xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_qp.c (revision 5eb61f6c6549f134a4f3bed4c164345d4f616bad)
1 /*-
2  * Copyright (c) 2013-2021, Mellanox Technologies. 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 enum {
471 	/* this is the first blue flame register in the array of bfregs assigned
472 	 * to a processes. Since we do not use it for blue flame but rather
473 	 * regular 64 bit doorbells, we do not need a lock for maintaiing
474 	 * "odd/even" order
475 	 */
476 	NUM_NON_BLUE_FLAME_BFREGS = 1,
477 };
478 
479 static int max_bfregs(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi)
480 {
481 	return get_num_static_uars(dev, bfregi) * MLX5_NON_FP_BFREGS_PER_UAR;
482 }
483 
484 static int num_med_bfreg(struct mlx5_ib_dev *dev,
485 			 struct mlx5_bfreg_info *bfregi)
486 {
487 	int n;
488 
489 	n = max_bfregs(dev, bfregi) - bfregi->num_low_latency_bfregs -
490 	    NUM_NON_BLUE_FLAME_BFREGS;
491 
492 	return n >= 0 ? n : 0;
493 }
494 
495 static int first_med_bfreg(struct mlx5_ib_dev *dev,
496 			   struct mlx5_bfreg_info *bfregi)
497 {
498 	return num_med_bfreg(dev, bfregi) ? 1 : -ENOMEM;
499 }
500 
501 static int first_hi_bfreg(struct mlx5_ib_dev *dev,
502 			  struct mlx5_bfreg_info *bfregi)
503 {
504 	int med;
505 
506 	med = num_med_bfreg(dev, bfregi);
507 	return ++med;
508 }
509 
510 static int alloc_high_class_bfreg(struct mlx5_ib_dev *dev,
511 				  struct mlx5_bfreg_info *bfregi)
512 {
513 	int i;
514 
515 	for (i = first_hi_bfreg(dev, bfregi); i < max_bfregs(dev, bfregi); i++) {
516 		if (!bfregi->count[i]) {
517 			bfregi->count[i]++;
518 			return i;
519 		}
520 	}
521 
522 	return -ENOMEM;
523 }
524 
525 static int alloc_med_class_bfreg(struct mlx5_ib_dev *dev,
526 				 struct mlx5_bfreg_info *bfregi)
527 {
528 	int minidx = first_med_bfreg(dev, bfregi);
529 	int i;
530 
531 	if (minidx < 0)
532 		return minidx;
533 
534 	for (i = minidx; i < first_hi_bfreg(dev, bfregi); i++) {
535 		if (bfregi->count[i] < bfregi->count[minidx])
536 			minidx = i;
537 		if (!bfregi->count[minidx])
538 			break;
539 	}
540 
541 	bfregi->count[minidx]++;
542 	return minidx;
543 }
544 
545 static int alloc_bfreg(struct mlx5_ib_dev *dev,
546 		       struct mlx5_bfreg_info *bfregi)
547 {
548 	int bfregn = -ENOMEM;
549 
550 	if (bfregi->lib_uar_dyn)
551 		return -EINVAL;
552 
553 	mutex_lock(&bfregi->lock);
554 	if (bfregi->ver >= 2) {
555 		bfregn = alloc_high_class_bfreg(dev, bfregi);
556 		if (bfregn < 0)
557 			bfregn = alloc_med_class_bfreg(dev, bfregi);
558 	}
559 
560 	if (bfregn < 0) {
561 		BUILD_BUG_ON(NUM_NON_BLUE_FLAME_BFREGS != 1);
562 		bfregn = 0;
563 		bfregi->count[bfregn]++;
564 	}
565 	mutex_unlock(&bfregi->lock);
566 
567 	return bfregn;
568 }
569 
570 void mlx5_ib_free_bfreg(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi, int bfregn)
571 {
572 	mutex_lock(&bfregi->lock);
573 	bfregi->count[bfregn]--;
574 	mutex_unlock(&bfregi->lock);
575 }
576 
577 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
578 {
579 	switch (state) {
580 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
581 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
582 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
583 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
584 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
585 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
586 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
587 	default:		return -1;
588 	}
589 }
590 
591 static int to_mlx5_st(enum ib_qp_type type)
592 {
593 	switch (type) {
594 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
595 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
596 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
597 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
598 	case IB_QPT_XRC_INI:
599 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
600 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
601 	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
602 	case IB_QPT_RAW_IPV6:		return MLX5_QP_ST_RAW_IPV6;
603 	case IB_QPT_RAW_PACKET:
604 	case IB_QPT_RAW_ETHERTYPE:	return MLX5_QP_ST_RAW_ETHERTYPE;
605 	case IB_QPT_MAX:
