xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_qp.c (revision ec4deee4e4f2aef1b97d9424f25d04e91fd7dc10)
1 /*-
2  * Copyright (c) 2013-2015, Mellanox Technologies, Ltd.  All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  * $FreeBSD$
26  */
27 
28 #include <linux/module.h>
29 #include <rdma/ib_umem.h>
30 #include <rdma/ib_cache.h>
31 #include <rdma/ib_user_verbs.h>
32 #include "mlx5_ib.h"
33 
34 /* not supported currently */
35 static int wq_signature;
36 
37 enum {
38 	MLX5_IB_ACK_REQ_FREQ	= 8,
39 };
40 
41 enum {
42 	MLX5_IB_DEFAULT_SCHED_QUEUE	= 0x83,
43 	MLX5_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
44 	MLX5_IB_LINK_TYPE_IB		= 0,
45 	MLX5_IB_LINK_TYPE_ETH		= 1
46 };
47 
48 enum {
49 	MLX5_IB_SQ_STRIDE	= 6,
50 };
51 
52 static const u32 mlx5_ib_opcode[] = {
53 	[IB_WR_SEND]				= MLX5_OPCODE_SEND,
54 	[IB_WR_LSO]				= MLX5_OPCODE_LSO,
55 	[IB_WR_SEND_WITH_IMM]			= MLX5_OPCODE_SEND_IMM,
56 	[IB_WR_RDMA_WRITE]			= MLX5_OPCODE_RDMA_WRITE,
57 	[IB_WR_RDMA_WRITE_WITH_IMM]		= MLX5_OPCODE_RDMA_WRITE_IMM,
58 	[IB_WR_RDMA_READ]			= MLX5_OPCODE_RDMA_READ,
59 	[IB_WR_ATOMIC_CMP_AND_SWP]		= MLX5_OPCODE_ATOMIC_CS,
60 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= MLX5_OPCODE_ATOMIC_FA,
61 	[IB_WR_SEND_WITH_INV]			= MLX5_OPCODE_SEND_INVAL,
62 	[IB_WR_LOCAL_INV]			= MLX5_OPCODE_UMR,
63 	[IB_WR_REG_MR]				= MLX5_OPCODE_UMR,
64 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= MLX5_OPCODE_ATOMIC_MASKED_CS,
65 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= MLX5_OPCODE_ATOMIC_MASKED_FA,
66 	[MLX5_IB_WR_UMR]			= MLX5_OPCODE_UMR,
67 };
68 
69 struct mlx5_wqe_eth_pad {
70 	u8 rsvd0[16];
71 };
72 
73 enum raw_qp_set_mask_map {
74 	MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID		= 1UL << 0,
75 };
76 
77 struct mlx5_modify_raw_qp_param {
78 	u16 operation;
79 
80 	u32 set_mask; /* raw_qp_set_mask_map */
81 	u8 rq_q_ctr_id;
82 };
83 
84 static void get_cqs(enum ib_qp_type qp_type,
85 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
86 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq);
87 
88 static int is_qp0(enum ib_qp_type qp_type)
89 {
90 	return qp_type == IB_QPT_SMI;
91 }
92 
93 static int is_sqp(enum ib_qp_type qp_type)
94 {
95 	return is_qp0(qp_type) || is_qp1(qp_type);
96 }
97 
98 static void *get_wqe(struct mlx5_ib_qp *qp, int offset)
99 {
100 	return mlx5_buf_offset(&qp->buf, offset);
101 }
102 
103 static void *get_recv_wqe(struct mlx5_ib_qp *qp, int n)
104 {
105 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
106 }
107 
108 void *mlx5_get_send_wqe(struct mlx5_ib_qp *qp, int n)
109 {
110 	return get_wqe(qp, qp->sq.offset + (n << MLX5_IB_SQ_STRIDE));
111 }
112 
113 /**
114  * mlx5_ib_read_user_wqe() - Copy a user-space WQE to kernel space.
115  *
116  * @qp: QP to copy from.
117  * @send: copy from the send queue when non-zero, use the receive queue
118  *	  otherwise.
119  * @wqe_index:  index to start copying from. For send work queues, the
120  *		wqe_index is in units of MLX5_SEND_WQE_BB.
121  *		For receive work queue, it is the number of work queue
122  *		element in the queue.
123  * @buffer: destination buffer.
124  * @length: maximum number of bytes to copy.
125  *
126  * Copies at least a single WQE, but may copy more data.
127  *
128  * Return: the number of bytes copied, or an error code.
129  */
130 int mlx5_ib_read_user_wqe(struct mlx5_ib_qp *qp, int send, int wqe_index,
131 			  void *buffer, u32 length,
132 			  struct mlx5_ib_qp_base *base)
133 {
134 	struct ib_device *ibdev = qp->ibqp.device;
135 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
136 	struct mlx5_ib_wq *wq = send ? &qp->sq : &qp->rq;
137 	size_t offset;
138 	size_t wq_end;
139 	struct ib_umem *umem = base->ubuffer.umem;
140 	u32 first_copy_length;
141 	int wqe_length;
142 	int ret;
143 
144 	if (wq->wqe_cnt == 0) {
145 		mlx5_ib_dbg(dev, "mlx5_ib_read_user_wqe for a QP with wqe_cnt == 0. qp_type: 0x%x\n",
146 			    qp->ibqp.qp_type);
147 		return -EINVAL;
148 	}
149 
150 	offset = wq->offset + ((wqe_index % wq->wqe_cnt) << wq->wqe_shift);
151 	wq_end = wq->offset + (wq->wqe_cnt << wq->wqe_shift);
152 
153 	if (send && length < sizeof(struct mlx5_wqe_ctrl_seg))
154 		return -EINVAL;
155 
156 	if (offset > umem->length ||
157 	    (send && offset + sizeof(struct mlx5_wqe_ctrl_seg) > umem->length))
158 		return -EINVAL;
159 
160 	first_copy_length = min_t(u32, offset + length, wq_end) - offset;
161 	ret = ib_umem_copy_from(buffer, umem, offset, first_copy_length);
162 	if (ret)
163 		return ret;
164 
165 	if (send) {
166 		struct mlx5_wqe_ctrl_seg *ctrl = buffer;
167 		int ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
168 
169 		wqe_length = ds * MLX5_WQE_DS_UNITS;
170 	} else {
171 		wqe_length = 1 << wq->wqe_shift;
172 	}
173 
174 	if (wqe_length <= first_copy_length)
175 		return first_copy_length;
176 
177 	ret = ib_umem_copy_from(buffer + first_copy_length, umem, wq->offset,
178 				wqe_length - first_copy_length);
179 	if (ret)
180 		return ret;
181 
182 	return wqe_length;
183 }
184 
185 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
186 {
187 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
188 	struct ib_event event;
189 
190 	if (type == MLX5_EVENT_TYPE_PATH_MIG) {
191 		/* This event is only valid for trans_qps */
192 		to_mibqp(qp)->port = to_mibqp(qp)->trans_qp.alt_port;
193 	}
194 
195 	if (ibqp->event_handler) {
196 		event.device     = ibqp->device;
197 		event.element.qp = ibqp;
198 		switch (type) {
199 		case MLX5_EVENT_TYPE_PATH_MIG:
200 			event.event = IB_EVENT_PATH_MIG;
201 			break;
202 		case MLX5_EVENT_TYPE_COMM_EST:
203 			event.event = IB_EVENT_COMM_EST;
204 			break;
205 		case MLX5_EVENT_TYPE_SQ_DRAINED:
206 			event.event = IB_EVENT_SQ_DRAINED;
207 			break;
208 		case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
209 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
210 			break;
211 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
212 			event.event = IB_EVENT_QP_FATAL;
213 			break;
214 		case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
215 			event.event = IB_EVENT_PATH_MIG_ERR;
216 			break;
217 		case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
218 			event.event = IB_EVENT_QP_REQ_ERR;
219 			break;
220 		case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
221 			event.event = IB_EVENT_QP_ACCESS_ERR;
222 			break;
223 		default:
224 			pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
225 			return;
226 		}
227 
228 		ibqp->event_handler(&event, ibqp->qp_context);
229 	}
230 }
231 
232 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
233 		       int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
234 {
235 	int wqe_size;
236 	int wq_size;
237 
238 	/* Sanity check RQ size before proceeding */
239 	if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
240 		return -EINVAL;
241 
242 	if (!has_rq) {
243 		qp->rq.max_gs = 0;
244 		qp->rq.wqe_cnt = 0;
245 		qp->rq.wqe_shift = 0;
246 		cap->max_recv_wr = 0;
247 		cap->max_recv_sge = 0;
248 	} else {
249 		if (ucmd) {
250 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
251 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
252 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
253 			qp->rq.max_post = qp->rq.wqe_cnt;
254 		} else {
255 			wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
256 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
257 			wqe_size = roundup_pow_of_two(wqe_size);
258 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
259 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
260 			qp->rq.wqe_cnt = wq_size / wqe_size;
261 			if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
262 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
263 					    wqe_size,
264 					    MLX5_CAP_GEN(dev->mdev,
265 							 max_wqe_sz_rq));
266 				return -EINVAL;
267 			}
268 			qp->rq.wqe_shift = ilog2(wqe_size);
269 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
270 			qp->rq.max_post = qp->rq.wqe_cnt;
271 		}
272 	}
273 
274 	return 0;
275 }
276 
277 static int sq_overhead(struct ib_qp_init_attr *attr)
278 {
279 	int size = 0;
280 
281 	switch (attr->qp_type) {
282 	case IB_QPT_XRC_INI:
283 		size += sizeof(struct mlx5_wqe_xrc_seg);
284 		/* fall through */
285 	case IB_QPT_RC:
286 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
287 			max(sizeof(struct mlx5_wqe_atomic_seg) +
288 			    sizeof(struct mlx5_wqe_raddr_seg),
289 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
290 			    sizeof(struct mlx5_mkey_seg));
291 		break;
292 
293 	case IB_QPT_XRC_TGT:
294 		return 0;
295 
296 	case IB_QPT_UC:
297 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
298 			max(sizeof(struct mlx5_wqe_raddr_seg),
299 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
300 			    sizeof(struct mlx5_mkey_seg));
301 		break;
302 
303 	case IB_QPT_UD:
304 		if (attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
305 			size += sizeof(struct mlx5_wqe_eth_pad) +
306 				sizeof(struct mlx5_wqe_eth_seg);
307 		/* fall through */
308 	case IB_QPT_SMI:
309 	case MLX5_IB_QPT_HW_GSI:
310 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
311 			sizeof(struct mlx5_wqe_datagram_seg);
312 		break;
313 
314 	case MLX5_IB_QPT_REG_UMR:
315 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
316 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
317 			sizeof(struct mlx5_mkey_seg);
318 		break;
319 
320 	default:
321 		return -EINVAL;
322 	}
323 
324 	return size;
325 }
326 
327 static int calc_send_wqe(struct ib_qp_init_attr *attr)
328 {
329 	int inl_size = 0;
330 	int size;
331 
332 	size = sq_overhead(attr);
333 	if (size < 0)
334 		return size;
335 
336 	if (attr->cap.max_inline_data) {
337 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
338 			attr->cap.max_inline_data;
339 	}
340 
341 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
342 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN &&
343 	    ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
344 			return MLX5_SIG_WQE_SIZE;
345 	else
346 		return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
347 }
348 
349 static int get_send_sge(struct ib_qp_init_attr *attr, int wqe_size)
350 {
351 	int max_sge;
352 
353 	if (attr->qp_type == IB_QPT_RC)
354 		max_sge = (min_t(int, wqe_size, 512) -
355 			   sizeof(struct mlx5_wqe_ctrl_seg) -
356 			   sizeof(struct mlx5_wqe_raddr_seg)) /
357 			sizeof(struct mlx5_wqe_data_seg);
358 	else if (attr->qp_type == IB_QPT_XRC_INI)
359 		max_sge = (min_t(int, wqe_size, 512) -
360 			   sizeof(struct mlx5_wqe_ctrl_seg) -
361 			   sizeof(struct mlx5_wqe_xrc_seg) -
362 			   sizeof(struct mlx5_wqe_raddr_seg)) /
363 			sizeof(struct mlx5_wqe_data_seg);
364 	else
365 		max_sge = (wqe_size - sq_overhead(attr)) /
366 			sizeof(struct mlx5_wqe_data_seg);
367 
368 	return min_t(int, max_sge, wqe_size - sq_overhead(attr) /
369 		     sizeof(struct mlx5_wqe_data_seg));
370 }
371 
372 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
373 			struct mlx5_ib_qp *qp)
374 {
375 	int wqe_size;
376 	int wq_size;
377 
378 	if (!attr->cap.max_send_wr)
379 		return 0;
380 
381 	wqe_size = calc_send_wqe(attr);
382 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
383 	if (wqe_size < 0)
384 		return wqe_size;
385 
386 	if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
387 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
388 			    wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
389 		return -EINVAL;
390 	}
391 
392 	qp->max_inline_data = wqe_size - sq_overhead(attr) -
393 			      sizeof(struct mlx5_wqe_inline_seg);
394 	attr->cap.max_inline_data = qp->max_inline_data;
395 
396 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN)
397 		qp->signature_en = true;
398 
399 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
400 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
401 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
402 		mlx5_ib_dbg(dev, "wqe count(%d) exceeds limits(%d)\n",
403 			    qp->sq.wqe_cnt,
404 			    1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
405 		return -ENOMEM;
406 	}
407 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
408 	qp->sq.max_gs = get_send_sge(attr, wqe_size);
409 	if (qp->sq.max_gs < attr->cap.max_send_sge)
410 		return -ENOMEM;
411 
412 	attr->cap.max_send_sge = qp->sq.max_gs;
413 	qp->sq.max_post = wq_size / wqe_size;
414 	attr->cap.max_send_wr = qp->sq.max_post;
415 
416 	return wq_size;
417 }
418 
419 static int set_user_buf_size(struct mlx5_ib_dev *dev,
420 			    struct mlx5_ib_qp *qp,
421 			    struct mlx5_ib_create_qp *ucmd,
422 			    struct mlx5_ib_qp_base *base,
423 			    struct ib_qp_init_attr *attr)
424 {
425 	int desc_sz = 1 << qp->sq.wqe_shift;
426 
427 	if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
428 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
429 			     desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
430 		return -EINVAL;
431 	}
432 
433 	if (ucmd->sq_wqe_count && ((1 << ilog2(ucmd->sq_wqe_count)) != ucmd->sq_wqe_count)) {
434 		mlx5_ib_warn(dev, "sq_wqe_count %d, sq_wqe_count %d\n",
435 			     ucmd->sq_wqe_count, ucmd->sq_wqe_count);
436 		return -EINVAL;
437 	}
438 
439 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
440 
441 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
442 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
443 			     qp->sq.wqe_cnt,
444 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
445 		return -EINVAL;
446 	}
447 
448 	if (attr->qp_type == IB_QPT_RAW_PACKET) {
449 		base->ubuffer.buf_size = qp->rq.wqe_cnt << qp->rq.wqe_shift;
450 		qp->raw_packet_qp.sq.ubuffer.buf_size = qp->sq.wqe_cnt << 6;
451 	} else {
452 		base->ubuffer.buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
453 					 (qp->sq.wqe_cnt << 6);
454 	}
455 
456 	return 0;
457 }
458 
459 static int qp_has_rq(struct ib_qp_init_attr *attr)
460 {
461 	if (attr->qp_type == IB_QPT_XRC_INI ||
462 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
463 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
464 	    !attr->cap.max_recv_wr)
465 		return 0;
466 
467 	return 1;
468 }
469 
470 static int first_med_uuar(void)
471 {
472 	return 1;
473 }
474 
475 static int next_uuar(int n)
476 {
477 	n++;
478 
479 	while (((n % 4) & 2))
480 		n++;
481 
482 	return n;
483 }
484 
485 static int num_med_uuar(struct mlx5_uuar_info *uuari)
486 {
487 	int n;
488 
489 	n = uuari->num_uars * MLX5_NON_FP_BF_REGS_PER_PAGE -
490 		uuari->num_low_latency_uuars - 1;
491 
492 	return n >= 0 ? n : 0;
493 }
494 
495 static int max_uuari(struct mlx5_uuar_info *uuari)
496 {
497 	return uuari->num_uars * 4;
498 }
499 
500 static int first_hi_uuar(struct mlx5_uuar_info *uuari)
501 {
502 	int med;
503 	int i;
504 	int t;
505 
506 	med = num_med_uuar(uuari);
507 	for (t = 0, i = first_med_uuar();; i = next_uuar(i)) {
508 		t++;
509 		if (t == med)
510 			return next_uuar(i);
511 	}
512 
513 	return 0;
514 }
515 
516 static int alloc_high_class_uuar(struct mlx5_uuar_info *uuari)
517 {
518 	int i;
519 
520 	for (i = first_hi_uuar(uuari); i < max_uuari(uuari); i = next_uuar(i)) {
521 		if (!test_bit(i, uuari->bitmap)) {
522 			set_bit(i, uuari->bitmap);
523 			uuari->count[i]++;
524 			return i;
525 		}
526 	}
527 
528 	return -ENOMEM;
529 }
530 
531 static int alloc_med_class_uuar(struct mlx5_uuar_info *uuari)
532 {
533 	int minidx = first_med_uuar();
534 	int i;
535 
536 	for (i = first_med_uuar(); i < first_hi_uuar(uuari); i = next_uuar(i)) {
537 		if (uuari->count[i] < uuari->count[minidx])
538 			minidx = i;
539 	}
540 
541 	uuari->count[minidx]++;
542 	return minidx;
543 }
544 
545 static int alloc_uuar(struct mlx5_uuar_info *uuari,
546 		      enum mlx5_ib_latency_class lat)
547 {
548 	int uuarn = -EINVAL;
549 
550 	mutex_lock(&uuari->lock);
551 	switch (lat) {
552 	case MLX5_IB_LATENCY_CLASS_LOW:
553 		uuarn = 0;
554 		uuari->count[uuarn]++;
555 		break;
556 
557 	case MLX5_IB_LATENCY_CLASS_MEDIUM:
558 		if (uuari->ver < 2)
559 			uuarn = -ENOMEM;
560 		else
561 			uuarn = alloc_med_class_uuar(uuari);
562 		break;
563 
564 	case MLX5_IB_LATENCY_CLASS_HIGH:
565 		if (uuari->ver < 2)
566 			uuarn = -ENOMEM;
567 		else
568 			uuarn = alloc_high_class_uuar(uuari);
569 		break;
570 
571 	case MLX5_IB_LATENCY_CLASS_FAST_PATH:
572 		uuarn = 2;
573 		break;
574 	}
575 	mutex_unlock(&uuari->lock);
576 
577 	return uuarn;
578 }
579 
580 static void free_med_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
581 {
582 	clear_bit(uuarn, uuari->bitmap);
583 	--uuari->count[uuarn];
584 }
585 
586 static void free_high_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
587 {
588 	clear_bit(uuarn, uuari->bitmap);
589 	--uuari->count[uuarn];
590 }
591 
592 static void free_uuar(struct mlx5_uuar_info *uuari, int uuarn)
593 {
594 	int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
595 	int high_uuar = nuuars - uuari->num_low_latency_uuars;
596 
597 	mutex_lock(&uuari->lock);
598 	if (uuarn == 0) {
599 		--uuari->count[uuarn];
600 		goto out;
601 	}
602 
603 	if (uuarn < high_uuar) {
604 		free_med_class_uuar(uuari, uuarn);
605 		goto out;
606 	}
607 
608 	free_high_class_uuar(uuari, uuarn);
609 
610 out:
611 	mutex_unlock(&uuari->lock);
612 }
613 
614 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
615 {
616 	switch (state) {
617 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
618 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
619 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
620 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
621 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
622 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
623 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
624 	default:		return -1;
625 	}
626 }
627 
628 static int to_mlx5_st(enum ib_qp_type type)
629 {
630 	switch (type) {
631 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
632 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
633 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
634 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
635 	case IB_QPT_XRC_INI:
636 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
637 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
638 	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
639 	case IB_QPT_RAW_IPV6:		return MLX5_QP_ST_RAW_IPV6;
640 	case IB_QPT_RAW_PACKET:
641 	case IB_QPT_RAW_ETHERTYPE:	return MLX5_QP_ST_RAW_ETHERTYPE;
642 	case IB_QPT_MAX:
643 	default:		return -EINVAL;
644 	}
645 }
646 
647 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
648 			     struct mlx5_ib_cq *recv_cq);
649 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
650 			       struct mlx5_ib_cq *recv_cq);
651 
652 static int uuarn_to_uar_index(struct mlx5_uuar_info *uuari, int uuarn)
653 {
654 	return uuari->uars[uuarn / MLX5_BF_REGS_PER_PAGE].index;
655 }
656 
657 static int mlx5_ib_umem_get(struct mlx5_ib_dev *dev,
658 			    struct ib_pd *pd,
659 			    unsigned long addr, size_t size,
660 			    struct ib_umem **umem,
661 			    int *npages, int *page_shift, int *ncont,
662 			    u32 *offset)
663 {
664 	int err;
665 
666 	*umem = ib_umem_get(pd->uobject->context, addr, size, 0, 0);
667 	if (IS_ERR(*umem)) {
668 		mlx5_ib_dbg(dev, "umem_get failed\n");
669 		return PTR_ERR(*umem);
670 	}
671 
672 	mlx5_ib_cont_pages(*umem, addr, npages, page_shift, ncont, NULL);
673 
674 	err = mlx5_ib_get_buf_offset(addr, *page_shift, offset);
675 	if (err) {
676 		mlx5_ib_warn(dev, "bad offset\n");
677 		goto err_umem;
678 	}
679 
680 	mlx5_ib_dbg(dev, "addr 0x%lx, size %zu, npages %d, page_shift %d, ncont %d, offset %d\n",
681 		    addr, size, *npages, *page_shift, *ncont, *offset);
682 
683 	return 0;
684 
685 err_umem:
