct ib_sig_domain *domain,
4648 		      u32 psv_idx, void **seg, int *size)
4649 {
4650 	struct mlx5_seg_set_psv *psv_seg = *seg;
4651 
4652 	memset(psv_seg, 0, sizeof(*psv_seg));
4653 	psv_seg->psv_num = cpu_to_be32(psv_idx);
4654 	switch (domain->sig_type) {
4655 	case IB_SIG_TYPE_NONE:
4656 		break;
4657 	case IB_SIG_TYPE_T10_DIF:
4658 		psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
4659 						     domain->sig.dif.app_tag);
4660 		psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
4661 		break;
4662 	default:
4663 		pr_err("Bad signature type (%d) is given.\n",
4664 		       domain->sig_type);
4665 		return -EINVAL;
4666 	}
4667 
4668 	*seg += sizeof(*psv_seg);
4669 	*size += sizeof(*psv_seg) / 16;
4670 
4671 	return 0;
4672 }
4673 
4674 static int set_reg_wr(struct mlx5_ib_qp *qp,
4675 		      const struct ib_reg_wr *wr,
4676 		      void **seg, int *size, void **cur_edge)
4677 {
4678 	struct mlx5_ib_mr *mr = to_mmr(wr->mr);
4679 	struct mlx5_ib_pd *pd = to_mpd(qp->ibqp.pd);
4680 	size_t mr_list_size = mr->ndescs * mr->desc_size;
4681 	bool umr_inline = mr_list_size <= MLX5_IB_SQ_UMR_INLINE_THRESHOLD;
4682 
4683 	if (unlikely(wr->wr.send_flags & IB_SEND_INLINE)) {
4684 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
4685 			     "Invalid IB_SEND_INLINE send flag\n");
4686 		return -EINVAL;
4687 	}
4688 
4689 	set_reg_umr_seg(*seg, mr, umr_inline);
4690 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4691 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4692 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4693 
4694 	set_reg_mkey_seg(*seg, mr, wr->key, wr->access);
4695 	*seg += sizeof(struct mlx5_mkey_seg);
4696 	*size += sizeof(struct mlx5_mkey_seg) / 16;
4697 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4698 
4699 	if (umr_inline) {
4700 		memcpy_send_wqe(&qp->sq, cur_edge, seg, size, mr->descs,
4701 				mr_list_size);
4702 		*size = ALIGN(*size, MLX5_SEND_WQE_BB >> 4);
4703 	} else {
4704 		set_reg_data_seg(*seg, mr, pd);
4705 		*seg += sizeof(struct mlx5_wqe_data_seg);
4706 		*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
4707 	}
4708 	return 0;
4709 }
4710 
4711 static void set_linv_wr(struct mlx5_ib_qp *qp, void **seg, int *size,
4712 			void **cur_edge)
4713 {
4714 	set_linv_umr_seg(*seg);
4715 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4716 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4717 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4718 	set_linv_mkey_seg(*seg);
4719 	*seg += sizeof(struct mlx5_mkey_seg);
4720 	*size += sizeof(struct mlx5_mkey_seg) / 16;
4721 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4722 }
4723 
4724 static void dump_wqe(struct mlx5_ib_qp *qp, u32 idx, int size_16)
4725 {
4726 	__be32 *p = NULL;
4727 	u32 tidx = idx;
4728 	int i, j;
4729 
4730 	pr_debug("dump WQE index %u:\n", idx);
4731 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
4732 		if ((i & 0xf) == 0) {
4733 			tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
4734 			p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, tidx);
4735 			pr_debug("WQBB at %p:\n", (void *)p);
4736 			j = 0;
4737 		}
4738 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
4739 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
4740 			 be32_to_cpu(p[j + 3]));
4741 	}
4742 }
4743 
4744 static int __begin_wqe(struct mlx5_ib_qp *qp, void **seg,
4745 		       struct mlx5_wqe_ctrl_seg **ctrl,
4746 		       const struct ib_send_wr *wr, unsigned int *idx,
4747 		       int *size, void **cur_edge, int nreq,
4748 		       bool send_signaled, bool solicited)
4749 {
4750 	if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)))
4751 		return -ENOMEM;
4752 
4753 	*idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
4754 	*seg = mlx5_frag_buf_get_wqe(&qp->sq.fbc, *idx);
4755 	*ctrl = *seg;
4756 	*(uint32_t *)(*seg + 8) = 0;
4757 	(*ctrl)->imm = send_ieth(wr);
4758 	(*ctrl)->fm_ce_se = qp->sq_signal_bits |
4759 		(send_signaled ? MLX5_WQE_CTRL_CQ_UPDATE : 0) |
4760 		(solicited ? MLX5_WQE_CTRL_SOLICITED : 0);
4761 
4762 	*seg += sizeof(**ctrl);
4763 	*size = sizeof(**ctrl) / 16;
4764 	*cur_edge = qp->sq.cur_edge;
4765 
4766 	return 0;
4767 }
4768 
4769 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
4770 		     struct mlx5_wqe_ctrl_seg **ctrl,
4771 		     const struct ib_send_wr *wr, unsigned *idx,
4772 		     int *size, void **cur_edge, int nreq)
4773 {
4774 	return __begin_wqe(qp, seg, ctrl, wr, idx, size, cur_edge, nreq,
4775 			   wr->send_flags & IB_SEND_SIGNALED,
4776 			   wr->send_flags & IB_SEND_SOLICITED);
4777 }
4778 
4779 static void finish_wqe(struct mlx5_ib_qp *qp,
4780 		       struct mlx5_wqe_ctrl_seg *ctrl,
4781 		       void *seg, u8 size, void *cur_edge,
4782 		       unsigned int idx, u64 wr_id, int nreq, u8 fence,
4783 		       u32 mlx5_opcode)
4784 {
4785 	u8 opmod = 0;
4786 
4787 	ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
4788 					     mlx5_opcode | ((u32)opmod << 24));
4789 	ctrl->qpn_ds = cpu_to_be32(size | (qp->trans_qp.base.mqp.qpn << 8));
4790 	ctrl->fm_ce_se |= fence;
4791 	if (unlikely(qp->wq_sig))
4792 		ctrl->signature = wq_sig(ctrl);
4793 
4794 	qp->sq.wrid[idx] = wr_id;
4795 	qp->sq.w_list[idx].opcode = mlx5_opcode;
4796 	qp->sq.wqe_head[idx] = qp->sq.head + nreq;
4797 	qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
4798 	qp->sq.w_list[idx].next = qp->sq.cur_post;
4799 
4800 	/* We save the edge which was possibly updated during the WQE
4801 	 * construction, into SQ's cache.
4802 	 */
4803 	seg = PTR_ALIGN(seg, MLX5_SEND_WQE_BB);
4804 	qp->sq.cur_edge = (unlikely(seg == cur_edge)) ?