606 	default:		return -EINVAL;
607 	}
608 }
609 
610 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
611 			     struct mlx5_ib_cq *recv_cq);
612 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
613 			       struct mlx5_ib_cq *recv_cq);
614 
615 int bfregn_to_uar_index(struct mlx5_ib_dev *dev,
616 			struct mlx5_bfreg_info *bfregi, u32 bfregn,
617 			bool dyn_bfreg)
618 {
619 	unsigned int bfregs_per_sys_page;
620 	u32 index_of_sys_page;
621 	u32 offset;
622 
623 	if (bfregi->lib_uar_dyn)
624 		return -EINVAL;
625 
626 	bfregs_per_sys_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k) *
627 				MLX5_NON_FP_BFREGS_PER_UAR;
628 	index_of_sys_page = bfregn / bfregs_per_sys_page;
629 
630 	if (dyn_bfreg) {
631 		index_of_sys_page += bfregi->num_static_sys_pages;
632 
633 		if (index_of_sys_page >= bfregi->num_sys_pages)
634 			return -EINVAL;
635 
636 		if (bfregn > bfregi->num_dyn_bfregs ||
637 		    bfregi->sys_pages[index_of_sys_page] == MLX5_IB_INVALID_UAR_INDEX) {
638 			mlx5_ib_dbg(dev, "Invalid dynamic uar index\n");
639 			return -EINVAL;
640 		}
641 	}
642 
643 	offset = bfregn % bfregs_per_sys_page / MLX5_NON_FP_BFREGS_PER_UAR;
644 	return bfregi->sys_pages[index_of_sys_page] + offset;
645 }
646 
647 static int mlx5_ib_umem_get(struct mlx5_ib_dev *dev,
648 			    struct ib_pd *pd,
649 			    unsigned long addr, size_t size,
650 			    struct ib_umem **umem,
651 			    int *npages, int *page_shift, int *ncont,
652 			    u32 *offset)
653 {
654 	int err;
655 
656 	*umem = ib_umem_get(pd->uobject->context, addr, size, 0, 0);
657 	if (IS_ERR(*umem)) {
658 		mlx5_ib_dbg(dev, "umem_get failed\n");
659 		return PTR_ERR(*umem);
660 	}
661 
662 	mlx5_ib_cont_pages(*umem, addr, 0, npages, page_shift, ncont, NULL);
663 
664 	err = mlx5_ib_get_buf_offset(addr, *page_shift, offset);
665 	if (err) {
666 		mlx5_ib_warn(dev, "bad offset\n");
667 		goto err_umem;
668 	}
669 
670 	mlx5_ib_dbg(dev, "addr 0x%lx, size %zu, npages %d, page_shift %d, ncont %d, offset %d\n",
671 		    addr, size, *npages, *page_shift, *ncont, *offset);
672 
673 	return 0;
674 
675 err_umem:
676 	ib_umem_release(*umem);
677 	*umem = NULL;
678 
679 	return err;
680 }
681 
682 static void destroy_user_rq(struct ib_pd *pd, struct mlx5_ib_rwq *rwq)
683 {
684 	struct mlx5_ib_ucontext *context;
685 
686 	context = to_mucontext(pd->uobject->context);
687 	mlx5_ib_db_unmap_user(context, &rwq->db);
688 	if (rwq->umem)
689 		ib_umem_release(rwq->umem);
690 }
691 
692 static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
693 			  struct mlx5_ib_rwq *rwq,
694 			  struct mlx5_ib_create_wq *ucmd)
695 {
696 	struct mlx5_ib_ucontext *context;
697 	int page_shift = 0;
698 	int npages;
699 	u32 offset = 0;
700 	int ncont = 0;
701 	int err;
702 
703 	if (!ucmd->buf_addr)
704 		return -EINVAL;
705 
706 	context = to_mucontext(pd->uobject->context);
707 	rwq->umem = ib_umem_get(pd->uobject->context, ucmd->buf_addr,
708 			       rwq->buf_size, 0, 0);
709 	if (IS_ERR(rwq->umem)) {
710 		mlx5_ib_dbg(dev, "umem_get failed\n");
711 		err = PTR_ERR(rwq->umem);
712 		return err;
713 	}
714 
715 	mlx5_ib_cont_pages(rwq->umem, ucmd->buf_addr, 0, &npages, &page_shift,
716 			   &ncont, NULL);
717 	err = mlx5_ib_get_buf_offset(ucmd->buf_addr, page_shift,
718 				     &rwq->rq_page_offset);
719 	if (err) {
720 		mlx5_ib_warn(dev, "bad offset\n");
721 		goto err_umem;
722 	}
723 
724 	rwq->rq_num_pas = ncont;
725 	rwq->page_shift = page_shift;
726 	rwq->log_page_size =  page_shift - MLX5_ADAPTER_PAGE_SHIFT;
727 	rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
728 
729 	mlx5_ib_dbg(dev, "addr 0x%llx, size %zd, npages %d, page_shift %d, ncont %d, offset %d\n",
730 		    (unsigned long long)ucmd->buf_addr, rwq->buf_size,
731 		    npages, page_shift, ncont, offset);
732 
733 	err = mlx5_ib_db_map_user(context, ucmd->db_addr, &rwq->db);
734 	if (err) {
735 		mlx5_ib_dbg(dev, "map failed\n");
736 		goto err_umem;
737 	}
738 
739 	rwq->create_type = MLX5_WQ_USER;
740 	return 0;
741 
742 err_umem:
743 	ib_umem_release(rwq->umem);
744 	return err;
745 }
746 
747 static int adjust_bfregn(struct mlx5_ib_dev *dev,