686 	ib_umem_release(*umem);
687 	*umem = NULL;
688 
689 	return err;
690 }
691 
692 static void destroy_user_rq(struct ib_pd *pd, struct mlx5_ib_rwq *rwq)
693 {
694 	struct mlx5_ib_ucontext *context;
695 
696 	context = to_mucontext(pd->uobject->context);
697 	mlx5_ib_db_unmap_user(context, &rwq->db);
698 	if (rwq->umem)
699 		ib_umem_release(rwq->umem);
700 }
701 
702 static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
703 			  struct mlx5_ib_rwq *rwq,
704 			  struct mlx5_ib_create_wq *ucmd)
705 {
706 	struct mlx5_ib_ucontext *context;
707 	int page_shift = 0;
708 	int npages;
709 	u32 offset = 0;
710 	int ncont = 0;
711 	int err;
712 
713 	if (!ucmd->buf_addr)
714 		return -EINVAL;
715 
716 	context = to_mucontext(pd->uobject->context);
717 	rwq->umem = ib_umem_get(pd->uobject->context, ucmd->buf_addr,
718 			       rwq->buf_size, 0, 0);
719 	if (IS_ERR(rwq->umem)) {
720 		mlx5_ib_dbg(dev, "umem_get failed\n");
721 		err = PTR_ERR(rwq->umem);
722 		return err;
723 	}
724 
725 	mlx5_ib_cont_pages(rwq->umem, ucmd->buf_addr, &npages, &page_shift,
726 			   &ncont, NULL);
727 	err = mlx5_ib_get_buf_offset(ucmd->buf_addr, page_shift,
728 				     &rwq->rq_page_offset);
729 	if (err) {
730 		mlx5_ib_warn(dev, "bad offset\n");
731 		goto err_umem;
732 	}
733 
734 	rwq->rq_num_pas = ncont;
735 	rwq->page_shift = page_shift;
736 	rwq->log_page_size =  page_shift - MLX5_ADAPTER_PAGE_SHIFT;
737 	rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
738 
739 	mlx5_ib_dbg(dev, "addr 0x%llx, size %zd, npages %d, page_shift %d, ncont %d, offset %d\n",
740 		    (unsigned long long)ucmd->buf_addr, rwq->buf_size,
741 		    npages, page_shift, ncont, offset);
742 
743 	err = mlx5_ib_db_map_user(context, ucmd->db_addr, &rwq->db);
744 	if (err) {
745 		mlx5_ib_dbg(dev, "map failed\n");
746 		goto err_umem;
747 	}
748 
749 	rwq->create_type = MLX5_WQ_USER;
750 	return 0;
751 
752 err_umem:
753 	ib_umem_release(rwq->umem);
754 	return err;
755 }
756 
757 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
758 			  struct mlx5_ib_qp *qp, struct ib_udata *udata,
759 			  struct ib_qp_init_attr *attr,
760 			  u32 **in,
761 			  struct mlx5_ib_create_qp_resp *resp, int *inlen,
762 			  struct mlx5_ib_qp_base *base)
763 {
764 	struct mlx5_ib_ucontext *context;
765 	struct mlx5_ib_create_qp ucmd;
766 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
767 	int page_shift = 0;
768 	int uar_index;
769 	int npages;
770 	u32 offset = 0;
771 	int uuarn;
772 	int ncont = 0;
773 	__be64 *pas;
774 	void *qpc;
775 	int err;
776 
777 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
778 	if (err) {
779 		mlx5_ib_dbg(dev, "copy failed\n");
780 		return err;
781 	}
782 
783 	context = to_mucontext(pd->uobject->context);
784 	/*
785 	 * TBD: should come from the verbs when we have the API
786 	 */
787 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
788 		/* In CROSS_CHANNEL CQ and QP must use the same UAR */
789 		uuarn = MLX5_CROSS_CHANNEL_UUAR;
790 	else {
791 		uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_HIGH);
792 		if (uuarn < 0) {
793 			mlx5_ib_dbg(dev, "failed to allocate low latency UUAR\n");
794 			mlx5_ib_dbg(dev, "reverting to medium latency\n");
795 			uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_MEDIUM);
796 			if (uuarn < 0) {
797 				mlx5_ib_dbg(dev, "failed to allocate medium latency UUAR\n");
798 				mlx5_ib_dbg(dev, "reverting to high latency\n");
799 				uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_LOW);
800 				if (uuarn < 0) {
801 					mlx5_ib_warn(dev, "uuar allocation failed\n");
802 					return uuarn;
803 				}
804 			}
805 		}
806 	}
807 
808 	uar_index = uuarn_to_uar_index(&context->uuari, uuarn);
809 	mlx5_ib_dbg(dev, "uuarn 0x%x, uar_index 0x%x\n", uuarn, uar_index);
810 
811 	qp->rq.offset = 0;
812 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
813 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
814 
815 	err = set_user_buf_size(dev, qp, &ucmd, base, attr);
816 	if (err)
817 		goto err_uuar;
818 
819 	if (ucmd.buf_addr && ubuffer->buf_size) {
820 		ubuffer->buf_addr = ucmd.buf_addr;
821 		err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr,
822 				       ubuffer->buf_size,
823 				       &ubuffer->umem, &npages, &page_shift,
824 				       &ncont, &offset);
825 		if (err)
826 			goto err_uuar;
827 	} else {
828 		ubuffer->umem = NULL;
829 	}
830 
831 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
832 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * ncont;
833 	*in = mlx5_vzalloc(*inlen);
834 	if (!*in) {
835 		err = -ENOMEM;
836 		goto err_umem;
837 	}
838 
839 	pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas);
840 	if (ubuffer->umem)
841 		mlx5_ib_populate_pas(dev, ubuffer->umem, page_shift, pas, 0);
842 
843 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
844 
845 	MLX5_SET(qpc, qpc, log_page_size, page_shift - MLX5_ADAPTER_PAGE_SHIFT);
846 	MLX5_SET(qpc, qpc, page_offset, offset);
847 
848 	MLX5_SET(qpc, qpc, uar_page, uar_index);
849 	resp->uuar_index = uuarn;
850 	qp->uuarn = uuarn;
851 
852 	err = mlx5_ib_db_map_user(context, ucmd.db_addr, &qp->db);
853 	if (err) {
854 		mlx5_ib_dbg(dev, "map failed\n");
855 		goto err_free;
856 	}
857 
858 	err = ib_copy_to_udata(udata, resp, sizeof(*resp));
859 	if (err) {
860 		mlx5_ib_dbg(dev, "copy failed\n");
861 		goto err_unmap;
862 	}
863 	qp->create_type = MLX5_QP_USER;
864 
865 	return 0;
866 
867 err_unmap:
868 	mlx5_ib_db_unmap_user(context, &qp->db);
869 
870 err_free:
871 	kvfree(*in);
872 
873 err_umem:
874 	if (ubuffer->umem)
875 		ib_umem_release(ubuffer->umem);
876 
877 err_uuar:
878 	free_uuar(&context->uuari, uuarn);
879 	return err;
880 }
881 
882 static void destroy_qp_user(struct ib_pd *pd, struct mlx5_ib_qp *qp,
883 			    struct mlx5_ib_qp_base *base)
884 {
885 	struct mlx5_ib_ucontext *context;
886 
887 	context = to_mucontext(pd->uobject->context);
888 	mlx5_ib_db_unmap_user(context, &qp->db);
889 	if (base->ubuffer.umem)
890 		ib_umem_release(base->ubuffer.umem);
891 	free_uuar(&context->uuari, qp->uuarn);
892 }
893 
894 static int create_kernel_qp(struct mlx5_ib_dev *dev,
895 			    struct ib_qp_init_attr *init_attr,
896 			    struct mlx5_ib_qp *qp,
897 			    u32 **in, int *inlen,
898 			    struct mlx5_ib_qp_base *base)
899 {
900 	enum mlx5_ib_latency_class lc = MLX5_IB_LATENCY_CLASS_LOW;
901 	struct mlx5_uuar_info *uuari;
902 	int uar_index;
903 	void *qpc;
904 	int uuarn;
905 	int err;
906 
907 	uuari = &dev->mdev->priv.uuari;
908 	if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN |
909 					IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
910 					IB_QP_CREATE_IPOIB_UD_LSO |
911 					mlx5_ib_create_qp_sqpn_qp1()))
912 		return -EINVAL;
913 
914 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
915 		lc = MLX5_IB_LATENCY_CLASS_FAST_PATH;
916 
917 	uuarn = alloc_uuar(uuari, lc);
918 	if (uuarn < 0) {
919 		mlx5_ib_dbg(dev, "\n");
920 		return -ENOMEM;
921 	}
922 
923 	qp->bf = &uuari->bfs[uuarn];
924 	uar_index = qp->bf->uar->index;
925 
926 	err = calc_sq_size(dev, init_attr, qp);
927 	if (err < 0) {
928 		mlx5_ib_dbg(dev, "err %d\n", err);
929 		goto err_uuar;
930 	}
931 
932 	qp->rq.offset = 0;
933 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
934 	base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
935 
936 	err = mlx5_buf_alloc(dev->mdev, base->ubuffer.buf_size,
937 	    2 * PAGE_SIZE, &qp->buf);
938 	if (err) {
939 		mlx5_ib_dbg(dev, "err %d\n", err);
940 		goto err_uuar;
941 	}
942 
943 	qp->sq.qend = mlx5_get_send_wqe(qp, qp->sq.wqe_cnt);
944 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
945 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * qp->buf.npages;
946 	*in = mlx5_vzalloc(*inlen);
947 	if (!*in) {
948 		err = -ENOMEM;
949 		goto err_buf;
950 	}
951 
952 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
953 	MLX5_SET(qpc, qpc, uar_page, uar_index);
954 	MLX5_SET(qpc, qpc, log_page_size, qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
955 
956 	/* Set "fast registration enabled" for all kernel QPs */
957 	MLX5_SET(qpc, qpc, fre, 1);
958 	MLX5_SET(qpc, qpc, rlky, 1);
959 
960 	if (init_attr->create_flags & mlx5_ib_create_qp_sqpn_qp1()) {
961 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
962 		qp->flags |= MLX5_IB_QP_SQPN_QP1;
963 	}
964 
965 	mlx5_fill_page_array(&qp->buf,
966 			     (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas));
967 
968 	err = mlx5_db_alloc(dev->mdev, &qp->db);
969 	if (err) {
970 		mlx5_ib_dbg(dev, "err %d\n", err);
971 		goto err_free;
972 	}
973 
974 	qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
975 	qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
976 	qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
977 	qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
978 	qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
979 
980 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
981 	    !qp->sq.w_list || !qp->sq.wqe_head) {
982 		err = -ENOMEM;
983 		goto err_wrid;
984 	}
985 	qp->create_type = MLX5_QP_KERNEL;
986 
987 	return 0;
988 
989 err_wrid:
990 	mlx5_db_free(dev->mdev, &qp->db);
991 	kfree(qp->sq.wqe_head);
992 	kfree(qp->sq.w_list);
993 	kfree(qp->sq.wrid);
994 	kfree(qp->sq.wr_data);
995 	kfree(qp->rq.wrid);
996 
997 err_free:
998 	kvfree(*in);
999 
1000 err_buf:
1001 	mlx5_buf_free(dev->mdev, &qp->buf);
1002 
1003 err_uuar:
1004 	free_uuar(&dev->mdev->priv.uuari, uuarn);
1005 	return err;
1006 }
1007 
1008 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1009 {
1010 	mlx5_db_free(dev->mdev, &qp->db);
1011 	kfree(qp->sq.wqe_head);
1012 	kfree(qp->sq.w_list);
1013 	kfree(qp->sq.wrid);
1014 	kfree(qp->sq.wr_data);
1015 	kfree(qp->rq.wrid);
1016 	mlx5_buf_free(dev->mdev, &qp->buf);
1017 	free_uuar(&dev->mdev->priv.uuari, qp->bf->uuarn);
1018 }
1019 
1020 static u32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
1021 {
1022 	if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
1023 	    (attr->qp_type == IB_QPT_XRC_INI))
1024 		return MLX5_SRQ_RQ;
1025 	else if (!qp->has_rq)
1026 		return MLX5_ZERO_LEN_RQ;
1027 	else
1028 		return MLX5_NON_ZERO_RQ;
1029 }
1030 
1031 static int is_connected(enum ib_qp_type qp_type)
1032 {
1033 	if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC)
1034 		return 1;
1035 
1036 	return 0;
1037 }
1038 
1039 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1040 				    struct mlx5_ib_sq *sq, u32 tdn)
1041 {
1042 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {0};
1043 	void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1044 
1045 	MLX5_SET(tisc, tisc, transport_domain, tdn);
1046 	return mlx5_core_create_tis(dev->mdev, in, sizeof(in), &sq->tisn);
1047 }
1048 
1049 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1050 				      struct mlx5_ib_sq *sq)
1051 {
1052 	mlx5_core_destroy_tis(dev->mdev, sq->tisn);
1053 }
1054 
1055 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1056 				   struct mlx5_ib_sq *sq, void *qpin,
1057 				   struct ib_pd *pd)
1058 {
1059 	struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1060 	__be64 *pas;
1061 	void *in;
1062 	void *sqc;
1063 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1064 	void *wq;
1065 	int inlen;
1066 	int err;
1067 	int page_shift = 0;
1068 	int npages;
1069 	int ncont = 0;
1070 	u32 offset = 0;
1071 
1072 	err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr, ubuffer->buf_size,
1073 			       &sq->ubuffer.umem, &npages, &page_shift,
1074 			       &ncont, &offset);
1075 	if (err)
1076 		return err;
1077 
1078 	inlen = MLX5_ST_SZ_BYTES(create_sq_in) + sizeof(u64) * ncont;
1079 	in = mlx5_vzalloc(inlen);
1080 	if (!in) {
1081 		err = -ENOMEM;
1082 		goto err_umem;
1083 	}
1084 
1085 	sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1086 	MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1087 	MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1088 	MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1089 	MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1090 	MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1091 	MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1092 
1093 	wq = MLX5_ADDR_OF(sqc, sqc, wq);
1094 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1095 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1096 	MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1097 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1098 	MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1099 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1100 	MLX5_SET(wq, wq, log_wq_pg_sz,  page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1101 	MLX5_SET(wq, wq, page_offset, offset);
1102 
1103 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1104 	mlx5_ib_populate_pas(dev, sq->ubuffer.umem, page_shift, pas, 0);
1105 
1106 	err = mlx5_core_create_sq_tracked(dev->mdev, in, inlen, &sq->base.mqp);
1107 
1108 	kvfree(in);
1109 
1110 	if (err)
1111 		goto err_umem;
1112 
1113 	return 0;
1114 
1115 err_umem:
1116 	ib_umem_release(sq->ubuffer.umem);
1117 	sq->ubuffer.umem = NULL;
1118 
1119 	return err;
1120 }
1121 
1122 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1123 				     struct mlx5_ib_sq *sq)
1124 {
1125 	mlx5_core_destroy_sq_tracked(dev->mdev, &sq->base.mqp);
1126 	ib_umem_release(sq->ubuffer.umem);
1127 }
1128 
1129 static int get_rq_pas_size(void *qpc)
1130 {
1131 	u32 log_page_size = MLX5_GET(qpc, qpc, log_page_size) + 12;
1132 	u32 log_rq_stride = MLX5_GET(qpc, qpc, log_rq_stride);
1133 	u32 log_rq_size   = MLX5_GET(qpc, qpc, log_rq_size);
1134 	u32 page_offset   = MLX5_GET(qpc, qpc, page_offset);
1135 	u32 po_quanta	  = 1 << (log_page_size - 6);
1136 	u32 rq_sz	  = 1 << (log_rq_size + 4 + log_rq_stride);
1137 	u32 page_size	  = 1 << log_page_size;
1138 	u32 rq_sz_po      = rq_sz + (page_offset * po_quanta);
1139 	u32 rq_num_pas	  = (rq_sz_po + page_size - 1) / page_size;
1140 
1141 	return rq_num_pas * sizeof(u64);
1142 }
1143 
1144 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1145 				   struct mlx5_ib_rq *rq, void *qpin)
1146 {
1147 	struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1148 	__be64 *pas;
1149 	__be64 *qp_pas;
1150 	void *in;
1151 	void *rqc;
1152 	void *wq;
1153 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1154 	int inlen;
1155 	int err;
1156 	u32 rq_pas_size = get_rq_pas_size(qpc);
1157 
1158 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + rq_pas_size;
1159 	in = mlx5_vzalloc(inlen);
1160 	if (!in)
1161 		return -ENOMEM;
1162 
1163 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1164 	MLX5_SET(rqc, rqc, vlan_strip_disable, 1);
1165 	MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_RQ_TYPE_MEMORY_RQ_INLINE);
1166 	MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1167 	MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1168 	MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1169 	MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1170 
1171 	if (mqp->flags & MLX5_IB_QP_CAP_SCATTER_FCS)
1172 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
1173 
1174 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
1175 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1176 	MLX5_SET(wq, wq, end_padding_mode,
1177 		 MLX5_GET(qpc, qpc, end_padding_mode));
1178 	MLX5_SET(wq, wq, page_offset, MLX5_GET(qpc, qpc, page_offset));
1179 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1180 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1181 	MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1182 	MLX5_SET(wq, wq, log_wq_pg_sz, MLX5_GET(qpc, qpc, log_page_size));
1183 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1184 
1185 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1186 	qp_pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, qpin, pas);
1187 	memcpy(pas, qp_pas, rq_pas_size);
1188 
1189 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rq->base.mqp);
1190 
1191 	kvfree(in);
1192 
1193 	return err;
1194 }
1195 
1196 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1197 				     struct mlx5_ib_rq *rq)
1198 {
1199 	mlx5_core_destroy_rq_tracked(dev->mdev, &rq->base.mqp);
1200 }
1201 
1202 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1203 				    struct mlx5_ib_rq *rq, u32 tdn)
1204 {
1205 	u32 *in;
1206 	void *tirc;
1207 	int inlen;
1208 	int err;
1209 
1210 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1211 	in = mlx5_vzalloc(inlen);
1212 	if (!in)
1213 		return -ENOMEM;
1214 
1215 	tirc = MLX5_ADDR_OF(create_tir_in, in, tir_context);
1216 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1217 	MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1218 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1219 
1220 	err = mlx5_core_create_tir(dev->mdev, in, inlen, &rq->tirn);
1221 
1222 	kvfree(in);
1223 
1224 	return err;
1225 }
1226 
1227 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1228 				      struct mlx5_ib_rq *rq)
1229 {
1230 	mlx5_core_destroy_tir(dev->mdev, rq->tirn);
1231 }
1232 
1233 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1234 				u32 *in,
1235 				struct ib_pd *pd)
1236 {
1237 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1238 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1239 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1240 	struct ib_uobject *uobj = pd->uobject;
1241 	struct ib_ucontext *ucontext = uobj->context;
1242 	struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1243 	int err;
1244 	u32 tdn = mucontext->tdn;
1245 
1246 	if (qp->sq.wqe_cnt) {
1247 		err = create_raw_packet_qp_tis(dev, sq, tdn);
1248 		if (err)
1249 			return err;
1250 
1251 		err = create_raw_packet_qp_sq(dev, sq, in, pd);
1252 		if (err)
1253 			goto err_destroy_tis;
1254 
1255 		sq->base.container_mibqp = qp;
1256 	}
1257 
1258 	if (qp->rq.wqe_cnt) {
1259 		rq->base.container_mibqp = qp;
1260 
1261 		err = create_raw_packet_qp_rq(dev, rq, in);
1262 		if (err)
1263 			goto err_destroy_sq;
1264 
1265 
1266 		err = create_raw_packet_qp_tir(dev, rq, tdn);
1267 		if (err)
1268 			goto err_destroy_rq;
1269 	}
1270 
1271 	qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1272 						     rq->base.mqp.qpn;
1273 
1274 	return 0;
1275 
1276 err_destroy_rq:
1277 	destroy_raw_packet_qp_rq(dev, rq);
1278 err_destroy_sq:
1279 	if (!qp->sq.wqe_cnt)
1280 		return err;
1281 	destroy_raw_packet_qp_sq(dev, sq);
1282 err_destroy_tis:
1283 	destroy_raw_packet_qp_tis(dev, sq);
1284 
1285 	return err;
1286 }
1287 
1288 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1289 				  struct mlx5_ib_qp *qp)
1290 {
1291 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1292 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1293 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1294 
1295 	if (qp->rq.wqe_cnt) {
1296 		destroy_raw_packet_qp_tir(dev, rq);
1297 		destroy_raw_packet_qp_rq(dev, rq);
1298 	}
1299 
1300 	if (qp->sq.wqe_cnt) {
1301 		destroy_raw_packet_qp_sq(dev, sq);
1302 		destroy_raw_packet_qp_tis(dev, sq);
1303 	}
1304 }
1305 
1306 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1307 				    struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1308 {
1309 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1310 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1311 
1312 	sq->sq = &qp->sq;
1313 	rq->rq = &qp->rq;
1314 	sq->doorbell = &qp->db;
1315 	rq->doorbell = &qp->db;
1316 }
1317 
1318 static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1319 {
1320 	mlx5_core_destroy_tir(dev->mdev, qp->rss_qp.tirn);
1321 }
1322 
1323 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1324 				 struct ib_pd *pd,
1325 				 struct ib_qp_init_attr *init_attr,
1326 				 struct ib_udata *udata)
