xref: /linux/drivers/infiniband/hw/mlx5/qp.c (revision b6ebbac51bedf9e98e837688bc838f400196da5e)
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 "mlx5_ib.h"
38 #include "user.h"
39 
40 /* not supported currently */
41 static int wq_signature;
42 
43 enum {
44 	MLX5_IB_ACK_REQ_FREQ	= 8,
45 };
46 
47 enum {
48 	MLX5_IB_DEFAULT_SCHED_QUEUE	= 0x83,
49 	MLX5_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
50 	MLX5_IB_LINK_TYPE_IB		= 0,
51 	MLX5_IB_LINK_TYPE_ETH		= 1
52 };
53 
54 enum {
55 	MLX5_IB_SQ_STRIDE	= 6,
56 	MLX5_IB_CACHE_LINE_SIZE	= 64,
57 };
58 
59 static const u32 mlx5_ib_opcode[] = {
60 	[IB_WR_SEND]				= MLX5_OPCODE_SEND,
61 	[IB_WR_LSO]				= MLX5_OPCODE_LSO,
62 	[IB_WR_SEND_WITH_IMM]			= MLX5_OPCODE_SEND_IMM,
63 	[IB_WR_RDMA_WRITE]			= MLX5_OPCODE_RDMA_WRITE,
64 	[IB_WR_RDMA_WRITE_WITH_IMM]		= MLX5_OPCODE_RDMA_WRITE_IMM,
65 	[IB_WR_RDMA_READ]			= MLX5_OPCODE_RDMA_READ,
66 	[IB_WR_ATOMIC_CMP_AND_SWP]		= MLX5_OPCODE_ATOMIC_CS,
67 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= MLX5_OPCODE_ATOMIC_FA,
68 	[IB_WR_SEND_WITH_INV]			= MLX5_OPCODE_SEND_INVAL,
69 	[IB_WR_LOCAL_INV]			= MLX5_OPCODE_UMR,
70 	[IB_WR_REG_MR]				= MLX5_OPCODE_UMR,
71 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= MLX5_OPCODE_ATOMIC_MASKED_CS,
72 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= MLX5_OPCODE_ATOMIC_MASKED_FA,
73 	[MLX5_IB_WR_UMR]			= MLX5_OPCODE_UMR,
74 };
75 
76 struct mlx5_wqe_eth_pad {
77 	u8 rsvd0[16];
78 };
79 
80 static void get_cqs(enum ib_qp_type qp_type,
81 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
82 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq);
83 
84 static int is_qp0(enum ib_qp_type qp_type)
85 {
86 	return qp_type == IB_QPT_SMI;
87 }
88 
89 static int is_sqp(enum ib_qp_type qp_type)
90 {
91 	return is_qp0(qp_type) || is_qp1(qp_type);
92 }
93 
94 static void *get_wqe(struct mlx5_ib_qp *qp, int offset)
95 {
96 	return mlx5_buf_offset(&qp->buf, offset);
97 }
98 
99 static void *get_recv_wqe(struct mlx5_ib_qp *qp, int n)
100 {
101 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
102 }
103 
104 void *mlx5_get_send_wqe(struct mlx5_ib_qp *qp, int n)
105 {
106 	return get_wqe(qp, qp->sq.offset + (n << MLX5_IB_SQ_STRIDE));
107 }
108 
109 /**
110  * mlx5_ib_read_user_wqe() - Copy a user-space WQE to kernel space.
111  *
112  * @qp: QP to copy from.
113  * @send: copy from the send queue when non-zero, use the receive queue
114  *	  otherwise.
115  * @wqe_index:  index to start copying from. For send work queues, the
116  *		wqe_index is in units of MLX5_SEND_WQE_BB.
117  *		For receive work queue, it is the number of work queue
118  *		element in the queue.
119  * @buffer: destination buffer.
120  * @length: maximum number of bytes to copy.
121  *
122  * Copies at least a single WQE, but may copy more data.
123  *
124  * Return: the number of bytes copied, or an error code.
125  */
126 int mlx5_ib_read_user_wqe(struct mlx5_ib_qp *qp, int send, int wqe_index,
127 			  void *buffer, u32 length,
128 			  struct mlx5_ib_qp_base *base)
129 {
130 	struct ib_device *ibdev = qp->ibqp.device;
131 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
132 	struct mlx5_ib_wq *wq = send ? &qp->sq : &qp->rq;
133 	size_t offset;
134 	size_t wq_end;
135 	struct ib_umem *umem = base->ubuffer.umem;
136 	u32 first_copy_length;
137 	int wqe_length;
138 	int ret;
139 
140 	if (wq->wqe_cnt == 0) {
141 		mlx5_ib_dbg(dev, "mlx5_ib_read_user_wqe for a QP with wqe_cnt == 0. qp_type: 0x%x\n",
142 			    qp->ibqp.qp_type);
143 		return -EINVAL;
144 	}
145 
146 	offset = wq->offset + ((wqe_index % wq->wqe_cnt) << wq->wqe_shift);
147 	wq_end = wq->offset + (wq->wqe_cnt << wq->wqe_shift);
148 
149 	if (send && length < sizeof(struct mlx5_wqe_ctrl_seg))
150 		return -EINVAL;
151 
152 	if (offset > umem->length ||
153 	    (send && offset + sizeof(struct mlx5_wqe_ctrl_seg) > umem->length))
154 		return -EINVAL;
155 
156 	first_copy_length = min_t(u32, offset + length, wq_end) - offset;
157 	ret = ib_umem_copy_from(buffer, umem, offset, first_copy_length);
158 	if (ret)
159 		return ret;
160 
161 	if (send) {
162 		struct mlx5_wqe_ctrl_seg *ctrl = buffer;
163 		int ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
164 
165 		wqe_length = ds * MLX5_WQE_DS_UNITS;
166 	} else {
167 		wqe_length = 1 << wq->wqe_shift;
168 	}
169 
170 	if (wqe_length <= first_copy_length)
171 		return first_copy_length;
172 
173 	ret = ib_umem_copy_from(buffer + first_copy_length, umem, wq->offset,
174 				wqe_length - first_copy_length);
175 	if (ret)
176 		return ret;
177 
178 	return wqe_length;
179 }
180 
181 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
182 {
183 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
184 	struct ib_event event;
185 
186 	if (type == MLX5_EVENT_TYPE_PATH_MIG) {
187 		/* This event is only valid for trans_qps */
188 		to_mibqp(qp)->port = to_mibqp(qp)->trans_qp.alt_port;
189 	}
190 
191 	if (ibqp->event_handler) {
192 		event.device     = ibqp->device;
193 		event.element.qp = ibqp;
194 		switch (type) {
195 		case MLX5_EVENT_TYPE_PATH_MIG:
196 			event.event = IB_EVENT_PATH_MIG;
197 			break;
198 		case MLX5_EVENT_TYPE_COMM_EST:
199 			event.event = IB_EVENT_COMM_EST;
200 			break;
201 		case MLX5_EVENT_TYPE_SQ_DRAINED:
202 			event.event = IB_EVENT_SQ_DRAINED;
203 			break;
204 		case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
205 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
206 			break;
207 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
208 			event.event = IB_EVENT_QP_FATAL;
209 			break;
210 		case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
211 			event.event = IB_EVENT_PATH_MIG_ERR;
212 			break;
213 		case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
214 			event.event = IB_EVENT_QP_REQ_ERR;
215 			break;
216 		case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
217 			event.event = IB_EVENT_QP_ACCESS_ERR;
218 			break;
219 		default:
220 			pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
221 			return;
222 		}
223 
224 		ibqp->event_handler(&event, ibqp->qp_context);
225 	}
226 }
227 
228 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
229 		       int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
230 {
231 	int wqe_size;
232 	int wq_size;
233 
234 	/* Sanity check RQ size before proceeding */
235 	if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
236 		return -EINVAL;
237 
238 	if (!has_rq) {
239 		qp->rq.max_gs = 0;
240 		qp->rq.wqe_cnt = 0;
241 		qp->rq.wqe_shift = 0;
242 		cap->max_recv_wr = 0;
243 		cap->max_recv_sge = 0;
244 	} else {
245 		if (ucmd) {
246 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
247 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
248 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
249 			qp->rq.max_post = qp->rq.wqe_cnt;
250 		} else {
251 			wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
252 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
253 			wqe_size = roundup_pow_of_two(wqe_size);
254 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
255 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
256 			qp->rq.wqe_cnt = wq_size / wqe_size;
257 			if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
258 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
259 					    wqe_size,
260 					    MLX5_CAP_GEN(dev->mdev,
261 							 max_wqe_sz_rq));
262 				return -EINVAL;
263 			}
264 			qp->rq.wqe_shift = ilog2(wqe_size);
265 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
266 			qp->rq.max_post = qp->rq.wqe_cnt;
267 		}
268 	}
269 
270 	return 0;
271 }
272 
273 static int sq_overhead(struct ib_qp_init_attr *attr)
274 {
275 	int size = 0;
276 
277 	switch (attr->qp_type) {
278 	case IB_QPT_XRC_INI:
279 		size += sizeof(struct mlx5_wqe_xrc_seg);
280 		/* fall through */
281 	case IB_QPT_RC:
282 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
283 			max(sizeof(struct mlx5_wqe_atomic_seg) +
284 			    sizeof(struct mlx5_wqe_raddr_seg),
285 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
286 			    sizeof(struct mlx5_mkey_seg));
287 		break;
288 
289 	case IB_QPT_XRC_TGT:
290 		return 0;
291 
292 	case IB_QPT_UC:
293 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
294 			max(sizeof(struct mlx5_wqe_raddr_seg),
295 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
296 			    sizeof(struct mlx5_mkey_seg));
297 		break;
298 
299 	case IB_QPT_UD:
300 		if (attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
301 			size += sizeof(struct mlx5_wqe_eth_pad) +
302 				sizeof(struct mlx5_wqe_eth_seg);
303 		/* fall through */
304 	case IB_QPT_SMI:
305 	case MLX5_IB_QPT_HW_GSI:
306 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
307 			sizeof(struct mlx5_wqe_datagram_seg);
308 		break;
309 
310 	case MLX5_IB_QPT_REG_UMR:
311 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
312 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
313 			sizeof(struct mlx5_mkey_seg);
314 		break;
315 
316 	default:
317 		return -EINVAL;
318 	}
319 
320 	return size;
321 }
322 
323 static int calc_send_wqe(struct ib_qp_init_attr *attr)
324 {
325 	int inl_size = 0;
326 	int size;
327 
328 	size = sq_overhead(attr);
329 	if (size < 0)
330 		return size;
331 
332 	if (attr->cap.max_inline_data) {
333 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
334 			attr->cap.max_inline_data;
335 	}
336 
337 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
338 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN &&
339 	    ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
340 			return MLX5_SIG_WQE_SIZE;
341 	else
342 		return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
343 }
344 
345 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
346 			struct mlx5_ib_qp *qp)
347 {
348 	int wqe_size;
349 	int wq_size;
350 
351 	if (!attr->cap.max_send_wr)
352 		return 0;
353 
354 	wqe_size = calc_send_wqe(attr);
355 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
356 	if (wqe_size < 0)
357 		return wqe_size;
358 
359 	if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
360 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
361 			    wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
362 		return -EINVAL;
363 	}
364 
365 	qp->max_inline_data = wqe_size - sq_overhead(attr) -
366 			      sizeof(struct mlx5_wqe_inline_seg);
367 	attr->cap.max_inline_data = qp->max_inline_data;
368 
369 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN)
370 		qp->signature_en = true;
371 
372 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
373 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
374 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
375 		mlx5_ib_dbg(dev, "wqe count(%d) exceeds limits(%d)\n",
376 			    qp->sq.wqe_cnt,
377 			    1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
378 		return -ENOMEM;
379 	}
380 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
381 	qp->sq.max_gs = attr->cap.max_send_sge;
382 	qp->sq.max_post = wq_size / wqe_size;
383 	attr->cap.max_send_wr = qp->sq.max_post;
384 
385 	return wq_size;
386 }
387 
388 static int set_user_buf_size(struct mlx5_ib_dev *dev,
389 			    struct mlx5_ib_qp *qp,
390 			    struct mlx5_ib_create_qp *ucmd,
391 			    struct mlx5_ib_qp_base *base,
392 			    struct ib_qp_init_attr *attr)
393 {
394 	int desc_sz = 1 << qp->sq.wqe_shift;
395 
396 	if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
397 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
398 			     desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
399 		return -EINVAL;
400 	}
401 
402 	if (ucmd->sq_wqe_count && ((1 << ilog2(ucmd->sq_wqe_count)) != ucmd->sq_wqe_count)) {
403 		mlx5_ib_warn(dev, "sq_wqe_count %d, sq_wqe_count %d\n",
404 			     ucmd->sq_wqe_count, ucmd->sq_wqe_count);
405 		return -EINVAL;
406 	}
407 
408 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
409 
410 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
411 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
412 			     qp->sq.wqe_cnt,
413 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
414 		return -EINVAL;
415 	}
416 
417 	if (attr->qp_type == IB_QPT_RAW_PACKET) {
418 		base->ubuffer.buf_size = qp->rq.wqe_cnt << qp->rq.wqe_shift;
419 		qp->raw_packet_qp.sq.ubuffer.buf_size = qp->sq.wqe_cnt << 6;
420 	} else {
421 		base->ubuffer.buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
422 					 (qp->sq.wqe_cnt << 6);
423 	}
424 
425 	return 0;
426 }
427 
428 static int qp_has_rq(struct ib_qp_init_attr *attr)
429 {
430 	if (attr->qp_type == IB_QPT_XRC_INI ||
431 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
432 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
433 	    !attr->cap.max_recv_wr)
434 		return 0;
435 
436 	return 1;
437 }
438 
439 static int first_med_uuar(void)
440 {
441 	return 1;
442 }
443 
444 static int next_uuar(int n)
445 {
446 	n++;
447 
448 	while (((n % 4) & 2))
449 		n++;
450 
451 	return n;
452 }
453 
454 static int num_med_uuar(struct mlx5_uuar_info *uuari)
455 {
456 	int n;
457 
458 	n = uuari->num_uars * MLX5_NON_FP_BF_REGS_PER_PAGE -
459 		uuari->num_low_latency_uuars - 1;
460 
461 	return n >= 0 ? n : 0;
462 }
463 
464 static int max_uuari(struct mlx5_uuar_info *uuari)
465 {
466 	return uuari->num_uars * 4;
467 }
468 
469 static int first_hi_uuar(struct mlx5_uuar_info *uuari)
470 {
471 	int med;
472 	int i;
473 	int t;
474 
475 	med = num_med_uuar(uuari);
476 	for (t = 0, i = first_med_uuar();; i = next_uuar(i)) {
477 		t++;
478 		if (t == med)
479 			return next_uuar(i);
480 	}
481 
482 	return 0;
483 }
484 
485 static int alloc_high_class_uuar(struct mlx5_uuar_info *uuari)
486 {
487 	int i;
488 
489 	for (i = first_hi_uuar(uuari); i < max_uuari(uuari); i = next_uuar(i)) {
490 		if (!test_bit(i, uuari->bitmap)) {
491 			set_bit(i, uuari->bitmap);
492 			uuari->count[i]++;
493 			return i;
494 		}
495 	}
496 
497 	return -ENOMEM;
498 }
499 
500 static int alloc_med_class_uuar(struct mlx5_uuar_info *uuari)
501 {
502 	int minidx = first_med_uuar();
503 	int i;
504 
505 	for (i = first_med_uuar(); i < first_hi_uuar(uuari); i = next_uuar(i)) {
506 		if (uuari->count[i] < uuari->count[minidx])
507 			minidx = i;
508 	}
509 
510 	uuari->count[minidx]++;
511 	return minidx;
512 }
513 
514 static int alloc_uuar(struct mlx5_uuar_info *uuari,
515 		      enum mlx5_ib_latency_class lat)
516 {
517 	int uuarn = -EINVAL;
518 
519 	mutex_lock(&uuari->lock);
520 	switch (lat) {
521 	case MLX5_IB_LATENCY_CLASS_LOW:
522 		uuarn = 0;
523 		uuari->count[uuarn]++;
524 		break;
525 
526 	case MLX5_IB_LATENCY_CLASS_MEDIUM:
527 		if (uuari->ver < 2)
528 			uuarn = -ENOMEM;
529 		else
530 			uuarn = alloc_med_class_uuar(uuari);
531 		break;
532 
533 	case MLX5_IB_LATENCY_CLASS_HIGH:
534 		if (uuari->ver < 2)
535 			uuarn = -ENOMEM;
536 		else
537 			uuarn = alloc_high_class_uuar(uuari);
538 		break;
539 
540 	case MLX5_IB_LATENCY_CLASS_FAST_PATH:
541 		uuarn = 2;
542 		break;
543 	}
544 	mutex_unlock(&uuari->lock);
545 
546 	return uuarn;
547 }
548 
549 static void free_med_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
550 {
551 	clear_bit(uuarn, uuari->bitmap);
552 	--uuari->count[uuarn];
553 }
554 
555 static void free_high_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
556 {
557 	clear_bit(uuarn, uuari->bitmap);
558 	--uuari->count[uuarn];
559 }
560 
561 static void free_uuar(struct mlx5_uuar_info *uuari, int uuarn)
562 {
563 	int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
564 	int high_uuar = nuuars - uuari->num_low_latency_uuars;
565 
566 	mutex_lock(&uuari->lock);
567 	if (uuarn == 0) {
568 		--uuari->count[uuarn];
569 		goto out;
570 	}
571 
572 	if (uuarn < high_uuar) {
573 		free_med_class_uuar(uuari, uuarn);
574 		goto out;
575 	}
576 
577 	free_high_class_uuar(uuari, uuarn);
578 
579 out:
580 	mutex_unlock(&uuari->lock);
581 }
582 
583 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
584 {
585 	switch (state) {
586 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
587 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
588 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
589 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
590 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
591 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
592 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
593 	default:		return -1;
594 	}
595 }
596 
597 static int to_mlx5_st(enum ib_qp_type type)
598 {
599 	switch (type) {
600 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
601 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
602 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
603 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
604 	case IB_QPT_XRC_INI:
605 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
606 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
607 	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
608 	case IB_QPT_RAW_IPV6:		return MLX5_QP_ST_RAW_IPV6;
609 	case IB_QPT_RAW_PACKET:
610 	case IB_QPT_RAW_ETHERTYPE:	return MLX5_QP_ST_RAW_ETHERTYPE;
611 	case IB_QPT_MAX:
612 	default:		return -EINVAL;
613 	}
614 }
615 
616 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
617 			     struct mlx5_ib_cq *recv_cq);
618 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
619 			       struct mlx5_ib_cq *recv_cq);
620 
621 static int uuarn_to_uar_index(struct mlx5_uuar_info *uuari, int uuarn)
622 {
623 	return uuari->uars[uuarn / MLX5_BF_REGS_PER_PAGE].index;
624 }
625 
626 static int mlx5_ib_umem_get(struct mlx5_ib_dev *dev,
627 			    struct ib_pd *pd,
628 			    unsigned long addr, size_t size,
629 			    struct ib_umem **umem,
630 			    int *npages, int *page_shift, int *ncont,
631 			    u32 *offset)
632 {
633 	int err;
634 
635 	*umem = ib_umem_get(pd->uobject->context, addr, size, 0, 0);
636 	if (IS_ERR(*umem)) {
637 		mlx5_ib_dbg(dev, "umem_get failed\n");
638 		return PTR_ERR(*umem);
639 	}
640 
641 	mlx5_ib_cont_pages(*umem, addr, npages, page_shift, ncont, NULL);
642 
643 	err = mlx5_ib_get_buf_offset(addr, *page_shift, offset);
644 	if (err) {
645 		mlx5_ib_warn(dev, "bad offset\n");
646 		goto err_umem;
647 	}
648 
649 	mlx5_ib_dbg(dev, "addr 0x%lx, size %zu, npages %d, page_shift %d, ncont %d, offset %d\n",
650 		    addr, size, *npages, *page_shift, *ncont, *offset);
651 
652 	return 0;
653 
654 err_umem:
655 	ib_umem_release(*umem);
656 	*umem = NULL;
657 
658 	return err;
659 }
660 
661 static void destroy_user_rq(struct ib_pd *pd, struct mlx5_ib_rwq *rwq)
662 {
663 	struct mlx5_ib_ucontext *context;
664 
665 	context = to_mucontext(pd->uobject->context);
666 	mlx5_ib_db_unmap_user(context, &rwq->db);
667 	if (rwq->umem)
668 		ib_umem_release(rwq->umem);
669 }
670 
671 static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
672 			  struct mlx5_ib_rwq *rwq,
673 			  struct mlx5_ib_create_wq *ucmd)
674 {
675 	struct mlx5_ib_ucontext *context;
676 	int page_shift = 0;
677 	int npages;
678 	u32 offset = 0;
679 	int ncont = 0;
680 	int err;
681 
682 	if (!ucmd->buf_addr)
683 		return -EINVAL;
684 
685 	context = to_mucontext(pd->uobject->context);
686 	rwq->umem = ib_umem_get(pd->uobject->context, ucmd->buf_addr,
687 			       rwq->buf_size, 0, 0);
688 	if (IS_ERR(rwq->umem)) {
689 		mlx5_ib_dbg(dev, "umem_get failed\n");
690 		err = PTR_ERR(rwq->umem);
691 		return err;
692 	}
693 
694 	mlx5_ib_cont_pages(rwq->umem, ucmd->buf_addr, &npages, &page_shift,
695 			   &ncont, NULL);
696 	err = mlx5_ib_get_buf_offset(ucmd->buf_addr, page_shift,
697 				     &rwq->rq_page_offset);
698 	if (err) {
699 		mlx5_ib_warn(dev, "bad offset\n");
700 		goto err_umem;
701 	}
702 
703 	rwq->rq_num_pas = ncont;
704 	rwq->page_shift = page_shift;
705 	rwq->log_page_size =  page_shift - MLX5_ADAPTER_PAGE_SHIFT;
706 	rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
707 
708 	mlx5_ib_dbg(dev, "addr 0x%llx, size %zd, npages %d, page_shift %d, ncont %d, offset %d\n",
709 		    (unsigned long long)ucmd->buf_addr, rwq->buf_size,
710 		    npages, page_shift, ncont, offset);
711 
712 	err = mlx5_ib_db_map_user(context, ucmd->db_addr, &rwq->db);
713 	if (err) {
714 		mlx5_ib_dbg(dev, "map failed\n");
715 		goto err_umem;
716 	}
717 
718 	rwq->create_type = MLX5_WQ_USER;
719 	return 0;
720 
721 err_umem:
722 	ib_umem_release(rwq->umem);
723 	return err;
724 }
725 
726 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
727 			  struct mlx5_ib_qp *qp, struct ib_udata *udata,
728 			  struct ib_qp_init_attr *attr,
729 			  struct mlx5_create_qp_mbox_in **in,
730 			  struct mlx5_ib_create_qp_resp *resp, int *inlen,
731 			  struct mlx5_ib_qp_base *base)
732 {
733 	struct mlx5_ib_ucontext *context;
734 	struct mlx5_ib_create_qp ucmd;
735 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
736 	int page_shift = 0;
737 	int uar_index;
738 	int npages;
739 	u32 offset = 0;
740 	int uuarn;
741 	int ncont = 0;
742 	int err;
743 
744 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
745 	if (err) {
746 		mlx5_ib_dbg(dev, "copy failed\n");
747 		return err;
748 	}
749 
750 	context = to_mucontext(pd->uobject->context);
751 	/*
752 	 * TBD: should come from the verbs when we have the API
753 	 */
754 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
755 		/* In CROSS_CHANNEL CQ and QP must use the same UAR */
756 		uuarn = MLX5_CROSS_CHANNEL_UUAR;
757 	else {
758 		uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_HIGH);
759 		if (uuarn < 0) {
760 			mlx5_ib_dbg(dev, "failed to allocate low latency UUAR\n");
761 			mlx5_ib_dbg(dev, "reverting to medium latency\n");
762 			uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_MEDIUM);
763 			if (uuarn < 0) {
764 				mlx5_ib_dbg(dev, "failed to allocate medium latency UUAR\n");
765 				mlx5_ib_dbg(dev, "reverting to high latency\n");
766 				uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_LOW);
767 				if (uuarn < 0) {
768 					mlx5_ib_warn(dev, "uuar allocation failed\n");
769 					return uuarn;
770 				}
771 			}
772 		}
773 	}
774 
775 	uar_index = uuarn_to_uar_index(&context->uuari, uuarn);
776 	mlx5_ib_dbg(dev, "uuarn 0x%x, uar_index 0x%x\n", uuarn, uar_index);
777 
778 	qp->rq.offset = 0;