748 			 struct mlx5_bfreg_info *bfregi, int bfregn)
749 {
750 	return bfregn / MLX5_NON_FP_BFREGS_PER_UAR * MLX5_BFREGS_PER_UAR +
751 				bfregn % MLX5_NON_FP_BFREGS_PER_UAR;
752 }
753 
754 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
755 			  struct mlx5_ib_qp *qp, struct ib_udata *udata,
756 			  struct ib_qp_init_attr *attr,
757 			  u32 **in,
758 			  struct mlx5_ib_create_qp_resp *resp, int *inlen,
759 			  struct mlx5_ib_qp_base *base)
760 {
761 	struct mlx5_ib_ucontext *context;
762 	struct mlx5_ib_create_qp ucmd;
763 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
764 	int page_shift = 0;
765 	int uar_index = 0;
766 	int npages;
767 	u32 offset = 0;
768 	int bfregn;
769 	int ncont = 0;
770 	__be64 *pas;
771 	void *qpc;
772 	int err;
773 	u32 uar_flags;
774 
775 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
776 	if (err) {
777 		mlx5_ib_dbg(dev, "copy failed\n");
778 		return err;
779 	}
780 
781 	context = to_mucontext(pd->uobject->context);
782 	uar_flags = ucmd.flags & (MLX5_QP_FLAG_UAR_PAGE_INDEX |
783 				  MLX5_QP_FLAG_BFREG_INDEX);
784 	switch (uar_flags) {
785 	case MLX5_QP_FLAG_UAR_PAGE_INDEX:
786 		uar_index = ucmd.bfreg_index;
787 		bfregn = MLX5_IB_INVALID_BFREG;
788 		break;
789 	case MLX5_QP_FLAG_BFREG_INDEX:
790 		uar_index = bfregn_to_uar_index(dev, &context->bfregi,
791 						ucmd.bfreg_index, true);
792 		if (uar_index < 0)
793 			return uar_index;
794 		bfregn = MLX5_IB_INVALID_BFREG;
795 		break;
796 	case 0:
797 		if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
798 			return -EINVAL;
799 		bfregn = alloc_bfreg(dev, &context->bfregi);
800 		if (bfregn < 0)
801 			return bfregn;
802 		break;
803 	default:
804 		return -EINVAL;
805 	}
806 
807 	mlx5_ib_dbg(dev, "bfregn 0x%x, uar_index 0x%x\n", bfregn, uar_index);
808 	if (bfregn != MLX5_IB_INVALID_BFREG)
809 		uar_index = bfregn_to_uar_index(dev, &context->bfregi, bfregn,
810 						false);
811 
812 	qp->rq.offset = 0;
813 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
814 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
815 
816 	err = set_user_buf_size(dev, qp, &ucmd, base, attr);
817 	if (err)
818 		goto err_bfreg;
819 
820 	if (ucmd.buf_addr && ubuffer->buf_size) {
821 		ubuffer->buf_addr = ucmd.buf_addr;
822 		err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr,
823 				       ubuffer->buf_size,
824 				       &ubuffer->umem, &npages, &page_shift,
825 				       &ncont, &offset);
826 		if (err)
827 			goto err_bfreg;
828 	} else {
829 		ubuffer->umem = NULL;
830 	}
831 
832 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
833 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * ncont;
834 	*in = mlx5_vzalloc(*inlen);
835 	if (!*in) {
836 		err = -ENOMEM;
837 		goto err_umem;
838 	}
839 
840 	pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas);
841 	if (ubuffer->umem)
842 		mlx5_ib_populate_pas(dev, ubuffer->umem, page_shift, pas, 0);
843 
844 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
845 
846 	MLX5_SET(qpc, qpc, log_page_size, page_shift - MLX5_ADAPTER_PAGE_SHIFT);
847 	MLX5_SET(qpc, qpc, page_offset, offset);
848 
849 	MLX5_SET(qpc, qpc, uar_page, uar_index);
850 	if (bfregn != MLX5_IB_INVALID_BFREG)
851 		resp->bfreg_index = adjust_bfregn(dev, &context->bfregi, bfregn);
852 	else
853 		resp->bfreg_index = MLX5_IB_INVALID_BFREG;
854 	qp->bfregn = bfregn;
855 
856 	err = mlx5_ib_db_map_user(context, ucmd.db_addr, &qp->db);
857 	if (err) {
858 		mlx5_ib_dbg(dev, "map failed\n");
859 		goto err_free;
860 	}
861 
862 	err = ib_copy_to_udata(udata, resp, sizeof(*resp));
863 	if (err) {
864 		mlx5_ib_dbg(dev, "copy failed\n");
865 		goto err_unmap;
866 	}
867 	qp->create_type = MLX5_QP_USER;
868 
869 	return 0;
870 
871 err_unmap:
872 	mlx5_ib_db_unmap_user(context, &qp->db);
873 
874 err_free:
875 	kvfree(*in);
876 
877 err_umem:
878 	if (ubuffer->umem)
879 		ib_umem_release(ubuffer->umem);
880 
881 err_bfreg:
882 	if (bfregn != MLX5_IB_INVALID_BFREG)
883 		mlx5_ib_free_bfreg(dev, &context->bfregi, bfregn);
884 	return err;
885 }
886 
887 static void destroy_qp_user(struct mlx5_ib_dev *dev, struct ib_pd *pd, struct mlx5_ib_qp *qp,
888 			    struct mlx5_ib_qp_base *base)