4805 			  get_sq_edge(&qp->sq, qp->sq.cur_post &
4806 				      (qp->sq.wqe_cnt - 1)) :
4807 			  cur_edge;
4808 }
4809 
4810 static int _mlx5_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
4811 			      const struct ib_send_wr **bad_wr, bool drain)
4812 {
4813 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
4814 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4815 	struct mlx5_core_dev *mdev = dev->mdev;
4816 	struct mlx5_ib_qp *qp;
4817 	struct mlx5_ib_mr *mr;
4818 	struct mlx5_wqe_xrc_seg *xrc;
4819 	struct mlx5_bf *bf;
4820 	void *cur_edge;
4821 	int uninitialized_var(size);
4822 	unsigned long flags;
4823 	unsigned idx;
4824 	int err = 0;
4825 	int num_sge;
4826 	void *seg;
4827 	int nreq;
4828 	int i;
4829 	u8 next_fence = 0;
4830 	u8 fence;
4831 
4832 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR &&
4833 		     !drain)) {
4834 		*bad_wr = wr;
4835 		return -EIO;
4836 	}
4837 
4838 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4839 		return mlx5_ib_gsi_post_send(ibqp, wr, bad_wr);
4840 
4841 	qp = to_mqp(ibqp);
4842 	bf = &qp->bf;
4843 
4844 	spin_lock_irqsave(&qp->sq.lock, flags);
4845 
4846 	for (nreq = 0; wr; nreq++, wr = wr->next) {
4847 		if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
4848 			mlx5_ib_warn(dev, "\n");
4849 			err = -EINVAL;
4850 			*bad_wr = wr;
4851 			goto out;
4852 		}
4853 
4854 		num_sge = wr->num_sge;
4855 		if (unlikely(num_sge > qp->sq.max_gs)) {
4856 			mlx5_ib_warn(dev, "\n");
4857 			err = -EINVAL;
4858 			*bad_wr = wr;
4859 			goto out;
4860 		}
4861 
4862 		err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, &cur_edge,
4863 				nreq);
4864 		if (err) {
4865 			mlx5_ib_warn(dev, "\n");
4866 			err = -ENOMEM;
4867 			*bad_wr = wr;
4868 			goto out;
4869 		}
4870 
4871 		if (wr->opcode == IB_WR_REG_MR) {
4872 			fence = dev->umr_fence;
4873 			next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
4874 		} else  {
4875 			if (wr->send_flags & IB_SEND_FENCE) {
4876 				if (qp->next_fence)
4877 					fence = MLX5_FENCE_MODE_SMALL_AND_FENCE;
4878 				else
4879 					fence = MLX5_FENCE_MODE_FENCE;
4880 			} else {
4881 				fence = qp->next_fence;
4882 			}
4883 		}
4884 
4885 		switch (ibqp->qp_type) {
4886 		case IB_QPT_XRC_INI:
4887 			xrc = seg;
4888 			seg += sizeof(*xrc);
4889 			size += sizeof(*xrc) / 16;
4890 			/* fall through */
4891 		case IB_QPT_RC:
4892 			switch (wr->opcode) {
4893 			case IB_WR_RDMA_READ:
4894 			case IB_WR_RDMA_WRITE:
4895 			case IB_WR_RDMA_WRITE_WITH_IMM:
4896 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
4897 					      rdma_wr(wr)->rkey);
4898 				seg += sizeof(struct mlx5_wqe_raddr_seg);
4899 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
4900 				break;
4901 
4902 			case IB_WR_ATOMIC_CMP_AND_SWP:
4903 			case IB_WR_ATOMIC_FETCH_AND_ADD:
4904 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
4905 				mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
4906 				err = -ENOSYS;
4907 				*bad_wr = wr;
4908 				goto out;
4909 
4910 			case IB_WR_LOCAL_INV:
4911 				qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
4912 				ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
4913 				set_linv_wr(qp, &seg, &size, &cur_edge);
4914 				num_sge = 0;
4915 				break;
4916 
4917 			case IB_WR_REG_MR:
4918 				qp->sq.wr_data[idx] = IB_WR_REG_MR;
4919 				ctrl->imm = cpu_to_be32(reg_wr(wr)->key);
4920 				err = set_reg_wr(qp, reg_wr(wr), &seg, &size,
4921 						 &cur_edge);
4922 				if (err) {
4923 					*bad_wr = wr;
4924 					goto out;
4925 				}
4926 				num_sge = 0;
4927 				break;
4928 
4929 			case IB_WR_REG_SIG_MR:
4930 				qp->sq.wr_data[idx] = IB_WR_REG_SIG_MR;
4931 				mr = to_mmr(sig_handover_wr(wr)->sig_mr);
4932 
4933 				ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
4934 				err = set_sig_umr_wr(wr, qp, &seg, &size,
4935 						     &cur_edge);
4936 				if (err) {
4937 					mlx5_ib_warn(dev, "\n");
4938 					*bad_wr = wr;
4939 					goto out;
4940 				}
4941 
4942 				finish_wqe(qp, ctrl, seg, size, cur_edge, idx,
4943 					   wr->wr_id, nreq, fence,
4944 					   MLX5_OPCODE_UMR);
4945 				/*
4946 				 * SET_PSV WQEs are not signaled and solicited
4947 				 * on error
4948 				 */
4949 				err = __begin_wqe(qp, &seg, &ctrl, wr, &idx,
4950 						  &size, &cur_edge, nreq, false,
4951 						  true);
4952 				if (err) {
4953 					mlx5_ib_warn(dev, "\n");
4954 					err = -ENOMEM;
4955 					*bad_wr = wr;
4956 					goto out;
4957 				}
4958 
4959 				err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->mem,
4960 						 mr->sig->psv_memory.psv_idx, &seg,
4961 						 &size);
4962 				if (err) {
4963 					mlx5_ib_warn(dev, "\n");
4964 					*bad_wr = wr;
4965 					goto out;
4966 				}
4967 
4968 				finish_wqe(qp, ctrl, seg, size, cur_edge, idx,
4969 					   wr->wr_id, nreq, fence,
4970 					   MLX5_OPCODE_SET_PSV);
4971 				err = __begin_wqe(qp, &seg, &ctrl, wr, &idx,
4972 						  &size, &cur_edge, nreq, false,
4973 						  true);
4974 				if (err) {
4975 					mlx5_ib_warn(dev, "\n");
4976 					err = -ENOMEM;
4977 					*bad_wr = wr;
4978 					goto out;
4979 				}
4980 
4981 				err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->wire,
4982 						 mr->sig->psv_wire.psv_idx, &seg,
4983 						 &size);
4984 				if (err) {
4985 					mlx5_ib_warn(dev, "\n");
4986 					*bad_wr = wr;
4987 					goto out;
4988 				}
4989 
4990 				finish_wqe(qp, ctrl, seg, size, cur_edge, idx,
4991 					   wr->wr_id, nreq, fence,
4992 					   MLX5_OPCODE_SET_PSV);
4993 				qp->next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
4994 				num_sge = 0;
4995 				goto skip_psv;
4996 
4997 			default:
4998 				break;
4999 			}
5000 			break;
5001 
5002 		case IB_QPT_UC:
5003 			switch (wr->opcode) {
5004 			case IB_WR_RDMA_WRITE:
5005 			case IB_WR_RDMA_WRITE_WITH_IMM:
5006 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
5007 					      rdma_wr(wr)->rkey);
5008 				seg  += sizeof(struct mlx5_wqe_raddr_seg);
5009 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
5010 				break;
5011 
5012 			default:
5013 				break;
5014 			}
5015 			break;
5016 