1327 {
1328 	struct ib_uobject *uobj = pd->uobject;
1329 	struct ib_ucontext *ucontext = uobj->context;
1330 	struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1331 	struct mlx5_ib_create_qp_resp resp = {};
1332 	int inlen;
1333 	int err;
1334 	u32 *in;
1335 	void *tirc;
1336 	void *hfso;
1337 	u32 selected_fields = 0;
1338 	size_t min_resp_len;
1339 	u32 tdn = mucontext->tdn;
1340 	struct mlx5_ib_create_qp_rss ucmd = {};
1341 	size_t required_cmd_sz;
1342 
1343 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1344 		return -EOPNOTSUPP;
1345 
1346 	if (init_attr->create_flags || init_attr->send_cq)
1347 		return -EINVAL;
1348 
1349 	min_resp_len = offsetof(typeof(resp), uuar_index) + sizeof(resp.uuar_index);
1350 	if (udata->outlen < min_resp_len)
1351 		return -EINVAL;
1352 
1353 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) + sizeof(ucmd.reserved1);
1354 	if (udata->inlen < required_cmd_sz) {
1355 		mlx5_ib_dbg(dev, "invalid inlen\n");
1356 		return -EINVAL;
1357 	}
1358 
1359 	if (udata->inlen > sizeof(ucmd) &&
1360 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
1361 				 udata->inlen - sizeof(ucmd))) {
1362 		mlx5_ib_dbg(dev, "inlen is not supported\n");
1363 		return -EOPNOTSUPP;
1364 	}
1365 
1366 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
1367 		mlx5_ib_dbg(dev, "copy failed\n");
1368 		return -EFAULT;
1369 	}
1370 
1371 	if (ucmd.comp_mask) {
1372 		mlx5_ib_dbg(dev, "invalid comp mask\n");
1373 		return -EOPNOTSUPP;
1374 	}
1375 
1376 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)) || ucmd.reserved1) {
1377 		mlx5_ib_dbg(dev, "invalid reserved\n");
1378 		return -EOPNOTSUPP;
1379 	}
1380 
1381 	err = ib_copy_to_udata(udata, &resp, min_resp_len);
1382 	if (err) {
1383 		mlx5_ib_dbg(dev, "copy failed\n");
1384 		return -EINVAL;
1385 	}
1386 
1387 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1388 	in = mlx5_vzalloc(inlen);
1389 	if (!in)
1390 		return -ENOMEM;
1391 
1392 	tirc = MLX5_ADDR_OF(create_tir_in, in, tir_context);
1393 	MLX5_SET(tirc, tirc, disp_type,
1394 		 MLX5_TIRC_DISP_TYPE_INDIRECT);
1395 	MLX5_SET(tirc, tirc, indirect_table,
1396 		 init_attr->rwq_ind_tbl->ind_tbl_num);
1397 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1398 
1399 	hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1400 	switch (ucmd.rx_hash_function) {
1401 	case MLX5_RX_HASH_FUNC_TOEPLITZ:
1402 	{
1403 		void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1404 		size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1405 
1406 		if (len != ucmd.rx_key_len) {
1407 			err = -EINVAL;
1408 			goto err;
1409 		}
1410 
1411 		MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FUNC_TOEPLITZ);
1412 		MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
1413 		memcpy(rss_key, ucmd.rx_hash_key, len);
1414 		break;
1415 	}
1416 	default:
1417 		err = -EOPNOTSUPP;
1418 		goto err;
1419 	}
1420 
1421 	if (!ucmd.rx_hash_fields_mask) {
1422 		/* special case when this TIR serves as steering entry without hashing */
1423 		if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1424 			goto create_tir;
1425 		err = -EINVAL;
1426 		goto err;
1427 	}
1428 
1429 	if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1430 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1431 	     ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1432 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1433 		err = -EINVAL;
1434 		goto err;
1435 	}
1436 
1437 	/* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1438 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1439 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1440 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1441 			 MLX5_L3_PROT_TYPE_IPV4);
1442 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1443 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1444 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1445 			 MLX5_L3_PROT_TYPE_IPV6);
1446 
1447 	if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1448 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP)) &&
1449 	     ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1450 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))) {
1451 		err = -EINVAL;
1452 		goto err;
1453 	}
1454 
1455 	/* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1456 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1457 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1458 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1459 			 MLX5_L4_PROT_TYPE_TCP);
1460 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1461 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1462 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1463 			 MLX5_L4_PROT_TYPE_UDP);
1464 
1465 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1466 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1467 		selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1468 
1469 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1470 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1471 		selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1472 
1473 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1474 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1475 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1476 
1477 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1478 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1479 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1480 
1481 	MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1482 
1483 create_tir:
1484 	err = mlx5_core_create_tir(dev->mdev, in, inlen, &qp->rss_qp.tirn);
1485 
1486 	if (err)
1487 		goto err;
1488 
1489 	kvfree(in);
1490 	/* qpn is reserved for that QP */
1491 	qp->trans_qp.base.mqp.qpn = 0;
1492 	qp->flags |= MLX5_IB_QP_RSS;
1493 	return 0;
1494 
1495 err:
1496 	kvfree(in);
1497 	return err;
1498 }
1499 
1500 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1501 			    struct ib_qp_init_attr *init_attr,
1502 			    struct ib_udata *udata, struct mlx5_ib_qp *qp)
1503 {
1504 	struct mlx5_ib_resources *devr = &dev->devr;
1505 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1506 	struct mlx5_core_dev *mdev = dev->mdev;
1507 	struct mlx5_ib_create_qp_resp resp;
1508 	struct mlx5_ib_cq *send_cq;
1509 	struct mlx5_ib_cq *recv_cq;
1510 	unsigned long flags;
1511 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
1512 	struct mlx5_ib_create_qp ucmd;
1513 	struct mlx5_ib_qp_base *base;
1514 	void *qpc;
1515 	u32 *in;
1516 	int err;
1517 
1518 	base = init_attr->qp_type == IB_QPT_RAW_PACKET ?
1519 	       &qp->raw_packet_qp.rq.base :
1520 	       &qp->trans_qp.base;
1521 
1522 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1523 		mlx5_ib_odp_create_qp(qp);
1524 
1525 	mutex_init(&qp->mutex);
1526 	spin_lock_init(&qp->sq.lock);
1527 	spin_lock_init(&qp->rq.lock);
1528 
1529 	if (init_attr->rwq_ind_tbl) {
1530 		if (!udata)
1531 			return -ENOSYS;
1532 
1533 		err = create_rss_raw_qp_tir(dev, qp, pd, init_attr, udata);
1534 		return err;
1535 	}
1536 
1537 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
1538 		if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
1539 			mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
1540 			return -EINVAL;
1541 		} else {
1542 			qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1543 		}
1544 	}
1545 
1546 	if (init_attr->create_flags &
1547 			(IB_QP_CREATE_CROSS_CHANNEL |
1548 			 IB_QP_CREATE_MANAGED_SEND |
1549 			 IB_QP_CREATE_MANAGED_RECV)) {
1550 		if (!MLX5_CAP_GEN(mdev, cd)) {
1551 			mlx5_ib_dbg(dev, "cross-channel isn't supported\n");
1552 			return -EINVAL;
1553 		}
1554 		if (init_attr->create_flags & IB_QP_CREATE_CROSS_CHANNEL)
1555 			qp->flags |= MLX5_IB_QP_CROSS_CHANNEL;
1556 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_SEND)
1557 			qp->flags |= MLX5_IB_QP_MANAGED_SEND;
1558 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_RECV)
1559 			qp->flags |= MLX5_IB_QP_MANAGED_RECV;
1560 	}
1561 
1562 	if (init_attr->qp_type == IB_QPT_UD &&
1563 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO))
1564 		if (!MLX5_CAP_GEN(mdev, ipoib_ipoib_offloads)) {
1565 			mlx5_ib_dbg(dev, "ipoib UD lso qp isn't supported\n");
1566 			return -EOPNOTSUPP;
1567 		}
1568 
1569 	if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1570 		if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
1571 			mlx5_ib_dbg(dev, "Scatter FCS is supported only for Raw Packet QPs");
1572 			return -EOPNOTSUPP;
1573 		}
1574 		if (!MLX5_CAP_GEN(dev->mdev, eth_net_offloads) ||
1575 		    !MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
1576 			mlx5_ib_dbg(dev, "Scatter FCS isn't supported\n");
1577 			return -EOPNOTSUPP;
1578 		}
1579 		qp->flags |= MLX5_IB_QP_CAP_SCATTER_FCS;
1580 	}
1581 
1582 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1583 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
1584 
1585 	if (pd && pd->uobject) {
1586 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
1587 			mlx5_ib_dbg(dev, "copy failed\n");
1588 			return -EFAULT;
1589 		}
1590 
1591 		err = get_qp_user_index(to_mucontext(pd->uobject->context),
1592 					&ucmd, udata->inlen, &uidx);
1593 		if (err)
1594 			return err;
1595 
1596 		qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
1597 		qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
1598 	} else {
1599 		qp->wq_sig = !!wq_signature;
1600 	}
1601 
1602 	qp->has_rq = qp_has_rq(init_attr);
1603 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
1604 			  qp, (pd && pd->uobject) ? &ucmd : NULL);
1605 	if (err) {
1606 		mlx5_ib_dbg(dev, "err %d\n", err);
1607 		return err;
1608 	}
1609 
1610 	if (pd) {
1611 		if (pd->uobject) {
1612 			__u32 max_wqes =
1613 				1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1614 			mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
1615 			if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
1616 			    ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
1617 				mlx5_ib_dbg(dev, "invalid rq params\n");
1618 				return -EINVAL;
1619 			}
1620 			if (ucmd.sq_wqe_count > max_wqes) {
1621 				mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
1622 					    ucmd.sq_wqe_count, max_wqes);
1623 				return -EINVAL;
1624 			}
1625 			if (init_attr->create_flags &
1626 			    mlx5_ib_create_qp_sqpn_qp1()) {
1627 				mlx5_ib_dbg(dev, "user-space is not allowed to create UD QPs spoofing as QP1\n");
1628 				return -EINVAL;
1629 			}
1630 			err = create_user_qp(dev, pd, qp, udata, init_attr, &in,
1631 					     &resp, &inlen, base);
1632 			if (err)
1633 				mlx5_ib_dbg(dev, "err %d\n", err);
1634 		} else {
1635 			err = create_kernel_qp(dev, init_attr, qp, &in, &inlen,
1636 					       base);
1637 			if (err)
1638 				mlx5_ib_dbg(dev, "err %d\n", err);
1639 		}
1640 
1641 		if (err)
1642 			return err;
1643 	} else {
1644 		in = mlx5_vzalloc(inlen);
1645 		if (!in)
1646 			return -ENOMEM;
1647 
1648 		qp->create_type = MLX5_QP_EMPTY;
1649 	}
1650 
1651 	if (is_sqp(init_attr->qp_type))
1652 		qp->port = init_attr->port_num;
1653 
1654 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1655 
1656 	MLX5_SET(qpc, qpc, st, to_mlx5_st(init_attr->qp_type));
1657 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
1658 
1659 	if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
1660 		MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
1661 	else
1662 		MLX5_SET(qpc, qpc, latency_sensitive, 1);
1663 
1664 
1665 	if (qp->wq_sig)
1666 		MLX5_SET(qpc, qpc, wq_signature, 1);
1667 
1668 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
1669 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
1670 
1671 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
1672 		MLX5_SET(qpc, qpc, cd_master, 1);
1673 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
1674 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
1675 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
1676 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
1677 
1678 	if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
1679 		int rcqe_sz;
1680 		int scqe_sz;
1681 
1682 		rcqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->recv_cq);
1683 		scqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->send_cq);
1684 
1685 		if (rcqe_sz == 128)
1686 			MLX5_SET(qpc, qpc, cs_res, MLX5_RES_SCAT_DATA64_CQE);
1687 		else
1688 			MLX5_SET(qpc, qpc, cs_res, MLX5_RES_SCAT_DATA32_CQE);
1689 
1690 		if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) {
1691 			if (scqe_sz == 128)
1692 				MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA64_CQE);
1693 			else
1694 				MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA32_CQE);
1695 		}
1696 	}
1697 
1698 	if (qp->rq.wqe_cnt) {
1699 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
1700 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
1701 	}
1702 
1703 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
1704 
1705 	if (qp->sq.wqe_cnt)
1706 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
1707 	else
1708 		MLX5_SET(qpc, qpc, no_sq, 1);
1709 
1710 	/* Set default resources */
1711 	switch (init_attr->qp_type) {
1712 	case IB_QPT_XRC_TGT:
1713 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1714 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(devr->c0)->mcq.cqn);
1715 		MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(devr->s0)->msrq.srqn);
1716 		MLX5_SET(qpc, qpc, xrcd, to_mxrcd(init_attr->xrcd)->xrcdn);
1717 		break;
1718 	case IB_QPT_XRC_INI:
1719 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1720 		MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
1721 		MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(devr->s0)->msrq.srqn);
1722 		break;
1723 	default:
1724 		if (init_attr->srq) {
1725 			MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x0)->xrcdn);
1726 			MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(init_attr->srq)->msrq.srqn);
1727 		} else {
1728 			MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
1729 			MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(devr->s1)->msrq.srqn);
1730 		}
1731 	}
1732 
1733 	if (init_attr->send_cq)
1734 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(init_attr->send_cq)->mcq.cqn);
1735 
1736 	if (init_attr->recv_cq)
1737 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(init_attr->recv_cq)->mcq.cqn);
1738 
1739 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
1740 
1741 	/* 0xffffff means we ask to work with cqe version 0 */
1742 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
1743 		MLX5_SET(qpc, qpc, user_index, uidx);
1744 
1745 	/* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
1746 	if (init_attr->qp_type == IB_QPT_UD &&
1747 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)) {
1748 		MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
1749 		qp->flags |= MLX5_IB_QP_LSO;
1750 	}
1751 
1752 	if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
1753 		qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd.sq_buf_addr;
1754 		raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
1755 		err = create_raw_packet_qp(dev, qp, in, pd);
1756 	} else {
1757 		err = mlx5_core_create_qp(dev->mdev, &base->mqp, in, inlen);
1758 	}
1759 
1760 	if (err) {
1761 		mlx5_ib_dbg(dev, "create qp failed\n");
1762 		goto err_create;
1763 	}
1764 
1765 	kvfree(in);
1766 
1767 	base->container_mibqp = qp;
1768 	base->mqp.event = mlx5_ib_qp_event;
1769 
1770 	get_cqs(init_attr->qp_type, init_attr->send_cq, init_attr->recv_cq,
1771 		&send_cq, &recv_cq);
1772 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1773 	mlx5_ib_lock_cqs(send_cq, recv_cq);
1774 	/* Maintain device to QPs access, needed for further handling via reset
1775 	 * flow
1776 	 */
1777 	list_add_tail(&qp->qps_list, &dev->qp_list);
1778 	/* Maintain CQ to QPs access, needed for further handling via reset flow
1779 	 */
1780 	if (send_cq)
1781 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
1782 	if (recv_cq)
1783 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
1784 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
1785 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1786 
1787 	return 0;
1788 
1789 err_create:
1790 	if (qp->create_type == MLX5_QP_USER)
1791 		destroy_qp_user(pd, qp, base);
1792 	else if (qp->create_type == MLX5_QP_KERNEL)
1793 		destroy_qp_kernel(dev, qp);
1794 
1795 	kvfree(in);
1796 	return err;
1797 }
1798 
1799 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1800 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1801 {
1802 	if (send_cq) {
1803 		if (recv_cq) {
1804 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1805 				spin_lock(&send_cq->lock);
1806 				spin_lock_nested(&recv_cq->lock,
1807 						 SINGLE_DEPTH_NESTING);
1808 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1809 				spin_lock(&send_cq->lock);
1810 				__acquire(&recv_cq->lock);
1811 			} else {
1812 				spin_lock(&recv_cq->lock);
1813 				spin_lock_nested(&send_cq->lock,
1814 						 SINGLE_DEPTH_NESTING);
1815 			}
1816 		} else {
1817 			spin_lock(&send_cq->lock);
1818 			__acquire(&recv_cq->lock);
1819 		}
1820 	} else if (recv_cq) {
1821 		spin_lock(&recv_cq->lock);
1822 		__acquire(&send_cq->lock);
1823 	} else {
1824 		__acquire(&send_cq->lock);
1825 		__acquire(&recv_cq->lock);
1826 	}
1827 }
1828 
1829 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1830 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1831 {
1832 	if (send_cq) {
1833 		if (recv_cq) {
1834 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1835 				spin_unlock(&recv_cq->lock);
1836 				spin_unlock(&send_cq->lock);
1837 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1838 				__release(&recv_cq->lock);
1839 				spin_unlock(&send_cq->lock);
1840 			} else {
1841 				spin_unlock(&send_cq->lock);
1842 				spin_unlock(&recv_cq->lock);
1843 			}
1844 		} else {
1845 			__release(&recv_cq->lock);
1846 			spin_unlock(&send_cq->lock);
1847 		}
1848 	} else if (recv_cq) {
1849 		__release(&send_cq->lock);
1850 		spin_unlock(&recv_cq->lock);
1851 	} else {
1852 		__release(&recv_cq->lock);
1853 		__release(&send_cq->lock);
1854 	}
1855 }
1856 
1857 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
1858 {
1859 	return to_mpd(qp->ibqp.pd);
1860 }
1861 
1862 static void get_cqs(enum ib_qp_type qp_type,
1863 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
1864 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
1865 {
1866 	switch (qp_type) {
1867 	case IB_QPT_XRC_TGT:
1868 		*send_cq = NULL;
1869 		*recv_cq = NULL;
1870 		break;
1871 	case MLX5_IB_QPT_REG_UMR:
1872 	case IB_QPT_XRC_INI:
1873 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
1874 		*recv_cq = NULL;
1875 		break;
1876 
1877 	case IB_QPT_SMI:
1878 	case MLX5_IB_QPT_HW_GSI:
1879 	case IB_QPT_RC:
1880 	case IB_QPT_UC:
1881 	case IB_QPT_UD:
1882 	case IB_QPT_RAW_IPV6:
1883 	case IB_QPT_RAW_ETHERTYPE:
1884 	case IB_QPT_RAW_PACKET:
1885 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
1886 		*recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
1887 		break;
1888 
1889 	case IB_QPT_MAX:
1890 	default:
1891 		*send_cq = NULL;
1892 		*recv_cq = NULL;
1893 		break;
1894 	}
1895 }
1896 
1897 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1898 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
1899 				u8 lag_tx_affinity);
1900 
1901 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1902 {
1903 	struct mlx5_ib_cq *send_cq, *recv_cq;
1904 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
1905 	unsigned long flags;
1906 	int err;
1907 
1908 	if (qp->ibqp.rwq_ind_tbl) {
1909 		destroy_rss_raw_qp_tir(dev, qp);
1910 		return;
1911 	}
1912 
1913 	base = qp->ibqp.qp_type == IB_QPT_RAW_PACKET ?