889 {
890 	struct mlx5_ib_ucontext *context;
891 
892 	context = to_mucontext(pd->uobject->context);
893 	mlx5_ib_db_unmap_user(context, &qp->db);
894 	if (base->ubuffer.umem)
895 		ib_umem_release(base->ubuffer.umem);
896 
897 	/*
898 	 * Free only the BFREGs which are handled by the kernel.
899 	 * BFREGs of UARs allocated dynamically are handled by user.
900 	 */
901 	if (qp->bfregn != MLX5_IB_INVALID_BFREG)
902 		mlx5_ib_free_bfreg(dev, &context->bfregi, qp->bfregn);
903 }
904 
905 static int create_kernel_qp(struct mlx5_ib_dev *dev,
906 			    struct ib_qp_init_attr *init_attr,
907 			    struct mlx5_ib_qp *qp,
908 			    u32 **in, int *inlen,
909 			    struct mlx5_ib_qp_base *base)
910 {
911 	int uar_index;
912 	void *qpc;
913 	int err;
914 
915 	if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN |
916 					IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
917 					IB_QP_CREATE_IPOIB_UD_LSO |
918 					MLX5_IB_QP_CREATE_SQPN_QP1 |
919 					MLX5_IB_QP_CREATE_WC_TEST))
920 		return -EINVAL;
921 
922 	spin_lock_init(&qp->bf.lock32);
923 
924 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
925 		qp->bf.bfreg = &dev->fp_bfreg;
926 	else if (init_attr->create_flags & MLX5_IB_QP_CREATE_WC_TEST)
927 		qp->bf.bfreg = &dev->wc_bfreg;
928 	else
929 		qp->bf.bfreg = &dev->bfreg;
930 
931 	/* We need to divide by two since each register is comprised of
932 	 * two buffers of identical size, namely odd and even
933 	 */
934 	qp->bf.buf_size = (1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size)) / 2;
935 	uar_index = qp->bf.bfreg->index;
936 
937 	err = calc_sq_size(dev, init_attr, qp);
938 	if (err < 0) {
939 		mlx5_ib_dbg(dev, "err %d\n", err);
940 		return err;
941 	}
942 
943 	qp->rq.offset = 0;
944 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
945 	base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
946 
947 	err = mlx5_buf_alloc(dev->mdev, base->ubuffer.buf_size,
948 	    2 * PAGE_SIZE, &qp->buf);
949 	if (err) {
950 		mlx5_ib_dbg(dev, "err %d\n", err);
951 		return err;
952 	}
953 
954 	qp->sq.qend = mlx5_get_send_wqe(qp, qp->sq.wqe_cnt);
955 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
956 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * qp->buf.npages;
957 	*in = mlx5_vzalloc(*inlen);
958 	if (!*in) {
959 		err = -ENOMEM;
960 		goto err_buf;
961 	}
962 
963 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
964 	MLX5_SET(qpc, qpc, uar_page, uar_index);
965 	MLX5_SET(qpc, qpc, ts_format, mlx5_get_qp_default_ts(dev->mdev));
966 	MLX5_SET(qpc, qpc, log_page_size, qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
967 
968 	/* Set "fast registration enabled" for all kernel QPs */
969 	MLX5_SET(qpc, qpc, fre, 1);
970 	MLX5_SET(qpc, qpc, rlky, 1);
971 
972 	if (init_attr->create_flags & MLX5_IB_QP_CREATE_SQPN_QP1) {
973 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
974 		qp->flags |= MLX5_IB_QP_SQPN_QP1;
975 	}
976 
977 	mlx5_fill_page_array(&qp->buf,
978 			     (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas));
979 
980 	err = mlx5_db_alloc(dev->mdev, &qp->db);
981 	if (err) {
982 		mlx5_ib_dbg(dev, "err %d\n", err);
983 		goto err_free;
984 	}
985 
986 	qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
987 	qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
988 	qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
989 	qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
990 	qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
991 
992 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
993 	    !qp->sq.w_list || !qp->sq.wqe_head) {
994 		err = -ENOMEM;
995 		goto err_wrid;
996 	}
997 	qp->create_type = MLX5_QP_KERNEL;
998 
999 	return 0;
1000 
1001 err_wrid:
1002 	kfree(qp->sq.wqe_head);
1003 	kfree(qp->sq.w_list);
1004 	kfree(qp->sq.wrid);
1005 	kfree(qp->sq.wr_data);
1006 	kfree(qp->rq.wrid);
1007 	mlx5_db_free(dev->mdev, &qp->db);
1008 
1009 err_free:
1010 	kvfree(*in);
1011 
1012 err_buf:
1013 	mlx5_buf_free(dev->mdev, &qp->buf);
1014 	return err;
1015 }
1016 
1017 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1018 {
1019 	kfree(qp->sq.wqe_head);
1020 	kfree(qp->sq.w_list);