5017 		case IB_QPT_SMI:
5018 			if (unlikely(!mdev->port_caps[qp->port - 1].has_smi)) {
5019 				mlx5_ib_warn(dev, "Send SMP MADs is not allowed\n");
5020 				err = -EPERM;
5021 				*bad_wr = wr;
5022 				goto out;
5023 			}
5024 			/* fall through */
5025 		case MLX5_IB_QPT_HW_GSI:
5026 			set_datagram_seg(seg, wr);
5027 			seg += sizeof(struct mlx5_wqe_datagram_seg);
5028 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
5029 			handle_post_send_edge(&qp->sq, &seg, size, &cur_edge);
5030 
5031 			break;
5032 		case IB_QPT_UD:
5033 			set_datagram_seg(seg, wr);
5034 			seg += sizeof(struct mlx5_wqe_datagram_seg);
5035 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
5036 			handle_post_send_edge(&qp->sq, &seg, size, &cur_edge);
5037 
5038 			/* handle qp that supports ud offload */
5039 			if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO) {
5040 				struct mlx5_wqe_eth_pad *pad;
5041 
5042 				pad = seg;
5043 				memset(pad, 0, sizeof(struct mlx5_wqe_eth_pad));
5044 				seg += sizeof(struct mlx5_wqe_eth_pad);
5045 				size += sizeof(struct mlx5_wqe_eth_pad) / 16;
5046 				set_eth_seg(wr, qp, &seg, &size, &cur_edge);
5047 				handle_post_send_edge(&qp->sq, &seg, size,
5048 						      &cur_edge);
5049 			}
5050 			break;
5051 		case MLX5_IB_QPT_REG_UMR:
5052 			if (wr->opcode != MLX5_IB_WR_UMR) {
5053 				err = -EINVAL;
5054 				mlx5_ib_warn(dev, "bad opcode\n");
5055 				goto out;
5056 			}
5057 			qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
5058 			ctrl->imm = cpu_to_be32(umr_wr(wr)->mkey);
5059 			err = set_reg_umr_segment(dev, seg, wr, !!(MLX5_CAP_GEN(mdev, atomic)));
5060 			if (unlikely(err))
5061 				goto out;
5062 			seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
5063 			size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
5064 			handle_post_send_edge(&qp->sq, &seg, size, &cur_edge);
5065 			set_reg_mkey_segment(seg, wr);
5066 			seg += sizeof(struct mlx5_mkey_seg);
5067 			size += sizeof(struct mlx5_mkey_seg) / 16;
5068 			handle_post_send_edge(&qp->sq, &seg, size, &cur_edge);
5069 			break;
5070 
5071 		default:
5072 			break;
5073 		}
5074 
5075 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
5076 			err = set_data_inl_seg(qp, wr, &seg, &size, &cur_edge);
5077 			if (unlikely(err)) {
5078 				mlx5_ib_warn(dev, "\n");
5079 				*bad_wr = wr;
5080 				goto out;
5081 			}
5082 		} else {
5083 			for (i = 0; i < num_sge; i++) {
5084 				handle_post_send_edge(&qp->sq, &seg, size,
5085 						      &cur_edge);
5086 				if (likely(wr->sg_list[i].length)) {
5087 					set_data_ptr_seg
5088 					((struct mlx5_wqe_data_seg *)seg,
5089 					 wr->sg_list + i);
5090 					size += sizeof(struct mlx5_wqe_data_seg) / 16;
5091 					seg += sizeof(struct mlx5_wqe_data_seg);
5092 				}
5093 			}
5094 		}
5095 
5096 		qp->next_fence = next_fence;
5097 		finish_wqe(qp, ctrl, seg, size, cur_edge, idx, wr->wr_id, nreq,
5098 			   fence, mlx5_ib_opcode[wr->opcode]);
5099 skip_psv:
5100 		if (0)
5101 			dump_wqe(qp, idx, size);
5102 	}
5103 
5104 out:
5105 	if (likely(nreq)) {
5106 		qp->sq.head += nreq;
5107 
5108 		/* Make sure that descriptors are written before
5109 		 * updating doorbell record and ringing the doorbell
5110 		 */
5111 		wmb();
5112 
5113 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
5114 
5115 		/* Make sure doorbell record is visible to the HCA before
5116 		 * we hit doorbell */
5117 		wmb();
5118 
5119 		/* currently we support only regular doorbells */
5120 		mlx5_write64((__be32 *)ctrl, bf->bfreg->map + bf->offset, NULL);
5121 		/* Make sure doorbells don't leak out of SQ spinlock
5122 		 * and reach the HCA out of order.
5123 		 */
5124 		mmiowb();
5125 		bf->offset ^= bf->buf_size;
5126 	}
5127 
5128 	spin_unlock_irqrestore(&qp->sq.lock, flags);
5129 
5130 	return err;
5131 }
5132 
5133 int mlx5_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
5134 		      const struct ib_send_wr **bad_wr)
5135 {
5136 	return _mlx5_ib_post_send(ibqp, wr, bad_wr, false);
5137 }
5138 
5139 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
5140 {
5141 	sig->signature = calc_sig(sig, size);
5142 }
5143 
5144 static int _mlx5_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
5145 		      const struct ib_recv_wr **bad_wr, bool drain)
5146 {
5147 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
5148 	struct mlx5_wqe_data_seg *scat;
5149 	struct mlx5_rwqe_sig *sig;
5150 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
5151 	struct mlx5_core_dev *mdev = dev->mdev;
5152 	unsigned long flags;
5153 	int err = 0;
5154 	int nreq;
5155 	int ind;
5156 	int i;
5157 
5158 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR &&
5159 		     !drain)) {
5160 		*bad_wr = wr;
5161 		return -EIO;
5162 	}
5163 
5164 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
5165 		return mlx5_ib_gsi_post_recv(ibqp, wr, bad_wr);
5166 
5167 	spin_lock_irqsave(&qp->rq.lock, flags);
5168 
5169 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
5170 
5171 	for (nreq = 0; wr; nreq++, wr = wr->next) {
5172 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
5173 			err = -ENOMEM;
5174 			*bad_wr = wr;
5175 			goto out;
5176 		}
5177 
5178 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
5179 			err = -EINVAL;
5180 			*bad_wr = wr;
5181 			goto out;
5182 		}
5183 
5184 		scat = mlx5_frag_buf_get_wqe(&qp->rq.fbc, ind);
5185 		if (qp->wq_sig)
5186 			scat++;
5187 
5188 		for (i = 0; i < wr->num_sge; i++)
5189 			set_data_ptr_seg(scat + i, wr->sg_list + i);
5190 
5191 		if (i < qp->rq.max_gs) {
5192 			scat[i].byte_count = 0;
5193 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
5194 			scat[i].addr       = 0;
5195 		}
5196 
5197 		if (qp->wq_sig) {
5198 			sig = (struct mlx5_rwqe_sig *)scat;
5199 			set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
5200 		}
5201 
5202 		qp->rq.wrid[ind] = wr->wr_id;
5203 
5204 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
5205 	}
5206 
5207 out:
5208 	if (likely(nreq)) {
5209 		qp->rq.head += nreq;
5210 
5211 		/* Make sure that descriptors are written before
5212 		 * doorbell record.