1914 	       &qp->raw_packet_qp.rq.base :
1915 	       &qp->trans_qp.base;
1916 
1917 	if (qp->state != IB_QPS_RESET) {
1918 		if (qp->ibqp.qp_type != IB_QPT_RAW_PACKET) {
1919 			mlx5_ib_qp_disable_pagefaults(qp);
1920 			err = mlx5_core_qp_modify(dev->mdev,
1921 						  MLX5_CMD_OP_2RST_QP, 0,
1922 						  NULL, &base->mqp);
1923 		} else {
1924 			struct mlx5_modify_raw_qp_param raw_qp_param = {
1925 				.operation = MLX5_CMD_OP_2RST_QP
1926 			};
1927 
1928 			err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
1929 		}
1930 		if (err)
1931 			mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
1932 				     base->mqp.qpn);
1933 	}
1934 
1935 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
1936 		&send_cq, &recv_cq);
1937 
1938 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1939 	mlx5_ib_lock_cqs(send_cq, recv_cq);
1940 	/* del from lists under both locks above to protect reset flow paths */
1941 	list_del(&qp->qps_list);
1942 	if (send_cq)
1943 		list_del(&qp->cq_send_list);
1944 
1945 	if (recv_cq)
1946 		list_del(&qp->cq_recv_list);
1947 
1948 	if (qp->create_type == MLX5_QP_KERNEL) {
1949 		__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
1950 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1951 		if (send_cq != recv_cq)
1952 			__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
1953 					   NULL);
1954 	}
1955 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
1956 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1957 
1958 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1959 		destroy_raw_packet_qp(dev, qp);
1960 	} else {
1961 		err = mlx5_core_destroy_qp(dev->mdev, &base->mqp);
1962 		if (err)
1963 			mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
1964 				     base->mqp.qpn);
1965 	}
1966 
1967 	if (qp->create_type == MLX5_QP_KERNEL)
1968 		destroy_qp_kernel(dev, qp);
1969 	else if (qp->create_type == MLX5_QP_USER)
1970 		destroy_qp_user(&get_pd(qp)->ibpd, qp, base);
1971 }
1972 
1973 static const char *ib_qp_type_str(enum ib_qp_type type)
1974 {
1975 	switch (type) {
1976 	case IB_QPT_SMI:
1977 		return "IB_QPT_SMI";
1978 	case IB_QPT_GSI:
1979 		return "IB_QPT_GSI";
1980 	case IB_QPT_RC:
1981 		return "IB_QPT_RC";
1982 	case IB_QPT_UC:
1983 		return "IB_QPT_UC";
1984 	case IB_QPT_UD:
1985 		return "IB_QPT_UD";
1986 	case IB_QPT_RAW_IPV6:
1987 		return "IB_QPT_RAW_IPV6";
1988 	case IB_QPT_RAW_ETHERTYPE:
1989 		return "IB_QPT_RAW_ETHERTYPE";
1990 	case IB_QPT_XRC_INI:
1991 		return "IB_QPT_XRC_INI";
1992 	case IB_QPT_XRC_TGT:
1993 		return "IB_QPT_XRC_TGT";
1994 	case IB_QPT_RAW_PACKET:
1995 		return "IB_QPT_RAW_PACKET";
1996 	case MLX5_IB_QPT_REG_UMR:
1997 		return "MLX5_IB_QPT_REG_UMR";
1998 	case IB_QPT_MAX:
1999 	default:
2000 		return "Invalid QP type";
2001 	}
2002 }
2003 
2004 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
2005 				struct ib_qp_init_attr *init_attr,
2006 				struct ib_udata *udata)
2007 {
2008 	struct mlx5_ib_dev *dev;
2009 	struct mlx5_ib_qp *qp;
2010 	u16 xrcdn = 0;
2011 	int err;
2012 
2013 	if (pd) {
2014 		dev = to_mdev(pd->device);
2015 
2016 		if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
2017 			if (!pd->uobject) {
2018 				mlx5_ib_dbg(dev, "Raw Packet QP is not supported for kernel consumers\n");
2019 				return ERR_PTR(-EINVAL);
2020 			} else if (!to_mucontext(pd->uobject->context)->cqe_version) {
2021 				mlx5_ib_dbg(dev, "Raw Packet QP is only supported for CQE version > 0\n");
2022 				return ERR_PTR(-EINVAL);
2023 			}
2024 		}
2025 	} else {
2026 		/* being cautious here */
2027 		if (init_attr->qp_type != IB_QPT_XRC_TGT &&
2028 		    init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
2029 			pr_warn("%s: no PD for transport %s\n", __func__,
2030 				ib_qp_type_str(init_attr->qp_type));
2031 			return ERR_PTR(-EINVAL);
2032 		}
2033 		dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
2034 	}
2035 
2036 	switch (init_attr->qp_type) {
2037 	case IB_QPT_XRC_TGT:
2038 	case IB_QPT_XRC_INI:
2039 		if (!MLX5_CAP_GEN(dev->mdev, xrc)) {
2040 			mlx5_ib_dbg(dev, "XRC not supported\n");
2041 			return ERR_PTR(-ENOSYS);
2042 		}
2043 		init_attr->recv_cq = NULL;
2044 		if (init_attr->qp_type == IB_QPT_XRC_TGT) {
2045 			xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
2046 			init_attr->send_cq = NULL;
2047 		}
2048 
2049 		/* fall through */
2050 	case IB_QPT_RAW_PACKET:
2051 	case IB_QPT_RC:
2052 	case IB_QPT_UC:
2053 	case IB_QPT_UD:
2054 	case IB_QPT_SMI:
2055 	case MLX5_IB_QPT_HW_GSI:
2056 	case MLX5_IB_QPT_REG_UMR:
2057 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2058 		if (!qp)
2059 			return ERR_PTR(-ENOMEM);
2060 
2061 		err = create_qp_common(dev, pd, init_attr, udata, qp);
2062 		if (err) {
2063 			mlx5_ib_dbg(dev, "create_qp_common failed\n");
2064 			kfree(qp);
2065 			return ERR_PTR(err);
2066 		}
2067 
2068 		if (is_qp0(init_attr->qp_type))
2069 			qp->ibqp.qp_num = 0;
2070 		else if (is_qp1(init_attr->qp_type))
2071 			qp->ibqp.qp_num = 1;
2072 		else
2073 			qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
2074 
2075 		mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
2076 			    qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
2077 			    init_attr->recv_cq ? to_mcq(init_attr->recv_cq)->mcq.cqn : -1,
2078 			    init_attr->send_cq ? to_mcq(init_attr->send_cq)->mcq.cqn : -1);
2079 
2080 		qp->trans_qp.xrcdn = xrcdn;
2081 
2082 		break;
2083 
2084 	case IB_QPT_GSI:
2085 		return mlx5_ib_gsi_create_qp(pd, init_attr);
2086 
2087 	case IB_QPT_RAW_IPV6:
2088 	case IB_QPT_RAW_ETHERTYPE:
2089 	case IB_QPT_MAX:
2090 	default:
2091 		mlx5_ib_dbg(dev, "unsupported qp type %d\n",
2092 			    init_attr->qp_type);
2093 		/* Don't support raw QPs */
2094 		return ERR_PTR(-EINVAL);
2095 	}
2096 
2097 	return &qp->ibqp;
2098 }
2099 
2100 int mlx5_ib_destroy_qp(struct ib_qp *qp)
2101 {
2102 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2103 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2104 
2105 	if (unlikely(qp->qp_type == IB_QPT_GSI))
2106 		return mlx5_ib_gsi_destroy_qp(qp);
2107 
2108 	destroy_qp_common(dev, mqp);
2109 
2110 	kfree(mqp);
2111 
2112 	return 0;
2113 }
2114 
2115 static __be32 to_mlx5_access_flags(struct mlx5_ib_qp *qp, const struct ib_qp_attr *attr,
2116 				   int attr_mask)
2117 {
2118 	u32 hw_access_flags = 0;
2119 	u8 dest_rd_atomic;
2120 	u32 access_flags;
2121 
2122 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2123 		dest_rd_atomic = attr->max_dest_rd_atomic;
2124 	else
2125 		dest_rd_atomic = qp->trans_qp.resp_depth;
2126 
2127 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2128 		access_flags = attr->qp_access_flags;
2129 	else
2130 		access_flags = qp->trans_qp.atomic_rd_en;
2131 
2132 	if (!dest_rd_atomic)
2133 		access_flags &= IB_ACCESS_REMOTE_WRITE;
2134 
2135 	if (access_flags & IB_ACCESS_REMOTE_READ)
2136 		hw_access_flags |= MLX5_QP_BIT_RRE;
2137 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
2138 		hw_access_flags |= (MLX5_QP_BIT_RAE | MLX5_ATOMIC_MODE_CX);
2139 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
2140 		hw_access_flags |= MLX5_QP_BIT_RWE;
2141 
2142 	return cpu_to_be32(hw_access_flags);
2143 }
2144 
2145 enum {
2146 	MLX5_PATH_FLAG_FL	= 1 << 0,
2147 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
2148 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
2149 };
2150 
2151 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
2152 {
2153 	if (rate == IB_RATE_PORT_CURRENT) {
2154 		return 0;
2155 	} else if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_600_GBPS) {
2156 		return -EINVAL;
2157 	} else {
2158 		while (rate != IB_RATE_2_5_GBPS &&
2159 		       !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
2160 			 MLX5_CAP_GEN(dev->mdev, stat_rate_support)))
2161 			--rate;
2162 	}
2163 
2164 	return rate + MLX5_STAT_RATE_OFFSET;
2165 }
2166 
2167 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
2168 				      struct mlx5_ib_sq *sq, u8 sl)
2169 {
2170 	void *in;
2171 	void *tisc;
2172 	int inlen;
2173 	int err;
2174 
2175 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2176 	in = mlx5_vzalloc(inlen);
2177 	if (!in)
2178 		return -ENOMEM;
2179 
2180 	MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
2181 
2182 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2183 	MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
2184 
2185 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2186 
2187 	kvfree(in);
2188 
2189 	return err;
2190 }
2191 
2192 static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,
2193 					 struct mlx5_ib_sq *sq, u8 tx_affinity)
2194 {
2195 	void *in;
2196 	void *tisc;
2197 	int inlen;
2198 	int err;
2199 
2200 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2201 	in = mlx5_vzalloc(inlen);
2202 	if (!in)
2203 		return -ENOMEM;
2204 
2205 	MLX5_SET(modify_tis_in, in, bitmask.lag_tx_port_affinity, 1);
2206 
2207 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2208 	MLX5_SET(tisc, tisc, lag_tx_port_affinity, tx_affinity);
2209 
2210 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2211 
2212 	kvfree(in);
2213 
2214 	return err;
2215 }
2216 
2217 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2218 			 const struct ib_ah_attr *ah,
2219 			 struct mlx5_qp_path *path, u8 port, int attr_mask,
2220 			 u32 path_flags, const struct ib_qp_attr *attr,
2221 			 bool alt)
2222 {
2223 	enum rdma_link_layer ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2224 	int err;
2225 	enum ib_gid_type gid_type;
2226 
2227 	if (attr_mask & IB_QP_PKEY_INDEX)
2228 		path->pkey_index = cpu_to_be16(alt ? attr->alt_pkey_index :
2229 						     attr->pkey_index);
2230 
2231 	if (ah->ah_flags & IB_AH_GRH) {
2232 		if (ah->grh.sgid_index >=
2233 		    dev->mdev->port_caps[port - 1].gid_table_len) {
2234 			pr_err("sgid_index (%u) too large. max is %d\n",
2235 			       ah->grh.sgid_index,
2236 			       dev->mdev->port_caps[port - 1].gid_table_len);
2237 			return -EINVAL;
2238 		}
2239 	}
2240 
2241 	if (ll == IB_LINK_LAYER_ETHERNET) {
2242 		if (!(ah->ah_flags & IB_AH_GRH))
2243 			return -EINVAL;
2244 		err = mlx5_get_roce_gid_type(dev, port, ah->grh.sgid_index,
2245 					     &gid_type);
2246 		if (err)
2247 			return err;
2248 		memcpy(path->rmac, ah->dmac, sizeof(ah->dmac));
2249 		path->udp_sport = mlx5_get_roce_udp_sport(dev, port,
2250 							  ah->grh.sgid_index);
2251 		path->dci_cfi_prio_sl = (ah->sl & 0x7) << 4;
2252 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
2253 			path->ecn_dscp = (ah->grh.traffic_class >> 2) & 0x3f;
2254 	} else {
2255 		path->fl_free_ar = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
2256 		path->fl_free_ar |=
2257 			(path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x40 : 0;
2258 		path->rlid = cpu_to_be16(ah->dlid);
2259 		path->grh_mlid = ah->src_path_bits & 0x7f;
2260 		if (ah->ah_flags & IB_AH_GRH)
2261 			path->grh_mlid	|= 1 << 7;
2262 		path->dci_cfi_prio_sl = ah->sl & 0xf;
2263 	}
2264 
2265 	if (ah->ah_flags & IB_AH_GRH) {
2266 		path->mgid_index = ah->grh.sgid_index;
2267 		path->hop_limit  = ah->grh.hop_limit;
2268 		path->tclass_flowlabel =
2269 			cpu_to_be32((ah->grh.traffic_class << 20) |
2270 				    (ah->grh.flow_label));
2271 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
2272 	}
2273 
2274 	err = ib_rate_to_mlx5(dev, ah->static_rate);
2275 	if (err < 0)
2276 		return err;
2277 	path->static_rate = err;
2278 	path->port = port;
2279 
2280 	if (attr_mask & IB_QP_TIMEOUT)
2281 		path->ackto_lt = (alt ? attr->alt_timeout : attr->timeout) << 3;
2282 
2283 	if ((qp->ibqp.qp_type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
2284 		return modify_raw_packet_eth_prio(dev->mdev,
2285 						  &qp->raw_packet_qp.sq,
2286 						  ah->sl & 0xf);
2287 
2288 	return 0;
2289 }
2290 
2291 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
2292 	[MLX5_QP_STATE_INIT] = {
2293 		[MLX5_QP_STATE_INIT] = {
2294 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
2295 					  MLX5_QP_OPTPAR_RAE		|
2296 					  MLX5_QP_OPTPAR_RWE		|
2297 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
2298 					  MLX5_QP_OPTPAR_PRI_PORT,
2299 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
2300 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
2301 					  MLX5_QP_OPTPAR_PRI_PORT,
2302 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
2303 					  MLX5_QP_OPTPAR_Q_KEY		|
2304 					  MLX5_QP_OPTPAR_PRI_PORT,
2305 		},
2306 		[MLX5_QP_STATE_RTR] = {
2307 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
2308 					  MLX5_QP_OPTPAR_RRE            |
2309 					  MLX5_QP_OPTPAR_RAE            |
2310 					  MLX5_QP_OPTPAR_RWE            |
2311 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2312 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
2313 					  MLX5_QP_OPTPAR_RWE            |
2314 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2315 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
2316 					  MLX5_QP_OPTPAR_Q_KEY,
2317 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
2318 					   MLX5_QP_OPTPAR_Q_KEY,
2319 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2320 					  MLX5_QP_OPTPAR_RRE            |
2321 					  MLX5_QP_OPTPAR_RAE            |
2322 					  MLX5_QP_OPTPAR_RWE            |
2323 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2324 		},
2325 	},
2326 	[MLX5_QP_STATE_RTR] = {
2327 		[MLX5_QP_STATE_RTS] = {
2328 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
2329 					  MLX5_QP_OPTPAR_RRE		|
2330 					  MLX5_QP_OPTPAR_RAE		|
2331 					  MLX5_QP_OPTPAR_RWE		|
2332 					  MLX5_QP_OPTPAR_PM_STATE	|
2333 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
2334 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
2335 					  MLX5_QP_OPTPAR_RWE		|
2336 					  MLX5_QP_OPTPAR_PM_STATE,
2337 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
2338 		},
2339 	},
2340 	[MLX5_QP_STATE_RTS] = {
2341 		[MLX5_QP_STATE_RTS] = {
2342 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
2343 					  MLX5_QP_OPTPAR_RAE		|
2344 					  MLX5_QP_OPTPAR_RWE		|
2345 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
2346 					  MLX5_QP_OPTPAR_PM_STATE	|
2347 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
2348 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
2349 					  MLX5_QP_OPTPAR_PM_STATE	|
2350 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
2351 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
2352 					  MLX5_QP_OPTPAR_SRQN		|
2353 					  MLX5_QP_OPTPAR_CQN_RCV,
2354 		},
2355 	},
2356 	[MLX5_QP_STATE_SQER] = {
2357 		[MLX5_QP_STATE_RTS] = {
2358 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
2359 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
2360 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
2361 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
2362 					   MLX5_QP_OPTPAR_RWE		|
2363 					   MLX5_QP_OPTPAR_RAE		|
2364 					   MLX5_QP_OPTPAR_RRE,
2365 		},
2366 	},
2367 };
2368 
2369 static int ib_nr_to_mlx5_nr(int ib_mask)
2370 {
2371 	switch (ib_mask) {
2372 	case IB_QP_STATE:
2373 		return 0;
2374 	case IB_QP_CUR_STATE:
2375 		return 0;
2376 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
2377 		return 0;
2378 	case IB_QP_ACCESS_FLAGS:
2379 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
2380 			MLX5_QP_OPTPAR_RAE;
2381 	case IB_QP_PKEY_INDEX:
2382 		return MLX5_QP_OPTPAR_PKEY_INDEX;
2383 	case IB_QP_PORT:
2384 		return MLX5_QP_OPTPAR_PRI_PORT;
2385 	case IB_QP_QKEY:
2386 		return MLX5_QP_OPTPAR_Q_KEY;
2387 	case IB_QP_AV:
2388 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
2389 			MLX5_QP_OPTPAR_PRI_PORT;
2390 	case IB_QP_PATH_MTU:
2391 		return 0;
2392 	case IB_QP_TIMEOUT:
2393 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
2394 	case IB_QP_RETRY_CNT:
2395 		return MLX5_QP_OPTPAR_RETRY_COUNT;
2396 	case IB_QP_RNR_RETRY:
2397 		return MLX5_QP_OPTPAR_RNR_RETRY;
2398 	case IB_QP_RQ_PSN:
2399 		return 0;
2400 	case IB_QP_MAX_QP_RD_ATOMIC:
2401 		return MLX5_QP_OPTPAR_SRA_MAX;
2402 	case IB_QP_ALT_PATH:
2403 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
2404 	case IB_QP_MIN_RNR_TIMER:
2405 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
2406 	case IB_QP_SQ_PSN:
2407 		return 0;
2408 	case IB_QP_MAX_DEST_RD_ATOMIC:
2409 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
2410 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
2411 	case IB_QP_PATH_MIG_STATE:
2412 		return MLX5_QP_OPTPAR_PM_STATE;
2413 	case IB_QP_CAP:
2414 		return 0;
2415 	case IB_QP_DEST_QPN:
2416 		return 0;
2417 	}
2418 	return 0;
2419 }
2420 
2421 static int ib_mask_to_mlx5_opt(int ib_mask)
2422 {
2423 	int result = 0;
2424 	int i;
2425 
2426 	for (i = 0; i < 8 * sizeof(int); i++) {
2427 		if ((1 << i) & ib_mask)
2428 			result |= ib_nr_to_mlx5_nr(1 << i);
2429 	}
2430 
2431 	return result;
2432 }
2433 
2434 static int modify_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
2435 				   struct mlx5_ib_rq *rq, int new_state,
2436 				   const struct mlx5_modify_raw_qp_param *raw_qp_param)
2437 {
2438 	void *in;
2439 	void *rqc;
2440 	int inlen;
2441 	int err;
2442 
2443 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
2444 	in = mlx5_vzalloc(inlen);
2445 	if (!in)
2446 		return -ENOMEM;
2447 
2448 	MLX5_SET(modify_rq_in, in, rqn, rq->base.mqp.qpn);
2449 	MLX5_SET(modify_rq_in, in, rq_state, rq->state);
2450 
2451 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
2452 	MLX5_SET(rqc, rqc, state, new_state);
2453 
2454 	if (raw_qp_param->set_mask & MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID) {
2455 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counters_set_id)) {
2456 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
2457 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_MODIFY_RQ_COUNTER_SET_ID);
2458 			MLX5_SET(rqc, rqc, counter_set_id, raw_qp_param->rq_q_ctr_id);
2459 		} else
2460 			pr_info_once("%s: RAW PACKET QP counters are not supported on current FW\n",
2461 				     dev->ib_dev.name);
2462 	}
2463 
2464 	err = mlx5_core_modify_rq(dev->mdev, in, inlen);
2465 	if (err)
2466 		goto out;
2467 
2468 	rq->state = new_state;
2469 
2470 out:
2471 	kvfree(in);
2472 	return err;
2473 }
2474 
2475 static int modify_raw_packet_qp_sq(struct mlx5_core_dev *dev,
2476 				   struct mlx5_ib_sq *sq, int new_state)
2477 {
2478 	void *in;
2479 	void *sqc;
2480 	int inlen;
2481 	int err;
2482 
2483 	inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
2484 	in = mlx5_vzalloc(inlen);
2485 	if (!in)
2486 		return -ENOMEM;
2487 
2488 	MLX5_SET(modify_sq_in, in, sqn, sq->base.mqp.qpn);
2489 	MLX5_SET(modify_sq_in, in, sq_state, sq->state);
2490 
2491 	sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
2492 	MLX5_SET(sqc, sqc, state, new_state);
2493 
2494 	err = mlx5_core_modify_sq(dev, in, inlen);
2495 	if (err)
2496 		goto out;
2497 
2498 	sq->state = new_state;
2499 
2500 out:
2501 	kvfree(in);
2502 	return err;
2503 }
2504 
2505 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2506 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
2507 				u8 tx_affinity)
2508 {
2509 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
2510 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
2511 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
2512 	int rq_state;
2513 	int sq_state;
2514 	int err;
2515 
2516 	switch (raw_qp_param->operation) {
2517 	case MLX5_CMD_OP_RST2INIT_QP:
2518 		rq_state = MLX5_RQC_STATE_RDY;
2519 		sq_state = MLX5_SQC_STATE_RDY;
2520 		break;
2521 	case MLX5_CMD_OP_2ERR_QP:
2522 		rq_state = MLX5_RQC_STATE_ERR;
2523 		sq_state = MLX5_SQC_STATE_ERR;
2524 		break;
2525 	case MLX5_CMD_OP_2RST_QP:
2526 		rq_state = MLX5_RQC_STATE_RST;
2527 		sq_state = MLX5_SQC_STATE_RST;
2528 		break;
2529 	case MLX5_CMD_OP_INIT2INIT_QP:
2530 	case MLX5_CMD_OP_INIT2RTR_QP:
2531 	case MLX5_CMD_OP_RTR2RTS_QP:
2532 	case MLX5_CMD_OP_RTS2RTS_QP:
2533 		if (raw_qp_param->set_mask)
2534 			return -EINVAL;
2535 		else
2536 			return 0;
2537 	default:
2538 		WARN_ON(1);
2539 		return -EINVAL;
2540 	}
2541 
2542 	if (qp->rq.wqe_cnt) {