1021 	kfree(qp->sq.wrid);
1022 	kfree(qp->sq.wr_data);
1023 	kfree(qp->rq.wrid);
1024 	mlx5_db_free(dev->mdev, &qp->db);
1025 	mlx5_buf_free(dev->mdev, &qp->buf);
1026 }
1027 
1028 static u32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
1029 {
1030 	if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
1031 	    (attr->qp_type == IB_QPT_XRC_INI))
1032 		return MLX5_SRQ_RQ;
1033 	else if (!qp->has_rq)
1034 		return MLX5_ZERO_LEN_RQ;
1035 	else
1036 		return MLX5_NON_ZERO_RQ;
1037 }
1038 
1039 static int is_connected(enum ib_qp_type qp_type)
1040 {
1041 	if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC)
1042 		return 1;
1043 
1044 	return 0;
1045 }
1046 
1047 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1048 				    struct mlx5_ib_sq *sq, u32 tdn)
1049 {
1050 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {0};
1051 	void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1052 
1053 	MLX5_SET(tisc, tisc, transport_domain, tdn);
1054 	return mlx5_core_create_tis(dev->mdev, in, sizeof(in), &sq->tisn);
1055 }
1056 
1057 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1058 				      struct mlx5_ib_sq *sq)
1059 {
1060 	mlx5_core_destroy_tis(dev->mdev, sq->tisn);
1061 }
1062 
1063 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1064 				   struct mlx5_ib_sq *sq, void *qpin,
1065 				   struct ib_pd *pd)
1066 {
1067 	struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1068 	__be64 *pas;
1069 	void *in;
1070 	void *sqc;
1071 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1072 	void *wq;
1073 	int inlen;
1074 	int err;
1075 	int page_shift = 0;
1076 	int npages;
1077 	int ncont = 0;
1078 	u32 offset = 0;
1079 	u8 ts_format;
1080 
1081 	ts_format = mlx5_get_sq_default_ts(dev->mdev);
1082 
1083 	err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr, ubuffer->buf_size,
1084 			       &sq->ubuffer.umem, &npages, &page_shift,
1085 			       &ncont, &offset);
1086 	if (err)
1087 		return err;
1088 
1089 	inlen = MLX5_ST_SZ_BYTES(create_sq_in) + sizeof(u64) * ncont;
1090 	in = mlx5_vzalloc(inlen);
1091 	if (!in) {
1092 		err = -ENOMEM;
1093 		goto err_umem;
1094 	}
1095 
1096 	sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1097 	MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1098 	MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1099 	MLX5_SET(sqc, sqc, ts_format, ts_format);
1100 	MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1101 	MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1102 	MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1103 	MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1104 
1105 	wq = MLX5_ADDR_OF(sqc, sqc, wq);
1106 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1107 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1108 	MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1109 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1110 	MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1111 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1112 	MLX5_SET(wq, wq, log_wq_pg_sz,  page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1113 	MLX5_SET(wq, wq, page_offset, offset);
1114 
1115 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1116 	mlx5_ib_populate_pas(dev, sq->ubuffer.umem, page_shift, pas, 0);
1117 
1118 	err = mlx5_core_create_sq_tracked(dev->mdev, in, inlen, &sq->base.mqp);
1119 
1120 	kvfree(in);
1121 
1122 	if (err)
1123 		goto err_umem;
1124 
1125 	return 0;
1126 
1127 err_umem:
1128 	ib_umem_release(sq->ubuffer.umem);
1129 	sq->ubuffer.umem = NULL;
1130 
1131 	return err;
1132 }
1133 
1134 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1135 				     struct mlx5_ib_sq *sq)
1136 {
1137 	mlx5_core_destroy_sq_tracked(dev->mdev, &sq->base.mqp);
1138 	ib_umem_release(sq->ubuffer.umem);
1139 }
1140 
1141 static int get_rq_pas_size(void *qpc)
1142 {
1143 	u32 log_page_size = MLX5_GET(qpc, qpc, log_page_size) + 12;
1144 	u32 log_rq_stride = MLX5_GET(qpc, qpc, log_rq_stride);
1145 	u32 log_rq_size   = MLX5_GET(qpc, qpc, log_rq_size);
1146 	u32 page_offset   = MLX5_GET(qpc, qpc, page_offset);
1147 	u32 po_quanta	  = 1 << (log_page_size - 6);