779 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
780 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
781 
782 	err = set_user_buf_size(dev, qp, &ucmd, base, attr);
783 	if (err)
784 		goto err_uuar;
785 
786 	if (ucmd.buf_addr && ubuffer->buf_size) {
787 		ubuffer->buf_addr = ucmd.buf_addr;
788 		err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr,
789 				       ubuffer->buf_size,
790 				       &ubuffer->umem, &npages, &page_shift,
791 				       &ncont, &offset);
792 		if (err)
793 			goto err_uuar;
794 	} else {
795 		ubuffer->umem = NULL;
796 	}
797 
798 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * ncont;
799 	*in = mlx5_vzalloc(*inlen);
800 	if (!*in) {
801 		err = -ENOMEM;
802 		goto err_umem;
803 	}
804 	if (ubuffer->umem)
805 		mlx5_ib_populate_pas(dev, ubuffer->umem, page_shift,
806 				     (*in)->pas, 0);
807 	(*in)->ctx.log_pg_sz_remote_qpn =
808 		cpu_to_be32((page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
809 	(*in)->ctx.params2 = cpu_to_be32(offset << 6);
810 
811 	(*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
812 	resp->uuar_index = uuarn;
813 	qp->uuarn = uuarn;
814 
815 	err = mlx5_ib_db_map_user(context, ucmd.db_addr, &qp->db);
816 	if (err) {
817 		mlx5_ib_dbg(dev, "map failed\n");
818 		goto err_free;
819 	}
820 
821 	err = ib_copy_to_udata(udata, resp, sizeof(*resp));
822 	if (err) {
823 		mlx5_ib_dbg(dev, "copy failed\n");
824 		goto err_unmap;
825 	}
826 	qp->create_type = MLX5_QP_USER;
827 
828 	return 0;
829 
830 err_unmap:
831 	mlx5_ib_db_unmap_user(context, &qp->db);
832 
833 err_free:
834 	kvfree(*in);
835 
836 err_umem:
837 	if (ubuffer->umem)
838 		ib_umem_release(ubuffer->umem);
839 
840 err_uuar:
841 	free_uuar(&context->uuari, uuarn);
842 	return err;
843 }
844 
845 static void destroy_qp_user(struct ib_pd *pd, struct mlx5_ib_qp *qp,
846 			    struct mlx5_ib_qp_base *base)
847 {
848 	struct mlx5_ib_ucontext *context;
849 
850 	context = to_mucontext(pd->uobject->context);
851 	mlx5_ib_db_unmap_user(context, &qp->db);
852 	if (base->ubuffer.umem)
853 		ib_umem_release(base->ubuffer.umem);
854 	free_uuar(&context->uuari, qp->uuarn);
855 }
856 
857 static int create_kernel_qp(struct mlx5_ib_dev *dev,
858 			    struct ib_qp_init_attr *init_attr,
859 			    struct mlx5_ib_qp *qp,
860 			    struct mlx5_create_qp_mbox_in **in, int *inlen,
861 			    struct mlx5_ib_qp_base *base)
862 {
863 	enum mlx5_ib_latency_class lc = MLX5_IB_LATENCY_CLASS_LOW;
864 	struct mlx5_uuar_info *uuari;
865 	int uar_index;
866 	int uuarn;
867 	int err;
868 
869 	uuari = &dev->mdev->priv.uuari;
870 	if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN |
871 					IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
872 					IB_QP_CREATE_IPOIB_UD_LSO |
873 					mlx5_ib_create_qp_sqpn_qp1()))
874 		return -EINVAL;
875 
876 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
877 		lc = MLX5_IB_LATENCY_CLASS_FAST_PATH;
878 
879 	uuarn = alloc_uuar(uuari, lc);
880 	if (uuarn < 0) {
881 		mlx5_ib_dbg(dev, "\n");
882 		return -ENOMEM;
883 	}
884 
885 	qp->bf = &uuari->bfs[uuarn];
886 	uar_index = qp->bf->uar->index;
887 
888 	err = calc_sq_size(dev, init_attr, qp);
889 	if (err < 0) {
890 		mlx5_ib_dbg(dev, "err %d\n", err);
891 		goto err_uuar;
892 	}
893 
894 	qp->rq.offset = 0;
895 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
896 	base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
897 
898 	err = mlx5_buf_alloc(dev->mdev, base->ubuffer.buf_size, &qp->buf);
899 	if (err) {
900 		mlx5_ib_dbg(dev, "err %d\n", err);
901 		goto err_uuar;
902 	}
903 
904 	qp->sq.qend = mlx5_get_send_wqe(qp, qp->sq.wqe_cnt);
905 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * qp->buf.npages;
906 	*in = mlx5_vzalloc(*inlen);
907 	if (!*in) {
908 		err = -ENOMEM;
909 		goto err_buf;
910 	}
911 	(*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
912 	(*in)->ctx.log_pg_sz_remote_qpn =
913 		cpu_to_be32((qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
914 	/* Set "fast registration enabled" for all kernel QPs */
915 	(*in)->ctx.params1 |= cpu_to_be32(1 << 11);
916 	(*in)->ctx.sq_crq_size |= cpu_to_be16(1 << 4);
917 
918 	if (init_attr->create_flags & mlx5_ib_create_qp_sqpn_qp1()) {
919 		(*in)->ctx.deth_sqpn = cpu_to_be32(1);
920 		qp->flags |= MLX5_IB_QP_SQPN_QP1;
921 	}
922 
923 	mlx5_fill_page_array(&qp->buf, (*in)->pas);
924 
925 	err = mlx5_db_alloc(dev->mdev, &qp->db);
926 	if (err) {
927 		mlx5_ib_dbg(dev, "err %d\n", err);
928 		goto err_free;
929 	}
930 
931 	qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
932 	qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
933 	qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
934 	qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
935 	qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
936 
937 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
938 	    !qp->sq.w_list || !qp->sq.wqe_head) {
939 		err = -ENOMEM;
940 		goto err_wrid;
941 	}
942 	qp->create_type = MLX5_QP_KERNEL;
943 
944 	return 0;
945 
946 err_wrid:
947 	mlx5_db_free(dev->mdev, &qp->db);
948 	kfree(qp->sq.wqe_head);
949 	kfree(qp->sq.w_list);
950 	kfree(qp->sq.wrid);
951 	kfree(qp->sq.wr_data);
952 	kfree(qp->rq.wrid);
953 
954 err_free:
955 	kvfree(*in);
956 
957 err_buf:
958 	mlx5_buf_free(dev->mdev, &qp->buf);
959 
960 err_uuar:
961 	free_uuar(&dev->mdev->priv.uuari, uuarn);
962 	return err;
963 }
964 
965 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
966 {
967 	mlx5_db_free(dev->mdev, &qp->db);
968 	kfree(qp->sq.wqe_head);
969 	kfree(qp->sq.w_list);
970 	kfree(qp->sq.wrid);
971 	kfree(qp->sq.wr_data);
972 	kfree(qp->rq.wrid);
973 	mlx5_buf_free(dev->mdev, &qp->buf);
974 	free_uuar(&dev->mdev->priv.uuari, qp->bf->uuarn);
975 }
976 
977 static __be32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
978 {
979 	if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
980 	    (attr->qp_type == IB_QPT_XRC_INI))
981 		return cpu_to_be32(MLX5_SRQ_RQ);
982 	else if (!qp->has_rq)
983 		return cpu_to_be32(MLX5_ZERO_LEN_RQ);
984 	else
985 		return cpu_to_be32(MLX5_NON_ZERO_RQ);
986 }
987 
988 static int is_connected(enum ib_qp_type qp_type)
989 {
990 	if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC)
991 		return 1;
992 
993 	return 0;
994 }
995 
996 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
997 				    struct mlx5_ib_sq *sq, u32 tdn)
998 {
999 	u32 in[MLX5_ST_SZ_DW(create_tis_in)];
1000 	void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1001 
1002 	memset(in, 0, sizeof(in));
1003 
1004 	MLX5_SET(tisc, tisc, transport_domain, tdn);
1005 
1006 	return mlx5_core_create_tis(dev->mdev, in, sizeof(in), &sq->tisn);
1007 }
1008 
1009 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1010 				      struct mlx5_ib_sq *sq)
1011 {
1012 	mlx5_core_destroy_tis(dev->mdev, sq->tisn);
1013 }
1014 
1015 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1016 				   struct mlx5_ib_sq *sq, void *qpin,
1017 				   struct ib_pd *pd)
1018 {
1019 	struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1020 	__be64 *pas;
1021 	void *in;
1022 	void *sqc;
1023 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1024 	void *wq;
1025 	int inlen;
1026 	int err;
1027 	int page_shift = 0;
1028 	int npages;
1029 	int ncont = 0;
1030 	u32 offset = 0;
1031 
1032 	err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr, ubuffer->buf_size,
1033 			       &sq->ubuffer.umem, &npages, &page_shift,
1034 			       &ncont, &offset);
1035 	if (err)
1036 		return err;
1037 
1038 	inlen = MLX5_ST_SZ_BYTES(create_sq_in) + sizeof(u64) * ncont;
1039 	in = mlx5_vzalloc(inlen);
1040 	if (!in) {
1041 		err = -ENOMEM;
1042 		goto err_umem;
1043 	}
1044 
1045 	sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1046 	MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1047 	MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1048 	MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1049 	MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1050 	MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1051 	MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1052 
1053 	wq = MLX5_ADDR_OF(sqc, sqc, wq);
1054 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1055 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1056 	MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1057 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1058 	MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1059 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1060 	MLX5_SET(wq, wq, log_wq_pg_sz,  page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1061 	MLX5_SET(wq, wq, page_offset, offset);
1062 
1063 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1064 	mlx5_ib_populate_pas(dev, sq->ubuffer.umem, page_shift, pas, 0);
1065 
1066 	err = mlx5_core_create_sq_tracked(dev->mdev, in, inlen, &sq->base.mqp);
1067 
1068 	kvfree(in);
1069 
1070 	if (err)
1071 		goto err_umem;
1072 
1073 	return 0;
1074 
1075 err_umem:
1076 	ib_umem_release(sq->ubuffer.umem);
1077 	sq->ubuffer.umem = NULL;
1078 
1079 	return err;
1080 }
1081 
1082 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1083 				     struct mlx5_ib_sq *sq)
1084 {
1085 	mlx5_core_destroy_sq_tracked(dev->mdev, &sq->base.mqp);
1086 	ib_umem_release(sq->ubuffer.umem);
1087 }
1088 
1089 static int get_rq_pas_size(void *qpc)
1090 {
1091 	u32 log_page_size = MLX5_GET(qpc, qpc, log_page_size) + 12;
1092 	u32 log_rq_stride = MLX5_GET(qpc, qpc, log_rq_stride);
1093 	u32 log_rq_size   = MLX5_GET(qpc, qpc, log_rq_size);
1094 	u32 page_offset   = MLX5_GET(qpc, qpc, page_offset);
1095 	u32 po_quanta	  = 1 << (log_page_size - 6);
1096 	u32 rq_sz	  = 1 << (log_rq_size + 4 + log_rq_stride);
1097 	u32 page_size	  = 1 << log_page_size;
1098 	u32 rq_sz_po      = rq_sz + (page_offset * po_quanta);
1099 	u32 rq_num_pas	  = (rq_sz_po + page_size - 1) / page_size;
1100 
1101 	return rq_num_pas * sizeof(u64);
1102 }
1103 
1104 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1105 				   struct mlx5_ib_rq *rq, void *qpin)
1106 {
1107 	struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1108 	__be64 *pas;
1109 	__be64 *qp_pas;
1110 	void *in;
1111 	void *rqc;
1112 	void *wq;
1113 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1114 	int inlen;
1115 	int err;
1116 	u32 rq_pas_size = get_rq_pas_size(qpc);
1117 
1118 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + rq_pas_size;
1119 	in = mlx5_vzalloc(inlen);
1120 	if (!in)
1121 		return -ENOMEM;
1122 
1123 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1124 	MLX5_SET(rqc, rqc, vsd, 1);
1125 	MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
1126 	MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1127 	MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1128 	MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1129 	MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1130 
1131 	if (mqp->flags & MLX5_IB_QP_CAP_SCATTER_FCS)
1132 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
1133 
1134 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
1135 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1136 	MLX5_SET(wq, wq, end_padding_mode,
1137 		 MLX5_GET(qpc, qpc, end_padding_mode));
1138 	MLX5_SET(wq, wq, page_offset, MLX5_GET(qpc, qpc, page_offset));
1139 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1140 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1141 	MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1142 	MLX5_SET(wq, wq, log_wq_pg_sz, MLX5_GET(qpc, qpc, log_page_size));
1143 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1144 
1145 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1146 	qp_pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, qpin, pas);
1147 	memcpy(pas, qp_pas, rq_pas_size);
1148 
1149 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rq->base.mqp);
1150 
1151 	kvfree(in);
1152 
1153 	return err;
1154 }
1155 
1156 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1157 				     struct mlx5_ib_rq *rq)
1158 {
1159 	mlx5_core_destroy_rq_tracked(dev->mdev, &rq->base.mqp);
1160 }
1161 
1162 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1163 				    struct mlx5_ib_rq *rq, u32 tdn)
1164 {
1165 	u32 *in;
1166 	void *tirc;
1167 	int inlen;
1168 	int err;
1169 
1170 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1171 	in = mlx5_vzalloc(inlen);
1172 	if (!in)
1173 		return -ENOMEM;
1174 
1175 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1176 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1177 	MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1178 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1179 
1180 	err = mlx5_core_create_tir(dev->mdev, in, inlen, &rq->tirn);
1181 
1182 	kvfree(in);
1183 
1184 	return err;
1185 }
1186 
1187 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1188 				      struct mlx5_ib_rq *rq)
1189 {
1190 	mlx5_core_destroy_tir(dev->mdev, rq->tirn);
1191 }
1192 
1193 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1194 				struct mlx5_create_qp_mbox_in *in,
1195 				struct ib_pd *pd)
1196 {
1197 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1198 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1199 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1200 	struct ib_uobject *uobj = pd->uobject;
1201 	struct ib_ucontext *ucontext = uobj->context;
1202 	struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1203 	int err;
1204 	u32 tdn = mucontext->tdn;
1205 
1206 	if (qp->sq.wqe_cnt) {
1207 		err = create_raw_packet_qp_tis(dev, sq, tdn);
1208 		if (err)
1209 			return err;
1210 
1211 		err = create_raw_packet_qp_sq(dev, sq, in, pd);
1212 		if (err)
1213 			goto err_destroy_tis;
1214 
1215 		sq->base.container_mibqp = qp;
1216 	}
1217 
1218 	if (qp->rq.wqe_cnt) {
1219 		rq->base.container_mibqp = qp;
1220 
1221 		err = create_raw_packet_qp_rq(dev, rq, in);
1222 		if (err)
1223 			goto err_destroy_sq;
1224 
1225 
1226 		err = create_raw_packet_qp_tir(dev, rq, tdn);
1227 		if (err)
1228 			goto err_destroy_rq;
1229 	}
1230 
1231 	qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1232 						     rq->base.mqp.qpn;
1233 
1234 	return 0;
1235 
1236 err_destroy_rq:
1237 	destroy_raw_packet_qp_rq(dev, rq);
1238 err_destroy_sq:
1239 	if (!qp->sq.wqe_cnt)
1240 		return err;
1241 	destroy_raw_packet_qp_sq(dev, sq);
1242 err_destroy_tis:
1243 	destroy_raw_packet_qp_tis(dev, sq);
1244 
1245 	return err;
1246 }
1247 
1248 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1249 				  struct mlx5_ib_qp *qp)
1250 {
1251 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1252 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1253 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1254 
1255 	if (qp->rq.wqe_cnt) {
1256 		destroy_raw_packet_qp_tir(dev, rq);
1257 		destroy_raw_packet_qp_rq(dev, rq);
1258 	}
1259 
1260 	if (qp->sq.wqe_cnt) {
1261 		destroy_raw_packet_qp_sq(dev, sq);
1262 		destroy_raw_packet_qp_tis(dev, sq);
1263 	}
1264 }
1265 
1266 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1267 				    struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1268 {
1269 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1270 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1271 
1272 	sq->sq = &qp->sq;
1273 	rq->rq = &qp->rq;
1274 	sq->doorbell = &qp->db;
1275 	rq->doorbell = &qp->db;
1276 }
1277 
1278 static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1279 {
1280 	mlx5_core_destroy_tir(dev->mdev, qp->rss_qp.tirn);
1281 }
1282 
1283 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1284 				 struct ib_pd *pd,
1285 				 struct ib_qp_init_attr *init_attr,
1286 				 struct ib_udata *udata)
1287 {
1288 	struct ib_uobject *uobj = pd->uobject;
1289 	struct ib_ucontext *ucontext = uobj->context;
1290 	struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1291 	struct mlx5_ib_create_qp_resp resp = {};
1292 	int inlen;
1293 	int err;
1294 	u32 *in;
1295 	void *tirc;
1296 	void *hfso;
1297 	u32 selected_fields = 0;
1298 	size_t min_resp_len;
1299 	u32 tdn = mucontext->tdn;
1300 	struct mlx5_ib_create_qp_rss ucmd = {};
1301 	size_t required_cmd_sz;
1302 
1303 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1304 		return -EOPNOTSUPP;
1305 
1306 	if (init_attr->create_flags || init_attr->send_cq)
1307 		return -EINVAL;
1308 
1309 	min_resp_len = offsetof(typeof(resp), uuar_index) + sizeof(resp.uuar_index);
1310 	if (udata->outlen < min_resp_len)
1311 		return -EINVAL;
1312 
1313 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) + sizeof(ucmd.reserved1);
1314 	if (udata->inlen < required_cmd_sz) {
1315 		mlx5_ib_dbg(dev, "invalid inlen\n");
1316 		return -EINVAL;
1317 	}
1318 
1319 	if (udata->inlen > sizeof(ucmd) &&
1320 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
1321 				 udata->inlen - sizeof(ucmd))) {
1322 		mlx5_ib_dbg(dev, "inlen is not supported\n");
1323 		return -EOPNOTSUPP;
1324 	}
1325 
1326 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
1327 		mlx5_ib_dbg(dev, "copy failed\n");
1328 		return -EFAULT;
1329 	}
1330 
1331 	if (ucmd.comp_mask) {
1332 		mlx5_ib_dbg(dev, "invalid comp mask\n");
1333 		return -EOPNOTSUPP;
1334 	}
1335 
1336 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)) || ucmd.reserved1) {
1337 		mlx5_ib_dbg(dev, "invalid reserved\n");
1338 		return -EOPNOTSUPP;
1339 	}
1340 
1341 	err = ib_copy_to_udata(udata, &resp, min_resp_len);
1342 	if (err) {
1343 		mlx5_ib_dbg(dev, "copy failed\n");
1344 		return -EINVAL;
1345 	}
1346 
1347 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1348 	in = mlx5_vzalloc(inlen);
1349 	if (!in)
1350 		return -ENOMEM;
1351 
1352 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1353 	MLX5_SET(tirc, tirc, disp_type,
1354 		 MLX5_TIRC_DISP_TYPE_INDIRECT);
1355 	MLX5_SET(tirc, tirc, indirect_table,
1356 		 init_attr->rwq_ind_tbl->ind_tbl_num);
1357 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1358 
1359 	hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1360 	switch (ucmd.rx_hash_function) {
1361 	case MLX5_RX_HASH_FUNC_TOEPLITZ:
1362 	{
1363 		void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1364 		size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1365 
1366 		if (len != ucmd.rx_key_len) {
1367 			err = -EINVAL;
1368 			goto err;
1369 		}
1370 
1371 		MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_TOEPLITZ);
1372 		MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
1373 		memcpy(rss_key, ucmd.rx_hash_key, len);
1374 		break;
1375 	}
1376 	default:
1377 		err = -EOPNOTSUPP;
1378 		goto err;
1379 	}
1380 
1381 	if (!ucmd.rx_hash_fields_mask) {
1382 		/* special case when this TIR serves as steering entry without hashing */
1383 		if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1384 			goto create_tir;
1385 		err = -EINVAL;
1386 		goto err;
1387 	}
1388 
1389 	if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1390 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1391 	     ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1392 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1393 		err = -EINVAL;
1394 		goto err;
1395 	}
1396 
1397 	/* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1398 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1399 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1400 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1401 			 MLX5_L3_PROT_TYPE_IPV4);
1402 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1403 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1404 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1405 			 MLX5_L3_PROT_TYPE_IPV6);
1406 
1407 	if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1408 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP)) &&
1409 	     ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1410 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))) {
1411 		err = -EINVAL;
1412 		goto err;
1413 	}
1414 
1415 	/* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1416 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1417 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1418 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1419 			 MLX5_L4_PROT_TYPE_TCP);
1420 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1421 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1422 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1423 			 MLX5_L4_PROT_TYPE_UDP);
1424 
1425 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1426 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1427 		selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1428 
1429 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1430 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1431 		selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1432 
1433 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1434 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1435 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1436 
1437 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1438 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1439 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1440 
1441 	MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1442 
1443 create_tir:
1444 	err = mlx5_core_create_tir(dev->mdev, in, inlen, &qp->rss_qp.tirn);
1445 
1446 	if (err)
1447 		goto err;
1448 
1449 	kvfree(in);
1450 	/* qpn is reserved for that QP */
1451 	qp->trans_qp.base.mqp.qpn = 0;
1452 	return 0;
1453 
1454 err:
1455 	kvfree(in);
1456 	return err;
1457 }
1458 
1459 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1460 			    struct ib_qp_init_attr *init_attr,
1461 			    struct ib_udata *udata, struct mlx5_ib_qp *qp)
1462 {
1463 	struct mlx5_ib_resources *devr = &dev->devr;
1464 	struct mlx5_core_dev *mdev = dev->mdev;
1465 	struct mlx5_ib_qp_base *base;
1466 	struct mlx5_ib_create_qp_resp resp;
1467 	struct mlx5_create_qp_mbox_in *in;
1468 	struct mlx5_ib_create_qp ucmd;
1469 	struct mlx5_ib_cq *send_cq;
1470 	struct mlx5_ib_cq *recv_cq;
1471 	unsigned long flags;
1472 	int inlen = sizeof(*in);
1473 	int err;
1474 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
1475 	void *qpc;
1476 
1477 	base = init_attr->qp_type == IB_QPT_RAW_PACKET ?