2543 		err = modify_raw_packet_qp_rq(dev, rq, rq_state, raw_qp_param);
2544 		if (err)
2545 			return err;
2546 	}
2547 
2548 	if (qp->sq.wqe_cnt) {
2549 		if (tx_affinity) {
2550 			err = modify_raw_packet_tx_affinity(dev->mdev, sq,
2551 							    tx_affinity);
2552 			if (err)
2553 				return err;
2554 		}
2555 
2556 		return modify_raw_packet_qp_sq(dev->mdev, sq, sq_state);
2557 	}
2558 
2559 	return 0;
2560 }
2561 
2562 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
2563 			       const struct ib_qp_attr *attr, int attr_mask,
2564 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
2565 {
2566 	static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
2567 		[MLX5_QP_STATE_RST] = {
2568 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2569 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2570 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_RST2INIT_QP,
2571 		},
2572 		[MLX5_QP_STATE_INIT]  = {
2573 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2574 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2575 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_INIT2INIT_QP,
2576 			[MLX5_QP_STATE_RTR]	= MLX5_CMD_OP_INIT2RTR_QP,
2577 		},
2578 		[MLX5_QP_STATE_RTR]   = {
2579 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2580 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2581 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTR2RTS_QP,
2582 		},
2583 		[MLX5_QP_STATE_RTS]   = {
2584 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2585 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2586 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTS2RTS_QP,
2587 		},
2588 		[MLX5_QP_STATE_SQD] = {
2589 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2590 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2591 		},
2592 		[MLX5_QP_STATE_SQER] = {
2593 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2594 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2595 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQERR2RTS_QP,
2596 		},
2597 		[MLX5_QP_STATE_ERR] = {
2598 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2599 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2600 		}
2601 	};
2602 
2603 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2604 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2605 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
2606 	struct mlx5_ib_cq *send_cq, *recv_cq;
2607 	struct mlx5_qp_context *context;
2608 	struct mlx5_ib_pd *pd;
2609 	struct mlx5_ib_port *mibport = NULL;
2610 	enum mlx5_qp_state mlx5_cur, mlx5_new;
2611 	enum mlx5_qp_optpar optpar;
2612 	int sqd_event;
2613 	int mlx5_st;
2614 	int err;
2615 	u16 op;
2616 
2617 	context = kzalloc(sizeof(*context), GFP_KERNEL);
2618 	if (!context)
2619 		return -ENOMEM;
2620 
2621 	err = to_mlx5_st(ibqp->qp_type);
2622 	if (err < 0) {
2623 		mlx5_ib_dbg(dev, "unsupported qp type %d\n", ibqp->qp_type);
2624 		goto out;
2625 	}
2626 
2627 	context->flags = cpu_to_be32(err << 16);
2628 
2629 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
2630 		context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2631 	} else {
2632 		switch (attr->path_mig_state) {
2633 		case IB_MIG_MIGRATED:
2634 			context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2635 			break;
2636 		case IB_MIG_REARM:
2637 			context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
2638 			break;
2639 		case IB_MIG_ARMED:
2640 			context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
2641 			break;
2642 		}
2643 	}
2644 
2645 	if (is_sqp(ibqp->qp_type)) {
2646 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
2647 	} else if (ibqp->qp_type == IB_QPT_UD ||
2648 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
2649 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
2650 	} else if (attr_mask & IB_QP_PATH_MTU) {
2651 		if (attr->path_mtu < IB_MTU_256 ||
2652 		    attr->path_mtu > IB_MTU_4096) {
2653 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
2654 			err = -EINVAL;
2655 			goto out;
2656 		}
2657 		context->mtu_msgmax = (attr->path_mtu << 5) |
2658 				      (u8)MLX5_CAP_GEN(dev->mdev, log_max_msg);
2659 	}
2660 
2661 	if (attr_mask & IB_QP_DEST_QPN)
2662 		context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
2663 
2664 	if (attr_mask & IB_QP_PKEY_INDEX)
2665 		context->pri_path.pkey_index = cpu_to_be16(attr->pkey_index);
2666 
2667 	/* todo implement counter_index functionality */
2668 
2669 	if (is_sqp(ibqp->qp_type))
2670 		context->pri_path.port = qp->port;
2671 
2672 	if (attr_mask & IB_QP_PORT)
2673 		context->pri_path.port = attr->port_num;
2674 
2675 	if (attr_mask & IB_QP_AV) {
2676 		err = mlx5_set_path(dev, qp, &attr->ah_attr, &context->pri_path,
2677 				    attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
2678 				    attr_mask, 0, attr, false);
2679 		if (err)
2680 			goto out;
2681 	}
2682 
2683 	if (attr_mask & IB_QP_TIMEOUT)
2684 		context->pri_path.ackto_lt |= attr->timeout << 3;
2685 
2686 	if (attr_mask & IB_QP_ALT_PATH) {
2687 		err = mlx5_set_path(dev, qp, &attr->alt_ah_attr,
2688 				    &context->alt_path,
2689 				    attr->alt_port_num,
2690 				    attr_mask | IB_QP_PKEY_INDEX | IB_QP_TIMEOUT,
2691 				    0, attr, true);
2692 		if (err)
2693 			goto out;
2694 	}
2695 
2696 	pd = get_pd(qp);
2697 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
2698 		&send_cq, &recv_cq);
2699 
2700 	context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
2701 	context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
2702 	context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
2703 	context->params1  = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
2704 
2705 	if (attr_mask & IB_QP_RNR_RETRY)
2706 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2707 
2708 	if (attr_mask & IB_QP_RETRY_CNT)
2709 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2710 
2711 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2712 		if (attr->max_rd_atomic)
2713 			context->params1 |=
2714 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2715 	}
2716 
2717 	if (attr_mask & IB_QP_SQ_PSN)
2718 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2719 
2720 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2721 		if (attr->max_dest_rd_atomic)
2722 			context->params2 |=
2723 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2724 	}
2725 
2726 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
2727 		context->params2 |= to_mlx5_access_flags(qp, attr, attr_mask);
2728 
2729 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
2730 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2731 
2732 	if (attr_mask & IB_QP_RQ_PSN)
2733 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2734 
2735 	if (attr_mask & IB_QP_QKEY)
2736 		context->qkey = cpu_to_be32(attr->qkey);
2737 
2738 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2739 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2740 
2741 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2742 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2743 		sqd_event = 1;
2744 	else
2745 		sqd_event = 0;
2746 
2747 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2748 		u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
2749 			       qp->port) - 1;
2750 		mibport = &dev->port[port_num];
2751 		context->qp_counter_set_usr_page |=
2752 			cpu_to_be32((u32)(mibport->q_cnt_id) << 24);
2753 	}
2754 
2755 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2756 		context->sq_crq_size |= cpu_to_be16(1 << 4);
2757 
2758 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
2759 		context->deth_sqpn = cpu_to_be32(1);
2760 
2761 	mlx5_cur = to_mlx5_state(cur_state);
2762 	mlx5_new = to_mlx5_state(new_state);
2763 	mlx5_st = to_mlx5_st(ibqp->qp_type);
2764 	if (mlx5_st < 0)
2765 		goto out;
2766 
2767 	/* If moving to a reset or error state, we must disable page faults on
2768 	 * this QP and flush all current page faults. Otherwise a stale page
2769 	 * fault may attempt to work on this QP after it is reset and moved
2770 	 * again to RTS, and may cause the driver and the device to get out of
2771 	 * sync. */
2772 	if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2773 	    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR) &&
2774 	    (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
2775 		mlx5_ib_qp_disable_pagefaults(qp);
2776 
2777 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
2778 	    !optab[mlx5_cur][mlx5_new])
2779 		goto out;
2780 
2781 	op = optab[mlx5_cur][mlx5_new];
2782 	optpar = ib_mask_to_mlx5_opt(attr_mask);
2783 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
2784 
2785 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
2786 		struct mlx5_modify_raw_qp_param raw_qp_param = {};
2787 
2788 		raw_qp_param.operation = op;
2789 		if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2790 			raw_qp_param.rq_q_ctr_id = mibport->q_cnt_id;
2791 			raw_qp_param.set_mask |= MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID;
2792 		}
2793 		err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
2794 	} else {
2795 		err = mlx5_core_qp_modify(dev->mdev, op, optpar, context,
2796 					  &base->mqp);
2797 	}
2798 
2799 	if (err)
2800 		goto out;
2801 
2802 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT &&
2803 	    (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
2804 		mlx5_ib_qp_enable_pagefaults(qp);
2805 
2806 	qp->state = new_state;
2807 
2808 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2809 		qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
2810 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2811 		qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
2812 	if (attr_mask & IB_QP_PORT)
2813 		qp->port = attr->port_num;
2814 	if (attr_mask & IB_QP_ALT_PATH)
2815 		qp->trans_qp.alt_port = attr->alt_port_num;
2816 
2817 	/*
2818 	 * If we moved a kernel QP to RESET, clean up all old CQ
2819 	 * entries and reinitialize the QP.
2820 	 */
2821 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
2822 		mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2823 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
2824 		if (send_cq != recv_cq)
2825 			mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
2826 
2827 		qp->rq.head = 0;
2828 		qp->rq.tail = 0;
2829 		qp->sq.head = 0;
2830 		qp->sq.tail = 0;
2831 		qp->sq.cur_post = 0;
2832 		qp->sq.last_poll = 0;
2833 		qp->db.db[MLX5_RCV_DBR] = 0;
2834 		qp->db.db[MLX5_SND_DBR] = 0;
2835 	}
2836 
2837 out:
2838 	kfree(context);
2839 	return err;
2840 }
2841 
2842 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2843 		      int attr_mask, struct ib_udata *udata)
2844 {
2845 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2846 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2847 	enum ib_qp_type qp_type;
2848 	enum ib_qp_state cur_state, new_state;
2849 	int err = -EINVAL;
2850 	int port;
2851 	enum rdma_link_layer ll = IB_LINK_LAYER_UNSPECIFIED;
2852 
2853 	if (ibqp->rwq_ind_tbl)
2854 		return -ENOSYS;
2855 
2856 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
2857 		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
2858 
2859 	qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ?
2860 		IB_QPT_GSI : ibqp->qp_type;
2861 
2862 	mutex_lock(&qp->mutex);
2863 
2864 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2865 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2866 
2867 	if (!(cur_state == new_state && cur_state == IB_QPS_RESET)) {
2868 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2869 		ll = dev->ib_dev.get_link_layer(&dev->ib_dev, port);
2870 	}
2871 
2872 	if (qp_type != MLX5_IB_QPT_REG_UMR &&
2873 	    !ib_modify_qp_is_ok(cur_state, new_state, qp_type, attr_mask, ll)) {
2874 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
2875 			    cur_state, new_state, ibqp->qp_type, attr_mask);
2876 		goto out;
2877 	}
2878 
2879 	if ((attr_mask & IB_QP_PORT) &&
2880 	    (attr->port_num == 0 ||
2881 	     attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports))) {
2882 		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
2883 			    attr->port_num, dev->num_ports);
2884 		goto out;
2885 	}
2886 
2887 	if (attr_mask & IB_QP_PKEY_INDEX) {
2888 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2889 		if (attr->pkey_index >=
2890 		    dev->mdev->port_caps[port - 1].pkey_table_len) {
2891 			mlx5_ib_dbg(dev, "invalid pkey index %d\n",
2892 				    attr->pkey_index);
2893 			goto out;
2894 		}
2895 	}
2896 
2897 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2898 	    attr->max_rd_atomic >
2899 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
2900 		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
2901 			    attr->max_rd_atomic);
2902 		goto out;
2903 	}
2904 
2905 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2906 	    attr->max_dest_rd_atomic >
2907 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
2908 		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
2909 			    attr->max_dest_rd_atomic);
2910 		goto out;
2911 	}
2912 
2913 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2914 		err = 0;
2915 		goto out;
2916 	}
2917 
2918 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2919 
2920 out:
2921 	mutex_unlock(&qp->mutex);
2922 	return err;
2923 }
2924 
2925 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2926 {
2927 	struct mlx5_ib_cq *cq;
2928 	unsigned cur;
2929 
2930 	cur = wq->head - wq->tail;
2931 	if (likely(cur + nreq < wq->max_post))
2932 		return 0;
2933 
2934 	cq = to_mcq(ib_cq);
2935 	spin_lock(&cq->lock);
2936 	cur = wq->head - wq->tail;
2937 	spin_unlock(&cq->lock);
2938 
2939 	return cur + nreq >= wq->max_post;
2940 }
2941 
2942 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
2943 					  u64 remote_addr, u32 rkey)
2944 {
2945 	rseg->raddr    = cpu_to_be64(remote_addr);
2946 	rseg->rkey     = cpu_to_be32(rkey);
2947 	rseg->reserved = 0;
2948 }
2949 
2950 static void *set_eth_seg(struct mlx5_wqe_eth_seg *eseg,
2951 			 struct ib_send_wr *wr, void *qend,
2952 			 struct mlx5_ib_qp *qp, int *size)
2953 {
2954 	void *seg = eseg;
2955 
2956 	memset(eseg, 0, sizeof(struct mlx5_wqe_eth_seg));
2957 
2958 	if (wr->send_flags & IB_SEND_IP_CSUM)
2959 		eseg->cs_flags = MLX5_ETH_WQE_L3_CSUM |
2960 				 MLX5_ETH_WQE_L4_CSUM;
2961 
2962 	seg += sizeof(struct mlx5_wqe_eth_seg);
2963 	*size += sizeof(struct mlx5_wqe_eth_seg) / 16;
2964 
2965 	if (wr->opcode == IB_WR_LSO) {
2966 		struct ib_ud_wr *ud_wr = container_of(wr, struct ib_ud_wr, wr);
2967 		int size_of_inl_hdr_start = sizeof(eseg->inline_hdr_start);
2968 		u64 left, leftlen, copysz;
2969 		void *pdata = ud_wr->header;
2970 
2971 		left = ud_wr->hlen;
2972 		eseg->mss = cpu_to_be16(ud_wr->mss);
2973 		eseg->inline_hdr_sz = cpu_to_be16(left);
2974 
2975 		/*
2976 		 * check if there is space till the end of queue, if yes,
2977 		 * copy all in one shot, otherwise copy till the end of queue,
2978 		 * rollback and than the copy the left
2979 		 */
2980 		leftlen = qend - (void *)eseg->inline_hdr_start;
2981 		copysz = min_t(u64, leftlen, left);
2982 
2983 		memcpy(seg - size_of_inl_hdr_start, pdata, copysz);
2984 
2985 		if (likely(copysz > size_of_inl_hdr_start)) {
2986 			seg += ALIGN(copysz - size_of_inl_hdr_start, 16);
2987 			*size += ALIGN(copysz - size_of_inl_hdr_start, 16) / 16;
2988 		}
2989 
2990 		if (unlikely(copysz < left)) { /* the last wqe in the queue */
2991 			seg = mlx5_get_send_wqe(qp, 0);
2992 			left -= copysz;
2993 			pdata += copysz;
2994 			memcpy(seg, pdata, left);
2995 			seg += ALIGN(left, 16);
2996 			*size += ALIGN(left, 16) / 16;
2997 		}
2998 	}
2999 
3000 	return seg;
3001 }
3002 
3003 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
3004 			     struct ib_send_wr *wr)
3005 {
3006 	memcpy(&dseg->av, &to_mah(ud_wr(wr)->ah)->av, sizeof(struct mlx5_av));
3007 	dseg->av.dqp_dct = cpu_to_be32(ud_wr(wr)->remote_qpn | MLX5_EXTENDED_UD_AV);
3008 	dseg->av.key.qkey.qkey = cpu_to_be32(ud_wr(wr)->remote_qkey);
3009 }
3010 
3011 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
3012 {
3013 	dseg->byte_count = cpu_to_be32(sg->length);
3014 	dseg->lkey       = cpu_to_be32(sg->lkey);
3015 	dseg->addr       = cpu_to_be64(sg->addr);
3016 }
3017 
3018 static __be16 get_klm_octo(int npages)
3019 {
3020 	return cpu_to_be16(ALIGN(npages, 8) / 2);
3021 }
3022 
3023 static __be64 frwr_mkey_mask(void)
3024 {
3025 	u64 result;
3026 
3027 	result = MLX5_MKEY_MASK_LEN		|
3028 		MLX5_MKEY_MASK_PAGE_SIZE	|
3029 		MLX5_MKEY_MASK_START_ADDR	|
3030 		MLX5_MKEY_MASK_EN_RINVAL	|
3031 		MLX5_MKEY_MASK_KEY		|
3032 		MLX5_MKEY_MASK_LR		|
3033 		MLX5_MKEY_MASK_LW		|
3034 		MLX5_MKEY_MASK_RR		|
3035 		MLX5_MKEY_MASK_RW		|
3036 		MLX5_MKEY_MASK_A		|
3037 		MLX5_MKEY_MASK_SMALL_FENCE	|
3038 		MLX5_MKEY_MASK_FREE;
3039 
3040 	return cpu_to_be64(result);
3041 }
3042 
3043 static __be64 sig_mkey_mask(void)
3044 {
3045 	u64 result;
3046 
3047 	result = MLX5_MKEY_MASK_LEN		|
3048 		MLX5_MKEY_MASK_PAGE_SIZE	|
3049 		MLX5_MKEY_MASK_START_ADDR	|
3050 		MLX5_MKEY_MASK_EN_SIGERR	|
3051 		MLX5_MKEY_MASK_EN_RINVAL	|
3052 		MLX5_MKEY_MASK_KEY		|
3053 		MLX5_MKEY_MASK_LR		|
3054 		MLX5_MKEY_MASK_LW		|
3055 		MLX5_MKEY_MASK_RR		|
3056 		MLX5_MKEY_MASK_RW		|
3057 		MLX5_MKEY_MASK_SMALL_FENCE	|
3058 		MLX5_MKEY_MASK_FREE		|
3059 		MLX5_MKEY_MASK_BSF_EN;
3060 
3061 	return cpu_to_be64(result);
3062 }
3063 
3064 static void set_reg_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr,
3065 				struct mlx5_ib_mr *mr)
3066 {
3067 	int ndescs = mr->ndescs;
3068 
3069 	memset(umr, 0, sizeof(*umr));
3070 
3071 	if (mr->access_mode == MLX5_ACCESS_MODE_KLM)
3072 		/* KLMs take twice the size of MTTs */
3073 		ndescs *= 2;
3074 
3075 	umr->flags = MLX5_UMR_CHECK_NOT_FREE;
3076 	umr->klm_octowords = get_klm_octo(ndescs);
3077 	umr->mkey_mask = frwr_mkey_mask();
3078 }
3079 
3080 static void set_linv_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr)
3081 {
3082 	memset(umr, 0, sizeof(*umr));
3083 	umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
3084 	umr->flags = 1 << 7;
3085 }
3086 
3087 static __be64 get_umr_reg_mr_mask(void)
3088 {
3089 	u64 result;
3090 
3091 	result = MLX5_MKEY_MASK_LEN		|
3092 		 MLX5_MKEY_MASK_PAGE_SIZE	|
3093 		 MLX5_MKEY_MASK_START_ADDR	|
3094 		 MLX5_MKEY_MASK_PD		|
3095 		 MLX5_MKEY_MASK_LR		|
3096 		 MLX5_MKEY_MASK_LW		|
3097 		 MLX5_MKEY_MASK_KEY		|
3098 		 MLX5_MKEY_MASK_RR		|
3099 		 MLX5_MKEY_MASK_RW		|
3100 		 MLX5_MKEY_MASK_A		|
3101 		 MLX5_MKEY_MASK_FREE;
3102 
3103 	return cpu_to_be64(result);
3104 }
3105 
3106 static __be64 get_umr_unreg_mr_mask(void)
3107 {
3108 	u64 result;
3109 
3110 	result = MLX5_MKEY_MASK_FREE;
3111 
3112 	return cpu_to_be64(result);
3113 }
3114 
3115 static __be64 get_umr_update_mtt_mask(void)
3116 {
3117 	u64 result;
3118 
3119 	result = MLX5_MKEY_MASK_FREE;
3120 
3121 	return cpu_to_be64(result);
3122 }
3123 
3124 static __be64 get_umr_update_translation_mask(void)
3125 {
3126 	u64 result;
3127 
3128 	result = MLX5_MKEY_MASK_LEN |
3129 		 MLX5_MKEY_MASK_PAGE_SIZE |
3130 		 MLX5_MKEY_MASK_START_ADDR |
3131 		 MLX5_MKEY_MASK_KEY |
3132 		 MLX5_MKEY_MASK_FREE;
3133 
3134 	return cpu_to_be64(result);
3135 }
3136 
3137 static __be64 get_umr_update_access_mask(void)
3138 {
3139 	u64 result;
3140 
3141 	result = MLX5_MKEY_MASK_LW |
3142 		 MLX5_MKEY_MASK_RR |
3143 		 MLX5_MKEY_MASK_RW |
3144 		 MLX5_MKEY_MASK_A |
3145 		 MLX5_MKEY_MASK_KEY |
3146 		 MLX5_MKEY_MASK_FREE;
3147 
3148 	return cpu_to_be64(result);
3149 }
3150 
3151 static __be64 get_umr_update_pd_mask(void)
3152 {
3153 	u64 result;
3154 
3155 	result = MLX5_MKEY_MASK_PD |
3156 		 MLX5_MKEY_MASK_KEY |
3157 		 MLX5_MKEY_MASK_FREE;
3158 
3159 	return cpu_to_be64(result);
3160 }
3161 
3162 static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
3163 				struct ib_send_wr *wr)
3164 {
3165 	struct mlx5_umr_wr *umrwr = umr_wr(wr);
3166 
3167 	memset(umr, 0, sizeof(*umr));
3168 
3169 	if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
3170 		umr->flags = MLX5_UMR_CHECK_FREE; /* fail if free */
3171 	else