1148 	u32 rq_sz	  = 1 << (log_rq_size + 4 + log_rq_stride);
1149 	u32 page_size	  = 1 << log_page_size;
1150 	u32 rq_sz_po      = rq_sz + (page_offset * po_quanta);
1151 	u32 rq_num_pas	  = (rq_sz_po + page_size - 1) / page_size;
1152 
1153 	return rq_num_pas * sizeof(u64);
1154 }
1155 
1156 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1157 				   struct mlx5_ib_rq *rq, void *qpin)
1158 {
1159 	struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1160 	__be64 *pas;
1161 	__be64 *qp_pas;
1162 	void *in;
1163 	void *rqc;
1164 	void *wq;
1165 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1166 	int inlen;
1167 	int err;
1168 	u32 rq_pas_size = get_rq_pas_size(qpc);
1169 	u8 ts_format;
1170 
1171 	ts_format = mlx5_get_rq_default_ts(dev->mdev);
1172 
1173 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + rq_pas_size;
1174 	in = mlx5_vzalloc(inlen);
1175 	if (!in)
1176 		return -ENOMEM;
1177 
1178 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1179 	MLX5_SET(rqc, rqc, vlan_strip_disable, 1);
1180 	MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_RQ_TYPE_MEMORY_RQ_INLINE);
1181 	MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1182 	MLX5_SET(rqc, rqc, ts_format, ts_format);
1183 	MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1184 	MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1185 	MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1186 
1187 	if (mqp->flags & MLX5_IB_QP_CAP_SCATTER_FCS)
1188 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
1189 
1190 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
1191 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1192 	MLX5_SET(wq, wq, end_padding_mode,
1193 		 MLX5_GET(qpc, qpc, end_padding_mode));
1194 	MLX5_SET(wq, wq, page_offset, MLX5_GET(qpc, qpc, page_offset));
1195 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1196 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1197 	MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1198 	MLX5_SET(wq, wq, log_wq_pg_sz, MLX5_GET(qpc, qpc, log_page_size));
1199 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1200 
1201 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1202 	qp_pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, qpin, pas);
1203 	memcpy(pas, qp_pas, rq_pas_size);
1204 
1205 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rq->base.mqp);
1206 
1207 	kvfree(in);
1208 
1209 	return err;
1210 }
1211 
1212 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1213 				     struct mlx5_ib_rq *rq)
1214 {
1215 	mlx5_core_destroy_rq_tracked(dev->mdev, &rq->base.mqp);
1216 }
1217 
1218 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1219 				    struct mlx5_ib_rq *rq, u32 tdn)
1220 {
1221 	u32 *in;
1222 	void *tirc;
1223 	int inlen;
1224 	int err;
1225 
1226 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1227 	in = mlx5_vzalloc(inlen);
1228 	if (!in)
1229 		return -ENOMEM;
1230 
1231 	tirc = MLX5_ADDR_OF(create_tir_in, in, tir_context);
1232 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1233 	MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1234 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1235 
1236 	err = mlx5_core_create_tir(dev->mdev, in, inlen, &rq->tirn);
1237 
1238 	kvfree(in);
1239 
1240 	return err;
1241 }
1242 
1243 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1244 				      struct mlx5_ib_rq *rq)
1245 {
1246 	mlx5_core_destroy_tir(dev->mdev, rq->tirn);
1247 }
1248 
1249 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1250 				u32 *in,
1251 				struct ib_pd *pd)
1252 {
1253 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1254 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1255 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1256 	struct ib_uobject *uobj = pd->uobject;
1257 	struct ib_ucontext *ucontext = uobj->context;
1258 	struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1259 	int err;
1260 	u32 tdn = mucontext->tdn;
1261 
1262 	if (qp->sq.wqe_cnt) {
1263 		err = create_raw_packet_qp_tis(dev, sq, tdn);
1264 		if (err)
1265 			return err;
1266 
1267 		err = create_raw_packet_qp_sq(dev, sq, in, pd);
1268 		if (err)
1269 			goto err_destroy_tis;