1478 	       &qp->raw_packet_qp.rq.base :
1479 	       &qp->trans_qp.base;
1480 
1481 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1482 		mlx5_ib_odp_create_qp(qp);
1483 
1484 	mutex_init(&qp->mutex);
1485 	spin_lock_init(&qp->sq.lock);
1486 	spin_lock_init(&qp->rq.lock);
1487 
1488 	if (init_attr->rwq_ind_tbl) {
1489 		if (!udata)
1490 			return -ENOSYS;
1491 
1492 		err = create_rss_raw_qp_tir(dev, qp, pd, init_attr, udata);
1493 		return err;
1494 	}
1495 
1496 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
1497 		if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
1498 			mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
1499 			return -EINVAL;
1500 		} else {
1501 			qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1502 		}
1503 	}
1504 
1505 	if (init_attr->create_flags &
1506 			(IB_QP_CREATE_CROSS_CHANNEL |
1507 			 IB_QP_CREATE_MANAGED_SEND |
1508 			 IB_QP_CREATE_MANAGED_RECV)) {
1509 		if (!MLX5_CAP_GEN(mdev, cd)) {
1510 			mlx5_ib_dbg(dev, "cross-channel isn't supported\n");
1511 			return -EINVAL;
1512 		}
1513 		if (init_attr->create_flags & IB_QP_CREATE_CROSS_CHANNEL)
1514 			qp->flags |= MLX5_IB_QP_CROSS_CHANNEL;
1515 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_SEND)
1516 			qp->flags |= MLX5_IB_QP_MANAGED_SEND;
1517 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_RECV)
1518 			qp->flags |= MLX5_IB_QP_MANAGED_RECV;
1519 	}
1520 
1521 	if (init_attr->qp_type == IB_QPT_UD &&
1522 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO))
1523 		if (!MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
1524 			mlx5_ib_dbg(dev, "ipoib UD lso qp isn't supported\n");
1525 			return -EOPNOTSUPP;
1526 		}
1527 
1528 	if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1529 		if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
1530 			mlx5_ib_dbg(dev, "Scatter FCS is supported only for Raw Packet QPs");
1531 			return -EOPNOTSUPP;
1532 		}
1533 		if (!MLX5_CAP_GEN(dev->mdev, eth_net_offloads) ||
1534 		    !MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
1535 			mlx5_ib_dbg(dev, "Scatter FCS isn't supported\n");
1536 			return -EOPNOTSUPP;
1537 		}
1538 		qp->flags |= MLX5_IB_QP_CAP_SCATTER_FCS;
1539 	}
1540 
1541 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1542 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
1543 
1544 	if (pd && pd->uobject) {
1545 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
1546 			mlx5_ib_dbg(dev, "copy failed\n");
1547 			return -EFAULT;
1548 		}
1549 
1550 		err = get_qp_user_index(to_mucontext(pd->uobject->context),
1551 					&ucmd, udata->inlen, &uidx);
1552 		if (err)
1553 			return err;
1554 
1555 		qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
1556 		qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
1557 	} else {
1558 		qp->wq_sig = !!wq_signature;
1559 	}
1560 
1561 	qp->has_rq = qp_has_rq(init_attr);
1562 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
1563 			  qp, (pd && pd->uobject) ? &ucmd : NULL);
1564 	if (err) {
1565 		mlx5_ib_dbg(dev, "err %d\n", err);
1566 		return err;
1567 	}
1568 
1569 	if (pd) {
1570 		if (pd->uobject) {
1571 			__u32 max_wqes =
1572 				1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1573 			mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
1574 			if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
1575 			    ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
1576 				mlx5_ib_dbg(dev, "invalid rq params\n");
1577 				return -EINVAL;
1578 			}
1579 			if (ucmd.sq_wqe_count > max_wqes) {
1580 				mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
1581 					    ucmd.sq_wqe_count, max_wqes);
1582 				return -EINVAL;
1583 			}
1584 			if (init_attr->create_flags &
1585 			    mlx5_ib_create_qp_sqpn_qp1()) {
1586 				mlx5_ib_dbg(dev, "user-space is not allowed to create UD QPs spoofing as QP1\n");
1587 				return -EINVAL;
1588 			}
1589 			err = create_user_qp(dev, pd, qp, udata, init_attr, &in,
1590 					     &resp, &inlen, base);
1591 			if (err)
1592 				mlx5_ib_dbg(dev, "err %d\n", err);
1593 		} else {
1594 			err = create_kernel_qp(dev, init_attr, qp, &in, &inlen,
1595 					       base);
1596 			if (err)
1597 				mlx5_ib_dbg(dev, "err %d\n", err);
1598 		}
1599 
1600 		if (err)
1601 			return err;
1602 	} else {
1603 		in = mlx5_vzalloc(sizeof(*in));
1604 		if (!in)
1605 			return -ENOMEM;
1606 
1607 		qp->create_type = MLX5_QP_EMPTY;
1608 	}
1609 
1610 	if (is_sqp(init_attr->qp_type))
1611 		qp->port = init_attr->port_num;
1612 
1613 	in->ctx.flags = cpu_to_be32(to_mlx5_st(init_attr->qp_type) << 16 |
1614 				    MLX5_QP_PM_MIGRATED << 11);
1615 
1616 	if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
1617 		in->ctx.flags_pd = cpu_to_be32(to_mpd(pd ? pd : devr->p0)->pdn);
1618 	else
1619 		in->ctx.flags_pd = cpu_to_be32(MLX5_QP_LAT_SENSITIVE);
1620 
1621 	if (qp->wq_sig)
1622 		in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_ENABLE_SIG);
1623 
1624 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
1625 		in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_BLOCK_MCAST);
1626 
1627 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
1628 		in->ctx.params2 |= cpu_to_be32(MLX5_QP_BIT_CC_MASTER);
1629 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
1630 		in->ctx.params2 |= cpu_to_be32(MLX5_QP_BIT_CC_SLAVE_SEND);
1631 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
1632 		in->ctx.params2 |= cpu_to_be32(MLX5_QP_BIT_CC_SLAVE_RECV);
1633 
1634 	if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
1635 		int rcqe_sz;
1636 		int scqe_sz;
1637 
1638 		rcqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->recv_cq);
1639 		scqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->send_cq);
1640 
1641 		if (rcqe_sz == 128)
1642 			in->ctx.cs_res = MLX5_RES_SCAT_DATA64_CQE;
1643 		else
1644 			in->ctx.cs_res = MLX5_RES_SCAT_DATA32_CQE;
1645 
1646 		if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) {
1647 			if (scqe_sz == 128)
1648 				in->ctx.cs_req = MLX5_REQ_SCAT_DATA64_CQE;
1649 			else
1650 				in->ctx.cs_req = MLX5_REQ_SCAT_DATA32_CQE;
1651 		}
1652 	}
1653 
1654 	if (qp->rq.wqe_cnt) {
1655 		in->ctx.rq_size_stride = (qp->rq.wqe_shift - 4);
1656 		in->ctx.rq_size_stride |= ilog2(qp->rq.wqe_cnt) << 3;
1657 	}
1658 
1659 	in->ctx.rq_type_srqn = get_rx_type(qp, init_attr);
1660 
1661 	if (qp->sq.wqe_cnt)
1662 		in->ctx.sq_crq_size |= cpu_to_be16(ilog2(qp->sq.wqe_cnt) << 11);
1663 	else
1664 		in->ctx.sq_crq_size |= cpu_to_be16(0x8000);
1665 
1666 	/* Set default resources */
1667 	switch (init_attr->qp_type) {
1668 	case IB_QPT_XRC_TGT:
1669 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
1670 		in->ctx.cqn_send = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
1671 		in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
1672 		in->ctx.xrcd = cpu_to_be32(to_mxrcd(init_attr->xrcd)->xrcdn);
1673 		break;
1674 	case IB_QPT_XRC_INI:
1675 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
1676 		in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
1677 		in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
1678 		break;
1679 	default:
1680 		if (init_attr->srq) {
1681 			in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x0)->xrcdn);
1682 			in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(init_attr->srq)->msrq.srqn);
1683 		} else {
1684 			in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
1685 			in->ctx.rq_type_srqn |=
1686 				cpu_to_be32(to_msrq(devr->s1)->msrq.srqn);
1687 		}
1688 	}
1689 
1690 	if (init_attr->send_cq)
1691 		in->ctx.cqn_send = cpu_to_be32(to_mcq(init_attr->send_cq)->mcq.cqn);
1692 
1693 	if (init_attr->recv_cq)
1694 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(init_attr->recv_cq)->mcq.cqn);
1695 
1696 	in->ctx.db_rec_addr = cpu_to_be64(qp->db.dma);
1697 
1698 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1) {
1699 		qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1700 		/* 0xffffff means we ask to work with cqe version 0 */
1701 		MLX5_SET(qpc, qpc, user_index, uidx);
1702 	}
1703 	/* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
1704 	if (init_attr->qp_type == IB_QPT_UD &&
1705 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)) {
1706 		qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1707 		MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
1708 		qp->flags |= MLX5_IB_QP_LSO;
1709 	}
1710 
1711 	if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
1712 		qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd.sq_buf_addr;
1713 		raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
1714 		err = create_raw_packet_qp(dev, qp, in, pd);
1715 	} else {
1716 		err = mlx5_core_create_qp(dev->mdev, &base->mqp, in, inlen);
1717 	}
1718 
1719 	if (err) {
1720 		mlx5_ib_dbg(dev, "create qp failed\n");
1721 		goto err_create;
1722 	}
1723 
1724 	kvfree(in);
1725 
1726 	base->container_mibqp = qp;
1727 	base->mqp.event = mlx5_ib_qp_event;
1728 
1729 	get_cqs(init_attr->qp_type, init_attr->send_cq, init_attr->recv_cq,
1730 		&send_cq, &recv_cq);
1731 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1732 	mlx5_ib_lock_cqs(send_cq, recv_cq);
1733 	/* Maintain device to QPs access, needed for further handling via reset
1734 	 * flow
1735 	 */
1736 	list_add_tail(&qp->qps_list, &dev->qp_list);
1737 	/* Maintain CQ to QPs access, needed for further handling via reset flow
1738 	 */
1739 	if (send_cq)
1740 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
1741 	if (recv_cq)
1742 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
1743 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
1744 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1745 
1746 	return 0;
1747 
1748 err_create:
1749 	if (qp->create_type == MLX5_QP_USER)
1750 		destroy_qp_user(pd, qp, base);
1751 	else if (qp->create_type == MLX5_QP_KERNEL)
1752 		destroy_qp_kernel(dev, qp);
1753 
1754 	kvfree(in);
1755 	return err;
1756 }
1757 
1758 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1759 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1760 {
1761 	if (send_cq) {
1762 		if (recv_cq) {
1763 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1764 				spin_lock(&send_cq->lock);
1765 				spin_lock_nested(&recv_cq->lock,
1766 						 SINGLE_DEPTH_NESTING);
1767 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1768 				spin_lock(&send_cq->lock);
1769 				__acquire(&recv_cq->lock);
1770 			} else {
1771 				spin_lock(&recv_cq->lock);
1772 				spin_lock_nested(&send_cq->lock,
1773 						 SINGLE_DEPTH_NESTING);
1774 			}
1775 		} else {
1776 			spin_lock(&send_cq->lock);
1777 			__acquire(&recv_cq->lock);
1778 		}
1779 	} else if (recv_cq) {
1780 		spin_lock(&recv_cq->lock);
1781 		__acquire(&send_cq->lock);
1782 	} else {
1783 		__acquire(&send_cq->lock);
1784 		__acquire(&recv_cq->lock);
1785 	}
1786 }
1787 
1788 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1789 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1790 {
1791 	if (send_cq) {
1792 		if (recv_cq) {
1793 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1794 				spin_unlock(&recv_cq->lock);
1795 				spin_unlock(&send_cq->lock);
1796 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1797 				__release(&recv_cq->lock);
1798 				spin_unlock(&send_cq->lock);
1799 			} else {
1800 				spin_unlock(&send_cq->lock);
1801 				spin_unlock(&recv_cq->lock);
1802 			}
1803 		} else {
1804 			__release(&recv_cq->lock);
1805 			spin_unlock(&send_cq->lock);
1806 		}
1807 	} else if (recv_cq) {
1808 		__release(&send_cq->lock);
1809 		spin_unlock(&recv_cq->lock);
1810 	} else {
1811 		__release(&recv_cq->lock);
1812 		__release(&send_cq->lock);
1813 	}
1814 }
1815 
1816 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
1817 {
1818 	return to_mpd(qp->ibqp.pd);
1819 }
1820 
1821 static void get_cqs(enum ib_qp_type qp_type,
1822 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
1823 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
1824 {
1825 	switch (qp_type) {
1826 	case IB_QPT_XRC_TGT:
1827 		*send_cq = NULL;
1828 		*recv_cq = NULL;
1829 		break;
1830 	case MLX5_IB_QPT_REG_UMR:
1831 	case IB_QPT_XRC_INI:
1832 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
1833 		*recv_cq = NULL;
1834 		break;
1835 
1836 	case IB_QPT_SMI:
1837 	case MLX5_IB_QPT_HW_GSI:
1838 	case IB_QPT_RC:
1839 	case IB_QPT_UC:
1840 	case IB_QPT_UD:
1841 	case IB_QPT_RAW_IPV6:
1842 	case IB_QPT_RAW_ETHERTYPE:
1843 	case IB_QPT_RAW_PACKET:
1844 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
1845 		*recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
1846 		break;
1847 
1848 	case IB_QPT_MAX:
1849 	default:
1850 		*send_cq = NULL;
1851 		*recv_cq = NULL;
1852 		break;
1853 	}
1854 }
1855 
1856 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1857 				u16 operation);
1858 
1859 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1860 {
1861 	struct mlx5_ib_cq *send_cq, *recv_cq;
1862 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
1863 	struct mlx5_modify_qp_mbox_in *in;
1864 	unsigned long flags;
1865 	int err;
1866 
1867 	if (qp->ibqp.rwq_ind_tbl) {
1868 		destroy_rss_raw_qp_tir(dev, qp);
1869 		return;
1870 	}
1871 
1872 	base = qp->ibqp.qp_type == IB_QPT_RAW_PACKET ?