3172 		umr->flags = MLX5_UMR_CHECK_NOT_FREE; /* fail if not free */
3173 
3174 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UNREG)) {
3175 		umr->klm_octowords = get_klm_octo(umrwr->npages);
3176 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT) {
3177 			umr->mkey_mask = get_umr_update_mtt_mask();
3178 			umr->bsf_octowords = get_klm_octo(umrwr->target.offset);
3179 			umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
3180 		}
3181 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION)
3182 			umr->mkey_mask |= get_umr_update_translation_mask();
3183 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_ACCESS)
3184 			umr->mkey_mask |= get_umr_update_access_mask();
3185 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_PD)
3186 			umr->mkey_mask |= get_umr_update_pd_mask();
3187 		if (!umr->mkey_mask)
3188 			umr->mkey_mask = get_umr_reg_mr_mask();
3189 	} else {
3190 		umr->mkey_mask = get_umr_unreg_mr_mask();
3191 	}
3192 
3193 	if (!wr->num_sge)
3194 		umr->flags |= MLX5_UMR_INLINE;
3195 }
3196 
3197 static u8 get_umr_flags(int acc)
3198 {
3199 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
3200 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
3201 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
3202 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
3203 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
3204 }
3205 
3206 static void set_reg_mkey_seg(struct mlx5_mkey_seg *seg,
3207 			     struct mlx5_ib_mr *mr,
3208 			     u32 key, int access)
3209 {
3210 	int ndescs = ALIGN(mr->ndescs, 8) >> 1;
3211 
3212 	memset(seg, 0, sizeof(*seg));
3213 
3214 	if (mr->access_mode == MLX5_ACCESS_MODE_MTT)
3215 		seg->log2_page_size = ilog2(mr->ibmr.page_size);
3216 	else if (mr->access_mode == MLX5_ACCESS_MODE_KLM)
3217 		/* KLMs take twice the size of MTTs */
3218 		ndescs *= 2;
3219 
3220 	seg->flags = get_umr_flags(access) | mr->access_mode;
3221 	seg->qpn_mkey7_0 = cpu_to_be32((key & 0xff) | 0xffffff00);
3222 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
3223 	seg->start_addr = cpu_to_be64(mr->ibmr.iova);
3224 	seg->len = cpu_to_be64(mr->ibmr.length);
3225 	seg->xlt_oct_size = cpu_to_be32(ndescs);
3226 }
3227 
3228 static void set_linv_mkey_seg(struct mlx5_mkey_seg *seg)
3229 {
3230 	memset(seg, 0, sizeof(*seg));
3231 	seg->status = MLX5_MKEY_STATUS_FREE;
3232 }
3233 
3234 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr)
3235 {
3236 	struct mlx5_umr_wr *umrwr = umr_wr(wr);
3237 
3238 	memset(seg, 0, sizeof(*seg));
3239 	if (wr->send_flags & MLX5_IB_SEND_UMR_UNREG) {
3240 		seg->status = MLX5_MKEY_STATUS_FREE;
3241 		return;
3242 	}
3243 
3244 	seg->flags = convert_access(umrwr->access_flags);
3245 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT)) {
3246 		if (umrwr->pd)
3247 			seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
3248 		seg->start_addr = cpu_to_be64(umrwr->target.virt_addr);
3249 	}
3250 	seg->len = cpu_to_be64(umrwr->length);
3251 	seg->log2_page_size = umrwr->page_shift;
3252 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
3253 				       mlx5_mkey_variant(umrwr->mkey));
3254 }
3255 
3256 static void set_reg_data_seg(struct mlx5_wqe_data_seg *dseg,
3257 			     struct mlx5_ib_mr *mr,
3258 			     struct mlx5_ib_pd *pd)
3259 {
3260 	int bcount = mr->desc_size * mr->ndescs;
3261 
3262 	dseg->addr = cpu_to_be64(mr->desc_map);
3263 	dseg->byte_count = cpu_to_be32(ALIGN(bcount, 64));
3264 	dseg->lkey = cpu_to_be32(pd->ibpd.local_dma_lkey);
3265 }
3266 
3267 static __be32 send_ieth(struct ib_send_wr *wr)
3268 {
3269 	switch (wr->opcode) {
3270 	case IB_WR_SEND_WITH_IMM:
3271 	case IB_WR_RDMA_WRITE_WITH_IMM:
3272 		return wr->ex.imm_data;
3273 
3274 	case IB_WR_SEND_WITH_INV:
3275 		return cpu_to_be32(wr->ex.invalidate_rkey);
3276 
3277 	default:
3278 		return 0;
3279 	}
3280 }
3281 
3282 static u8 calc_sig(void *wqe, int size)
3283 {
3284 	u8 *p = wqe;
3285 	u8 res = 0;
3286 	int i;
3287 
3288 	for (i = 0; i < size; i++)
3289 		res ^= p[i];
3290 
3291 	return ~res;
3292 }
3293 
3294 static u8 wq_sig(void *wqe)
3295 {
3296 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
3297 }
3298 
3299 static int set_data_inl_seg(struct mlx5_ib_qp *qp, struct ib_send_wr *wr,
3300 			    void *wqe, int *sz)
3301 {
3302 	struct mlx5_wqe_inline_seg *seg;
3303 	void *qend = qp->sq.qend;
3304 	void *addr;
3305 	int inl = 0;
3306 	int copy;
3307 	int len;
3308 	int i;
3309 
3310 	seg = wqe;
3311 	wqe += sizeof(*seg);
3312 	for (i = 0; i < wr->num_sge; i++) {
3313 		addr = (void *)(unsigned long)(wr->sg_list[i].addr);
3314 		len  = wr->sg_list[i].length;
3315 		inl += len;
3316 
3317 		if (unlikely(inl > qp->max_inline_data))
3318 			return -ENOMEM;
3319 
3320 		if (unlikely(wqe + len > qend)) {
3321 			copy = qend - wqe;
3322 			memcpy(wqe, addr, copy);
3323 			addr += copy;
3324 			len -= copy;
3325 			wqe = mlx5_get_send_wqe(qp, 0);
3326 		}
3327 		memcpy(wqe, addr, len);
3328 		wqe += len;
3329 	}
3330 
3331 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
3332 
3333 	*sz = ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
3334 
3335 	return 0;
3336 }
3337 
3338 static u16 prot_field_size(enum ib_signature_type type)
3339 {
3340 	switch (type) {
3341 	case IB_SIG_TYPE_T10_DIF:
3342 		return MLX5_DIF_SIZE;
3343 	default:
3344 		return 0;
3345 	}
3346 }
3347 
3348 static u8 bs_selector(int block_size)
3349 {
3350 	switch (block_size) {
3351 	case 512:	    return 0x1;
3352 	case 520:	    return 0x2;
3353 	case 4096:	    return 0x3;
3354 	case 4160:	    return 0x4;
3355 	case 1073741824:    return 0x5;
3356 	default:	    return 0;
3357 	}
3358 }
3359 
3360 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
3361 			      struct mlx5_bsf_inl *inl)
3362 {
3363 	/* Valid inline section and allow BSF refresh */
3364 	inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
3365 				       MLX5_BSF_REFRESH_DIF);
3366 	inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
3367 	inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
3368 	/* repeating block */
3369 	inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
3370 	inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
3371 			MLX5_DIF_CRC : MLX5_DIF_IPCS;
3372 
3373 	if (domain->sig.dif.ref_remap)
3374 		inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
3375 
3376 	if (domain->sig.dif.app_escape) {
3377 		if (domain->sig.dif.ref_escape)
3378 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
3379 		else
3380 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
3381 	}
3382 
3383 	inl->dif_app_bitmask_check =
3384 		cpu_to_be16(domain->sig.dif.apptag_check_mask);
3385 }
3386 
3387 static int mlx5_set_bsf(struct ib_mr *sig_mr,
3388 			struct ib_sig_attrs *sig_attrs,
3389 			struct mlx5_bsf *bsf, u32 data_size)
3390 {
3391 	struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
3392 	struct mlx5_bsf_basic *basic = &bsf->basic;
3393 	struct ib_sig_domain *mem = &sig_attrs->mem;
3394 	struct ib_sig_domain *wire = &sig_attrs->wire;
3395 
3396 	memset(bsf, 0, sizeof(*bsf));
3397 
3398 	/* Basic + Extended + Inline */
3399 	basic->bsf_size_sbs = 1 << 7;
3400 	/* Input domain check byte mask */
3401 	basic->check_byte_mask = sig_attrs->check_mask;
3402 	basic->raw_data_size = cpu_to_be32(data_size);
3403 
3404 	/* Memory domain */
3405 	switch (sig_attrs->mem.sig_type) {
3406 	case IB_SIG_TYPE_NONE:
3407 		break;
3408 	case IB_SIG_TYPE_T10_DIF:
3409 		basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
3410 		basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
3411 		mlx5_fill_inl_bsf(mem, &bsf->m_inl);
3412 		break;
3413 	default:
3414 		return -EINVAL;
3415 	}
3416 
3417 	/* Wire domain */
3418 	switch (sig_attrs->wire.sig_type) {
3419 	case IB_SIG_TYPE_NONE:
3420 		break;
3421 	case IB_SIG_TYPE_T10_DIF:
3422 		if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
3423 		    mem->sig_type == wire->sig_type) {
3424 			/* Same block structure */
3425 			basic->bsf_size_sbs |= 1 << 4;
3426 			if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
3427 				basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
3428 			if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
3429 				basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
3430 			if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
3431 				basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
3432 		} else
3433 			basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
3434 
3435 		basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
3436 		mlx5_fill_inl_bsf(wire, &bsf->w_inl);
3437 		break;
3438 	default:
3439 		return -EINVAL;
3440 	}
3441 
3442 	return 0;
3443 }
3444 
3445 static int set_sig_data_segment(struct ib_sig_handover_wr *wr,
3446 				struct mlx5_ib_qp *qp, void **seg, int *size)
3447 {
3448 	struct ib_sig_attrs *sig_attrs = wr->sig_attrs;
3449 	struct ib_mr *sig_mr = wr->sig_mr;
3450 	struct mlx5_bsf *bsf;
3451 	u32 data_len = wr->wr.sg_list->length;
3452 	u32 data_key = wr->wr.sg_list->lkey;
3453 	u64 data_va = wr->wr.sg_list->addr;
3454 	int ret;
3455 	int wqe_size;
3456 
3457 	if (!wr->prot ||
3458 	    (data_key == wr->prot->lkey &&
3459 	     data_va == wr->prot->addr &&
3460 	     data_len == wr->prot->length)) {
3461 		/**
3462 		 * Source domain doesn't contain signature information
3463 		 * or data and protection are interleaved in memory.
3464 		 * So need construct:
3465 		 *                  ------------------
3466 		 *                 |     data_klm     |
3467 		 *                  ------------------
3468 		 *                 |       BSF        |
3469 		 *                  ------------------
3470 		 **/
3471 		struct mlx5_klm *data_klm = *seg;
3472 
3473 		data_klm->bcount = cpu_to_be32(data_len);
3474 		data_klm->key = cpu_to_be32(data_key);
3475 		data_klm->va = cpu_to_be64(data_va);
3476 		wqe_size = ALIGN(sizeof(*data_klm), 64);
3477 	} else {
3478 		/**
3479 		 * Source domain contains signature information
3480 		 * So need construct a strided block format:
3481 		 *               ---------------------------
3482 		 *              |     stride_block_ctrl     |
3483 		 *               ---------------------------
3484 		 *              |          data_klm         |
3485 		 *               ---------------------------
3486 		 *              |          prot_klm         |
3487 		 *               ---------------------------
3488 		 *              |             BSF           |
3489 		 *               ---------------------------
3490 		 **/
3491 		struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
3492 		struct mlx5_stride_block_entry *data_sentry;
3493 		struct mlx5_stride_block_entry *prot_sentry;
3494 		u32 prot_key = wr->prot->lkey;
3495 		u64 prot_va = wr->prot->addr;
3496 		u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
3497 		int prot_size;
3498 
3499 		sblock_ctrl = *seg;
3500 		data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
3501 		prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
3502 
3503 		prot_size = prot_field_size(sig_attrs->mem.sig_type);
3504 		if (!prot_size) {
3505 			pr_err("Bad block size given: %u\n", block_size);
3506 			return -EINVAL;
3507 		}
3508 		sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
3509 							    prot_size);
3510 		sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
3511 		sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
3512 		sblock_ctrl->num_entries = cpu_to_be16(2);
3513 
3514 		data_sentry->bcount = cpu_to_be16(block_size);
3515 		data_sentry->key = cpu_to_be32(data_key);
3516 		data_sentry->va = cpu_to_be64(data_va);
3517 		data_sentry->stride = cpu_to_be16(block_size);
3518 
3519 		prot_sentry->bcount = cpu_to_be16(prot_size);
3520 		prot_sentry->key = cpu_to_be32(prot_key);
3521 		prot_sentry->va = cpu_to_be64(prot_va);
3522 		prot_sentry->stride = cpu_to_be16(prot_size);
3523 
3524 		wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
3525 				 sizeof(*prot_sentry), 64);
3526 	}
3527 
3528 	*seg += wqe_size;
3529 	*size += wqe_size / 16;
3530 	if (unlikely((*seg == qp->sq.qend)))
3531 		*seg = mlx5_get_send_wqe(qp, 0);
3532 
3533 	bsf = *seg;
3534 	ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
3535 	if (ret)
3536 		return -EINVAL;
3537 
3538 	*seg += sizeof(*bsf);
3539 	*size += sizeof(*bsf) / 16;
3540 	if (unlikely((*seg == qp->sq.qend)))
3541 		*seg = mlx5_get_send_wqe(qp, 0);
3542 
3543 	return 0;
3544 }
3545 
3546 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
3547 				 struct ib_sig_handover_wr *wr, u32 nelements,
3548 				 u32 length, u32 pdn)
3549 {
3550 	struct ib_mr *sig_mr = wr->sig_mr;
3551 	u32 sig_key = sig_mr->rkey;
3552 	u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
3553 
3554 	memset(seg, 0, sizeof(*seg));
3555 
3556 	seg->flags = get_umr_flags(wr->access_flags) |
3557 				   MLX5_ACCESS_MODE_KLM;
3558 	seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
3559 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
3560 				    MLX5_MKEY_BSF_EN | pdn);
3561 	seg->len = cpu_to_be64(length);
3562 	seg->xlt_oct_size = cpu_to_be32(be16_to_cpu(get_klm_octo(nelements)));
3563 	seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
3564 }
3565 
3566 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
3567 				u32 nelements)
3568 {
3569 	memset(umr, 0, sizeof(*umr));
3570 
3571 	umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
3572 	umr->klm_octowords = get_klm_octo(nelements);
3573 	umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
3574 	umr->mkey_mask = sig_mkey_mask();
3575 }
3576 
3577 
3578 static int set_sig_umr_wr(struct ib_send_wr *send_wr, struct mlx5_ib_qp *qp,
3579 			  void **seg, int *size)
3580 {
3581 	struct ib_sig_handover_wr *wr = sig_handover_wr(send_wr);
3582 	struct mlx5_ib_mr *sig_mr = to_mmr(wr->sig_mr);
3583 	u32 pdn = get_pd(qp)->pdn;
3584 	u32 klm_oct_size;
3585 	int region_len, ret;
3586 
3587 	if (unlikely(wr->wr.num_sge != 1) ||
3588 	    unlikely(wr->access_flags & IB_ACCESS_REMOTE_ATOMIC) ||
3589 	    unlikely(!sig_mr->sig) || unlikely(!qp->signature_en) ||
3590 	    unlikely(!sig_mr->sig->sig_status_checked))
3591 		return -EINVAL;
3592 
3593 	/* length of the protected region, data + protection */
3594 	region_len = wr->wr.sg_list->length;
3595 	if (wr->prot &&
3596 	    (wr->prot->lkey != wr->wr.sg_list->lkey  ||
3597 	     wr->prot->addr != wr->wr.sg_list->addr  ||
3598 	     wr->prot->length != wr->wr.sg_list->length))
3599 		region_len += wr->prot->length;
3600 
3601 	/**
3602 	 * KLM octoword size - if protection was provided
3603 	 * then we use strided block format (3 octowords),
3604 	 * else we use single KLM (1 octoword)
3605 	 **/
3606 	klm_oct_size = wr->prot ? 3 : 1;
3607 
3608 	set_sig_umr_segment(*seg, klm_oct_size);
3609 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3610 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3611 	if (unlikely((*seg == qp->sq.qend)))
3612 		*seg = mlx5_get_send_wqe(qp, 0);
3613 
3614 	set_sig_mkey_segment(*seg, wr, klm_oct_size, region_len, pdn);
3615 	*seg += sizeof(struct mlx5_mkey_seg);
3616 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3617 	if (unlikely((*seg == qp->sq.qend)))
3618 		*seg = mlx5_get_send_wqe(qp, 0);
3619 
3620 	ret = set_sig_data_segment(wr, qp, seg, size);
3621 	if (ret)
3622 		return ret;
3623 
3624 	sig_mr->sig->sig_status_checked = false;
3625 	return 0;
3626 }
3627 
3628 static int set_psv_wr(struct ib_sig_domain *domain,
3629 		      u32 psv_idx, void **seg, int *size)
3630 {
3631 	struct mlx5_seg_set_psv *psv_seg = *seg;
3632 
3633 	memset(psv_seg, 0, sizeof(*psv_seg));
3634 	psv_seg->psv_num = cpu_to_be32(psv_idx);
3635 	switch (domain->sig_type) {
3636 	case IB_SIG_TYPE_NONE:
3637 		break;
3638 	case IB_SIG_TYPE_T10_DIF:
3639 		psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
3640 						     domain->sig.dif.app_tag);
3641 		psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
3642 		break;
3643 	default:
3644 		pr_err("Bad signature type given.\n");
3645 		return 1;
3646 	}
3647 
3648 	*seg += sizeof(*psv_seg);
3649 	*size += sizeof(*psv_seg) / 16;
3650 
3651 	return 0;
3652 }
3653 
3654 static int set_reg_wr(struct mlx5_ib_qp *qp,
3655 		      struct ib_reg_wr *wr,
3656 		      void **seg, int *size)
3657 {
3658 	struct mlx5_ib_mr *mr = to_mmr(wr->mr);
3659 	struct mlx5_ib_pd *pd = to_mpd(qp->ibqp.pd);
3660 
3661 	if (unlikely(wr->wr.send_flags & IB_SEND_INLINE)) {
3662 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
3663 			     "Invalid IB_SEND_INLINE send flag\n");
3664 		return -EINVAL;
3665 	}
3666 
3667 	set_reg_umr_seg(*seg, mr);
3668 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3669 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3670 	if (unlikely((*seg == qp->sq.qend)))
3671 		*seg = mlx5_get_send_wqe(qp, 0);
3672 
3673 	set_reg_mkey_seg(*seg, mr, wr->key, wr->access);
3674 	*seg += sizeof(struct mlx5_mkey_seg);
3675 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3676 	if (unlikely((*seg == qp->sq.qend)))
3677 		*seg = mlx5_get_send_wqe(qp, 0);
3678 
3679 	set_reg_data_seg(*seg, mr, pd);
3680 	*seg += sizeof(struct mlx5_wqe_data_seg);
3681 	*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
3682 
3683 	return 0;
3684 }
3685 
3686 static void set_linv_wr(struct mlx5_ib_qp *qp, void **seg, int *size)
3687 {
3688 	set_linv_umr_seg(*seg);
3689 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3690 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3691 	if (unlikely((*seg == qp->sq.qend)))
3692 		*seg = mlx5_get_send_wqe(qp, 0);
3693 	set_linv_mkey_seg(*seg);
3694 	*seg += sizeof(struct mlx5_mkey_seg);
3695 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3696 	if (unlikely((*seg == qp->sq.qend)))
3697 		*seg = mlx5_get_send_wqe(qp, 0);
3698 }
3699 
3700 static void dump_wqe(struct mlx5_ib_qp *qp, int idx, int size_16)
3701 {
3702 	__be32 *p = NULL;
3703 	int tidx = idx;
3704 	int i, j;
3705 
3706 	pr_debug("dump wqe at %p\n", mlx5_get_send_wqe(qp, tidx));
3707 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
3708 		if ((i & 0xf) == 0) {
3709 			void *buf = mlx5_get_send_wqe(qp, tidx);
3710 			tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
3711 			p = buf;
3712 			j = 0;
3713 		}
3714 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
3715 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
3716 			 be32_to_cpu(p[j + 3]));
3717 	}
3718 }
3719 
3720 static void mlx5_bf_copy(u64 __iomem *dst, u64 *src,
3721 			 unsigned bytecnt, struct mlx5_ib_qp *qp)
3722 {
3723 	while (bytecnt > 0) {
3724 		__iowrite64_copy(dst++, src++, 8);
3725 		__iowrite64_copy(dst++, src++, 8);
3726 		__iowrite64_copy(dst++, src++, 8);
3727 		__iowrite64_copy(dst++, src++, 8);
3728 		__iowrite64_copy(dst++, src++, 8);
3729 		__iowrite64_copy(dst++, src++, 8);
3730 		__iowrite64_copy(dst++, src++, 8);
3731 		__iowrite64_copy(dst++, src++, 8);
3732 		bytecnt -= 64;
3733 		if (unlikely(src == qp->sq.qend))
3734 			src = mlx5_get_send_wqe(qp, 0);
3735 	}
3736 }
3737 
3738 static u8 get_fence(u8 fence, struct ib_send_wr *wr)
3739 {
3740 	if (unlikely(wr->opcode == IB_WR_LOCAL_INV &&
3741 		     wr->send_flags & IB_SEND_FENCE))
3742 		return MLX5_FENCE_MODE_STRONG_ORDERING;
3743 
3744 	if (unlikely(fence)) {
3745 		if (wr->send_flags & IB_SEND_FENCE)
3746 			return MLX5_FENCE_MODE_SMALL_AND_FENCE;
3747 		else
3748 			return fence;
3749 	} else if (unlikely(wr->send_flags & IB_SEND_FENCE)) {
3750 		return MLX5_FENCE_MODE_FENCE;
3751 	}
3752 
3753 	return 0;
3754 }
3755 
3756 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
3757 		     struct mlx5_wqe_ctrl_seg **ctrl,
3758 		     struct ib_send_wr *wr, unsigned *idx,
3759 		     int *size, int nreq)
3760 {
3761 	if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)))
3762 		return -ENOMEM;
3763 
3764 	*idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
3765 	*seg = mlx5_get_send_wqe(qp, *idx);
3766 	*ctrl = *seg;
3767 	*(uint32_t *)(*seg + 8) = 0;
3768 	(*ctrl)->imm = send_ieth(wr);
3769 	(*ctrl)->fm_ce_se = qp->sq_signal_bits |
3770 		(wr->send_flags & IB_SEND_SIGNALED ?