1270 
1271 		sq->base.container_mibqp = qp;
1272 	}
1273 
1274 	if (qp->rq.wqe_cnt) {
1275 		rq->base.container_mibqp = qp;
1276 
1277 		err = create_raw_packet_qp_rq(dev, rq, in);
1278 		if (err)
1279 			goto err_destroy_sq;
1280 
1281 
1282 		err = create_raw_packet_qp_tir(dev, rq, tdn);
1283 		if (err)
1284 			goto err_destroy_rq;
1285 	}
1286 
1287 	qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1288 						     rq->base.mqp.qpn;
1289 
1290 	return 0;
1291 
1292 err_destroy_rq:
1293 	destroy_raw_packet_qp_rq(dev, rq);
1294 err_destroy_sq:
1295 	if (!qp->sq.wqe_cnt)
1296 		return err;
1297 	destroy_raw_packet_qp_sq(dev, sq);
1298 err_destroy_tis:
1299 	destroy_raw_packet_qp_tis(dev, sq);
1300 
1301 	return err;
1302 }
1303 
1304 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1305 				  struct mlx5_ib_qp *qp)
1306 {
1307 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1308 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1309 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1310 
1311 	if (qp->rq.wqe_cnt) {
1312 		destroy_raw_packet_qp_tir(dev, rq);
1313 		destroy_raw_packet_qp_rq(dev, rq);
1314 	}
1315 
1316 	if (qp->sq.wqe_cnt) {
1317 		destroy_raw_packet_qp_sq(dev, sq);
1318 		destroy_raw_packet_qp_tis(dev, sq);
1319 	}
1320 }
1321 
1322 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1323 				    struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1324 {
1325 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1326 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1327 
1328 	sq->sq = &qp->sq;
1329 	rq->rq = &qp->rq;
1330 	sq->doorbell = &qp->db;
1331 	rq->doorbell = &qp->db;
1332 }
1333 
1334 static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1335 {
1336 	mlx5_core_destroy_tir(dev->mdev, qp->rss_qp.tirn);
1337 }
1338 
1339 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1340 				 struct ib_pd *pd,
1341 				 struct ib_qp_init_attr *init_attr,
1342 				 struct ib_udata *udata)
1343 {
1344 	struct ib_uobject *uobj = pd->uobject;
1345 	struct ib_ucontext *ucontext = uobj->context;
1346 	struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1347 	struct mlx5_ib_create_qp_resp resp = {};
1348 	int inlen;
1349 	int err;
1350 	u32 *in;
1351 	void *tirc;
1352 	void *hfso;
1353 	u32 selected_fields = 0;
1354 	size_t min_resp_len;
1355 	u32 tdn = mucontext->tdn;
1356 	struct mlx5_ib_create_qp_rss ucmd = {};
1357 	size_t required_cmd_sz;
1358 
1359 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1360 		return -EOPNOTSUPP;
1361 
1362 	if (init_attr->create_flags || init_attr->send_cq)
1363 		return -EINVAL;
1364 
1365 	min_resp_len = offsetof(typeof(resp), bfreg_index) + sizeof(resp.bfreg_index);
1366 	if (udata->outlen < min_resp_len)
1367 		return -EINVAL;
1368 
1369 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) + sizeof(ucmd.reserved1);
1370 	if (udata->inlen < required_cmd_sz) {
1371 		mlx5_ib_dbg(dev, "invalid inlen\n");
1372 		return -EINVAL;
1373 	}
1374 
1375 	if (udata->inlen > sizeof(ucmd) &&
1376 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
1377 				 udata->inlen - sizeof(ucmd))) {
1378 		mlx5_ib_dbg(dev, "inlen is not supported\n");
1379 		return -EOPNOTSUPP;
1380 	}
1381 
1382 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
1383 		mlx5_ib_dbg(dev, "copy failed\n");
1384 		return -EFAULT;
1385 	}
1386 
1387 	if (ucmd.comp_mask) {
1388 		mlx5_ib_dbg(dev, "invalid comp mask\n");
1389 		return -EOPNOTSUPP;
1390 	}
1391 
1392 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)) || ucmd.reserved1) {
1393 		mlx5_ib_dbg(dev, "invalid reserved\n");
1394 		return -EOPNOTSUPP;
1395 	}
1396 
1397 	err = ib_copy_to_udata(udata, &resp, min_resp_len);
1398 	if (err) {
1399 		mlx5_ib_dbg(dev, "copy failed\n");
1400 		return -EINVAL;
1401 	}
1402 
1403 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1404 	in = mlx5_vzalloc(inlen);
1405 	if (!in)
1406 		return -ENOMEM;
1407 
1408 	tirc = MLX5_ADDR_OF(create_tir_in, in, tir_context);
1409 	MLX5_SET(tirc, tirc, disp_type,
1410 		 MLX5_TIRC_DISP_TYPE_INDIRECT);
1411 	MLX5_SET(tirc, tirc, indirect_table,
1412 		 init_attr->rwq_ind_tbl->ind_tbl_num);