1873 	       &qp->raw_packet_qp.rq.base :
1874 	       &qp->trans_qp.base;
1875 
1876 	in = kzalloc(sizeof(*in), GFP_KERNEL);
1877 	if (!in)
1878 		return;
1879 
1880 	if (qp->state != IB_QPS_RESET) {
1881 		if (qp->ibqp.qp_type != IB_QPT_RAW_PACKET) {
1882 			mlx5_ib_qp_disable_pagefaults(qp);
1883 			err = mlx5_core_qp_modify(dev->mdev,
1884 						  MLX5_CMD_OP_2RST_QP, in, 0,
1885 						  &base->mqp);
1886 		} else {
1887 			err = modify_raw_packet_qp(dev, qp,
1888 						   MLX5_CMD_OP_2RST_QP);
1889 		}
1890 		if (err)
1891 			mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
1892 				     base->mqp.qpn);
1893 	}
1894 
1895 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
1896 		&send_cq, &recv_cq);
1897 
1898 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1899 	mlx5_ib_lock_cqs(send_cq, recv_cq);
1900 	/* del from lists under both locks above to protect reset flow paths */
1901 	list_del(&qp->qps_list);
1902 	if (send_cq)
1903 		list_del(&qp->cq_send_list);
1904 
1905 	if (recv_cq)
1906 		list_del(&qp->cq_recv_list);
1907 
1908 	if (qp->create_type == MLX5_QP_KERNEL) {
1909 		__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
1910 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1911 		if (send_cq != recv_cq)
1912 			__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
1913 					   NULL);
1914 	}
1915 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
1916 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1917 
1918 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1919 		destroy_raw_packet_qp(dev, qp);
1920 	} else {
1921 		err = mlx5_core_destroy_qp(dev->mdev, &base->mqp);
1922 		if (err)
1923 			mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
1924 				     base->mqp.qpn);
1925 	}
1926 
1927 	kfree(in);
1928 
1929 	if (qp->create_type == MLX5_QP_KERNEL)
1930 		destroy_qp_kernel(dev, qp);
1931 	else if (qp->create_type == MLX5_QP_USER)
1932 		destroy_qp_user(&get_pd(qp)->ibpd, qp, base);
1933 }
1934 
1935 static const char *ib_qp_type_str(enum ib_qp_type type)
1936 {
1937 	switch (type) {
1938 	case IB_QPT_SMI:
1939 		return "IB_QPT_SMI";
1940 	case IB_QPT_GSI:
1941 		return "IB_QPT_GSI";
1942 	case IB_QPT_RC:
1943 		return "IB_QPT_RC";
1944 	case IB_QPT_UC:
1945 		return "IB_QPT_UC";
1946 	case IB_QPT_UD:
1947 		return "IB_QPT_UD";
1948 	case IB_QPT_RAW_IPV6:
1949 		return "IB_QPT_RAW_IPV6";
1950 	case IB_QPT_RAW_ETHERTYPE:
1951 		return "IB_QPT_RAW_ETHERTYPE";
1952 	case IB_QPT_XRC_INI:
1953 		return "IB_QPT_XRC_INI";
1954 	case IB_QPT_XRC_TGT:
1955 		return "IB_QPT_XRC_TGT";
1956 	case IB_QPT_RAW_PACKET:
1957 		return "IB_QPT_RAW_PACKET";
1958 	case MLX5_IB_QPT_REG_UMR:
1959 		return "MLX5_IB_QPT_REG_UMR";
1960 	case IB_QPT_MAX:
1961 	default:
1962 		return "Invalid QP type";
1963 	}
1964 }
1965 
1966 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
1967 				struct ib_qp_init_attr *init_attr,
1968 				struct ib_udata *udata)
1969 {
1970 	struct mlx5_ib_dev *dev;
1971 	struct mlx5_ib_qp *qp;
1972 	u16 xrcdn = 0;
1973 	int err;
1974 
1975 	if (pd) {
1976 		dev = to_mdev(pd->device);
1977 
1978 		if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
1979 			if (!pd->uobject) {
1980 				mlx5_ib_dbg(dev, "Raw Packet QP is not supported for kernel consumers\n");
1981 				return ERR_PTR(-EINVAL);
1982 			} else if (!to_mucontext(pd->uobject->context)->cqe_version) {
1983 				mlx5_ib_dbg(dev, "Raw Packet QP is only supported for CQE version > 0\n");
1984 				return ERR_PTR(-EINVAL);
1985 			}
1986 		}
1987 	} else {
1988 		/* being cautious here */
1989 		if (init_attr->qp_type != IB_QPT_XRC_TGT &&
1990 		    init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
1991 			pr_warn("%s: no PD for transport %s\n", __func__,
1992 				ib_qp_type_str(init_attr->qp_type));
1993 			return ERR_PTR(-EINVAL);
1994 		}
1995 		dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
1996 	}
1997 
1998 	switch (init_attr->qp_type) {
1999 	case IB_QPT_XRC_TGT:
2000 	case IB_QPT_XRC_INI:
2001 		if (!MLX5_CAP_GEN(dev->mdev, xrc)) {
2002 			mlx5_ib_dbg(dev, "XRC not supported\n");
2003 			return ERR_PTR(-ENOSYS);
2004 		}
2005 		init_attr->recv_cq = NULL;
2006 		if (init_attr->qp_type == IB_QPT_XRC_TGT) {
2007 			xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
2008 			init_attr->send_cq = NULL;
2009 		}
2010 
2011 		/* fall through */
2012 	case IB_QPT_RAW_PACKET:
2013 	case IB_QPT_RC:
2014 	case IB_QPT_UC:
2015 	case IB_QPT_UD:
2016 	case IB_QPT_SMI:
2017 	case MLX5_IB_QPT_HW_GSI:
2018 	case MLX5_IB_QPT_REG_UMR:
2019 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2020 		if (!qp)
2021 			return ERR_PTR(-ENOMEM);
2022 
2023 		err = create_qp_common(dev, pd, init_attr, udata, qp);
2024 		if (err) {
2025 			mlx5_ib_dbg(dev, "create_qp_common failed\n");
2026 			kfree(qp);
2027 			return ERR_PTR(err);
2028 		}
2029 
2030 		if (is_qp0(init_attr->qp_type))
2031 			qp->ibqp.qp_num = 0;
2032 		else if (is_qp1(init_attr->qp_type))
2033 			qp->ibqp.qp_num = 1;
2034 		else
2035 			qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
2036 
2037 		mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
2038 			    qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
2039 			    to_mcq(init_attr->recv_cq)->mcq.cqn,
2040 			    to_mcq(init_attr->send_cq)->mcq.cqn);
2041 
2042 		qp->trans_qp.xrcdn = xrcdn;
2043 
2044 		break;
2045 
2046 	case IB_QPT_GSI:
2047 		return mlx5_ib_gsi_create_qp(pd, init_attr);
2048 
2049 	case IB_QPT_RAW_IPV6:
2050 	case IB_QPT_RAW_ETHERTYPE:
2051 	case IB_QPT_MAX:
2052 	default:
2053 		mlx5_ib_dbg(dev, "unsupported qp type %d\n",
2054 			    init_attr->qp_type);
2055 		/* Don't support raw QPs */
2056 		return ERR_PTR(-EINVAL);
2057 	}
2058 
2059 	return &qp->ibqp;
2060 }
2061 
2062 int mlx5_ib_destroy_qp(struct ib_qp *qp)
2063 {
2064 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2065 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2066 
2067 	if (unlikely(qp->qp_type == IB_QPT_GSI))
2068 		return mlx5_ib_gsi_destroy_qp(qp);
2069 
2070 	destroy_qp_common(dev, mqp);
2071 
2072 	kfree(mqp);
2073 
2074 	return 0;
2075 }
2076 
2077 static __be32 to_mlx5_access_flags(struct mlx5_ib_qp *qp, const struct ib_qp_attr *attr,
2078 				   int attr_mask)
2079 {
2080 	u32 hw_access_flags = 0;
2081 	u8 dest_rd_atomic;
2082 	u32 access_flags;
2083 
2084 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2085 		dest_rd_atomic = attr->max_dest_rd_atomic;
2086 	else
2087 		dest_rd_atomic = qp->trans_qp.resp_depth;
2088 
2089 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2090 		access_flags = attr->qp_access_flags;
2091 	else
2092 		access_flags = qp->trans_qp.atomic_rd_en;
2093 
2094 	if (!dest_rd_atomic)
2095 		access_flags &= IB_ACCESS_REMOTE_WRITE;
2096 
2097 	if (access_flags & IB_ACCESS_REMOTE_READ)
2098 		hw_access_flags |= MLX5_QP_BIT_RRE;
2099 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
2100 		hw_access_flags |= (MLX5_QP_BIT_RAE | MLX5_ATOMIC_MODE_CX);
2101 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
2102 		hw_access_flags |= MLX5_QP_BIT_RWE;
2103 
2104 	return cpu_to_be32(hw_access_flags);
2105 }
2106 
2107 enum {
2108 	MLX5_PATH_FLAG_FL	= 1 << 0,
2109 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
2110 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
2111 };
2112 
2113 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
2114 {
2115 	if (rate == IB_RATE_PORT_CURRENT) {
2116 		return 0;
2117 	} else if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_300_GBPS) {
2118 		return -EINVAL;
2119 	} else {
2120 		while (rate != IB_RATE_2_5_GBPS &&
2121 		       !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
2122 			 MLX5_CAP_GEN(dev->mdev, stat_rate_support)))
2123 			--rate;
2124 	}
2125 
2126 	return rate + MLX5_STAT_RATE_OFFSET;
2127 }
2128 
2129 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
2130 				      struct mlx5_ib_sq *sq, u8 sl)
2131 {
2132 	void *in;
2133 	void *tisc;
2134 	int inlen;
2135 	int err;
2136 
2137 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2138 	in = mlx5_vzalloc(inlen);
2139 	if (!in)
2140 		return -ENOMEM;
2141 
2142 	MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
2143 
2144 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2145 	MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
2146 
2147 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2148 
2149 	kvfree(in);
2150 
2151 	return err;
2152 }
2153 
2154 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2155 			 const struct ib_ah_attr *ah,
2156 			 struct mlx5_qp_path *path, u8 port, int attr_mask,
2157 			 u32 path_flags, const struct ib_qp_attr *attr,
2158 			 bool alt)
2159 {
2160 	enum rdma_link_layer ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2161 	int err;
2162 
2163 	if (attr_mask & IB_QP_PKEY_INDEX)
2164 		path->pkey_index = cpu_to_be16(alt ? attr->alt_pkey_index :
2165 						     attr->pkey_index);
2166 
2167 	if (ah->ah_flags & IB_AH_GRH) {
2168 		if (ah->grh.sgid_index >=
2169 		    dev->mdev->port_caps[port - 1].gid_table_len) {
2170 			pr_err("sgid_index (%u) too large. max is %d\n",
2171 			       ah->grh.sgid_index,
2172 			       dev->mdev->port_caps[port - 1].gid_table_len);
2173 			return -EINVAL;
2174 		}
2175 	}
2176 
2177 	if (ll == IB_LINK_LAYER_ETHERNET) {
2178 		if (!(ah->ah_flags & IB_AH_GRH))
2179 			return -EINVAL;
2180 		memcpy(path->rmac, ah->dmac, sizeof(ah->dmac));
2181 		path->udp_sport = mlx5_get_roce_udp_sport(dev, port,
2182 							  ah->grh.sgid_index);
2183 		path->dci_cfi_prio_sl = (ah->sl & 0x7) << 4;
2184 	} else {
2185 		path->fl_free_ar = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
2186 		path->fl_free_ar |=
2187 			(path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x40 : 0;
2188 		path->rlid = cpu_to_be16(ah->dlid);
2189 		path->grh_mlid = ah->src_path_bits & 0x7f;
2190 		if (ah->ah_flags & IB_AH_GRH)
2191 			path->grh_mlid	|= 1 << 7;
2192 		path->dci_cfi_prio_sl = ah->sl & 0xf;
2193 	}
2194 
2195 	if (ah->ah_flags & IB_AH_GRH) {
2196 		path->mgid_index = ah->grh.sgid_index;
2197 		path->hop_limit  = ah->grh.hop_limit;
2198 		path->tclass_flowlabel =
2199 			cpu_to_be32((ah->grh.traffic_class << 20) |
2200 				    (ah->grh.flow_label));
2201 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
2202 	}
2203 
2204 	err = ib_rate_to_mlx5(dev, ah->static_rate);
2205 	if (err < 0)
2206 		return err;
2207 	path->static_rate = err;
2208 	path->port = port;
2209 
2210 	if (attr_mask & IB_QP_TIMEOUT)
2211 		path->ackto_lt = (alt ? attr->alt_timeout : attr->timeout) << 3;
2212 
2213 	if ((qp->ibqp.qp_type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
2214 		return modify_raw_packet_eth_prio(dev->mdev,
2215 						  &qp->raw_packet_qp.sq,
2216 						  ah->sl & 0xf);
2217 
2218 	return 0;
2219 }
2220 
2221 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
2222 	[MLX5_QP_STATE_INIT] = {
2223 		[MLX5_QP_STATE_INIT] = {
2224 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
2225 					  MLX5_QP_OPTPAR_RAE		|
2226 					  MLX5_QP_OPTPAR_RWE		|
2227 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
2228 					  MLX5_QP_OPTPAR_PRI_PORT,
2229 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
2230 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
2231 					  MLX5_QP_OPTPAR_PRI_PORT,
2232 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
2233 					  MLX5_QP_OPTPAR_Q_KEY		|
2234 					  MLX5_QP_OPTPAR_PRI_PORT,
2235 		},
2236 		[MLX5_QP_STATE_RTR] = {
2237 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
2238 					  MLX5_QP_OPTPAR_RRE            |
2239 					  MLX5_QP_OPTPAR_RAE            |
2240 					  MLX5_QP_OPTPAR_RWE            |
2241 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2242 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
2243 					  MLX5_QP_OPTPAR_RWE            |
2244 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2245 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
2246 					  MLX5_QP_OPTPAR_Q_KEY,
2247 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
2248 					   MLX5_QP_OPTPAR_Q_KEY,
2249 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2250 					  MLX5_QP_OPTPAR_RRE            |
2251 					  MLX5_QP_OPTPAR_RAE            |
2252 					  MLX5_QP_OPTPAR_RWE            |
2253 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2254 		},
2255 	},
2256 	[MLX5_QP_STATE_RTR] = {
2257 		[MLX5_QP_STATE_RTS] = {
2258 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
2259 					  MLX5_QP_OPTPAR_RRE		|
2260 					  MLX5_QP_OPTPAR_RAE		|
2261 					  MLX5_QP_OPTPAR_RWE		|
2262 					  MLX5_QP_OPTPAR_PM_STATE	|
2263 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
2264 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
2265 					  MLX5_QP_OPTPAR_RWE		|
2266 					  MLX5_QP_OPTPAR_PM_STATE,
2267 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
2268 		},
2269 	},
2270 	[MLX5_QP_STATE_RTS] = {
2271 		[MLX5_QP_STATE_RTS] = {
2272 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
2273 					  MLX5_QP_OPTPAR_RAE		|
2274 					  MLX5_QP_OPTPAR_RWE		|
2275 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
2276 					  MLX5_QP_OPTPAR_PM_STATE	|
2277 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
2278 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
2279 					  MLX5_QP_OPTPAR_PM_STATE	|
2280 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
2281 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
2282 					  MLX5_QP_OPTPAR_SRQN		|
2283 					  MLX5_QP_OPTPAR_CQN_RCV,
2284 		},
2285 	},
2286 	[MLX5_QP_STATE_SQER] = {
2287 		[MLX5_QP_STATE_RTS] = {
2288 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
2289 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
2290 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
2291 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
2292 					   MLX5_QP_OPTPAR_RWE		|
2293 					   MLX5_QP_OPTPAR_RAE		|
2294 					   MLX5_QP_OPTPAR_RRE,
2295 		},
2296 	},
2297 };
2298 
2299 static int ib_nr_to_mlx5_nr(int ib_mask)
2300 {
2301 	switch (ib_mask) {
2302 	case IB_QP_STATE:
2303 		return 0;
2304 	case IB_QP_CUR_STATE:
2305 		return 0;
2306 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
2307 		return 0;
2308 	case IB_QP_ACCESS_FLAGS:
2309 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
2310 			MLX5_QP_OPTPAR_RAE;
2311 	case IB_QP_PKEY_INDEX:
2312 		return MLX5_QP_OPTPAR_PKEY_INDEX;
2313 	case IB_QP_PORT:
2314 		return MLX5_QP_OPTPAR_PRI_PORT;
2315 	case IB_QP_QKEY:
2316 		return MLX5_QP_OPTPAR_Q_KEY;
2317 	case IB_QP_AV:
2318 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
2319 			MLX5_QP_OPTPAR_PRI_PORT;
2320 	case IB_QP_PATH_MTU:
2321 		return 0;
2322 	case IB_QP_TIMEOUT:
2323 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
2324 	case IB_QP_RETRY_CNT:
2325 		return MLX5_QP_OPTPAR_RETRY_COUNT;
2326 	case IB_QP_RNR_RETRY:
2327 		return MLX5_QP_OPTPAR_RNR_RETRY;
2328 	case IB_QP_RQ_PSN:
2329 		return 0;
2330 	case IB_QP_MAX_QP_RD_ATOMIC:
2331 		return MLX5_QP_OPTPAR_SRA_MAX;
2332 	case IB_QP_ALT_PATH:
2333 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
2334 	case IB_QP_MIN_RNR_TIMER:
2335 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
2336 	case IB_QP_SQ_PSN:
2337 		return 0;
2338 	case IB_QP_MAX_DEST_RD_ATOMIC:
2339 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
2340 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
2341 	case IB_QP_PATH_MIG_STATE:
2342 		return MLX5_QP_OPTPAR_PM_STATE;
2343 	case IB_QP_CAP:
2344 		return 0;
2345 	case IB_QP_DEST_QPN:
2346 		return 0;
2347 	}
2348 	return 0;
2349 }
2350 
2351 static int ib_mask_to_mlx5_opt(int ib_mask)
2352 {
2353 	int result = 0;
2354 	int i;
2355 
2356 	for (i = 0; i < 8 * sizeof(int); i++) {
2357 		if ((1 << i) & ib_mask)
2358 			result |= ib_nr_to_mlx5_nr(1 << i);
2359 	}
2360 
2361 	return result;
2362 }
2363 
2364 static int modify_raw_packet_qp_rq(struct mlx5_core_dev *dev,
2365 				   struct mlx5_ib_rq *rq, int new_state)
2366 {
2367 	void *in;
2368 	void *rqc;
2369 	int inlen;
2370 	int err;
2371 
2372 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
2373 	in = mlx5_vzalloc(inlen);
2374 	if (!in)
2375 		return -ENOMEM;
2376 
2377 	MLX5_SET(modify_rq_in, in, rq_state, rq->state);
2378 
2379 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
2380 	MLX5_SET(rqc, rqc, state, new_state);
2381 
2382 	err = mlx5_core_modify_rq(dev, rq->base.mqp.qpn, in, inlen);
2383 	if (err)
2384 		goto out;
2385 
2386 	rq->state = new_state;
2387 
2388 out:
2389 	kvfree(in);
2390 	return err;
2391 }
2392 
2393 static int modify_raw_packet_qp_sq(struct mlx5_core_dev *dev,
2394 				   struct mlx5_ib_sq *sq, int new_state)
2395 {
2396 	void *in;
2397 	void *sqc;
2398 	int inlen;
2399 	int err;
2400 
2401 	inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
2402 	in = mlx5_vzalloc(inlen);
2403 	if (!in)
2404 		return -ENOMEM;
2405 
2406 	MLX5_SET(modify_sq_in, in, sq_state, sq->state);
2407 
2408 	sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
2409 	MLX5_SET(sqc, sqc, state, new_state);
2410 
2411 	err = mlx5_core_modify_sq(dev, sq->base.mqp.qpn, in, inlen);
2412 	if (err)
2413 		goto out;
2414 
2415 	sq->state = new_state;
2416 
2417 out:
2418 	kvfree(in);
2419 	return err;
2420 }
2421 
2422 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2423 				u16 operation)
2424 {
2425 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
2426 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
2427 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
2428 	int rq_state;
2429 	int sq_state;
2430 	int err;
2431 
2432 	switch (operation) {
2433 	case MLX5_CMD_OP_RST2INIT_QP:
2434 		rq_state = MLX5_RQC_STATE_RDY;
2435 		sq_state = MLX5_SQC_STATE_RDY;
2436 		break;
2437 	case MLX5_CMD_OP_2ERR_QP:
2438 		rq_state = MLX5_RQC_STATE_ERR;
2439 		sq_state = MLX5_SQC_STATE_ERR;
2440 		break;
2441 	case MLX5_CMD_OP_2RST_QP:
2442 		rq_state = MLX5_RQC_STATE_RST;
2443 		sq_state = MLX5_SQC_STATE_RST;
2444 		break;
2445 	case MLX5_CMD_OP_INIT2INIT_QP:
2446 	case MLX5_CMD_OP_INIT2RTR_QP:
2447 	case MLX5_CMD_OP_RTR2RTS_QP:
2448 	case MLX5_CMD_OP_RTS2RTS_QP:
2449 		/* Nothing to do here... */
2450 		return 0;
2451 	default:
2452 		WARN_ON(1);
2453 		return -EINVAL;
2454 	}
2455 
2456 	if (qp->rq.wqe_cnt) {
2457 		err =  modify_raw_packet_qp_rq(dev->mdev, rq, rq_state);
2458 		if (err)
2459 			return err;
2460 	}
2461 
2462 	if (qp->sq.wqe_cnt)
2463 		return modify_raw_packet_qp_sq(dev->mdev, sq, sq_state);
2464 
2465 	return 0;
2466 }
2467 
2468 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
2469 			       const struct ib_qp_attr *attr, int attr_mask,
2470 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
2471 {
2472 	static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
2473 		[MLX5_QP_STATE_RST] = {
2474 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2475 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2476 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_RST2INIT_QP,
2477 		},
2478 		[MLX5_QP_STATE_INIT]  = {
2479 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2480 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2481 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_INIT2INIT_QP,
2482 			[MLX5_QP_STATE_RTR]	= MLX5_CMD_OP_INIT2RTR_QP,
2483 		},
2484 		[MLX5_QP_STATE_RTR]   = {