3771 		 MLX5_WQE_CTRL_CQ_UPDATE : 0) |
3772 		(wr->send_flags & IB_SEND_SOLICITED ?
3773 		 MLX5_WQE_CTRL_SOLICITED : 0);
3774 
3775 	*seg += sizeof(**ctrl);
3776 	*size = sizeof(**ctrl) / 16;
3777 
3778 	return 0;
3779 }
3780 
3781 static void finish_wqe(struct mlx5_ib_qp *qp,
3782 		       struct mlx5_wqe_ctrl_seg *ctrl,
3783 		       u8 size, unsigned idx, u64 wr_id,
3784 		       int nreq, u8 fence, u8 next_fence,
3785 		       u32 mlx5_opcode)
3786 {
3787 	u8 opmod = 0;
3788 
3789 	ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
3790 					     mlx5_opcode | ((u32)opmod << 24));
3791 	ctrl->qpn_ds = cpu_to_be32(size | (qp->trans_qp.base.mqp.qpn << 8));
3792 	ctrl->fm_ce_se |= fence;
3793 	qp->fm_cache = next_fence;
3794 	if (unlikely(qp->wq_sig))
3795 		ctrl->signature = wq_sig(ctrl);
3796 
3797 	qp->sq.wrid[idx] = wr_id;
3798 	qp->sq.w_list[idx].opcode = mlx5_opcode;
3799 	qp->sq.wqe_head[idx] = qp->sq.head + nreq;
3800 	qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
3801 	qp->sq.w_list[idx].next = qp->sq.cur_post;
3802 }
3803 
3804 
3805 int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
3806 		      struct ib_send_wr **bad_wr)
3807 {
3808 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
3809 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3810 	struct mlx5_core_dev *mdev = dev->mdev;
3811 	struct mlx5_ib_qp *qp;
3812 	struct mlx5_ib_mr *mr;
3813 	struct mlx5_wqe_data_seg *dpseg;
3814 	struct mlx5_wqe_xrc_seg *xrc;
3815 	struct mlx5_bf *bf;
3816 	int uninitialized_var(size);
3817 	void *qend;
3818 	unsigned long flags;
3819 	unsigned idx;
3820 	int err = 0;
3821 	int inl = 0;
3822 	int num_sge;
3823 	void *seg;
3824 	int nreq;
3825 	int i;
3826 	u8 next_fence = 0;
3827 	u8 fence;
3828 
3829 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
3830 		return mlx5_ib_gsi_post_send(ibqp, wr, bad_wr);
3831 
3832 	qp = to_mqp(ibqp);
3833 	bf = qp->bf;
3834 	qend = qp->sq.qend;
3835 
3836 	spin_lock_irqsave(&qp->sq.lock, flags);
3837 
3838 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
3839 		err = -EIO;
3840 		*bad_wr = wr;
3841 		nreq = 0;
3842 		goto out;
3843 	}
3844 
3845 	for (nreq = 0; wr; nreq++, wr = wr->next) {
3846 		if (unlikely(wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
3847 			mlx5_ib_warn(dev, "\n");
3848 			err = -EINVAL;
3849 			*bad_wr = wr;
3850 			goto out;
3851 		}
3852 
3853 		fence = qp->fm_cache;
3854 		num_sge = wr->num_sge;
3855 		if (unlikely(num_sge > qp->sq.max_gs)) {
3856 			mlx5_ib_warn(dev, "\n");
3857 			err = -EINVAL;
3858 			*bad_wr = wr;
3859 			goto out;
3860 		}
3861 
3862 		err = begin_wqe(qp, &seg, &ctrl, wr, &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 		switch (ibqp->qp_type) {
3871 		case IB_QPT_XRC_INI:
3872 			xrc = seg;
3873 			seg += sizeof(*xrc);
3874 			size += sizeof(*xrc) / 16;
3875 			/* fall through */
3876 		case IB_QPT_RC:
3877 			switch (wr->opcode) {
3878 			case IB_WR_RDMA_READ:
3879 			case IB_WR_RDMA_WRITE:
3880 			case IB_WR_RDMA_WRITE_WITH_IMM:
3881 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
3882 					      rdma_wr(wr)->rkey);
3883 				seg += sizeof(struct mlx5_wqe_raddr_seg);
3884 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
3885 				break;
3886 
3887 			case IB_WR_ATOMIC_CMP_AND_SWP:
3888 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3889 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3890 				mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
3891 				err = -ENOSYS;
3892 				*bad_wr = wr;
3893 				goto out;
3894 
3895 			case IB_WR_LOCAL_INV:
3896 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3897 				qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
3898 				ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
3899 				set_linv_wr(qp, &seg, &size);
3900 				num_sge = 0;
3901 				break;
3902 
3903 			case IB_WR_REG_MR:
3904 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3905 				qp->sq.wr_data[idx] = IB_WR_REG_MR;
3906 				ctrl->imm = cpu_to_be32(reg_wr(wr)->key);
3907 				err = set_reg_wr(qp, reg_wr(wr), &seg, &size);
3908 				if (err) {
3909 					*bad_wr = wr;
3910 					goto out;
3911 				}
3912 				num_sge = 0;
3913 				break;
3914 
3915 			case IB_WR_REG_SIG_MR:
3916 				qp->sq.wr_data[idx] = IB_WR_REG_SIG_MR;
3917 				mr = to_mmr(sig_handover_wr(wr)->sig_mr);
3918 
3919 				ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
3920 				err = set_sig_umr_wr(wr, qp, &seg, &size);
3921 				if (err) {
3922 					mlx5_ib_warn(dev, "\n");
3923 					*bad_wr = wr;
3924 					goto out;
3925 				}
3926 
3927 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3928 					   nreq, get_fence(fence, wr),
3929 					   next_fence, MLX5_OPCODE_UMR);
3930 				/*
3931 				 * SET_PSV WQEs are not signaled and solicited
3932 				 * on error
3933 				 */
3934 				wr->send_flags &= ~IB_SEND_SIGNALED;
3935 				wr->send_flags |= IB_SEND_SOLICITED;
3936 				err = begin_wqe(qp, &seg, &ctrl, wr,
3937 						&idx, &size, nreq);
3938 				if (err) {
3939 					mlx5_ib_warn(dev, "\n");
3940 					err = -ENOMEM;
3941 					*bad_wr = wr;
3942 					goto out;
3943 				}
3944 
3945 				err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->mem,
3946 						 mr->sig->psv_memory.psv_idx, &seg,
3947 						 &size);
3948 				if (err) {
3949 					mlx5_ib_warn(dev, "\n");
3950 					*bad_wr = wr;
3951 					goto out;
3952 				}
3953 
3954 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3955 					   nreq, get_fence(fence, wr),
3956 					   next_fence, MLX5_OPCODE_SET_PSV);
3957 				err = begin_wqe(qp, &seg, &ctrl, wr,
3958 						&idx, &size, nreq);
3959 				if (err) {
3960 					mlx5_ib_warn(dev, "\n");
3961 					err = -ENOMEM;
3962 					*bad_wr = wr;
3963 					goto out;
3964 				}
3965 
3966 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3967 				err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->wire,
3968 						 mr->sig->psv_wire.psv_idx, &seg,
3969 						 &size);
3970 				if (err) {
3971 					mlx5_ib_warn(dev, "\n");
3972 					*bad_wr = wr;
3973 					goto out;
3974 				}
3975 
3976 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3977 					   nreq, get_fence(fence, wr),
3978 					   next_fence, MLX5_OPCODE_SET_PSV);
3979 				num_sge = 0;
3980 				goto skip_psv;
3981 
3982 			default:
3983 				break;
3984 			}
3985 			break;
3986 
3987 		case IB_QPT_UC:
3988 			switch (wr->opcode) {
3989 			case IB_WR_RDMA_WRITE:
3990 			case IB_WR_RDMA_WRITE_WITH_IMM:
3991 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
3992 					      rdma_wr(wr)->rkey);
3993 				seg  += sizeof(struct mlx5_wqe_raddr_seg);
3994 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
3995 				break;
3996 
3997 			default:
3998 				break;
3999 			}
4000 			break;
4001 
4002 		case IB_QPT_SMI:
4003 		case MLX5_IB_QPT_HW_GSI:
4004 			set_datagram_seg(seg, wr);
4005 			seg += sizeof(struct mlx5_wqe_datagram_seg);
4006 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
4007 			if (unlikely((seg == qend)))
4008 				seg = mlx5_get_send_wqe(qp, 0);
4009 			break;
4010 		case IB_QPT_UD:
4011 			set_datagram_seg(seg, wr);
4012 			seg += sizeof(struct mlx5_wqe_datagram_seg);
4013 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
4014 
4015 			if (unlikely((seg == qend)))
4016 				seg = mlx5_get_send_wqe(qp, 0);
4017 
4018 			/* handle qp that supports ud offload */
4019 			if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO) {
4020 				struct mlx5_wqe_eth_pad *pad;
4021 
4022 				pad = seg;
4023 				memset(pad, 0, sizeof(struct mlx5_wqe_eth_pad));
4024 				seg += sizeof(struct mlx5_wqe_eth_pad);
4025 				size += sizeof(struct mlx5_wqe_eth_pad) / 16;
4026 
4027 				seg = set_eth_seg(seg, wr, qend, qp, &size);
4028 
4029 				if (unlikely((seg == qend)))
4030 					seg = mlx5_get_send_wqe(qp, 0);
4031 			}
4032 			break;
4033 		case MLX5_IB_QPT_REG_UMR:
4034 			if (wr->opcode != MLX5_IB_WR_UMR) {
4035 				err = -EINVAL;
4036 				mlx5_ib_warn(dev, "bad opcode\n");
4037 				goto out;
4038 			}
4039 			qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
4040 			ctrl->imm = cpu_to_be32(umr_wr(wr)->mkey);
4041 			set_reg_umr_segment(seg, wr);
4042 			seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4043 			size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4044 			if (unlikely((seg == qend)))
4045 				seg = mlx5_get_send_wqe(qp, 0);
4046 			set_reg_mkey_segment(seg, wr);
4047 			seg += sizeof(struct mlx5_mkey_seg);
4048 			size += sizeof(struct mlx5_mkey_seg) / 16;
4049 			if (unlikely((seg == qend)))
4050 				seg = mlx5_get_send_wqe(qp, 0);
4051 			break;
4052 
4053 		default:
4054 			break;
4055 		}
4056 
4057 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
4058 			int uninitialized_var(sz);
4059 
4060 			err = set_data_inl_seg(qp, wr, seg, &sz);
4061 			if (unlikely(err)) {
4062 				mlx5_ib_warn(dev, "\n");
4063 				*bad_wr = wr;
4064 				goto out;
4065 			}
4066 			inl = 1;
4067 			size += sz;
4068 		} else {
4069 			dpseg = seg;
4070 			for (i = 0; i < num_sge; i++) {
4071 				if (unlikely(dpseg == qend)) {
4072 					seg = mlx5_get_send_wqe(qp, 0);
4073 					dpseg = seg;
4074 				}
4075 				if (likely(wr->sg_list[i].length)) {
4076 					set_data_ptr_seg(dpseg, wr->sg_list + i);
4077 					size += sizeof(struct mlx5_wqe_data_seg) / 16;
4078 					dpseg++;
4079 				}
4080 			}
4081 		}
4082 
4083 		finish_wqe(qp, ctrl, size, idx, wr->wr_id, nreq,
4084 			   get_fence(fence, wr), next_fence,
4085 			   mlx5_ib_opcode[wr->opcode]);
4086 skip_psv:
4087 		if (0)
4088 			dump_wqe(qp, idx, size);
4089 	}
4090 
4091 out:
4092 	if (likely(nreq)) {
4093 		qp->sq.head += nreq;
4094 
4095 		/* Make sure that descriptors are written before
4096 		 * updating doorbell record and ringing the doorbell
4097 		 */
4098 		wmb();
4099 
4100 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
4101 
4102 		/* Make sure doorbell record is visible to the HCA before
4103 		 * we hit doorbell */
4104 		wmb();
4105 
4106 		if (bf->need_lock)
4107 			spin_lock(&bf->lock);
4108 		else
4109 			__acquire(&bf->lock);
4110 
4111 		/* TBD enable WC */
4112 		if (0 && nreq == 1 && bf->uuarn && inl && size > 1 && size <= bf->buf_size / 16) {
4113 			mlx5_bf_copy(bf->reg + bf->offset, (u64 *)ctrl, ALIGN(size * 16, 64), qp);
4114 			/* wc_wmb(); */
4115 		} else {
4116 			mlx5_write64((__be32 *)ctrl, bf->regreg + bf->offset,
4117 				     MLX5_GET_DOORBELL_LOCK(&bf->lock32));
4118 			/* Make sure doorbells don't leak out of SQ spinlock
4119 			 * and reach the HCA out of order.
4120 			 */
4121 			mmiowb();
4122 		}
4123 		bf->offset ^= bf->buf_size;
4124 		if (bf->need_lock)
4125 			spin_unlock(&bf->lock);
4126 		else
4127 			__release(&bf->lock);
4128 	}
4129 
4130 	spin_unlock_irqrestore(&qp->sq.lock, flags);
4131 
4132 	return err;
4133 }
4134 
4135 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
4136 {
4137 	sig->signature = calc_sig(sig, size);
4138 }
4139 
4140 int mlx5_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
4141 		      struct ib_recv_wr **bad_wr)
4142 {
4143 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4144 	struct mlx5_wqe_data_seg *scat;
4145 	struct mlx5_rwqe_sig *sig;
4146 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4147 	struct mlx5_core_dev *mdev = dev->mdev;
4148 	unsigned long flags;
4149 	int err = 0;
4150 	int nreq;
4151 	int ind;
4152 	int i;
4153 
4154 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4155 		return mlx5_ib_gsi_post_recv(ibqp, wr, bad_wr);
4156 
4157 	spin_lock_irqsave(&qp->rq.lock, flags);
4158 
4159 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
4160 		err = -EIO;
4161 		*bad_wr = wr;
4162 		nreq = 0;
4163 		goto out;
4164 	}
4165 
4166 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
4167 
4168 	for (nreq = 0; wr; nreq++, wr = wr->next) {
4169 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
4170 			err = -ENOMEM;
4171 			*bad_wr = wr;
4172 			goto out;
4173 		}
4174 
4175 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
4176 			err = -EINVAL;
4177 			*bad_wr = wr;
4178 			goto out;
4179 		}
4180 
4181 		scat = get_recv_wqe(qp, ind);
4182 		if (qp->wq_sig)
4183 			scat++;
4184 
4185 		for (i = 0; i < wr->num_sge; i++)
4186 			set_data_ptr_seg(scat + i, wr->sg_list + i);
4187 
4188 		if (i < qp->rq.max_gs) {
4189 			scat[i].byte_count = 0;
4190 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
4191 			scat[i].addr       = 0;
4192 		}
4193 
4194 		if (qp->wq_sig) {
4195 			sig = (struct mlx5_rwqe_sig *)scat;
4196 			set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
4197 		}
4198 
4199 		qp->rq.wrid[ind] = wr->wr_id;
4200 
4201 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
4202 	}
4203 
4204 out:
4205 	if (likely(nreq)) {
4206 		qp->rq.head += nreq;
4207 
4208 		/* Make sure that descriptors are written before
4209 		 * doorbell record.