5213 		 */
5214 		wmb();
5215 
5216 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
5217 	}
5218 
5219 	spin_unlock_irqrestore(&qp->rq.lock, flags);
5220 
5221 	return err;
5222 }
5223 
5224 int mlx5_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
5225 		      const struct ib_recv_wr **bad_wr)
5226 {
5227 	return _mlx5_ib_post_recv(ibqp, wr, bad_wr, false);
5228 }
5229 
5230 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
5231 {
5232 	switch (mlx5_state) {
5233 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
5234 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
5235 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
5236 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
5237 	case MLX5_QP_STATE_SQ_DRAINING:
5238 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
5239 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
5240 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
5241 	default:		     return -1;
5242 	}
5243 }
5244 
5245 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
5246 {
5247 	switch (mlx5_mig_state) {
5248 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
5249 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
5250 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
5251 	default: return -1;
5252 	}
5253 }
5254 
5255 static int to_ib_qp_access_flags(int mlx5_flags)
5256 {
5257 	int ib_flags = 0;
5258 
5259 	if (mlx5_flags & MLX5_QP_BIT_RRE)
5260 		ib_flags |= IB_ACCESS_REMOTE_READ;
5261 	if (mlx5_flags & MLX5_QP_BIT_RWE)
5262 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
5263 	if (mlx5_flags & MLX5_QP_BIT_RAE)
5264 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
5265 
5266 	return ib_flags;
5267 }
5268 
5269 static void to_rdma_ah_attr(struct mlx5_ib_dev *ibdev,
5270 			    struct rdma_ah_attr *ah_attr,
5271 			    struct mlx5_qp_path *path)
5272 {
5273 
5274 	memset(ah_attr, 0, sizeof(*ah_attr));
5275 
5276 	if (!path->port || path->port > ibdev->num_ports)
5277 		return;
5278 
5279 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, path->port);
5280 
5281 	rdma_ah_set_port_num(ah_attr, path->port);
5282 	rdma_ah_set_sl(ah_attr, path->dci_cfi_prio_sl & 0xf);
5283 
5284 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
5285 	rdma_ah_set_path_bits(ah_attr, path->grh_mlid & 0x7f);
5286 	rdma_ah_set_static_rate(ah_attr,
5287 				path->static_rate ? path->static_rate - 5 : 0);
5288 	if (path->grh_mlid & (1 << 7)) {
5289 		u32 tc_fl = be32_to_cpu(path->tclass_flowlabel);
5290 
5291 		rdma_ah_set_grh(ah_attr, NULL,
5292 				tc_fl & 0xfffff,
5293 				path->mgid_index,
5294 				path->hop_limit,
5295 				(tc_fl >> 20) & 0xff);
5296 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
5297 	}
5298 }
5299 
5300 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
5301 					struct mlx5_ib_sq *sq,
5302 					u8 *sq_state)
5303 {
5304 	int err;
5305 
5306 	err = mlx5_core_query_sq_state(dev->mdev, sq->base.mqp.qpn, sq_state);
5307 	if (err)
5308 		goto out;
5309 	sq->state = *sq_state;
5310 
5311 out:
5312 	return err;
5313 }
5314 
5315 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
5316 					struct mlx5_ib_rq *rq,
5317 					u8 *rq_state)
5318 {
5319 	void *out;
5320 	void *rqc;
5321 	int inlen;
5322 	int err;
5323 
5324 	inlen = MLX5_ST_SZ_BYTES(query_rq_out);
5325 	out = kvzalloc(inlen, GFP_KERNEL);
5326 	if (!out)
5327 		return -ENOMEM;
5328 
5329 	err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
5330 	if (err)
5331 		goto out;
5332 
5333 	rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
5334 	*rq_state = MLX5_GET(rqc, rqc, state);
5335 	rq->state = *rq_state;
5336 
5337 out:
5338 	kvfree(out);
5339 	return err;
5340 }
5341 
5342 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
5343 				  struct mlx5_ib_qp *qp, u8 *qp_state)
5344 {
5345 	static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
5346 		[MLX5_RQC_STATE_RST] = {
5347 			[MLX5_SQC_STATE_RST]	= IB_QPS_RESET,
5348 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
5349 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE_BAD,
5350 			[MLX5_SQ_STATE_NA]	= IB_QPS_RESET,
5351 		},
5352 		[MLX5_RQC_STATE_RDY] = {
5353 			[MLX5_SQC_STATE_RST]	= MLX5_QP_STATE_BAD,
5354 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
5355 			[MLX5_SQC_STATE_ERR]	= IB_QPS_SQE,
5356 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE,
5357 		},
5358 		[MLX5_RQC_STATE_ERR] = {
5359 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE_BAD,
5360 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
5361 			[MLX5_SQC_STATE_ERR]	= IB_QPS_ERR,
5362 			[MLX5_SQ_STATE_NA]	= IB_QPS_ERR,
5363 		},
5364 		[MLX5_RQ_STATE_NA] = {
5365 			[MLX5_SQC_STATE_RST]    = IB_QPS_RESET,
5366 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
5367 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE,
5368 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE_BAD,
5369 		},
5370 	};
5371 
5372 	*qp_state = sqrq_trans[rq_state][sq_state];
5373 
5374 	if (*qp_state == MLX5_QP_STATE_BAD) {
5375 		WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
5376 		     qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
5377 		     qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
5378 		return -EINVAL;
5379 	}
5380 
5381 	if (*qp_state == MLX5_QP_STATE)
5382 		*qp_state = qp->state;
5383 
5384 	return 0;
5385 }
5386 
5387 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
5388 				     struct mlx5_ib_qp *qp,
5389 				     u8 *raw_packet_qp_state)
5390 {
5391 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
5392 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
5393 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
5394 	int err;
5395 	u8 sq_state = MLX5_SQ_STATE_NA;
5396 	u8 rq_state = MLX5_RQ_STATE_NA;
5397 
5398 	if (qp->sq.wqe_cnt) {
5399 		err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
5400 		if (err)
5401 			return err;
5402 	}
5403 
5404 	if (qp->rq.wqe_cnt) {
5405 		err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
5406 		if (err)
5407 			return err;
5408 	}
5409 
5410 	return sqrq_state_to_qp_state(sq_state, rq_state, qp,
5411 				      raw_packet_qp_state);
5412 }
5413 
5414 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
5415 			 struct ib_qp_attr *qp_attr)
5416 {
5417 	int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
5418 	struct mlx5_qp_context *context;
5419 	int mlx5_state;
5420 	u32 *outb;
5421 	int err = 0;
5422 
5423 	outb = kzalloc(outlen, GFP_KERNEL);
5424 	if (!outb)
5425 		return -ENOMEM;
5426 
5427 	err = mlx5_core_qp_query(dev->mdev, &qp->trans_qp.base.mqp, outb,
5428 				 outlen);
5429 	if (err)
5430 		goto out;
5431 
5432 	/* FIXME: use MLX5_GET rather than mlx5_qp_context manual struct */
5433 	context = (struct mlx5_qp_context *)MLX5_ADDR_OF(query_qp_out, outb, qpc);
5434 
5435 	mlx5_state = be32_to_cpu(context->flags) >> 28;
5436 
5437 	qp->state		     = to_ib_qp_state(mlx5_state);
5438 	qp_attr->path_mtu	     = context->mtu_msgmax >> 5;
5439 	qp_attr->path_mig_state	     =
5440 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
5441 	qp_attr->qkey		     = be32_to_cpu(context->qkey);
5442 	qp_attr->rq_psn		     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
5443 	qp_attr->sq_psn		     = be32_to_cpu(context->next_send_psn) & 0xffffff;
5444 	qp_attr->dest_qp_num	     = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
5445 	qp_attr->qp_access_flags     =
5446 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
5447 
5448 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
5449 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
5450 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
5451 		qp_attr->alt_pkey_index =
5452 			be16_to_cpu(context->alt_path.pkey_index);
5453 		qp_attr->alt_port_num	=
5454 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
5455 	}
5456 
5457 	qp_attr->pkey_index = be16_to_cpu(context->pri_path.pkey_index);
5458 	qp_attr->port_num = context->pri_path.port;
5459 
5460 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
5461 	qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
5462 
5463 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
5464 
5465 	qp_attr->max_dest_rd_atomic =
5466 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
5467 	qp_attr->min_rnr_timer	    =
5468 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
5469 	qp_attr->timeout	    = context->pri_path.ackto_lt >> 3;
5470 	qp_attr->retry_cnt	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
5471 	qp_attr->rnr_retry	    = (be32_to_cpu(context->params1) >> 13) & 0x7;
5472 	qp_attr->alt_timeout	    = context->alt_path.ackto_lt >> 3;