2485 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2486 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2487 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTR2RTS_QP,
2488 		},
2489 		[MLX5_QP_STATE_RTS]   = {
2490 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2491 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2492 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTS2RTS_QP,
2493 		},
2494 		[MLX5_QP_STATE_SQD] = {
2495 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2496 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2497 		},
2498 		[MLX5_QP_STATE_SQER] = {
2499 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2500 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2501 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQERR2RTS_QP,
2502 		},
2503 		[MLX5_QP_STATE_ERR] = {
2504 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2505 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2506 		}
2507 	};
2508 
2509 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2510 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2511 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
2512 	struct mlx5_ib_cq *send_cq, *recv_cq;
2513 	struct mlx5_qp_context *context;
2514 	struct mlx5_modify_qp_mbox_in *in;
2515 	struct mlx5_ib_pd *pd;
2516 	enum mlx5_qp_state mlx5_cur, mlx5_new;
2517 	enum mlx5_qp_optpar optpar;
2518 	int sqd_event;
2519 	int mlx5_st;
2520 	int err;
2521 	u16 op;
2522 
2523 	in = kzalloc(sizeof(*in), GFP_KERNEL);
2524 	if (!in)
2525 		return -ENOMEM;
2526 
2527 	context = &in->ctx;
2528 	err = to_mlx5_st(ibqp->qp_type);
2529 	if (err < 0) {
2530 		mlx5_ib_dbg(dev, "unsupported qp type %d\n", ibqp->qp_type);
2531 		goto out;
2532 	}
2533 
2534 	context->flags = cpu_to_be32(err << 16);
2535 
2536 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
2537 		context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2538 	} else {
2539 		switch (attr->path_mig_state) {
2540 		case IB_MIG_MIGRATED:
2541 			context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2542 			break;
2543 		case IB_MIG_REARM:
2544 			context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
2545 			break;
2546 		case IB_MIG_ARMED:
2547 			context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
2548 			break;
2549 		}
2550 	}
2551 
2552 	if (is_sqp(ibqp->qp_type)) {
2553 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
2554 	} else if (ibqp->qp_type == IB_QPT_UD ||
2555 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
2556 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
2557 	} else if (attr_mask & IB_QP_PATH_MTU) {
2558 		if (attr->path_mtu < IB_MTU_256 ||
2559 		    attr->path_mtu > IB_MTU_4096) {
2560 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
2561 			err = -EINVAL;
2562 			goto out;
2563 		}
2564 		context->mtu_msgmax = (attr->path_mtu << 5) |
2565 				      (u8)MLX5_CAP_GEN(dev->mdev, log_max_msg);
2566 	}
2567 
2568 	if (attr_mask & IB_QP_DEST_QPN)
2569 		context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
2570 
2571 	if (attr_mask & IB_QP_PKEY_INDEX)
2572 		context->pri_path.pkey_index = cpu_to_be16(attr->pkey_index);
2573 
2574 	/* todo implement counter_index functionality */
2575 
2576 	if (is_sqp(ibqp->qp_type))
2577 		context->pri_path.port = qp->port;
2578 
2579 	if (attr_mask & IB_QP_PORT)
2580 		context->pri_path.port = attr->port_num;
2581 
2582 	if (attr_mask & IB_QP_AV) {
2583 		err = mlx5_set_path(dev, qp, &attr->ah_attr, &context->pri_path,
2584 				    attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
2585 				    attr_mask, 0, attr, false);
2586 		if (err)
2587 			goto out;
2588 	}
2589 
2590 	if (attr_mask & IB_QP_TIMEOUT)
2591 		context->pri_path.ackto_lt |= attr->timeout << 3;
2592 
2593 	if (attr_mask & IB_QP_ALT_PATH) {
2594 		err = mlx5_set_path(dev, qp, &attr->alt_ah_attr,
2595 				    &context->alt_path,
2596 				    attr->alt_port_num,
2597 				    attr_mask | IB_QP_PKEY_INDEX | IB_QP_TIMEOUT,
2598 				    0, attr, true);
2599 		if (err)
2600 			goto out;
2601 	}
2602 
2603 	pd = get_pd(qp);
2604 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
2605 		&send_cq, &recv_cq);
2606 
2607 	context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
2608 	context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
2609 	context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
2610 	context->params1  = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
2611 
2612 	if (attr_mask & IB_QP_RNR_RETRY)
2613 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2614 
2615 	if (attr_mask & IB_QP_RETRY_CNT)
2616 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2617 
2618 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2619 		if (attr->max_rd_atomic)
2620 			context->params1 |=
2621 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2622 	}
2623 
2624 	if (attr_mask & IB_QP_SQ_PSN)
2625 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2626 
2627 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2628 		if (attr->max_dest_rd_atomic)
2629 			context->params2 |=
2630 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2631 	}
2632 
2633 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
2634 		context->params2 |= to_mlx5_access_flags(qp, attr, attr_mask);
2635 
2636 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
2637 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2638 
2639 	if (attr_mask & IB_QP_RQ_PSN)
2640 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2641 
2642 	if (attr_mask & IB_QP_QKEY)
2643 		context->qkey = cpu_to_be32(attr->qkey);
2644 
2645 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2646 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2647 
2648 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2649 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2650 		sqd_event = 1;
2651 	else
2652 		sqd_event = 0;
2653 
2654 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2655 		u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
2656 			       qp->port) - 1;
2657 		struct mlx5_ib_port *mibport = &dev->port[port_num];
2658 
2659 		context->qp_counter_set_usr_page |=
2660 			cpu_to_be32((u32)(mibport->q_cnt_id) << 24);
2661 	}
2662 
2663 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2664 		context->sq_crq_size |= cpu_to_be16(1 << 4);
2665 
2666 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
2667 		context->deth_sqpn = cpu_to_be32(1);
2668 
2669 	mlx5_cur = to_mlx5_state(cur_state);
2670 	mlx5_new = to_mlx5_state(new_state);
2671 	mlx5_st = to_mlx5_st(ibqp->qp_type);
2672 	if (mlx5_st < 0)
2673 		goto out;
2674 
2675 	/* If moving to a reset or error state, we must disable page faults on
2676 	 * this QP and flush all current page faults. Otherwise a stale page
2677 	 * fault may attempt to work on this QP after it is reset and moved
2678 	 * again to RTS, and may cause the driver and the device to get out of
2679 	 * sync. */
2680 	if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2681 	    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR) &&
2682 	    (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
2683 		mlx5_ib_qp_disable_pagefaults(qp);
2684 
2685 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
2686 	    !optab[mlx5_cur][mlx5_new])
2687 		goto out;
2688 
2689 	op = optab[mlx5_cur][mlx5_new];
2690 	optpar = ib_mask_to_mlx5_opt(attr_mask);
2691 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
2692 	in->optparam = cpu_to_be32(optpar);
2693 
2694 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET)
2695 		err = modify_raw_packet_qp(dev, qp, op);
2696 	else
2697 		err = mlx5_core_qp_modify(dev->mdev, op, in, sqd_event,
2698 					  &base->mqp);
2699 	if (err)
2700 		goto out;
2701 
2702 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT &&
2703 	    (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
2704 		mlx5_ib_qp_enable_pagefaults(qp);
2705 
2706 	qp->state = new_state;
2707 
2708 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2709 		qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
2710 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2711 		qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
2712 	if (attr_mask & IB_QP_PORT)
2713 		qp->port = attr->port_num;
2714 	if (attr_mask & IB_QP_ALT_PATH)
2715 		qp->trans_qp.alt_port = attr->alt_port_num;
2716 
2717 	/*
2718 	 * If we moved a kernel QP to RESET, clean up all old CQ
2719 	 * entries and reinitialize the QP.
2720 	 */
2721 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
2722 		mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2723 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
2724 		if (send_cq != recv_cq)
2725 			mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
2726 
2727 		qp->rq.head = 0;
2728 		qp->rq.tail = 0;
2729 		qp->sq.head = 0;
2730 		qp->sq.tail = 0;
2731 		qp->sq.cur_post = 0;
2732 		qp->sq.last_poll = 0;
2733 		qp->db.db[MLX5_RCV_DBR] = 0;
2734 		qp->db.db[MLX5_SND_DBR] = 0;
2735 	}
2736 
2737 out:
2738 	kfree(in);
2739 	return err;
2740 }
2741 
2742 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2743 		      int attr_mask, struct ib_udata *udata)
2744 {
2745 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2746 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2747 	enum ib_qp_type qp_type;
2748 	enum ib_qp_state cur_state, new_state;
2749 	int err = -EINVAL;
2750 	int port;
2751 	enum rdma_link_layer ll = IB_LINK_LAYER_UNSPECIFIED;
2752 
2753 	if (ibqp->rwq_ind_tbl)
2754 		return -ENOSYS;
2755 
2756 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
2757 		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
2758 
2759 	qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ?
2760 		IB_QPT_GSI : ibqp->qp_type;
2761 
2762 	mutex_lock(&qp->mutex);
2763 
2764 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2765 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2766 
2767 	if (!(cur_state == new_state && cur_state == IB_QPS_RESET)) {
2768 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2769 		ll = dev->ib_dev.get_link_layer(&dev->ib_dev, port);
2770 	}
2771 
2772 	if (qp_type != MLX5_IB_QPT_REG_UMR &&
2773 	    !ib_modify_qp_is_ok(cur_state, new_state, qp_type, attr_mask, ll)) {
2774 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
2775 			    cur_state, new_state, ibqp->qp_type, attr_mask);
2776 		goto out;
2777 	}
2778 
2779 	if ((attr_mask & IB_QP_PORT) &&
2780 	    (attr->port_num == 0 ||
2781 	     attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports))) {
2782 		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
2783 			    attr->port_num, dev->num_ports);
2784 		goto out;
2785 	}
2786 
2787 	if (attr_mask & IB_QP_PKEY_INDEX) {
2788 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2789 		if (attr->pkey_index >=
2790 		    dev->mdev->port_caps[port - 1].pkey_table_len) {
2791 			mlx5_ib_dbg(dev, "invalid pkey index %d\n",
2792 				    attr->pkey_index);
2793 			goto out;
2794 		}
2795 	}
2796 
2797 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2798 	    attr->max_rd_atomic >
2799 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
2800 		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
2801 			    attr->max_rd_atomic);
2802 		goto out;
2803 	}
2804 
2805 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2806 	    attr->max_dest_rd_atomic >
2807 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
2808 		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
2809 			    attr->max_dest_rd_atomic);
2810 		goto out;
2811 	}
2812 
2813 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2814 		err = 0;
2815 		goto out;
2816 	}
2817 
2818 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2819 
2820 out:
2821 	mutex_unlock(&qp->mutex);
2822 	return err;
2823 }
2824 
2825 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2826 {
2827 	struct mlx5_ib_cq *cq;
2828 	unsigned cur;
2829 
2830 	cur = wq->head - wq->tail;
2831 	if (likely(cur + nreq < wq->max_post))
2832 		return 0;
2833 
2834 	cq = to_mcq(ib_cq);
2835 	spin_lock(&cq->lock);
2836 	cur = wq->head - wq->tail;
2837 	spin_unlock(&cq->lock);
2838 
2839 	return cur + nreq >= wq->max_post;
2840 }
2841 
2842 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
2843 					  u64 remote_addr, u32 rkey)
2844 {
2845 	rseg->raddr    = cpu_to_be64(remote_addr);
2846 	rseg->rkey     = cpu_to_be32(rkey);
2847 	rseg->reserved = 0;
2848 }
2849 
2850 static void *set_eth_seg(struct mlx5_wqe_eth_seg *eseg,
2851 			 struct ib_send_wr *wr, void *qend,
2852 			 struct mlx5_ib_qp *qp, int *size)
2853 {
2854 	void *seg = eseg;
2855 
2856 	memset(eseg, 0, sizeof(struct mlx5_wqe_eth_seg));
2857 
2858 	if (wr->send_flags & IB_SEND_IP_CSUM)
2859 		eseg->cs_flags = MLX5_ETH_WQE_L3_CSUM |
2860 				 MLX5_ETH_WQE_L4_CSUM;
2861 
2862 	seg += sizeof(struct mlx5_wqe_eth_seg);
2863 	*size += sizeof(struct mlx5_wqe_eth_seg) / 16;
2864 
2865 	if (wr->opcode == IB_WR_LSO) {
2866 		struct ib_ud_wr *ud_wr = container_of(wr, struct ib_ud_wr, wr);
2867 		int size_of_inl_hdr_start = sizeof(eseg->inline_hdr_start);
2868 		u64 left, leftlen, copysz;
2869 		void *pdata = ud_wr->header;
2870 
2871 		left = ud_wr->hlen;
2872 		eseg->mss = cpu_to_be16(ud_wr->mss);
2873 		eseg->inline_hdr_sz = cpu_to_be16(left);
2874 
2875 		/*
2876 		 * check if there is space till the end of queue, if yes,
2877 		 * copy all in one shot, otherwise copy till the end of queue,
2878 		 * rollback and than the copy the left
2879 		 */
2880 		leftlen = qend - (void *)eseg->inline_hdr_start;
2881 		copysz = min_t(u64, leftlen, left);
2882 
2883 		memcpy(seg - size_of_inl_hdr_start, pdata, copysz);
2884 
2885 		if (likely(copysz > size_of_inl_hdr_start)) {
2886 			seg += ALIGN(copysz - size_of_inl_hdr_start, 16);
2887 			*size += ALIGN(copysz - size_of_inl_hdr_start, 16) / 16;
2888 		}
2889 
2890 		if (unlikely(copysz < left)) { /* the last wqe in the queue */
2891 			seg = mlx5_get_send_wqe(qp, 0);
2892 			left -= copysz;
2893 			pdata += copysz;
2894 			memcpy(seg, pdata, left);
2895 			seg += ALIGN(left, 16);
2896 			*size += ALIGN(left, 16) / 16;
2897 		}
2898 	}
2899 
2900 	return seg;
2901 }
2902 
2903 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
2904 			     struct ib_send_wr *wr)
2905 {
2906 	memcpy(&dseg->av, &to_mah(ud_wr(wr)->ah)->av, sizeof(struct mlx5_av));
2907 	dseg->av.dqp_dct = cpu_to_be32(ud_wr(wr)->remote_qpn | MLX5_EXTENDED_UD_AV);
2908 	dseg->av.key.qkey.qkey = cpu_to_be32(ud_wr(wr)->remote_qkey);
2909 }
2910 
2911 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
2912 {
2913 	dseg->byte_count = cpu_to_be32(sg->length);
2914 	dseg->lkey       = cpu_to_be32(sg->lkey);
2915 	dseg->addr       = cpu_to_be64(sg->addr);
2916 }
2917 
2918 static __be16 get_klm_octo(int npages)
2919 {
2920 	return cpu_to_be16(ALIGN(npages, 8) / 2);
2921 }
2922 
2923 static __be64 frwr_mkey_mask(void)
2924 {
2925 	u64 result;
2926 
2927 	result = MLX5_MKEY_MASK_LEN		|
2928 		MLX5_MKEY_MASK_PAGE_SIZE	|
2929 		MLX5_MKEY_MASK_START_ADDR	|
2930 		MLX5_MKEY_MASK_EN_RINVAL	|
2931 		MLX5_MKEY_MASK_KEY		|
2932 		MLX5_MKEY_MASK_LR		|
2933 		MLX5_MKEY_MASK_LW		|
2934 		MLX5_MKEY_MASK_RR		|
2935 		MLX5_MKEY_MASK_RW		|
2936 		MLX5_MKEY_MASK_A		|
2937 		MLX5_MKEY_MASK_SMALL_FENCE	|
2938 		MLX5_MKEY_MASK_FREE;
2939 
2940 	return cpu_to_be64(result);
2941 }
2942 
2943 static __be64 sig_mkey_mask(void)
2944 {
2945 	u64 result;
2946 
2947 	result = MLX5_MKEY_MASK_LEN		|
2948 		MLX5_MKEY_MASK_PAGE_SIZE	|
2949 		MLX5_MKEY_MASK_START_ADDR	|
2950 		MLX5_MKEY_MASK_EN_SIGERR	|
2951 		MLX5_MKEY_MASK_EN_RINVAL	|
2952 		MLX5_MKEY_MASK_KEY		|
2953 		MLX5_MKEY_MASK_LR		|
2954 		MLX5_MKEY_MASK_LW		|
2955 		MLX5_MKEY_MASK_RR		|
2956 		MLX5_MKEY_MASK_RW		|
2957 		MLX5_MKEY_MASK_SMALL_FENCE	|
2958 		MLX5_MKEY_MASK_FREE		|
2959 		MLX5_MKEY_MASK_BSF_EN;
2960 
2961 	return cpu_to_be64(result);
2962 }
2963 
2964 static void set_reg_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr,
2965 				struct mlx5_ib_mr *mr)
2966 {
2967 	int ndescs = mr->ndescs;
2968 
2969 	memset(umr, 0, sizeof(*umr));
2970 
2971 	if (mr->access_mode == MLX5_ACCESS_MODE_KLM)
2972 		/* KLMs take twice the size of MTTs */
2973 		ndescs *= 2;
2974 
2975 	umr->flags = MLX5_UMR_CHECK_NOT_FREE;
2976 	umr->klm_octowords = get_klm_octo(ndescs);
2977 	umr->mkey_mask = frwr_mkey_mask();
2978 }
2979 
2980 static void set_linv_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr)
2981 {
2982 	memset(umr, 0, sizeof(*umr));
2983 	umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
2984 	umr->flags = 1 << 7;
2985 }
2986 
2987 static __be64 get_umr_reg_mr_mask(void)
2988 {
2989 	u64 result;
2990 
2991 	result = MLX5_MKEY_MASK_LEN		|
2992 		 MLX5_MKEY_MASK_PAGE_SIZE	|
2993 		 MLX5_MKEY_MASK_START_ADDR	|
2994 		 MLX5_MKEY_MASK_PD		|
2995 		 MLX5_MKEY_MASK_LR		|
2996 		 MLX5_MKEY_MASK_LW		|
2997 		 MLX5_MKEY_MASK_KEY		|
2998 		 MLX5_MKEY_MASK_RR		|
2999 		 MLX5_MKEY_MASK_RW		|
3000 		 MLX5_MKEY_MASK_A		|
3001 		 MLX5_MKEY_MASK_FREE;
3002 
3003 	return cpu_to_be64(result);
3004 }
3005 
3006 static __be64 get_umr_unreg_mr_mask(void)
3007 {
3008 	u64 result;
3009 
3010 	result = MLX5_MKEY_MASK_FREE;
3011 
3012 	return cpu_to_be64(result);
3013 }
3014 
3015 static __be64 get_umr_update_mtt_mask(void)
3016 {
3017 	u64 result;
3018 
3019 	result = MLX5_MKEY_MASK_FREE;
3020 
3021 	return cpu_to_be64(result);
3022 }
3023 
3024 static __be64 get_umr_update_translation_mask(void)
3025 {
3026 	u64 result;
3027 
3028 	result = MLX5_MKEY_MASK_LEN |
3029 		 MLX5_MKEY_MASK_PAGE_SIZE |
3030 		 MLX5_MKEY_MASK_START_ADDR |
3031 		 MLX5_MKEY_MASK_KEY |
3032 		 MLX5_MKEY_MASK_FREE;
3033 
3034 	return cpu_to_be64(result);
3035 }
3036 
3037 static __be64 get_umr_update_access_mask(void)
3038 {
3039 	u64 result;
3040 
3041 	result = MLX5_MKEY_MASK_LW |
3042 		 MLX5_MKEY_MASK_RR |
3043 		 MLX5_MKEY_MASK_RW |
3044 		 MLX5_MKEY_MASK_A |
3045 		 MLX5_MKEY_MASK_KEY |
3046 		 MLX5_MKEY_MASK_FREE;
3047 
3048 	return cpu_to_be64(result);
3049 }
3050 
3051 static __be64 get_umr_update_pd_mask(void)
3052 {
3053 	u64 result;
3054 
3055 	result = MLX5_MKEY_MASK_PD |
3056 		 MLX5_MKEY_MASK_KEY |
3057 		 MLX5_MKEY_MASK_FREE;
3058 
3059 	return cpu_to_be64(result);
3060 }
3061 
3062 static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
3063 				struct ib_send_wr *wr)
3064 {
3065 	struct mlx5_umr_wr *umrwr = umr_wr(wr);
3066 
3067 	memset(umr, 0, sizeof(*umr));
3068 
3069 	if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
3070 		umr->flags = MLX5_UMR_CHECK_FREE; /* fail if free */
3071 	else
3072 		umr->flags = MLX5_UMR_CHECK_NOT_FREE; /* fail if not free */
3073 
3074 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UNREG)) {
3075 		umr->klm_octowords = get_klm_octo(umrwr->npages);
3076 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT) {
3077 			umr->mkey_mask = get_umr_update_mtt_mask();
3078 			umr->bsf_octowords = get_klm_octo(umrwr->target.offset);
3079 			umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
3080 		}
3081 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION)
3082 			umr->mkey_mask |= get_umr_update_translation_mask();
3083 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_ACCESS)
3084 			umr->mkey_mask |= get_umr_update_access_mask();
3085 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_PD)
3086 			umr->mkey_mask |= get_umr_update_pd_mask();