4210 		 */
4211 		wmb();
4212 
4213 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
4214 	}
4215 
4216 	spin_unlock_irqrestore(&qp->rq.lock, flags);
4217 
4218 	return err;
4219 }
4220 
4221 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
4222 {
4223 	switch (mlx5_state) {
4224 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
4225 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
4226 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
4227 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
4228 	case MLX5_QP_STATE_SQ_DRAINING:
4229 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
4230 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
4231 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
4232 	default:		     return -1;
4233 	}
4234 }
4235 
4236 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
4237 {
4238 	switch (mlx5_mig_state) {
4239 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
4240 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
4241 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
4242 	default: return -1;
4243 	}
4244 }
4245 
4246 static int to_ib_qp_access_flags(int mlx5_flags)
4247 {
4248 	int ib_flags = 0;
4249 
4250 	if (mlx5_flags & MLX5_QP_BIT_RRE)
4251 		ib_flags |= IB_ACCESS_REMOTE_READ;
4252 	if (mlx5_flags & MLX5_QP_BIT_RWE)
4253 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
4254 	if (mlx5_flags & MLX5_QP_BIT_RAE)
4255 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
4256 
4257 	return ib_flags;
4258 }
4259 
4260 static void to_ib_ah_attr(struct mlx5_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
4261 				struct mlx5_qp_path *path)
4262 {
4263 	struct mlx5_core_dev *dev = ibdev->mdev;
4264 
4265 	memset(ib_ah_attr, 0, sizeof(*ib_ah_attr));
4266 	ib_ah_attr->port_num	  = path->port;
4267 
4268 	if (ib_ah_attr->port_num == 0 ||
4269 	    ib_ah_attr->port_num > MLX5_CAP_GEN(dev, num_ports))
4270 		return;
4271 
4272 	ib_ah_attr->sl = path->dci_cfi_prio_sl & 0xf;
4273 
4274 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
4275 	ib_ah_attr->src_path_bits = path->grh_mlid & 0x7f;
4276 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
4277 	ib_ah_attr->ah_flags      = (path->grh_mlid & (1 << 7)) ? IB_AH_GRH : 0;
4278 	if (ib_ah_attr->ah_flags) {
4279 		ib_ah_attr->grh.sgid_index = path->mgid_index;
4280 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
4281 		ib_ah_attr->grh.traffic_class =
4282 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
4283 		ib_ah_attr->grh.flow_label =
4284 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
4285 		memcpy(ib_ah_attr->grh.dgid.raw,
4286 		       path->rgid, sizeof(ib_ah_attr->grh.dgid.raw));
4287 	}
4288 }
4289 
4290 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
4291 					struct mlx5_ib_sq *sq,
4292 					u8 *sq_state)
4293 {
4294 	void *out;
4295 	void *sqc;
4296 	int inlen;
4297 	int err;
4298 
4299 	inlen = MLX5_ST_SZ_BYTES(query_sq_out);
4300 	out = mlx5_vzalloc(inlen);
4301 	if (!out)
4302 		return -ENOMEM;
4303 
4304 	err = mlx5_core_query_sq(dev->mdev, sq->base.mqp.qpn, out);
4305 	if (err)
4306 		goto out;
4307 
4308 	sqc = MLX5_ADDR_OF(query_sq_out, out, sq_context);
4309 	*sq_state = MLX5_GET(sqc, sqc, state);
4310 	sq->state = *sq_state;
4311 
4312 out:
4313 	kvfree(out);
4314 	return err;
4315 }
4316 
4317 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
4318 					struct mlx5_ib_rq *rq,
4319 					u8 *rq_state)
4320 {
4321 	void *out;
4322 	void *rqc;
4323 	int inlen;
4324 	int err;
4325 
4326 	inlen = MLX5_ST_SZ_BYTES(query_rq_out);
4327 	out = mlx5_vzalloc(inlen);
4328 	if (!out)
4329 		return -ENOMEM;
4330 
4331 	err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
4332 	if (err)
4333 		goto out;
4334 
4335 	rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
4336 	*rq_state = MLX5_GET(rqc, rqc, state);
4337 	rq->state = *rq_state;
4338 
4339 out:
4340 	kvfree(out);
4341 	return err;
4342 }
4343 
4344 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
4345 				  struct mlx5_ib_qp *qp, u8 *qp_state)
4346 {
4347 	static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
4348 		[MLX5_RQC_STATE_RST] = {
4349 			[MLX5_SQC_STATE_RST]	= IB_QPS_RESET,
4350 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4351 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE_BAD,
4352 			[MLX5_SQ_STATE_NA]	= IB_QPS_RESET,
4353 		},
4354 		[MLX5_RQC_STATE_RDY] = {
4355 			[MLX5_SQC_STATE_RST]	= MLX5_QP_STATE_BAD,
4356 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4357 			[MLX5_SQC_STATE_ERR]	= IB_QPS_SQE,
4358 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE,
4359 		},
4360 		[MLX5_RQC_STATE_ERR] = {
4361 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE_BAD,
4362 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4363 			[MLX5_SQC_STATE_ERR]	= IB_QPS_ERR,
4364 			[MLX5_SQ_STATE_NA]	= IB_QPS_ERR,
4365 		},
4366 		[MLX5_RQ_STATE_NA] = {
4367 			[MLX5_SQC_STATE_RST]    = IB_QPS_RESET,
4368 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4369 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE,
4370 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE_BAD,
4371 		},
4372 	};
4373 
4374 	*qp_state = sqrq_trans[rq_state][sq_state];
4375 
4376 	if (*qp_state == MLX5_QP_STATE_BAD) {
4377 		WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
4378 		     qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
4379 		     qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
4380 		return -EINVAL;
4381 	}
4382 
4383 	if (*qp_state == MLX5_QP_STATE)
4384 		*qp_state = qp->state;
4385 
4386 	return 0;
4387 }
4388 
4389 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
4390 				     struct mlx5_ib_qp *qp,
4391 				     u8 *raw_packet_qp_state)
4392 {
4393 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
4394 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
4395 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
4396 	int err;
4397 	u8 sq_state = MLX5_SQ_STATE_NA;
4398 	u8 rq_state = MLX5_RQ_STATE_NA;
4399 
4400 	if (qp->sq.wqe_cnt) {
4401 		err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
4402 		if (err)
4403 			return err;
4404 	}
4405 
4406 	if (qp->rq.wqe_cnt) {
4407 		err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
4408 		if (err)
4409 			return err;
4410 	}
4411 
4412 	return sqrq_state_to_qp_state(sq_state, rq_state, qp,
4413 				      raw_packet_qp_state);
4414 }
4415 
4416 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
4417 			 struct ib_qp_attr *qp_attr)
4418 {
4419 	int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
4420 	struct mlx5_qp_context *context;
4421 	int mlx5_state;
4422 	u32 *outb;
4423 	int err = 0;
4424 
4425 	outb = kzalloc(outlen, GFP_KERNEL);
4426 	if (!outb)
4427 		return -ENOMEM;
4428 
4429 	err = mlx5_core_qp_query(dev->mdev, &qp->trans_qp.base.mqp, outb,
4430 				 outlen);
4431 	if (err)
4432 		goto out;
4433 
4434 	/* FIXME: use MLX5_GET rather than mlx5_qp_context manual struct */
4435 	context = (struct mlx5_qp_context *)MLX5_ADDR_OF(query_qp_out, outb, qpc);
4436 
4437 	mlx5_state = be32_to_cpu(context->flags) >> 28;
4438 
4439 	qp->state		     = to_ib_qp_state(mlx5_state);
4440 	qp_attr->path_mtu	     = context->mtu_msgmax >> 5;
4441 	qp_attr->path_mig_state	     =
4442 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
4443 	qp_attr->qkey		     = be32_to_cpu(context->qkey);
4444 	qp_attr->rq_psn		     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
4445 	qp_attr->sq_psn		     = be32_to_cpu(context->next_send_psn) & 0xffffff;
4446 	qp_attr->dest_qp_num	     = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
4447 	qp_attr->qp_access_flags     =
4448 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
4449 
4450 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4451 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
4452 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
4453 		qp_attr->alt_pkey_index =
4454 			be16_to_cpu(context->alt_path.pkey_index);
4455 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
4456 	}
4457 
4458 	qp_attr->pkey_index = be16_to_cpu(context->pri_path.pkey_index);
4459 	qp_attr->port_num = context->pri_path.port;
4460 
4461 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4462 	qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
4463 
4464 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
4465 
4466 	qp_attr->max_dest_rd_atomic =
4467 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
4468 	qp_attr->min_rnr_timer	    =
4469 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
4470 	qp_attr->timeout	    = context->pri_path.ackto_lt >> 3;
4471 	qp_attr->retry_cnt	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
4472 	qp_attr->rnr_retry	    = (be32_to_cpu(context->params1) >> 13) & 0x7;
4473 	qp_attr->alt_timeout	    = context->alt_path.ackto_lt >> 3;
4474 
4475 out:
4476 	kfree(outb);
4477 	return err;
4478 }
4479 
4480 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
4481 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
4482 {
4483 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4484 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4485 	int err = 0;
4486 	u8 raw_packet_qp_state;
4487 
4488 	if (ibqp->rwq_ind_tbl)
4489 		return -ENOSYS;
4490 
4491 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4492 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
4493 					    qp_init_attr);
4494 
4495 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4496 	/*
4497 	 * Wait for any outstanding page faults, in case the user frees memory
4498 	 * based upon this query's result.
4499 	 */
4500 	flush_workqueue(mlx5_ib_page_fault_wq);
4501 #endif
4502 
4503 	mutex_lock(&qp->mutex);
4504 
4505 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
4506 		err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
4507 		if (err)
4508 			goto out;
4509 		qp->state = raw_packet_qp_state;
4510 		qp_attr->port_num = 1;
4511 	} else {
4512 		err = query_qp_attr(dev, qp, qp_attr);
4513 		if (err)
4514 			goto out;
4515 	}
4516 
4517 	qp_attr->qp_state	     = qp->state;
4518 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4519 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4520 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4521 
4522 	if (!ibqp->uobject) {
4523 		qp_attr->cap.max_send_wr  = qp->sq.max_post;
4524 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4525 		qp_init_attr->qp_context = ibqp->qp_context;
4526 	} else {
4527 		qp_attr->cap.max_send_wr  = 0;
4528 		qp_attr->cap.max_send_sge = 0;
4529 	}
4530 
4531 	qp_init_attr->qp_type = ibqp->qp_type;
4532 	qp_init_attr->recv_cq = ibqp->recv_cq;
4533 	qp_init_attr->send_cq = ibqp->send_cq;
4534 	qp_init_attr->srq = ibqp->srq;
4535 	qp_attr->cap.max_inline_data = qp->max_inline_data;
4536 
4537 	qp_init_attr->cap	     = qp_attr->cap;
4538 
4539 	qp_init_attr->create_flags = 0;
4540 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4541 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4542 
4543 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
4544 		qp_init_attr->create_flags |= IB_QP_CREATE_CROSS_CHANNEL;
4545 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
4546 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_SEND;
4547 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
4548 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_RECV;
4549 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
4550 		qp_init_attr->create_flags |= mlx5_ib_create_qp_sqpn_qp1();
4551 
4552 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
4553 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4554 
4555 out:
4556 	mutex_unlock(&qp->mutex);
4557 	return err;
4558 }
4559 
4560 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
4561 					  struct ib_ucontext *context,
4562 					  struct ib_udata *udata)
4563 {
4564 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4565 	struct mlx5_ib_xrcd *xrcd;
4566 	int err;
4567 
4568 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
4569 		return ERR_PTR(-ENOSYS);
4570 
4571 	xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
4572 	if (!xrcd)
4573 		return ERR_PTR(-ENOMEM);
4574 
4575 	err = mlx5_core_xrcd_alloc(dev->mdev, &xrcd->xrcdn);
4576 	if (err) {
4577 		kfree(xrcd);
4578 		return ERR_PTR(-ENOMEM);
4579 	}
4580 
4581 	return &xrcd->ibxrcd;
4582 }
4583 
4584 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
4585 {
4586 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
4587 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
4588 	int err;
4589 
4590 	err = mlx5_core_xrcd_dealloc(dev->mdev, xrcdn);
4591 	if (err) {
4592 		mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
4593 		return err;
4594 	}
4595 
4596 	kfree(xrcd);
4597 
4598 	return 0;
4599 }
4600 
4601 static void mlx5_ib_wq_event(struct mlx5_core_qp *core_qp, int type)
4602 {
4603 	struct mlx5_ib_rwq *rwq = to_mibrwq(core_qp);
4604 	struct mlx5_ib_dev *dev = to_mdev(rwq->ibwq.device);
4605 	struct ib_event event;
4606 
4607 	if (rwq->ibwq.event_handler) {
4608 		event.device     = rwq->ibwq.device;
4609 		event.element.wq = &rwq->ibwq;
4610 		switch (type) {
4611 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
4612 			event.event = IB_EVENT_WQ_FATAL;
4613 			break;
4614 		default:
4615 			mlx5_ib_warn(dev, "Unexpected event type %d on WQ %06x\n", type, core_qp->qpn);
4616 			return;
4617 		}
4618 
4619 		rwq->ibwq.event_handler(&event, rwq->ibwq.wq_context);
4620 	}
4621 }
4622 
4623 static int  create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
4624 		      struct ib_wq_init_attr *init_attr)
4625 {
4626 	struct mlx5_ib_dev *dev;
4627 	__be64 *rq_pas0;
4628 	void *in;
4629 	void *rqc;
4630 	void *wq;
4631 	int inlen;
4632 	int err;
4633 
4634 	dev = to_mdev(pd->device);
4635 
4636 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
4637 	in = mlx5_vzalloc(inlen);
4638 	if (!in)
4639 		return -ENOMEM;
4640 
4641 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
4642 	MLX5_SET(rqc,  rqc, mem_rq_type,
4643 		 MLX5_RQC_RQ_TYPE_MEMORY_RQ_INLINE);
4644 	MLX5_SET(rqc, rqc, user_index, rwq->user_index);
4645 	MLX5_SET(rqc,  rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
4646 	MLX5_SET(rqc,  rqc, state, MLX5_RQC_STATE_RST);
4647 	MLX5_SET(rqc,  rqc, flush_in_error_en, 1);
4648 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
4649 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
4650 	MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
4651 	MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
4652 	MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
4653 	MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
4654 	MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
4655 	MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
4656 	MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
4657 	MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
4658 	rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
4659 	mlx5_ib_populate_pas(dev, rwq->umem, rwq->page_shift, rq_pas0, 0);
4660 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rwq->core_qp);
4661 	kvfree(in);
4662 	return err;
4663 }
4664 
4665 static int set_user_rq_size(struct mlx5_ib_dev *dev,
4666 			    struct ib_wq_init_attr *wq_init_attr,
4667 			    struct mlx5_ib_create_wq *ucmd,
4668 			    struct mlx5_ib_rwq *rwq)
4669 {
4670 	/* Sanity check RQ size before proceeding */
4671 	if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
4672 		return -EINVAL;
4673 
4674 	if (!ucmd->rq_wqe_count)
4675 		return -EINVAL;
4676 
4677 	rwq->wqe_count = ucmd->rq_wqe_count;
4678 	rwq->wqe_shift = ucmd->rq_wqe_shift;
4679 	rwq->buf_size = (rwq->wqe_count << rwq->wqe_shift);
4680 	rwq->log_rq_stride = rwq->wqe_shift;
4681 	rwq->log_rq_size = ilog2(rwq->wqe_count);
4682 	return 0;
4683 }
4684 
4685 static int prepare_user_rq(struct ib_pd *pd,
4686 			   struct ib_wq_init_attr *init_attr,
4687 			   struct ib_udata *udata,
4688 			   struct mlx5_ib_rwq *rwq)
4689 {
4690 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
4691 	struct mlx5_ib_create_wq ucmd = {};
4692 	int err;
4693 	size_t required_cmd_sz;
4694 
4695 	required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
4696 	if (udata->inlen < required_cmd_sz) {
4697 		mlx5_ib_dbg(dev, "invalid inlen\n");
4698 		return -EINVAL;
4699 	}
4700 
4701 	if (udata->inlen > sizeof(ucmd) &&
4702 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4703 				 udata->inlen - sizeof(ucmd))) {
4704 		mlx5_ib_dbg(dev, "inlen is not supported\n");
4705 		return -EOPNOTSUPP;
4706 	}
4707 
4708 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
4709 		mlx5_ib_dbg(dev, "copy failed\n");
4710 		return -EFAULT;
4711 	}
4712 
4713 	if (ucmd.comp_mask) {
4714 		mlx5_ib_dbg(dev, "invalid comp mask\n");
4715 		return -EOPNOTSUPP;
4716 	}
4717 
4718 	if (ucmd.reserved) {
4719 		mlx5_ib_dbg(dev, "invalid reserved\n");
4720 		return -EOPNOTSUPP;
4721 	}
4722 
4723 	err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
4724 	if (err) {
4725 		mlx5_ib_dbg(dev, "err %d\n", err);
4726 		return err;
4727 	}
4728 
4729 	err = create_user_rq(dev, pd, rwq, &ucmd);
4730 	if (err) {
4731 		mlx5_ib_dbg(dev, "err %d\n", err);
4732 		if (err)
4733 			return err;
4734 	}
4735 
4736 	rwq->user_index = ucmd.user_index;
4737 	return 0;
4738 }
4739 
4740 struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
4741 				struct ib_wq_init_attr *init_attr,
4742 				struct ib_udata *udata)
4743 {
4744 	struct mlx5_ib_dev *dev;
4745 	struct mlx5_ib_rwq *rwq;
4746 	struct mlx5_ib_create_wq_resp resp = {};
4747 	size_t min_resp_len;
4748 	int err;
4749 
4750 	if (!udata)
4751 		return ERR_PTR(-ENOSYS);
4752 
4753 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4754 	if (udata->outlen && udata->outlen < min_resp_len)
4755 		return ERR_PTR(-EINVAL);
4756 
4757 	dev = to_mdev(pd->device);
4758 	switch (init_attr->wq_type) {
4759 	case IB_WQT_RQ:
4760 		rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
4761 		if (!rwq)
4762 			return ERR_PTR(-ENOMEM);
4763 		err = prepare_user_rq(pd, init_attr, udata, rwq);
4764 		if (err)
4765 			goto err;
4766 		err = create_rq(rwq, pd, init_attr);
4767 		if (err)
4768 			goto err_user_rq;
4769 		break;
4770 	default:
4771 		mlx5_ib_dbg(dev, "unsupported wq type %d\n",
4772 			    init_attr->wq_type);
4773 		return ERR_PTR(-EINVAL);
4774 	}
4775 
4776 	rwq->ibwq.wq_num = rwq->core_qp.qpn;
4777 	rwq->ibwq.state = IB_WQS_RESET;
4778 	if (udata->outlen) {
4779 		resp.response_length = offsetof(typeof(resp), response_length) +
4780 				sizeof(resp.response_length);
4781 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4782 		if (err)
4783 			goto err_copy;
4784 	}
4785 
4786 	rwq->core_qp.event = mlx5_ib_wq_event;
4787 	rwq->ibwq.event_handler = init_attr->event_handler;
4788 	return &rwq->ibwq;
4789 
4790 err_copy:
4791 	mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
4792 err_user_rq:
4793 	destroy_user_rq(pd, rwq);
4794 err:
4795 	kfree(rwq);
4796 	return ERR_PTR(err);
4797 }
4798 
4799 int mlx5_ib_destroy_wq(struct ib_wq *wq)
4800 {
4801 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
4802 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
4803 
4804 	mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
4805 	destroy_user_rq(wq->pd, rwq);
4806 	kfree(rwq);
4807 
4808 	return 0;
4809 }
4810 
4811 struct ib_rwq_ind_table *mlx5_ib_create_rwq_ind_table(struct ib_device *device,
4812 						      struct ib_rwq_ind_table_init_attr *init_attr,
4813 						      struct ib_udata *udata)
4814 {
4815 	struct mlx5_ib_dev *dev = to_mdev(device);
4816 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl;
4817 	int sz = 1 << init_attr->log_ind_tbl_size;
4818 	struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
4819 	size_t min_resp_len;
4820 	int inlen;
4821 	int err;
4822 	int i;
4823 	u32 *in;
4824 	void *rqtc;
4825 
4826 	if (udata->inlen > 0 &&
4827 	    !ib_is_udata_cleared(udata, 0,
4828 				 udata->inlen))
4829 		return ERR_PTR(-EOPNOTSUPP);
4830 
4831 	if (init_attr->log_ind_tbl_size >
4832 	    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size)) {
4833 		mlx5_ib_dbg(dev, "log_ind_tbl_size = %d is bigger than supported = %d\n",
4834 			    init_attr->log_ind_tbl_size,
4835 			    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size));
4836 		return ERR_PTR(-EINVAL);
4837 	}
4838 
4839 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4840 	if (udata->outlen && udata->outlen < min_resp_len)
4841 		return ERR_PTR(-EINVAL);
4842 
4843 	rwq_ind_tbl = kzalloc(sizeof(*rwq_ind_tbl), GFP_KERNEL);
4844 	if (!rwq_ind_tbl)
4845 		return ERR_PTR(-ENOMEM);
4846 
4847 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
4848 	in = mlx5_vzalloc(inlen);
4849 	if (!in) {
4850 		err = -ENOMEM;
4851 		goto err;
4852 	}
4853 
4854 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
4855 
4856 	MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
4857 	MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
4858 
4859 	for (i = 0; i < sz; i++)
4860 		MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
4861 
4862 	err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
4863 	kvfree(in);
4864 
4865 	if (err)
4866 		goto err;
4867 
4868 	rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
4869 	if (udata->outlen) {
4870 		resp.response_length = offsetof(typeof(resp), response_length) +
4871 					sizeof(resp.response_length);
4872 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4873 		if (err)
4874 			goto err_copy;
4875 	}
4876 
4877 	return &rwq_ind_tbl->ib_rwq_ind_tbl;
4878 
4879 err_copy:
4880 	mlx5_core_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn);
4881 err:
4882 	kfree(rwq_ind_tbl);
4883 	return ERR_PTR(err);
4884 }
4885 
4886 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4887 {
4888 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
4889 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
4890 
4891 	mlx5_core_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn);
4892 
4893 	kfree(rwq_ind_tbl);
4894 	return 0;
4895 }
4896 
4897 int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
4898 		      u32 wq_attr_mask, struct ib_udata *udata)
4899 {
4900 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
4901 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
4902 	struct mlx5_ib_modify_wq ucmd = {};
4903 	size_t required_cmd_sz;
4904 	int curr_wq_state;
4905 	int wq_state;
4906 	int inlen;
4907 	int err;
4908 	void *rqc;
4909 	void *in;
4910 
4911 	required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
4912 	if (udata->inlen < required_cmd_sz)
4913 		return -EINVAL;
4914 
4915 	if (udata->inlen > sizeof(ucmd) &&
4916 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4917 				 udata->inlen - sizeof(ucmd)))
4918 		return -EOPNOTSUPP;
4919 
4920 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4921 		return -EFAULT;
4922 
4923 	if (ucmd.comp_mask || ucmd.reserved)
4924 		return -EOPNOTSUPP;
4925 
4926 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
4927 	in = mlx5_vzalloc(inlen);
4928 	if (!in)
4929 		return -ENOMEM;
4930 
4931 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
4932 
4933 	MLX5_SET(modify_rq_in, in, rqn, rwq->core_qp.qpn);
4934 	curr_wq_state = (wq_attr_mask & IB_WQ_CUR_STATE) ?
4935 		wq_attr->curr_wq_state : wq->state;
4936 	wq_state = (wq_attr_mask & IB_WQ_STATE) ?
4937 		wq_attr->wq_state : curr_wq_state;
4938 	if (curr_wq_state == IB_WQS_ERR)
4939 		curr_wq_state = MLX5_RQC_STATE_ERR;
4940 	if (wq_state == IB_WQS_ERR)
4941 		wq_state = MLX5_RQC_STATE_ERR;
4942 	MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
4943 	MLX5_SET(rqc, rqc, state, wq_state);
4944 
4945 	err = mlx5_core_modify_rq(dev->mdev, in, inlen);
4946 	kvfree(in);
4947 	if (!err)
4948 		rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;
4949 
4950 	return err;
4951 }
4952