5473 
5474 out:
5475 	kfree(outb);
5476 	return err;
5477 }
5478 
5479 static int mlx5_ib_dct_query_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *mqp,
5480 				struct ib_qp_attr *qp_attr, int qp_attr_mask,
5481 				struct ib_qp_init_attr *qp_init_attr)
5482 {
5483 	struct mlx5_core_dct	*dct = &mqp->dct.mdct;
5484 	u32 *out;
5485 	u32 access_flags = 0;
5486 	int outlen = MLX5_ST_SZ_BYTES(query_dct_out);
5487 	void *dctc;
5488 	int err;
5489 	int supported_mask = IB_QP_STATE |
5490 			     IB_QP_ACCESS_FLAGS |
5491 			     IB_QP_PORT |
5492 			     IB_QP_MIN_RNR_TIMER |
5493 			     IB_QP_AV |
5494 			     IB_QP_PATH_MTU |
5495 			     IB_QP_PKEY_INDEX;
5496 
5497 	if (qp_attr_mask & ~supported_mask)
5498 		return -EINVAL;
5499 	if (mqp->state != IB_QPS_RTR)
5500 		return -EINVAL;
5501 
5502 	out = kzalloc(outlen, GFP_KERNEL);
5503 	if (!out)
5504 		return -ENOMEM;
5505 
5506 	err = mlx5_core_dct_query(dev->mdev, dct, out, outlen);
5507 	if (err)
5508 		goto out;
5509 
5510 	dctc = MLX5_ADDR_OF(query_dct_out, out, dct_context_entry);
5511 
5512 	if (qp_attr_mask & IB_QP_STATE)
5513 		qp_attr->qp_state = IB_QPS_RTR;
5514 
5515 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
5516 		if (MLX5_GET(dctc, dctc, rre))
5517 			access_flags |= IB_ACCESS_REMOTE_READ;
5518 		if (MLX5_GET(dctc, dctc, rwe))
5519 			access_flags |= IB_ACCESS_REMOTE_WRITE;
5520 		if (MLX5_GET(dctc, dctc, rae))
5521 			access_flags |= IB_ACCESS_REMOTE_ATOMIC;
5522 		qp_attr->qp_access_flags = access_flags;
5523 	}
5524 
5525 	if (qp_attr_mask & IB_QP_PORT)
5526 		qp_attr->port_num = MLX5_GET(dctc, dctc, port);
5527 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER)
5528 		qp_attr->min_rnr_timer = MLX5_GET(dctc, dctc, min_rnr_nak);
5529 	if (qp_attr_mask & IB_QP_AV) {
5530 		qp_attr->ah_attr.grh.traffic_class = MLX5_GET(dctc, dctc, tclass);
5531 		qp_attr->ah_attr.grh.flow_label = MLX5_GET(dctc, dctc, flow_label);
5532 		qp_attr->ah_attr.grh.sgid_index = MLX5_GET(dctc, dctc, my_addr_index);
5533 		qp_attr->ah_attr.grh.hop_limit = MLX5_GET(dctc, dctc, hop_limit);
5534 	}
5535 	if (qp_attr_mask & IB_QP_PATH_MTU)
5536 		qp_attr->path_mtu = MLX5_GET(dctc, dctc, mtu);
5537 	if (qp_attr_mask & IB_QP_PKEY_INDEX)
5538 		qp_attr->pkey_index = MLX5_GET(dctc, dctc, pkey_index);
5539 out:
5540 	kfree(out);
5541 	return err;
5542 }
5543 
5544 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
5545 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
5546 {
5547 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
5548 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
5549 	int err = 0;
5550 	u8 raw_packet_qp_state;
5551 
5552 	if (ibqp->rwq_ind_tbl)
5553 		return -ENOSYS;
5554 
5555 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
5556 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
5557 					    qp_init_attr);
5558 
5559 	/* Not all of output fields are applicable, make sure to zero them */
5560 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
5561 	memset(qp_attr, 0, sizeof(*qp_attr));
5562 
5563 	if (unlikely(qp->qp_sub_type == MLX5_IB_QPT_DCT))
5564 		return mlx5_ib_dct_query_qp(dev, qp, qp_attr,
5565 					    qp_attr_mask, qp_init_attr);
5566 
5567 	mutex_lock(&qp->mutex);
5568 
5569 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
5570 	    qp->flags & MLX5_IB_QP_UNDERLAY) {
5571 		err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
5572 		if (err)
5573 			goto out;
5574 		qp->state = raw_packet_qp_state;
5575 		qp_attr->port_num = 1;
5576 	} else {
5577 		err = query_qp_attr(dev, qp, qp_attr);
5578 		if (err)
5579 			goto out;
5580 	}
5581 
5582 	qp_attr->qp_state	     = qp->state;
5583 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
5584 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
5585 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
5586 
5587 	if (!ibqp->uobject) {
5588 		qp_attr->cap.max_send_wr  = qp->sq.max_post;
5589 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
5590 		qp_init_attr->qp_context = ibqp->qp_context;
5591 	} else {
5592 		qp_attr->cap.max_send_wr  = 0;
5593 		qp_attr->cap.max_send_sge = 0;
5594 	}
5595 
5596 	qp_init_attr->qp_type = ibqp->qp_type;
5597 	qp_init_attr->recv_cq = ibqp->recv_cq;
5598 	qp_init_attr->send_cq = ibqp->send_cq;
5599 	qp_init_attr->srq = ibqp->srq;
5600 	qp_attr->cap.max_inline_data = qp->max_inline_data;
5601 
5602 	qp_init_attr->cap	     = qp_attr->cap;
5603 
5604 	qp_init_attr->create_flags = 0;
5605 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
5606 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
5607 
5608 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
5609 		qp_init_attr->create_flags |= IB_QP_CREATE_CROSS_CHANNEL;
5610 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
5611 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_SEND;
5612 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
5613 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_RECV;
5614 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
5615 		qp_init_attr->create_flags |= mlx5_ib_create_qp_sqpn_qp1();
5616 
5617 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
5618 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
5619 
5620 out:
5621 	mutex_unlock(&qp->mutex);
5622 	return err;
5623 }
5624 
5625 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
5626 					  struct ib_ucontext *context,
5627 					  struct ib_udata *udata)
5628 {
5629 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5630 	struct mlx5_ib_xrcd *xrcd;
5631 	int err;
5632 
5633 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
5634 		return ERR_PTR(-ENOSYS);
5635 
5636 	xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
5637 	if (!xrcd)
5638 		return ERR_PTR(-ENOMEM);
5639 
5640 	err = mlx5_cmd_xrcd_alloc(dev->mdev, &xrcd->xrcdn, 0);
5641 	if (err) {
5642 		kfree(xrcd);
5643 		return ERR_PTR(-ENOMEM);
5644 	}
5645 
5646 	return &xrcd->ibxrcd;
5647 }
5648 
5649 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
5650 {
5651 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
5652 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
5653 	int err;
5654 
5655 	err = mlx5_cmd_xrcd_dealloc(dev->mdev, xrcdn, 0);
5656 	if (err)
5657 		mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
5658 
5659 	kfree(xrcd);
5660 	return 0;
5661 }
5662 
5663 static void mlx5_ib_wq_event(struct mlx5_core_qp *core_qp, int type)
5664 {
5665 	struct mlx5_ib_rwq *rwq = to_mibrwq(core_qp);
5666 	struct mlx5_ib_dev *dev = to_mdev(rwq->ibwq.device);
5667 	struct ib_event event;
5668 
5669 	if (rwq->ibwq.event_handler) {
5670 		event.device     = rwq->ibwq.device;
5671 		event.element.wq = &rwq->ibwq;
5672 		switch (type) {
5673 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
5674 			event.event = IB_EVENT_WQ_FATAL;
5675 			break;
5676 		default:
5677 			mlx5_ib_warn(dev, "Unexpected event type %d on WQ %06x\n", type, core_qp->qpn);
5678 			return;
5679 		}
5680 
5681 		rwq->ibwq.event_handler(&event, rwq->ibwq.wq_context);
5682 	}
5683 }
5684 
5685 static int set_delay_drop(struct mlx5_ib_dev *dev)
5686 {
5687 	int err = 0;
5688 
5689 	mutex_lock(&dev->delay_drop.lock);
5690 	if (dev->delay_drop.activate)
5691 		goto out;
5692 
5693 	err = mlx5_core_set_delay_drop(dev->mdev, dev->delay_drop.timeout);
5694 	if (err)
5695 		goto out;
5696 
5697 	dev->delay_drop.activate = true;
5698 out:
5699 	mutex_unlock(&dev->delay_drop.lock);
5700 
5701 	if (!err)
5702 		atomic_inc(&dev->delay_drop.rqs_cnt);
5703 	return err;
5704 }
5705 
5706 static int  create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
5707 		      struct ib_wq_init_attr *init_attr)
5708 {
5709 	struct mlx5_ib_dev *dev;
5710 	int has_net_offloads;
5711 	__be64 *rq_pas0;
5712 	void *in;
5713 	void *rqc;
5714 	void *wq;
5715 	int inlen;
5716 	int err;
5717 
5718 	dev = to_mdev(pd->device);
5719 
5720 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
5721 	in = kvzalloc(inlen, GFP_KERNEL);
5722 	if (!in)
5723 		return -ENOMEM;
5724 
5725 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
5726 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
5727 	MLX5_SET(rqc,  rqc, mem_rq_type,
5728 		 MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
5729 	MLX5_SET(rqc, rqc, user_index, rwq->user_index);
5730 	MLX5_SET(rqc,  rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
5731 	MLX5_SET(rqc,  rqc, state, MLX5_RQC_STATE_RST);
5732 	MLX5_SET(rqc,  rqc, flush_in_error_en, 1);
5733 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
5734 	MLX5_SET(wq, wq, wq_type,
5735 		 rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ ?