3087 		if (!umr->mkey_mask)
3088 			umr->mkey_mask = get_umr_reg_mr_mask();
3089 	} else {
3090 		umr->mkey_mask = get_umr_unreg_mr_mask();
3091 	}
3092 
3093 	if (!wr->num_sge)
3094 		umr->flags |= MLX5_UMR_INLINE;
3095 }
3096 
3097 static u8 get_umr_flags(int acc)
3098 {
3099 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
3100 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
3101 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
3102 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
3103 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
3104 }
3105 
3106 static void set_reg_mkey_seg(struct mlx5_mkey_seg *seg,
3107 			     struct mlx5_ib_mr *mr,
3108 			     u32 key, int access)
3109 {
3110 	int ndescs = ALIGN(mr->ndescs, 8) >> 1;
3111 
3112 	memset(seg, 0, sizeof(*seg));
3113 
3114 	if (mr->access_mode == MLX5_ACCESS_MODE_MTT)
3115 		seg->log2_page_size = ilog2(mr->ibmr.page_size);
3116 	else if (mr->access_mode == MLX5_ACCESS_MODE_KLM)
3117 		/* KLMs take twice the size of MTTs */
3118 		ndescs *= 2;
3119 
3120 	seg->flags = get_umr_flags(access) | mr->access_mode;
3121 	seg->qpn_mkey7_0 = cpu_to_be32((key & 0xff) | 0xffffff00);
3122 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
3123 	seg->start_addr = cpu_to_be64(mr->ibmr.iova);
3124 	seg->len = cpu_to_be64(mr->ibmr.length);
3125 	seg->xlt_oct_size = cpu_to_be32(ndescs);
3126 }
3127 
3128 static void set_linv_mkey_seg(struct mlx5_mkey_seg *seg)
3129 {
3130 	memset(seg, 0, sizeof(*seg));
3131 	seg->status = MLX5_MKEY_STATUS_FREE;
3132 }
3133 
3134 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr)
3135 {
3136 	struct mlx5_umr_wr *umrwr = umr_wr(wr);
3137 
3138 	memset(seg, 0, sizeof(*seg));
3139 	if (wr->send_flags & MLX5_IB_SEND_UMR_UNREG) {
3140 		seg->status = MLX5_MKEY_STATUS_FREE;
3141 		return;
3142 	}
3143 
3144 	seg->flags = convert_access(umrwr->access_flags);
3145 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT)) {
3146 		if (umrwr->pd)
3147 			seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
3148 		seg->start_addr = cpu_to_be64(umrwr->target.virt_addr);
3149 	}
3150 	seg->len = cpu_to_be64(umrwr->length);
3151 	seg->log2_page_size = umrwr->page_shift;
3152 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
3153 				       mlx5_mkey_variant(umrwr->mkey));
3154 }
3155 
3156 static void set_reg_data_seg(struct mlx5_wqe_data_seg *dseg,
3157 			     struct mlx5_ib_mr *mr,
3158 			     struct mlx5_ib_pd *pd)
3159 {
3160 	int bcount = mr->desc_size * mr->ndescs;
3161 
3162 	dseg->addr = cpu_to_be64(mr->desc_map);
3163 	dseg->byte_count = cpu_to_be32(ALIGN(bcount, 64));
3164 	dseg->lkey = cpu_to_be32(pd->ibpd.local_dma_lkey);
3165 }
3166 
3167 static __be32 send_ieth(struct ib_send_wr *wr)
3168 {
3169 	switch (wr->opcode) {
3170 	case IB_WR_SEND_WITH_IMM:
3171 	case IB_WR_RDMA_WRITE_WITH_IMM:
3172 		return wr->ex.imm_data;
3173 
3174 	case IB_WR_SEND_WITH_INV:
3175 		return cpu_to_be32(wr->ex.invalidate_rkey);
3176 
3177 	default:
3178 		return 0;
3179 	}
3180 }
3181 
3182 static u8 calc_sig(void *wqe, int size)
3183 {
3184 	u8 *p = wqe;
3185 	u8 res = 0;
3186 	int i;
3187 
3188 	for (i = 0; i < size; i++)
3189 		res ^= p[i];
3190 
3191 	return ~res;
3192 }
3193 
3194 static u8 wq_sig(void *wqe)
3195 {
3196 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
3197 }
3198 
3199 static int set_data_inl_seg(struct mlx5_ib_qp *qp, struct ib_send_wr *wr,
3200 			    void *wqe, int *sz)
3201 {
3202 	struct mlx5_wqe_inline_seg *seg;
3203 	void *qend = qp->sq.qend;
3204 	void *addr;
3205 	int inl = 0;
3206 	int copy;
3207 	int len;
3208 	int i;
3209 
3210 	seg = wqe;
3211 	wqe += sizeof(*seg);
3212 	for (i = 0; i < wr->num_sge; i++) {
3213 		addr = (void *)(unsigned long)(wr->sg_list[i].addr);
3214 		len  = wr->sg_list[i].length;
3215 		inl += len;
3216 
3217 		if (unlikely(inl > qp->max_inline_data))
3218 			return -ENOMEM;
3219 
3220 		if (unlikely(wqe + len > qend)) {
3221 			copy = qend - wqe;
3222 			memcpy(wqe, addr, copy);
3223 			addr += copy;
3224 			len -= copy;
3225 			wqe = mlx5_get_send_wqe(qp, 0);
3226 		}
3227 		memcpy(wqe, addr, len);
3228 		wqe += len;
3229 	}
3230 
3231 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
3232 
3233 	*sz = ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
3234 
3235 	return 0;
3236 }
3237 
3238 static u16 prot_field_size(enum ib_signature_type type)
3239 {
3240 	switch (type) {
3241 	case IB_SIG_TYPE_T10_DIF:
3242 		return MLX5_DIF_SIZE;
3243 	default:
3244 		return 0;
3245 	}
3246 }
3247 
3248 static u8 bs_selector(int block_size)
3249 {
3250 	switch (block_size) {
3251 	case 512:	    return 0x1;
3252 	case 520:	    return 0x2;
3253 	case 4096:	    return 0x3;
3254 	case 4160:	    return 0x4;
3255 	case 1073741824:    return 0x5;
3256 	default:	    return 0;
3257 	}
3258 }
3259 
3260 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
3261 			      struct mlx5_bsf_inl *inl)
3262 {
3263 	/* Valid inline section and allow BSF refresh */
3264 	inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
3265 				       MLX5_BSF_REFRESH_DIF);
3266 	inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
3267 	inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
3268 	/* repeating block */
3269 	inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
3270 	inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
3271 			MLX5_DIF_CRC : MLX5_DIF_IPCS;
3272 
3273 	if (domain->sig.dif.ref_remap)
3274 		inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
3275 
3276 	if (domain->sig.dif.app_escape) {
3277 		if (domain->sig.dif.ref_escape)
3278 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
3279 		else
3280 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
3281 	}
3282 
3283 	inl->dif_app_bitmask_check =
3284 		cpu_to_be16(domain->sig.dif.apptag_check_mask);
3285 }
3286 
3287 static int mlx5_set_bsf(struct ib_mr *sig_mr,
3288 			struct ib_sig_attrs *sig_attrs,
3289 			struct mlx5_bsf *bsf, u32 data_size)
3290 {
3291 	struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
3292 	struct mlx5_bsf_basic *basic = &bsf->basic;
3293 	struct ib_sig_domain *mem = &sig_attrs->mem;
3294 	struct ib_sig_domain *wire = &sig_attrs->wire;
3295 
3296 	memset(bsf, 0, sizeof(*bsf));
3297 
3298 	/* Basic + Extended + Inline */
3299 	basic->bsf_size_sbs = 1 << 7;
3300 	/* Input domain check byte mask */
3301 	basic->check_byte_mask = sig_attrs->check_mask;
3302 	basic->raw_data_size = cpu_to_be32(data_size);
3303 
3304 	/* Memory domain */
3305 	switch (sig_attrs->mem.sig_type) {
3306 	case IB_SIG_TYPE_NONE:
3307 		break;
3308 	case IB_SIG_TYPE_T10_DIF:
3309 		basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
3310 		basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
3311 		mlx5_fill_inl_bsf(mem, &bsf->m_inl);
3312 		break;
3313 	default:
3314 		return -EINVAL;
3315 	}
3316 
3317 	/* Wire domain */
3318 	switch (sig_attrs->wire.sig_type) {
3319 	case IB_SIG_TYPE_NONE:
3320 		break;
3321 	case IB_SIG_TYPE_T10_DIF:
3322 		if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
3323 		    mem->sig_type == wire->sig_type) {
3324 			/* Same block structure */
3325 			basic->bsf_size_sbs |= 1 << 4;
3326 			if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
3327 				basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
3328 			if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
3329 				basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
3330 			if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
3331 				basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
3332 		} else
3333 			basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
3334 
3335 		basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
3336 		mlx5_fill_inl_bsf(wire, &bsf->w_inl);
3337 		break;
3338 	default:
3339 		return -EINVAL;
3340 	}
3341 
3342 	return 0;
3343 }
3344 
3345 static int set_sig_data_segment(struct ib_sig_handover_wr *wr,
3346 				struct mlx5_ib_qp *qp, void **seg, int *size)
3347 {
3348 	struct ib_sig_attrs *sig_attrs = wr->sig_attrs;
3349 	struct ib_mr *sig_mr = wr->sig_mr;
3350 	struct mlx5_bsf *bsf;
3351 	u32 data_len = wr->wr.sg_list->length;
3352 	u32 data_key = wr->wr.sg_list->lkey;
3353 	u64 data_va = wr->wr.sg_list->addr;
3354 	int ret;
3355 	int wqe_size;
3356 
3357 	if (!wr->prot ||
3358 	    (data_key == wr->prot->lkey &&
3359 	     data_va == wr->prot->addr &&
3360 	     data_len == wr->prot->length)) {
3361 		/**
3362 		 * Source domain doesn't contain signature information
3363 		 * or data and protection are interleaved in memory.
3364 		 * So need construct:
3365 		 *                  ------------------
3366 		 *                 |     data_klm     |
3367 		 *                  ------------------
3368 		 *                 |       BSF        |
3369 		 *                  ------------------
3370 		 **/
3371 		struct mlx5_klm *data_klm = *seg;
3372 
3373 		data_klm->bcount = cpu_to_be32(data_len);
3374 		data_klm->key = cpu_to_be32(data_key);
3375 		data_klm->va = cpu_to_be64(data_va);
3376 		wqe_size = ALIGN(sizeof(*data_klm), 64);
3377 	} else {
3378 		/**
3379 		 * Source domain contains signature information
3380 		 * So need construct a strided block format:
3381 		 *               ---------------------------
3382 		 *              |     stride_block_ctrl     |
3383 		 *               ---------------------------
3384 		 *              |          data_klm         |
3385 		 *               ---------------------------
3386 		 *              |          prot_klm         |
3387 		 *               ---------------------------
3388 		 *              |             BSF           |
3389 		 *               ---------------------------
3390 		 **/
3391 		struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
3392 		struct mlx5_stride_block_entry *data_sentry;
3393 		struct mlx5_stride_block_entry *prot_sentry;
3394 		u32 prot_key = wr->prot->lkey;
3395 		u64 prot_va = wr->prot->addr;
3396 		u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
3397 		int prot_size;
3398 
3399 		sblock_ctrl = *seg;
3400 		data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
3401 		prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
3402 
3403 		prot_size = prot_field_size(sig_attrs->mem.sig_type);
3404 		if (!prot_size) {
3405 			pr_err("Bad block size given: %u\n", block_size);
3406 			return -EINVAL;
3407 		}
3408 		sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
3409 							    prot_size);
3410 		sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
3411 		sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
3412 		sblock_ctrl->num_entries = cpu_to_be16(2);
3413 
3414 		data_sentry->bcount = cpu_to_be16(block_size);
3415 		data_sentry->key = cpu_to_be32(data_key);
3416 		data_sentry->va = cpu_to_be64(data_va);
3417 		data_sentry->stride = cpu_to_be16(block_size);
3418 
3419 		prot_sentry->bcount = cpu_to_be16(prot_size);
3420 		prot_sentry->key = cpu_to_be32(prot_key);
3421 		prot_sentry->va = cpu_to_be64(prot_va);
3422 		prot_sentry->stride = cpu_to_be16(prot_size);
3423 
3424 		wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
3425 				 sizeof(*prot_sentry), 64);
3426 	}
3427 
3428 	*seg += wqe_size;
3429 	*size += wqe_size / 16;
3430 	if (unlikely((*seg == qp->sq.qend)))
3431 		*seg = mlx5_get_send_wqe(qp, 0);
3432 
3433 	bsf = *seg;
3434 	ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
3435 	if (ret)
3436 		return -EINVAL;
3437 
3438 	*seg += sizeof(*bsf);
3439 	*size += sizeof(*bsf) / 16;
3440 	if (unlikely((*seg == qp->sq.qend)))
3441 		*seg = mlx5_get_send_wqe(qp, 0);
3442 
3443 	return 0;
3444 }
3445 
3446 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
3447 				 struct ib_sig_handover_wr *wr, u32 nelements,
3448 				 u32 length, u32 pdn)
3449 {
3450 	struct ib_mr *sig_mr = wr->sig_mr;
3451 	u32 sig_key = sig_mr->rkey;
3452 	u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
3453 
3454 	memset(seg, 0, sizeof(*seg));
3455 
3456 	seg->flags = get_umr_flags(wr->access_flags) |
3457 				   MLX5_ACCESS_MODE_KLM;
3458 	seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
3459 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
3460 				    MLX5_MKEY_BSF_EN | pdn);
3461 	seg->len = cpu_to_be64(length);
3462 	seg->xlt_oct_size = cpu_to_be32(be16_to_cpu(get_klm_octo(nelements)));
3463 	seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
3464 }
3465 
3466 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
3467 				u32 nelements)
3468 {
3469 	memset(umr, 0, sizeof(*umr));
3470 
3471 	umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
3472 	umr->klm_octowords = get_klm_octo(nelements);
3473 	umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
3474 	umr->mkey_mask = sig_mkey_mask();
3475 }
3476 
3477 
3478 static int set_sig_umr_wr(struct ib_send_wr *send_wr, struct mlx5_ib_qp *qp,
3479 			  void **seg, int *size)
3480 {
3481 	struct ib_sig_handover_wr *wr = sig_handover_wr(send_wr);
3482 	struct mlx5_ib_mr *sig_mr = to_mmr(wr->sig_mr);
3483 	u32 pdn = get_pd(qp)->pdn;
3484 	u32 klm_oct_size;
3485 	int region_len, ret;
3486 
3487 	if (unlikely(wr->wr.num_sge != 1) ||
3488 	    unlikely(wr->access_flags & IB_ACCESS_REMOTE_ATOMIC) ||
3489 	    unlikely(!sig_mr->sig) || unlikely(!qp->signature_en) ||
3490 	    unlikely(!sig_mr->sig->sig_status_checked))
3491 		return -EINVAL;
3492 
3493 	/* length of the protected region, data + protection */
3494 	region_len = wr->wr.sg_list->length;
3495 	if (wr->prot &&
3496 	    (wr->prot->lkey != wr->wr.sg_list->lkey  ||
3497 	     wr->prot->addr != wr->wr.sg_list->addr  ||
3498 	     wr->prot->length != wr->wr.sg_list->length))
3499 		region_len += wr->prot->length;
3500 
3501 	/**
3502 	 * KLM octoword size - if protection was provided
3503 	 * then we use strided block format (3 octowords),
3504 	 * else we use single KLM (1 octoword)
3505 	 **/
3506 	klm_oct_size = wr->prot ? 3 : 1;
3507 
3508 	set_sig_umr_segment(*seg, klm_oct_size);
3509 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3510 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3511 	if (unlikely((*seg == qp->sq.qend)))
3512 		*seg = mlx5_get_send_wqe(qp, 0);
3513 
3514 	set_sig_mkey_segment(*seg, wr, klm_oct_size, region_len, pdn);
3515 	*seg += sizeof(struct mlx5_mkey_seg);
3516 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3517 	if (unlikely((*seg == qp->sq.qend)))
3518 		*seg = mlx5_get_send_wqe(qp, 0);
3519 
3520 	ret = set_sig_data_segment(wr, qp, seg, size);
3521 	if (ret)
3522 		return ret;
3523 
3524 	sig_mr->sig->sig_status_checked = false;
3525 	return 0;
3526 }
3527 
3528 static int set_psv_wr(struct ib_sig_domain *domain,
3529 		      u32 psv_idx, void **seg, int *size)
3530 {
3531 	struct mlx5_seg_set_psv *psv_seg = *seg;
3532 
3533 	memset(psv_seg, 0, sizeof(*psv_seg));
3534 	psv_seg->psv_num = cpu_to_be32(psv_idx);
3535 	switch (domain->sig_type) {
3536 	case IB_SIG_TYPE_NONE:
3537 		break;
3538 	case IB_SIG_TYPE_T10_DIF:
3539 		psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
3540 						     domain->sig.dif.app_tag);
3541 		psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
3542 		break;
3543 	default:
3544 		pr_err("Bad signature type given.\n");
3545 		return 1;
3546 	}
3547 
3548 	*seg += sizeof(*psv_seg);
3549 	*size += sizeof(*psv_seg) / 16;
3550 
3551 	return 0;
3552 }
3553 
3554 static int set_reg_wr(struct mlx5_ib_qp *qp,
3555 		      struct ib_reg_wr *wr,
3556 		      void **seg, int *size)
3557 {
3558 	struct mlx5_ib_mr *mr = to_mmr(wr->mr);
3559 	struct mlx5_ib_pd *pd = to_mpd(qp->ibqp.pd);
3560 
3561 	if (unlikely(wr->wr.send_flags & IB_SEND_INLINE)) {
3562 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
3563 			     "Invalid IB_SEND_INLINE send flag\n");
3564 		return -EINVAL;
3565 	}
3566 
3567 	set_reg_umr_seg(*seg, mr);
3568 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3569 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3570 	if (unlikely((*seg == qp->sq.qend)))
3571 		*seg = mlx5_get_send_wqe(qp, 0);
3572 
3573 	set_reg_mkey_seg(*seg, mr, wr->key, wr->access);
3574 	*seg += sizeof(struct mlx5_mkey_seg);
3575 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3576 	if (unlikely((*seg == qp->sq.qend)))
3577 		*seg = mlx5_get_send_wqe(qp, 0);
3578 
3579 	set_reg_data_seg(*seg, mr, pd);
3580 	*seg += sizeof(struct mlx5_wqe_data_seg);
3581 	*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
3582 
3583 	return 0;
3584 }
3585 
3586 static void set_linv_wr(struct mlx5_ib_qp *qp, void **seg, int *size)
3587 {
3588 	set_linv_umr_seg(*seg);
3589 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3590 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3591 	if (unlikely((*seg == qp->sq.qend)))
3592 		*seg = mlx5_get_send_wqe(qp, 0);
3593 	set_linv_mkey_seg(*seg);
3594 	*seg += sizeof(struct mlx5_mkey_seg);
3595 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3596 	if (unlikely((*seg == qp->sq.qend)))
3597 		*seg = mlx5_get_send_wqe(qp, 0);
3598 }
3599 
3600 static void dump_wqe(struct mlx5_ib_qp *qp, int idx, int size_16)
3601 {
3602 	__be32 *p = NULL;
3603 	int tidx = idx;
3604 	int i, j;
3605 
3606 	pr_debug("dump wqe at %p\n", mlx5_get_send_wqe(qp, tidx));
3607 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
3608 		if ((i & 0xf) == 0) {
3609 			void *buf = mlx5_get_send_wqe(qp, tidx);
3610 			tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
3611 			p = buf;
3612 			j = 0;
3613 		}
3614 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
3615 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
3616 			 be32_to_cpu(p[j + 3]));
3617 	}
3618 }
3619 
3620 static void mlx5_bf_copy(u64 __iomem *dst, u64 *src,
3621 			 unsigned bytecnt, struct mlx5_ib_qp *qp)
3622 {
3623 	while (bytecnt > 0) {
3624 		__iowrite64_copy(dst++, src++, 8);
3625 		__iowrite64_copy(dst++, src++, 8);
3626 		__iowrite64_copy(dst++, src++, 8);
3627 		__iowrite64_copy(dst++, src++, 8);
3628 		__iowrite64_copy(dst++, src++, 8);
3629 		__iowrite64_copy(dst++, src++, 8);
3630 		__iowrite64_copy(dst++, src++, 8);
3631 		__iowrite64_copy(dst++, src++, 8);
3632 		bytecnt -= 64;
3633 		if (unlikely(src == qp->sq.qend))
3634 			src = mlx5_get_send_wqe(qp, 0);
3635 	}
3636 }
3637 
3638 static u8 get_fence(u8 fence, struct ib_send_wr *wr)
3639 {
3640 	if (unlikely(wr->opcode == IB_WR_LOCAL_INV &&
3641 		     wr->send_flags & IB_SEND_FENCE))
3642 		return MLX5_FENCE_MODE_STRONG_ORDERING;
3643 
3644 	if (unlikely(fence)) {
3645 		if (wr->send_flags & IB_SEND_FENCE)
3646 			return MLX5_FENCE_MODE_SMALL_AND_FENCE;
3647 		else
3648 			return fence;
3649 	} else if (unlikely(wr->send_flags & IB_SEND_FENCE)) {
3650 		return MLX5_FENCE_MODE_FENCE;
3651 	}
3652 
3653 	return 0;
3654 }
3655 
3656 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
3657 		     struct mlx5_wqe_ctrl_seg **ctrl,
3658 		     struct ib_send_wr *wr, unsigned *idx,
3659 		     int *size, int nreq)
3660 {
3661 	int err = 0;
3662 
3663 	if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq))) {
3664 		err = -ENOMEM;
3665 		return err;
3666 	}
3667 
3668 	*idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
3669 	*seg = mlx5_get_send_wqe(qp, *idx);
3670 	*ctrl = *seg;
3671 	*(uint32_t *)(*seg + 8) = 0;
3672 	(*ctrl)->imm = send_ieth(wr);
3673 	(*ctrl)->fm_ce_se = qp->sq_signal_bits |
3674 		(wr->send_flags & IB_SEND_SIGNALED ?