5736 		 MLX5_WQ_TYPE_CYCLIC_STRIDING_RQ : MLX5_WQ_TYPE_CYCLIC);
5737 	if (init_attr->create_flags & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
5738 		if (!MLX5_CAP_GEN(dev->mdev, end_pad)) {
5739 			mlx5_ib_dbg(dev, "Scatter end padding is not supported\n");
5740 			err = -EOPNOTSUPP;
5741 			goto out;
5742 		} else {
5743 			MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
5744 		}
5745 	}
5746 	MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
5747 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ) {
5748 		MLX5_SET(wq, wq, two_byte_shift_en, rwq->two_byte_shift_en);
5749 		MLX5_SET(wq, wq, log_wqe_stride_size,
5750 			 rwq->single_stride_log_num_of_bytes -
5751 			 MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES);
5752 		MLX5_SET(wq, wq, log_wqe_num_of_strides, rwq->log_num_strides -
5753 			 MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES);
5754 	}
5755 	MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
5756 	MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
5757 	MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
5758 	MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
5759 	MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
5760 	MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
5761 	has_net_offloads = MLX5_CAP_GEN(dev->mdev, eth_net_offloads);
5762 	if (init_attr->create_flags & IB_WQ_FLAGS_CVLAN_STRIPPING) {
5763 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
5764 			mlx5_ib_dbg(dev, "VLAN offloads are not supported\n");
5765 			err = -EOPNOTSUPP;
5766 			goto out;
5767 		}
5768 	} else {
5769 		MLX5_SET(rqc, rqc, vsd, 1);
5770 	}
5771 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS) {
5772 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, scatter_fcs))) {
5773 			mlx5_ib_dbg(dev, "Scatter FCS is not supported\n");
5774 			err = -EOPNOTSUPP;
5775 			goto out;
5776 		}
5777 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
5778 	}
5779 	if (init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
5780 		if (!(dev->ib_dev.attrs.raw_packet_caps &
5781 		      IB_RAW_PACKET_CAP_DELAY_DROP)) {
5782 			mlx5_ib_dbg(dev, "Delay drop is not supported\n");
5783 			err = -EOPNOTSUPP;
5784 			goto out;
5785 		}
5786 		MLX5_SET(rqc, rqc, delay_drop_en, 1);
5787 	}
5788 	rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
5789 	mlx5_ib_populate_pas(dev, rwq->umem, rwq->page_shift, rq_pas0, 0);
5790 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rwq->core_qp);
5791 	if (!err && init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
5792 		err = set_delay_drop(dev);
5793 		if (err) {
5794 			mlx5_ib_warn(dev, "Failed to enable delay drop err=%d\n",
5795 				     err);
5796 			mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
5797 		} else {
5798 			rwq->create_flags |= MLX5_IB_WQ_FLAGS_DELAY_DROP;
5799 		}
5800 	}
5801 out:
5802 	kvfree(in);
5803 	return err;
5804 }
5805 
5806 static int set_user_rq_size(struct mlx5_ib_dev *dev,
5807 			    struct ib_wq_init_attr *wq_init_attr,
5808 			    struct mlx5_ib_create_wq *ucmd,
5809 			    struct mlx5_ib_rwq *rwq)
5810 {
5811 	/* Sanity check RQ size before proceeding */
5812 	if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
5813 		return -EINVAL;
5814 
5815 	if (!ucmd->rq_wqe_count)
5816 		return -EINVAL;
5817 
5818 	rwq->wqe_count = ucmd->rq_wqe_count;
5819 	rwq->wqe_shift = ucmd->rq_wqe_shift;
5820 	if (check_shl_overflow(rwq->wqe_count, rwq->wqe_shift, &rwq->buf_size))
5821 		return -EINVAL;
5822 
5823 	rwq->log_rq_stride = rwq->wqe_shift;
5824 	rwq->log_rq_size = ilog2(rwq->wqe_count);
5825 	return 0;
5826 }
5827 
5828 static int prepare_user_rq(struct ib_pd *pd,
5829 			   struct ib_wq_init_attr *init_attr,
5830 			   struct ib_udata *udata,
5831 			   struct mlx5_ib_rwq *rwq)
5832 {
5833 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
5834 	struct mlx5_ib_create_wq ucmd = {};
5835 	int err;
5836 	size_t required_cmd_sz;
5837 
5838 	required_cmd_sz = offsetof(typeof(ucmd), single_stride_log_num_of_bytes)
5839 		+ sizeof(ucmd.single_stride_log_num_of_bytes);
5840 	if (udata->inlen < required_cmd_sz) {
5841 		mlx5_ib_dbg(dev, "invalid inlen\n");
5842 		return -EINVAL;
5843 	}
5844 
5845 	if (udata->inlen > sizeof(ucmd) &&
5846 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
5847 				 udata->inlen - sizeof(ucmd))) {
5848 		mlx5_ib_dbg(dev, "inlen is not supported\n");
5849 		return -EOPNOTSUPP;
5850 	}
5851 
5852 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
5853 		mlx5_ib_dbg(dev, "copy failed\n");
5854 		return -EFAULT;
5855 	}
5856 
5857 	if (ucmd.comp_mask & (~MLX5_IB_CREATE_WQ_STRIDING_RQ)) {
5858 		mlx5_ib_dbg(dev, "invalid comp mask\n");
5859 		return -EOPNOTSUPP;
5860 	} else if (ucmd.comp_mask & MLX5_IB_CREATE_WQ_STRIDING_RQ) {
5861 		if (!MLX5_CAP_GEN(dev->mdev, striding_rq)) {
5862 			mlx5_ib_dbg(dev, "Striding RQ is not supported\n");
5863 			return -EOPNOTSUPP;
5864 		}
5865 		if ((ucmd.single_stride_log_num_of_bytes <
5866 		    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES) ||
5867 		    (ucmd.single_stride_log_num_of_bytes >
5868 		     MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES)) {
5869 			mlx5_ib_dbg(dev, "Invalid log stride size (%u. Range is %u - %u)\n",
5870 				    ucmd.single_stride_log_num_of_bytes,
5871 				    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES,
5872 				    MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES);
5873 			return -EINVAL;
5874 		}
5875 		if ((ucmd.single_wqe_log_num_of_strides >
5876 		    MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES) ||
5877 		     (ucmd.single_wqe_log_num_of_strides <
5878 			MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES)) {
5879 			mlx5_ib_dbg(dev, "Invalid log num strides (%u. Range is %u - %u)\n",
5880 				    ucmd.single_wqe_log_num_of_strides,
5881 				    MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES,
5882 				    MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES);
5883 			return -EINVAL;
5884 		}
5885 		rwq->single_stride_log_num_of_bytes =
5886 			ucmd.single_stride_log_num_of_bytes;
5887 		rwq->log_num_strides = ucmd.single_wqe_log_num_of_strides;
5888 		rwq->two_byte_shift_en = !!ucmd.two_byte_shift_en;
5889 		rwq->create_flags |= MLX5_IB_WQ_FLAGS_STRIDING_RQ;
5890 	}
5891 
5892 	err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