3675 		 MLX5_WQE_CTRL_CQ_UPDATE : 0) |
3676 		(wr->send_flags & IB_SEND_SOLICITED ?
3677 		 MLX5_WQE_CTRL_SOLICITED : 0);
3678 
3679 	*seg += sizeof(**ctrl);
3680 	*size = sizeof(**ctrl) / 16;
3681 
3682 	return err;
3683 }
3684 
3685 static void finish_wqe(struct mlx5_ib_qp *qp,
3686 		       struct mlx5_wqe_ctrl_seg *ctrl,
3687 		       u8 size, unsigned idx, u64 wr_id,
3688 		       int nreq, u8 fence, u8 next_fence,
3689 		       u32 mlx5_opcode)
3690 {
3691 	u8 opmod = 0;
3692 
3693 	ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
3694 					     mlx5_opcode | ((u32)opmod << 24));
3695 	ctrl->qpn_ds = cpu_to_be32(size | (qp->trans_qp.base.mqp.qpn << 8));
3696 	ctrl->fm_ce_se |= fence;
3697 	qp->fm_cache = next_fence;
3698 	if (unlikely(qp->wq_sig))
3699 		ctrl->signature = wq_sig(ctrl);
3700 
3701 	qp->sq.wrid[idx] = wr_id;
3702 	qp->sq.w_list[idx].opcode = mlx5_opcode;
3703 	qp->sq.wqe_head[idx] = qp->sq.head + nreq;
3704 	qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
3705 	qp->sq.w_list[idx].next = qp->sq.cur_post;
3706 }
3707 
3708 
3709 int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
3710 		      struct ib_send_wr **bad_wr)
3711 {
3712 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
3713 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3714 	struct mlx5_core_dev *mdev = dev->mdev;
3715 	struct mlx5_ib_qp *qp;
3716 	struct mlx5_ib_mr *mr;
3717 	struct mlx5_wqe_data_seg *dpseg;
3718 	struct mlx5_wqe_xrc_seg *xrc;
3719 	struct mlx5_bf *bf;
3720 	int uninitialized_var(size);
3721 	void *qend;
3722 	unsigned long flags;
3723 	unsigned idx;
3724 	int err = 0;
3725 	int inl = 0;
3726 	int num_sge;
3727 	void *seg;
3728 	int nreq;
3729 	int i;
3730 	u8 next_fence = 0;
3731 	u8 fence;
3732 
3733 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
3734 		return mlx5_ib_gsi_post_send(ibqp, wr, bad_wr);
3735 
3736 	qp = to_mqp(ibqp);
3737 	bf = qp->bf;
3738 	qend = qp->sq.qend;
3739 
3740 	spin_lock_irqsave(&qp->sq.lock, flags);
3741 
3742 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
3743 		err = -EIO;
3744 		*bad_wr = wr;
3745 		nreq = 0;
3746 		goto out;
3747 	}
3748 
3749 	for (nreq = 0; wr; nreq++, wr = wr->next) {
3750 		if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
3751 			mlx5_ib_warn(dev, "\n");
3752 			err = -EINVAL;
3753 			*bad_wr = wr;
3754 			goto out;
3755 		}
3756 
3757 		fence = qp->fm_cache;
3758 		num_sge = wr->num_sge;
3759 		if (unlikely(num_sge > qp->sq.max_gs)) {
3760 			mlx5_ib_warn(dev, "\n");
3761 			err = -ENOMEM;
3762 			*bad_wr = wr;
3763 			goto out;
3764 		}
3765 
3766 		err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, nreq);
3767 		if (err) {
3768 			mlx5_ib_warn(dev, "\n");
3769 			err = -ENOMEM;
3770 			*bad_wr = wr;
3771 			goto out;
3772 		}
3773 
3774 		switch (ibqp->qp_type) {
3775 		case IB_QPT_XRC_INI:
3776 			xrc = seg;
3777 			seg += sizeof(*xrc);
3778 			size += sizeof(*xrc) / 16;
3779 			/* fall through */
3780 		case IB_QPT_RC:
3781 			switch (wr->opcode) {
3782 			case IB_WR_RDMA_READ:
3783 			case IB_WR_RDMA_WRITE:
3784 			case IB_WR_RDMA_WRITE_WITH_IMM:
3785 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
3786 					      rdma_wr(wr)->rkey);
3787 				seg += sizeof(struct mlx5_wqe_raddr_seg);
3788 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
3789 				break;
3790 
3791 			case IB_WR_ATOMIC_CMP_AND_SWP:
3792 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3793 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3794 				mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
3795 				err = -ENOSYS;
3796 				*bad_wr = wr;
3797 				goto out;
3798 
3799 			case IB_WR_LOCAL_INV:
3800 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3801 				qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
3802 				ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
3803 				set_linv_wr(qp, &seg, &size);
3804 				num_sge = 0;
3805 				break;
3806 
3807 			case IB_WR_REG_MR:
3808 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3809 				qp->sq.wr_data[idx] = IB_WR_REG_MR;
3810 				ctrl->imm = cpu_to_be32(reg_wr(wr)->key);
3811 				err = set_reg_wr(qp, reg_wr(wr), &seg, &size);
3812 				if (err) {
3813 					*bad_wr = wr;
3814 					goto out;
3815 				}
3816 				num_sge = 0;
3817 				break;
3818 
3819 			case IB_WR_REG_SIG_MR:
3820 				qp->sq.wr_data[idx] = IB_WR_REG_SIG_MR;
3821 				mr = to_mmr(sig_handover_wr(wr)->sig_mr);
3822 
3823 				ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
3824 				err = set_sig_umr_wr(wr, qp, &seg, &size);
3825 				if (err) {
3826 					mlx5_ib_warn(dev, "\n");
3827 					*bad_wr = wr;
3828 					goto out;
3829 				}
3830 
3831 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3832 					   nreq, get_fence(fence, wr),
3833 					   next_fence, MLX5_OPCODE_UMR);
3834 				/*
3835 				 * SET_PSV WQEs are not signaled and solicited
3836 				 * on error
3837 				 */
3838 				wr->send_flags &= ~IB_SEND_SIGNALED;
3839 				wr->send_flags |= IB_SEND_SOLICITED;
3840 				err = begin_wqe(qp, &seg, &ctrl, wr,
3841 						&idx, &size, nreq);
3842 				if (err) {
3843 					mlx5_ib_warn(dev, "\n");
3844 					err = -ENOMEM;
3845 					*bad_wr = wr;
3846 					goto out;
3847 				}
3848 
3849 				err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->mem,
3850 						 mr->sig->psv_memory.psv_idx, &seg,
3851 						 &size);
3852 				if (err) {
3853 					mlx5_ib_warn(dev, "\n");
3854 					*bad_wr = wr;
3855 					goto out;
3856 				}
3857 
3858 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3859 					   nreq, get_fence(fence, wr),
3860 					   next_fence, MLX5_OPCODE_SET_PSV);
3861 				err = begin_wqe(qp, &seg, &ctrl, wr,
3862 						&idx, &size, nreq);
3863 				if (err) {
3864 					mlx5_ib_warn(dev, "\n");
3865 					err = -ENOMEM;
3866 					*bad_wr = wr;
3867 					goto out;
3868 				}
3869 
3870 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3871 				err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->wire,
3872 						 mr->sig->psv_wire.psv_idx, &seg,
3873 						 &size);
3874 				if (err) {
3875 					mlx5_ib_warn(dev, "\n");
3876 					*bad_wr = wr;
3877 					goto out;
3878 				}
3879 
3880 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3881 					   nreq, get_fence(fence, wr),
3882 					   next_fence, MLX5_OPCODE_SET_PSV);
3883 				num_sge = 0;
3884 				goto skip_psv;
3885 
3886 			default:
3887 				break;
3888 			}
3889 			break;
3890 
3891 		case IB_QPT_UC:
3892 			switch (wr->opcode) {
3893 			case IB_WR_RDMA_WRITE:
3894 			case IB_WR_RDMA_WRITE_WITH_IMM:
3895 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
3896 					      rdma_wr(wr)->rkey);
3897 				seg  += sizeof(struct mlx5_wqe_raddr_seg);
3898 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
3899 				break;
3900 
3901 			default:
3902 				break;
3903 			}
3904 			break;
3905 
3906 		case IB_QPT_SMI:
3907 		case MLX5_IB_QPT_HW_GSI:
3908 			set_datagram_seg(seg, wr);
3909 			seg += sizeof(struct mlx5_wqe_datagram_seg);
3910 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
3911 			if (unlikely((seg == qend)))
3912 				seg = mlx5_get_send_wqe(qp, 0);
3913 			break;
3914 		case IB_QPT_UD:
3915 			set_datagram_seg(seg, wr);
3916 			seg += sizeof(struct mlx5_wqe_datagram_seg);
3917 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
3918 
3919 			if (unlikely((seg == qend)))
3920 				seg = mlx5_get_send_wqe(qp, 0);
3921 
3922 			/* handle qp that supports ud offload */
3923 			if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO) {
3924 				struct mlx5_wqe_eth_pad *pad;
3925 
3926 				pad = seg;
3927 				memset(pad, 0, sizeof(struct mlx5_wqe_eth_pad));
3928 				seg += sizeof(struct mlx5_wqe_eth_pad);
3929 				size += sizeof(struct mlx5_wqe_eth_pad) / 16;
3930 
3931 				seg = set_eth_seg(seg, wr, qend, qp, &size);
3932 
3933 				if (unlikely((seg == qend)))
3934 					seg = mlx5_get_send_wqe(qp, 0);
3935 			}
3936 			break;
3937 		case MLX5_IB_QPT_REG_UMR:
3938 			if (wr->opcode != MLX5_IB_WR_UMR) {
3939 				err = -EINVAL;
3940 				mlx5_ib_warn(dev, "bad opcode\n");
3941 				goto out;
3942 			}
3943 			qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
3944 			ctrl->imm = cpu_to_be32(umr_wr(wr)->mkey);
3945 			set_reg_umr_segment(seg, wr);
3946 			seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3947 			size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3948 			if (unlikely((seg == qend)))
3949 				seg = mlx5_get_send_wqe(qp, 0);
3950 			set_reg_mkey_segment(seg, wr);
3951 			seg += sizeof(struct mlx5_mkey_seg);
3952 			size += sizeof(struct mlx5_mkey_seg) / 16;
3953 			if (unlikely((seg == qend)))
3954 				seg = mlx5_get_send_wqe(qp, 0);
3955 			break;
3956 
3957 		default:
3958 			break;
3959 		}
3960 
3961 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
3962 			int uninitialized_var(sz);
3963 
3964 			err = set_data_inl_seg(qp, wr, seg, &sz);
3965 			if (unlikely(err)) {
3966 				mlx5_ib_warn(dev, "\n");
3967 				*bad_wr = wr;
3968 				goto out;
3969 			}
3970 			inl = 1;
3971 			size += sz;
3972 		} else {
3973 			dpseg = seg;
3974 			for (i = 0; i < num_sge; i++) {
3975 				if (unlikely(dpseg == qend)) {
3976 					seg = mlx5_get_send_wqe(qp, 0);
3977 					dpseg = seg;
3978 				}
3979 				if (likely(wr->sg_list[i].length)) {
3980 					set_data_ptr_seg(dpseg, wr->sg_list + i);
3981 					size += sizeof(struct mlx5_wqe_data_seg) / 16;
3982 					dpseg++;
3983 				}
3984 			}
3985 		}
3986 
3987 		finish_wqe(qp, ctrl, size, idx, wr->wr_id, nreq,
3988 			   get_fence(fence, wr), next_fence,
3989 			   mlx5_ib_opcode[wr->opcode]);
3990 skip_psv:
3991 		if (0)
3992 			dump_wqe(qp, idx, size);
3993 	}
3994 
3995 out:
3996 	if (likely(nreq)) {
3997 		qp->sq.head += nreq;
3998 
3999 		/* Make sure that descriptors are written before
4000 		 * updating doorbell record and ringing the doorbell
4001 		 */
4002 		wmb();
4003 
4004 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
4005 
4006 		/* Make sure doorbell record is visible to the HCA before
4007 		 * we hit doorbell */
4008 		wmb();
4009 
4010 		if (bf->need_lock)
4011 			spin_lock(&bf->lock);
4012 		else
4013 			__acquire(&bf->lock);
4014 
4015 		/* TBD enable WC */
4016 		if (0 && nreq == 1 && bf->uuarn && inl && size > 1 && size <= bf->buf_size / 16) {
4017 			mlx5_bf_copy(bf->reg + bf->offset, (u64 *)ctrl, ALIGN(size * 16, 64), qp);
4018 			/* wc_wmb(); */
4019 		} else {
4020 			mlx5_write64((__be32 *)ctrl, bf->regreg + bf->offset,
4021 				     MLX5_GET_DOORBELL_LOCK(&bf->lock32));
4022 			/* Make sure doorbells don't leak out of SQ spinlock
4023 			 * and reach the HCA out of order.
4024 			 */
4025 			mmiowb();
4026 		}
4027 		bf->offset ^= bf->buf_size;
4028 		if (bf->need_lock)
4029 			spin_unlock(&bf->lock);
4030 		else
4031 			__release(&bf->lock);
4032 	}
4033 
4034 	spin_unlock_irqrestore(&qp->sq.lock, flags);
4035 
4036 	return err;
4037 }
4038 
4039 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
4040 {
4041 	sig->signature = calc_sig(sig, size);
4042 }
4043 
4044 int mlx5_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
4045 		      struct ib_recv_wr **bad_wr)
4046 {
4047 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4048 	struct mlx5_wqe_data_seg *scat;
4049 	struct mlx5_rwqe_sig *sig;
4050 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4051 	struct mlx5_core_dev *mdev = dev->mdev;
4052 	unsigned long flags;
4053 	int err = 0;
4054 	int nreq;
4055 	int ind;
4056 	int i;
4057 
4058 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4059 		return mlx5_ib_gsi_post_recv(ibqp, wr, bad_wr);
4060 
4061 	spin_lock_irqsave(&qp->rq.lock, flags);
4062 
4063 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
4064 		err = -EIO;
4065 		*bad_wr = wr;
4066 		nreq = 0;
4067 		goto out;
4068 	}
4069 
4070 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
4071 
4072 	for (nreq = 0; wr; nreq++, wr = wr->next) {
4073 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
4074 			err = -ENOMEM;
4075 			*bad_wr = wr;
4076 			goto out;
4077 		}
4078 
4079 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
4080 			err = -EINVAL;
4081 			*bad_wr = wr;
4082 			goto out;
4083 		}
4084 
4085 		scat = get_recv_wqe(qp, ind);
4086 		if (qp->wq_sig)
4087 			scat++;
4088 
4089 		for (i = 0; i < wr->num_sge; i++)
4090 			set_data_ptr_seg(scat + i, wr->sg_list + i);
4091 
4092 		if (i < qp->rq.max_gs) {
4093 			scat[i].byte_count = 0;
4094 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
4095 			scat[i].addr       = 0;
4096 		}
4097 
4098 		if (qp->wq_sig) {
4099 			sig = (struct mlx5_rwqe_sig *)scat;
4100 			set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
4101 		}
4102 
4103 		qp->rq.wrid[ind] = wr->wr_id;
4104 
4105 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
4106 	}
4107 
4108 out:
4109 	if (likely(nreq)) {
4110 		qp->rq.head += nreq;
4111 
4112 		/* Make sure that descriptors are written before
4113 		 * doorbell record.