5893 	if (err) {
5894 		mlx5_ib_dbg(dev, "err %d\n", err);
5895 		return err;
5896 	}
5897 
5898 	err = create_user_rq(dev, pd, udata, rwq, &ucmd);
5899 	if (err) {
5900 		mlx5_ib_dbg(dev, "err %d\n", err);
5901 		return err;
5902 	}
5903 
5904 	rwq->user_index = ucmd.user_index;
5905 	return 0;
5906 }
5907 
5908 struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
5909 				struct ib_wq_init_attr *init_attr,
5910 				struct ib_udata *udata)
5911 {
5912 	struct mlx5_ib_dev *dev;
5913 	struct mlx5_ib_rwq *rwq;
5914 	struct mlx5_ib_create_wq_resp resp = {};
5915 	size_t min_resp_len;
5916 	int err;
5917 
5918 	if (!udata)
5919 		return ERR_PTR(-ENOSYS);
5920 
5921 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
5922 	if (udata->outlen && udata->outlen < min_resp_len)
5923 		return ERR_PTR(-EINVAL);
5924 
5925 	dev = to_mdev(pd->device);
5926 	switch (init_attr->wq_type) {
5927 	case IB_WQT_RQ:
5928 		rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
5929 		if (!rwq)
5930 			return ERR_PTR(-ENOMEM);
5931 		err = prepare_user_rq(pd, init_attr, udata, rwq);
5932 		if (err)
5933 			goto err;
5934 		err = create_rq(rwq, pd, init_attr);
5935 		if (err)
5936 			goto err_user_rq;
5937 		break;
5938 	default:
5939 		mlx5_ib_dbg(dev, "unsupported wq type %d\n",
5940 			    init_attr->wq_type);
5941 		return ERR_PTR(-EINVAL);
5942 	}
5943 
5944 	rwq->ibwq.wq_num = rwq->core_qp.qpn;
5945 	rwq->ibwq.state = IB_WQS_RESET;
5946 	if (udata->outlen) {
5947 		resp.response_length = offsetof(typeof(resp), response_length) +
5948 				sizeof(resp.response_length);
5949 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
5950 		if (err)
5951 			goto err_copy;
5952 	}
5953 
5954 	rwq->core_qp.event = mlx5_ib_wq_event;
5955 	rwq->ibwq.event_handler = init_attr->event_handler;
5956 	return &rwq->ibwq;
5957 
5958 err_copy:
5959 	mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
5960 err_user_rq:
5961 	destroy_user_rq(dev, pd, rwq);
5962 err:
5963 	kfree(rwq);
5964 	return ERR_PTR(err);
5965 }
5966 
5967 int mlx5_ib_destroy_wq(struct ib_wq *wq)
5968 {
5969 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
5970 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
5971 
5972 	mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
5973 	destroy_user_rq(dev, wq->pd, rwq);
5974 	kfree(rwq);
5975 
5976 	return 0;
5977 }
5978 
5979 struct ib_rwq_ind_table *mlx5_ib_create_rwq_ind_table(struct ib_device *device,
5980 						      struct ib_rwq_ind_table_init_attr *init_attr,
5981 						      struct ib_udata *udata)
5982 {
5983 	struct mlx5_ib_dev *dev = to_mdev(device);
5984 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl;
5985 	int sz = 1 << init_attr->log_ind_tbl_size;
5986 	struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
5987 	size_t min_resp_len;
5988 	int inlen;
5989 	int err;
5990 	int i;
5991 	u32 *in;
5992 	void *rqtc;
5993 
5994 	if (udata->inlen > 0 &&
5995 	    !ib_is_udata_cleared(udata, 0,
5996 				 udata->inlen))
5997 		return ERR_PTR(-EOPNOTSUPP);
5998 
5999 	if (init_attr->log_ind_tbl_size >
6000 	    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size)) {
6001 		mlx5_ib_dbg(dev, "log_ind_tbl_size = %d is bigger than supported = %d\n",
6002 			    init_attr->log_ind_tbl_size,
6003 			    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size));
6004 		return ERR_PTR(-EINVAL);
6005 	}
6006 
6007 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
6008 	if (udata->outlen && udata->outlen < min_resp_len)
6009 		return ERR_PTR(-EINVAL);
6010 
6011 	rwq_ind_tbl = kzalloc(sizeof(*rwq_ind_tbl), GFP_KERNEL);
6012 	if (!rwq_ind_tbl)
6013 		return ERR_PTR(-ENOMEM);
6014 
6015 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
6016 	in = kvzalloc(inlen, GFP_KERNEL);
6017 	if (!in) {
6018 		err = -ENOMEM;
6019 		goto err;
6020 	}
6021 
6022 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
6023 
6024 	MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
6025 	MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
6026 
6027 	for (i = 0; i < sz; i++)
6028 		MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
6029 
6030 	rwq_ind_tbl->uid = to_mpd(init_attr->ind_tbl[0]->pd)->uid;
6031 	MLX5_SET(create_rqt_in, in, uid, rwq_ind_tbl->uid);
6032 
6033 	err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
6034 	kvfree(in);
6035 
6036 	if (err)
6037 		goto err;
6038 
6039 	rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
6040 	if (udata->outlen) {
6041 		resp.response_length = offsetof(typeof(resp), response_length) +
6042 					sizeof(resp.response_length);
6043 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
6044 		if (err)
6045 			goto err_copy;
6046 	}
6047 
6048 	return &rwq_ind_tbl->ib_rwq_ind_tbl;
6049 
6050 err_copy:
6051 	mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
6052 err:
6053 	kfree(rwq_ind_tbl);
6054 	return ERR_PTR(err);
6055 }
6056 
6057 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
6058 {
6059 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
6060 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
6061 
6062 	mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
6063 
6064 	kfree(rwq_ind_tbl);
6065 	return 0;
6066 }
6067 
6068 int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
6069 		      u32 wq_attr_mask, struct ib_udata *udata)
6070 {
6071 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
6072 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
6073 	struct mlx5_ib_modify_wq ucmd = {};
6074 	size_t required_cmd_sz;
6075 	int curr_wq_state;
6076 	int wq_state;
6077 	int inlen;
6078 	int err;
6079 	void *rqc;
6080 	void *in;
6081 
6082 	required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
6083 	if (udata->inlen < required_cmd_sz)
6084 		return -EINVAL;
6085 
6086 	if (udata->inlen > sizeof(ucmd) &&
6087 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
6088 				 udata->inlen - sizeof(ucmd)))
6089 		return -EOPNOTSUPP;
6090 
6091 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
6092 		return -EFAULT;
6093 
6094 	if (ucmd.comp_mask || ucmd.reserved)
6095 		return -EOPNOTSUPP;
6096 
6097 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
6098 	in = kvzalloc(inlen, GFP_KERNEL);
6099 	if (!in)
6100 		return -ENOMEM;
6101 
6102 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
6103 
6104 	curr_wq_state = (wq_attr_mask & IB_WQ_CUR_STATE) ?
6105 		wq_attr->curr_wq_state : wq->state;
6106 	wq_state = (wq_attr_mask & IB_WQ_STATE) ?
6107 		wq_attr->wq_state : curr_wq_state;
6108 	if (curr_wq_state == IB_WQS_ERR)
6109 		curr_wq_state = MLX5_RQC_STATE_ERR;
6110 	if (wq_state == IB_WQS_ERR)
6111 		wq_state = MLX5_RQC_STATE_ERR;
6112 	MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
6113 	MLX5_SET(modify_rq_in, in, uid, to_mpd(wq->pd)->uid);
6114 	MLX5_SET(rqc, rqc, state, wq_state);
6115 
6116 	if (wq_attr_mask & IB_WQ_FLAGS) {
6117 		if (wq_attr->flags_mask & IB_WQ_FLAGS_CVLAN_STRIPPING) {
6118 			if (!(MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
6119 			      MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
6120 				mlx5_ib_dbg(dev, "VLAN offloads are not "
6121 					    "supported\n");
6122 				err = -EOPNOTSUPP;
6123 				goto out;
6124 			}
6125 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
6126 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD);
6127 			MLX5_SET(rqc, rqc, vsd,
6128 				 (wq_attr->flags & IB_WQ_FLAGS_CVLAN_STRIPPING) ? 0 : 1);
6129 		}
6130 
6131 		if (wq_attr->flags_mask & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
6132 			mlx5_ib_dbg(dev, "Modifying scatter end padding is not supported\n");
6133 			err = -EOPNOTSUPP;
6134 			goto out;
6135 		}
6136 	}
6137 
6138 	if (curr_wq_state == IB_WQS_RESET && wq_state == IB_WQS_RDY) {
6139 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
6140 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
6141 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
6142 			MLX5_SET(rqc, rqc, counter_set_id,
6143 				 dev->port->cnts.set_id);
6144 		} else
6145 			dev_info_once(
6146 				&dev->ib_dev.dev,
6147 				"Receive WQ counters are not supported on current FW\n");
6148 	}
6149 
6150 	err = mlx5_core_modify_rq(dev->mdev, rwq->core_qp.qpn, in, inlen);
6151 	if (!err)
6152 		rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;
6153 
6154 out:
6155 	kvfree(in);
6156 	return err;
6157 }
6158 
6159 struct mlx5_ib_drain_cqe {
6160 	struct ib_cqe cqe;
6161 	struct completion done;
6162 };
6163 
6164 static void mlx5_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
6165 {
6166 	struct mlx5_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
6167 						     struct mlx5_ib_drain_cqe,
6168 						     cqe);
6169 
6170 	complete(&cqe->done);
6171 }
6172 
6173 /* This function returns only once the drained WR was completed */
6174 static void handle_drain_completion(struct ib_cq *cq,
6175 				    struct mlx5_ib_drain_cqe *sdrain,
6176 				    struct mlx5_ib_dev *dev)
6177 {
6178 	struct mlx5_core_dev *mdev = dev->mdev;
6179 
6180 	if (cq->poll_ctx == IB_POLL_DIRECT) {
6181 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
6182 			ib_process_cq_direct(cq, -1);
6183 		return;
6184 	}
6185 
6186 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
6187 		struct mlx5_ib_cq *mcq = to_mcq(cq);
6188 		bool triggered = false;
6189 		unsigned long flags;
6190 
6191 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
6192 		/* Make sure that the CQ handler won't run if wasn't run yet */
6193 		if (!mcq->mcq.reset_notify_added)
6194 			mcq->mcq.reset_notify_added = 1;
6195 		else
6196 			triggered = true;
6197 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
6198 
6199 		if (triggered) {
6200 			/* Wait for any scheduled/running task to be ended */
6201 			switch (cq->poll_ctx) {
6202 			case IB_POLL_SOFTIRQ:
6203 				irq_poll_disable(&cq->iop);
6204 				irq_poll_enable(&cq->iop);
6205 				break;
6206 			case IB_POLL_WORKQUEUE:
6207 				cancel_work_sync(&cq->work);
6208 				break;
6209 			default:
6210 				WARN_ON_ONCE(1);
6211 			}
6212 		}
6213 
6214 		/* Run the CQ handler - this makes sure that the drain WR will
6215 		 * be processed if wasn't processed yet.
6216 		 */
6217 		mcq->mcq.comp(&mcq->mcq);
6218 	}
6219 
6220 	wait_for_completion(&sdrain->done);
6221 }
6222 
6223 void mlx5_ib_drain_sq(struct ib_qp *qp)
6224 {
6225 	struct ib_cq *cq = qp->send_cq;
6226 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
6227 	struct mlx5_ib_drain_cqe sdrain;
6228 	const struct ib_send_wr *bad_swr;
6229 	struct ib_rdma_wr swr = {
6230 		.wr = {
6231 			.next = NULL,
6232 			{ .wr_cqe	= &sdrain.cqe, },
6233 			.opcode	= IB_WR_RDMA_WRITE,
6234 		},
6235 	};
6236 	int ret;
6237 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
6238 	struct mlx5_core_dev *mdev = dev->mdev;
6239 
6240 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
6241 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
6242 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
6243 		return;
6244 	}
6245 
6246 	sdrain.cqe.done = mlx5_ib_drain_qp_done;
6247 	init_completion(&sdrain.done);
6248 
6249 	ret = _mlx5_ib_post_send(qp, &swr.wr, &bad_swr, true);
6250 	if (ret) {
6251 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
6252 		return;
6253 	}
6254 
6255 	handle_drain_completion(cq, &sdrain, dev);
6256 }
6257 
6258 void mlx5_ib_drain_rq(struct ib_qp *qp)
6259 {
6260 	struct ib_cq *cq = qp->recv_cq;
6261 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
6262 	struct mlx5_ib_drain_cqe rdrain;
6263 	struct ib_recv_wr rwr = {};
6264 	const struct ib_recv_wr *bad_rwr;
6265 	int ret;
6266 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
6267 	struct mlx5_core_dev *mdev = dev->mdev;
6268 
6269 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
6270 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
6271 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
6272 		return;
6273 	}
6274 
6275 	rwr.wr_cqe = &rdrain.cqe;
6276 	rdrain.cqe.done = mlx5_ib_drain_qp_done;
6277 	init_completion(&rdrain.done);
6278 
6279 	ret = _mlx5_ib_post_recv(qp, &rwr, &bad_rwr, true);
6280 	if (ret) {
6281 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
6282 		return;
6283 	}
6284 
6285 	handle_drain_completion(cq, &rdrain, dev);
6286 }
6287