4114 		 */
4115 		wmb();
4116 
4117 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
4118 	}
4119 
4120 	spin_unlock_irqrestore(&qp->rq.lock, flags);
4121 
4122 	return err;
4123 }
4124 
4125 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
4126 {
4127 	switch (mlx5_state) {
4128 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
4129 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
4130 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
4131 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
4132 	case MLX5_QP_STATE_SQ_DRAINING:
4133 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
4134 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
4135 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
4136 	default:		     return -1;
4137 	}
4138 }
4139 
4140 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
4141 {
4142 	switch (mlx5_mig_state) {
4143 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
4144 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
4145 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
4146 	default: return -1;
4147 	}
4148 }
4149 
4150 static int to_ib_qp_access_flags(int mlx5_flags)
4151 {
4152 	int ib_flags = 0;
4153 
4154 	if (mlx5_flags & MLX5_QP_BIT_RRE)
4155 		ib_flags |= IB_ACCESS_REMOTE_READ;
4156 	if (mlx5_flags & MLX5_QP_BIT_RWE)
4157 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
4158 	if (mlx5_flags & MLX5_QP_BIT_RAE)
4159 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
4160 
4161 	return ib_flags;
4162 }
4163 
4164 static void to_ib_ah_attr(struct mlx5_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
4165 				struct mlx5_qp_path *path)
4166 {
4167 	struct mlx5_core_dev *dev = ibdev->mdev;
4168 
4169 	memset(ib_ah_attr, 0, sizeof(*ib_ah_attr));
4170 	ib_ah_attr->port_num	  = path->port;
4171 
4172 	if (ib_ah_attr->port_num == 0 ||
4173 	    ib_ah_attr->port_num > MLX5_CAP_GEN(dev, num_ports))
4174 		return;
4175 
4176 	ib_ah_attr->sl = path->dci_cfi_prio_sl & 0xf;
4177 
4178 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
4179 	ib_ah_attr->src_path_bits = path->grh_mlid & 0x7f;
4180 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
4181 	ib_ah_attr->ah_flags      = (path->grh_mlid & (1 << 7)) ? IB_AH_GRH : 0;
4182 	if (ib_ah_attr->ah_flags) {
4183 		ib_ah_attr->grh.sgid_index = path->mgid_index;
4184 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
4185 		ib_ah_attr->grh.traffic_class =
4186 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
4187 		ib_ah_attr->grh.flow_label =
4188 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
4189 		memcpy(ib_ah_attr->grh.dgid.raw,
4190 		       path->rgid, sizeof(ib_ah_attr->grh.dgid.raw));
4191 	}
4192 }
4193 
4194 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
4195 					struct mlx5_ib_sq *sq,
4196 					u8 *sq_state)
4197 {
4198 	void *out;
4199 	void *sqc;
4200 	int inlen;
4201 	int err;
4202 
4203 	inlen = MLX5_ST_SZ_BYTES(query_sq_out);
4204 	out = mlx5_vzalloc(inlen);
4205 	if (!out)
4206 		return -ENOMEM;
4207 
4208 	err = mlx5_core_query_sq(dev->mdev, sq->base.mqp.qpn, out);
4209 	if (err)
4210 		goto out;
4211 
4212 	sqc = MLX5_ADDR_OF(query_sq_out, out, sq_context);
4213 	*sq_state = MLX5_GET(sqc, sqc, state);
4214 	sq->state = *sq_state;
4215 
4216 out:
4217 	kvfree(out);
4218 	return err;
4219 }
4220 
4221 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
4222 					struct mlx5_ib_rq *rq,
4223 					u8 *rq_state)
4224 {
4225 	void *out;
4226 	void *rqc;
4227 	int inlen;
4228 	int err;
4229 
4230 	inlen = MLX5_ST_SZ_BYTES(query_rq_out);
4231 	out = mlx5_vzalloc(inlen);
4232 	if (!out)
4233 		return -ENOMEM;
4234 
4235 	err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
4236 	if (err)
4237 		goto out;
4238 
4239 	rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
4240 	*rq_state = MLX5_GET(rqc, rqc, state);
4241 	rq->state = *rq_state;
4242 
4243 out:
4244 	kvfree(out);
4245 	return err;
4246 }
4247 
4248 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
4249 				  struct mlx5_ib_qp *qp, u8 *qp_state)
4250 {
4251 	static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
4252 		[MLX5_RQC_STATE_RST] = {
4253 			[MLX5_SQC_STATE_RST]	= IB_QPS_RESET,
4254 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4255 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE_BAD,
4256 			[MLX5_SQ_STATE_NA]	= IB_QPS_RESET,
4257 		},
4258 		[MLX5_RQC_STATE_RDY] = {
4259 			[MLX5_SQC_STATE_RST]	= MLX5_QP_STATE_BAD,
4260 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4261 			[MLX5_SQC_STATE_ERR]	= IB_QPS_SQE,
4262 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE,
4263 		},
4264 		[MLX5_RQC_STATE_ERR] = {
4265 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE_BAD,
4266 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4267 			[MLX5_SQC_STATE_ERR]	= IB_QPS_ERR,
4268 			[MLX5_SQ_STATE_NA]	= IB_QPS_ERR,
4269 		},
4270 		[MLX5_RQ_STATE_NA] = {
4271 			[MLX5_SQC_STATE_RST]    = IB_QPS_RESET,
4272 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4273 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE,
4274 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE_BAD,
4275 		},
4276 	};
4277 
4278 	*qp_state = sqrq_trans[rq_state][sq_state];
4279 
4280 	if (*qp_state == MLX5_QP_STATE_BAD) {
4281 		WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
4282 		     qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
4283 		     qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
4284 		return -EINVAL;
4285 	}
4286 
4287 	if (*qp_state == MLX5_QP_STATE)
4288 		*qp_state = qp->state;
4289 
4290 	return 0;
4291 }
4292 
4293 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
4294 				     struct mlx5_ib_qp *qp,
4295 				     u8 *raw_packet_qp_state)
4296 {
4297 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
4298 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
4299 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
4300 	int err;
4301 	u8 sq_state = MLX5_SQ_STATE_NA;
4302 	u8 rq_state = MLX5_RQ_STATE_NA;
4303 
4304 	if (qp->sq.wqe_cnt) {
4305 		err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
4306 		if (err)
4307 			return err;
4308 	}
4309 
4310 	if (qp->rq.wqe_cnt) {
4311 		err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
4312 		if (err)
4313 			return err;
4314 	}
4315 
4316 	return sqrq_state_to_qp_state(sq_state, rq_state, qp,
4317 				      raw_packet_qp_state);
4318 }
4319 
4320 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
4321 			 struct ib_qp_attr *qp_attr)
4322 {
4323 	struct mlx5_query_qp_mbox_out *outb;
4324 	struct mlx5_qp_context *context;
4325 	int mlx5_state;
4326 	int err = 0;
4327 
4328 	outb = kzalloc(sizeof(*outb), GFP_KERNEL);
4329 	if (!outb)
4330 		return -ENOMEM;
4331 
4332 	context = &outb->ctx;
4333 	err = mlx5_core_qp_query(dev->mdev, &qp->trans_qp.base.mqp, outb,
4334 				 sizeof(*outb));
4335 	if (err)
4336 		goto out;
4337 
4338 	mlx5_state = be32_to_cpu(context->flags) >> 28;
4339 
4340 	qp->state		     = to_ib_qp_state(mlx5_state);
4341 	qp_attr->path_mtu	     = context->mtu_msgmax >> 5;
4342 	qp_attr->path_mig_state	     =
4343 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
4344 	qp_attr->qkey		     = be32_to_cpu(context->qkey);
4345 	qp_attr->rq_psn		     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
4346 	qp_attr->sq_psn		     = be32_to_cpu(context->next_send_psn) & 0xffffff;
4347 	qp_attr->dest_qp_num	     = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
4348 	qp_attr->qp_access_flags     =
4349 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
4350 
4351 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4352 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
4353 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
4354 		qp_attr->alt_pkey_index =
4355 			be16_to_cpu(context->alt_path.pkey_index);
4356 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
4357 	}
4358 
4359 	qp_attr->pkey_index = be16_to_cpu(context->pri_path.pkey_index);
4360 	qp_attr->port_num = context->pri_path.port;
4361 
4362 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4363 	qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
4364 
4365 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
4366 
4367 	qp_attr->max_dest_rd_atomic =
4368 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
4369 	qp_attr->min_rnr_timer	    =
4370 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
4371 	qp_attr->timeout	    = context->pri_path.ackto_lt >> 3;
4372 	qp_attr->retry_cnt	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
4373 	qp_attr->rnr_retry	    = (be32_to_cpu(context->params1) >> 13) & 0x7;
4374 	qp_attr->alt_timeout	    = context->alt_path.ackto_lt >> 3;
4375 
4376 out:
4377 	kfree(outb);
4378 	return err;
4379 }
4380 
4381 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
4382 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
4383 {
4384 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4385 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4386 	int err = 0;
4387 	u8 raw_packet_qp_state;
4388 
4389 	if (ibqp->rwq_ind_tbl)
4390 		return -ENOSYS;
4391 
4392 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4393 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
4394 					    qp_init_attr);
4395 
4396 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4397 	/*
4398 	 * Wait for any outstanding page faults, in case the user frees memory
4399 	 * based upon this query's result.
4400 	 */
4401 	flush_workqueue(mlx5_ib_page_fault_wq);
4402 #endif
4403 
4404 	mutex_lock(&qp->mutex);
4405 
4406 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
4407 		err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
4408 		if (err)
4409 			goto out;
4410 		qp->state = raw_packet_qp_state;
4411 		qp_attr->port_num = 1;
4412 	} else {
4413 		err = query_qp_attr(dev, qp, qp_attr);
4414 		if (err)
4415 			goto out;
4416 	}
4417 
4418 	qp_attr->qp_state	     = qp->state;
4419 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4420 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4421 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4422 
4423 	if (!ibqp->uobject) {
4424 		qp_attr->cap.max_send_wr  = qp->sq.max_post;
4425 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4426 		qp_init_attr->qp_context = ibqp->qp_context;
4427 	} else {
4428 		qp_attr->cap.max_send_wr  = 0;
4429 		qp_attr->cap.max_send_sge = 0;
4430 	}
4431 
4432 	qp_init_attr->qp_type = ibqp->qp_type;
4433 	qp_init_attr->recv_cq = ibqp->recv_cq;
4434 	qp_init_attr->send_cq = ibqp->send_cq;
4435 	qp_init_attr->srq = ibqp->srq;
4436 	qp_attr->cap.max_inline_data = qp->max_inline_data;
4437 
4438 	qp_init_attr->cap	     = qp_attr->cap;
4439 
4440 	qp_init_attr->create_flags = 0;
4441 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4442 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4443 
4444 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
4445 		qp_init_attr->create_flags |= IB_QP_CREATE_CROSS_CHANNEL;
4446 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
4447 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_SEND;
4448 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
4449 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_RECV;
4450 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
4451 		qp_init_attr->create_flags |= mlx5_ib_create_qp_sqpn_qp1();
4452 
4453 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
4454 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4455 
4456 out:
4457 	mutex_unlock(&qp->mutex);
4458 	return err;
4459 }
4460 
4461 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
4462 					  struct ib_ucontext *context,
4463 					  struct ib_udata *udata)
4464 {
4465 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4466 	struct mlx5_ib_xrcd *xrcd;
4467 	int err;
4468 
4469 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
4470 		return ERR_PTR(-ENOSYS);
4471 
4472 	xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
4473 	if (!xrcd)
4474 		return ERR_PTR(-ENOMEM);
4475 
4476 	err = mlx5_core_xrcd_alloc(dev->mdev, &xrcd->xrcdn);
4477 	if (err) {
4478 		kfree(xrcd);
4479 		return ERR_PTR(-ENOMEM);
4480 	}
4481 
4482 	return &xrcd->ibxrcd;
4483 }
4484 
4485 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
4486 {
4487 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
4488 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
4489 	int err;
4490 
4491 	err = mlx5_core_xrcd_dealloc(dev->mdev, xrcdn);
4492 	if (err) {
4493 		mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
4494 		return err;
4495 	}
4496 
4497 	kfree(xrcd);
4498 
4499 	return 0;
4500 }
4501 
4502 static int  create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
4503 		      struct ib_wq_init_attr *init_attr)
4504 {
4505 	struct mlx5_ib_dev *dev;
4506 	__be64 *rq_pas0;
4507 	void *in;
4508 	void *rqc;
4509 	void *wq;
4510 	int inlen;
4511 	int err;
4512 
4513 	dev = to_mdev(pd->device);
4514 
4515 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
4516 	in = mlx5_vzalloc(inlen);
4517 	if (!in)
4518 		return -ENOMEM;
4519 
4520 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
4521 	MLX5_SET(rqc,  rqc, mem_rq_type,
4522 		 MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
4523 	MLX5_SET(rqc, rqc, user_index, rwq->user_index);
4524 	MLX5_SET(rqc,  rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
4525 	MLX5_SET(rqc,  rqc, state, MLX5_RQC_STATE_RST);
4526 	MLX5_SET(rqc,  rqc, flush_in_error_en, 1);
4527 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
4528 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
4529 	MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
4530 	MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
4531 	MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
4532 	MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
4533 	MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
4534 	MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
4535 	MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
4536 	MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
4537 	rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
4538 	mlx5_ib_populate_pas(dev, rwq->umem, rwq->page_shift, rq_pas0, 0);
4539 	err = mlx5_core_create_rq(dev->mdev, in, inlen, &rwq->rqn);
4540 	kvfree(in);
4541 	return err;
4542 }
4543 
4544 static int set_user_rq_size(struct mlx5_ib_dev *dev,
4545 			    struct ib_wq_init_attr *wq_init_attr,
4546 			    struct mlx5_ib_create_wq *ucmd,
4547 			    struct mlx5_ib_rwq *rwq)
4548 {
4549 	/* Sanity check RQ size before proceeding */
4550 	if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
4551 		return -EINVAL;
4552 
4553 	if (!ucmd->rq_wqe_count)
4554 		return -EINVAL;
4555 
4556 	rwq->wqe_count = ucmd->rq_wqe_count;
4557 	rwq->wqe_shift = ucmd->rq_wqe_shift;
4558 	rwq->buf_size = (rwq->wqe_count << rwq->wqe_shift);
4559 	rwq->log_rq_stride = rwq->wqe_shift;
4560 	rwq->log_rq_size = ilog2(rwq->wqe_count);
4561 	return 0;
4562 }
4563 
4564 static int prepare_user_rq(struct ib_pd *pd,
4565 			   struct ib_wq_init_attr *init_attr,
4566 			   struct ib_udata *udata,
4567 			   struct mlx5_ib_rwq *rwq)
4568 {
4569 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
4570 	struct mlx5_ib_create_wq ucmd = {};
4571 	int err;
4572 	size_t required_cmd_sz;
4573 
4574 	required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
4575 	if (udata->inlen < required_cmd_sz) {
4576 		mlx5_ib_dbg(dev, "invalid inlen\n");
4577 		return -EINVAL;
4578 	}
4579 
4580 	if (udata->inlen > sizeof(ucmd) &&
4581 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4582 				 udata->inlen - sizeof(ucmd))) {
4583 		mlx5_ib_dbg(dev, "inlen is not supported\n");
4584 		return -EOPNOTSUPP;
4585 	}
4586 
4587 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
4588 		mlx5_ib_dbg(dev, "copy failed\n");
4589 		return -EFAULT;
4590 	}
4591 
4592 	if (ucmd.comp_mask) {
4593 		mlx5_ib_dbg(dev, "invalid comp mask\n");
4594 		return -EOPNOTSUPP;
4595 	}
4596 
4597 	if (ucmd.reserved) {
4598 		mlx5_ib_dbg(dev, "invalid reserved\n");
4599 		return -EOPNOTSUPP;
4600 	}
4601 
4602 	err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
4603 	if (err) {
4604 		mlx5_ib_dbg(dev, "err %d\n", err);
4605 		return err;
4606 	}
4607 
4608 	err = create_user_rq(dev, pd, rwq, &ucmd);
4609 	if (err) {
4610 		mlx5_ib_dbg(dev, "err %d\n", err);
4611 		if (err)
4612 			return err;
4613 	}
4614 
4615 	rwq->user_index = ucmd.user_index;
4616 	return 0;
4617 }
4618 
4619 struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
4620 				struct ib_wq_init_attr *init_attr,
4621 				struct ib_udata *udata)
4622 {
4623 	struct mlx5_ib_dev *dev;
4624 	struct mlx5_ib_rwq *rwq;
4625 	struct mlx5_ib_create_wq_resp resp = {};
4626 	size_t min_resp_len;
4627 	int err;
4628 
4629 	if (!udata)
4630 		return ERR_PTR(-ENOSYS);
4631 
4632 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4633 	if (udata->outlen && udata->outlen < min_resp_len)
4634 		return ERR_PTR(-EINVAL);
4635 
4636 	dev = to_mdev(pd->device);
4637 	switch (init_attr->wq_type) {
4638 	case IB_WQT_RQ:
4639 		rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
4640 		if (!rwq)
4641 			return ERR_PTR(-ENOMEM);
4642 		err = prepare_user_rq(pd, init_attr, udata, rwq);
4643 		if (err)
4644 			goto err;
4645 		err = create_rq(rwq, pd, init_attr);
4646 		if (err)
4647 			goto err_user_rq;
4648 		break;
4649 	default:
4650 		mlx5_ib_dbg(dev, "unsupported wq type %d\n",
4651 			    init_attr->wq_type);
4652 		return ERR_PTR(-EINVAL);
4653 	}
4654 
4655 	rwq->ibwq.wq_num = rwq->rqn;
4656 	rwq->ibwq.state = IB_WQS_RESET;
4657 	if (udata->outlen) {
4658 		resp.response_length = offsetof(typeof(resp), response_length) +
4659 				sizeof(resp.response_length);
4660 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4661 		if (err)
4662 			goto err_copy;
4663 	}
4664 
4665 	return &rwq->ibwq;
4666 
4667 err_copy:
4668 	mlx5_core_destroy_rq(dev->mdev, rwq->rqn);
4669 err_user_rq:
4670 	destroy_user_rq(pd, rwq);
4671 err:
4672 	kfree(rwq);
4673 	return ERR_PTR(err);
4674 }
4675 
4676 int mlx5_ib_destroy_wq(struct ib_wq *wq)
4677 {
4678 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
4679 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
4680 
4681 	mlx5_core_destroy_rq(dev->mdev, rwq->rqn);
4682 	destroy_user_rq(wq->pd, rwq);
4683 	kfree(rwq);
4684 
4685 	return 0;
4686 }
4687 
4688 struct ib_rwq_ind_table *mlx5_ib_create_rwq_ind_table(struct ib_device *device,
4689 						      struct ib_rwq_ind_table_init_attr *init_attr,
4690 						      struct ib_udata *udata)
4691 {
4692 	struct mlx5_ib_dev *dev = to_mdev(device);
4693 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl;
4694 	int sz = 1 << init_attr->log_ind_tbl_size;
4695 	struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
4696 	size_t min_resp_len;
4697 	int inlen;
4698 	int err;
4699 	int i;
4700 	u32 *in;
4701 	void *rqtc;
4702 
4703 	if (udata->inlen > 0 &&
4704 	    !ib_is_udata_cleared(udata, 0,
4705 				 udata->inlen))
4706 		return ERR_PTR(-EOPNOTSUPP);
4707 
4708 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4709 	if (udata->outlen && udata->outlen < min_resp_len)
4710 		return ERR_PTR(-EINVAL);
4711 
4712 	rwq_ind_tbl = kzalloc(sizeof(*rwq_ind_tbl), GFP_KERNEL);
4713 	if (!rwq_ind_tbl)
4714 		return ERR_PTR(-ENOMEM);
4715 
4716 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
4717 	in = mlx5_vzalloc(inlen);
4718 	if (!in) {
4719 		err = -ENOMEM;
4720 		goto err;
4721 	}
4722 
4723 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
4724 
4725 	MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
4726 	MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
4727 
4728 	for (i = 0; i < sz; i++)
4729 		MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
4730 
4731 	err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
4732 	kvfree(in);
4733 
4734 	if (err)
4735 		goto err;
4736 
4737 	rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
4738 	if (udata->outlen) {
4739 		resp.response_length = offsetof(typeof(resp), response_length) +
4740 					sizeof(resp.response_length);
4741 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4742 		if (err)
4743 			goto err_copy;
4744 	}
4745 
4746 	return &rwq_ind_tbl->ib_rwq_ind_tbl;
4747 
4748 err_copy:
4749 	mlx5_core_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn);
4750 err:
4751 	kfree(rwq_ind_tbl);
4752 	return ERR_PTR(err);
4753 }
4754 
4755 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4756 {
4757 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
4758 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
4759 
4760 	mlx5_core_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn);
4761 
4762 	kfree(rwq_ind_tbl);
4763 	return 0;
4764 }
4765 
4766 int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
4767 		      u32 wq_attr_mask, struct ib_udata *udata)
4768 {
4769 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
4770 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
4771 	struct mlx5_ib_modify_wq ucmd = {};
4772 	size_t required_cmd_sz;
4773 	int curr_wq_state;
4774 	int wq_state;
4775 	int inlen;
4776 	int err;
4777 	void *rqc;
4778 	void *in;
4779 
4780 	required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
4781 	if (udata->inlen < required_cmd_sz)
4782 		return -EINVAL;
4783 
4784 	if (udata->inlen > sizeof(ucmd) &&
4785 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4786 				 udata->inlen - sizeof(ucmd)))
4787 		return -EOPNOTSUPP;
4788 
4789 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4790 		return -EFAULT;
4791 
4792 	if (ucmd.comp_mask || ucmd.reserved)
4793 		return -EOPNOTSUPP;
4794 
4795 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
4796 	in = mlx5_vzalloc(inlen);
4797 	if (!in)
4798 		return -ENOMEM;
4799 
4800 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
4801 
4802 	curr_wq_state = (wq_attr_mask & IB_WQ_CUR_STATE) ?
4803 		wq_attr->curr_wq_state : wq->state;
4804 	wq_state = (wq_attr_mask & IB_WQ_STATE) ?
4805 		wq_attr->wq_state : curr_wq_state;
4806 	if (curr_wq_state == IB_WQS_ERR)
4807 		curr_wq_state = MLX5_RQC_STATE_ERR;
4808 	if (wq_state == IB_WQS_ERR)
4809 		wq_state = MLX5_RQC_STATE_ERR;
4810 	MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
4811 	MLX5_SET(rqc, rqc, state, wq_state);
4812 
4813 	err = mlx5_core_modify_rq(dev->mdev, rwq->rqn, in, inlen);
4814 	kvfree(in);
4815 	if (!err)
4816 		rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;
4817 
4818 	return err;
4819 }
4820