1413 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1414 
1415 	hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1416 	switch (ucmd.rx_hash_function) {
1417 	case MLX5_RX_HASH_FUNC_TOEPLITZ:
1418 	{
1419 		void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1420 		size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1421 
1422 		if (len != ucmd.rx_key_len) {
1423 			err = -EINVAL;
1424 			goto err;
1425 		}
1426 
1427 		MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FUNC_TOEPLITZ);
1428 		MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
1429 		memcpy(rss_key, ucmd.rx_hash_key, len);
1430 		break;
1431 	}
1432 	default:
1433 		err = -EOPNOTSUPP;
1434 		goto err;
1435 	}
1436 
1437 	if (!ucmd.rx_hash_fields_mask) {
1438 		/* special case when this TIR serves as steering entry without hashing */
1439 		if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1440 			goto create_tir;
1441 		err = -EINVAL;
1442 		goto err;
1443 	}
1444 
1445 	if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1446 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1447 	     ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1448 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1449 		err = -EINVAL;
1450 		goto err;
1451 	}
1452 
1453 	/* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1454 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1455 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1456 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1457 			 MLX5_L3_PROT_TYPE_IPV4);
1458 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1459 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1460 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1461 			 MLX5_L3_PROT_TYPE_IPV6);
1462 
1463 	if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1464 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP)) &&
1465 	     ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1466 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))) {
1467 		err = -EINVAL;
1468 		goto err;
1469 	}
1470 
1471 	/* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1472 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1473 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1474 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1475 			 MLX5_L4_PROT_TYPE_TCP);
1476 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1477 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1478 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1479 			 MLX5_L4_PROT_TYPE_UDP);
1480 
1481 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1482 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1483 		selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1484 
1485 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1486 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1487 		selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1488 
1489 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1490 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1491 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1492 
1493 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1494 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1495 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1496 
1497 	MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1498 
1499 create_tir:
1500 	err = mlx5_core_create_tir(dev->mdev, in, inlen, &qp->rss_qp.tirn);
1501 
1502 	if (err)
1503 		goto err;
1504 
1505 	kvfree(in);
1506 	/* qpn is reserved for that QP */
1507 	qp->trans_qp.base.mqp.qpn = 0;
1508 	qp->flags |= MLX5_IB_QP_RSS;
1509 	return 0;
1510 
1511 err:
1512 	kvfree(in);
1513 	return err;
1514 }
1515 
1516 static int atomic_size_to_mode(int size_mask)
1517 {
1518 	/* driver does not support atomic_size > 256B
1519 	 * and does not know how to translate bigger sizes
1520 	 */
1521 	int supported_size_mask = size_mask & 0x1ff;
1522 	int log_max_size;
1523 
1524 	if (!supported_size_mask)
1525 		return -EOPNOTSUPP;
1526 
1527 	log_max_size = __fls(supported_size_mask);
1528 
1529 	if (log_max_size > 3)
1530 		return log_max_size;
1531 
1532 	return MLX5_ATOMIC_MODE_8B;
1533 }
1534 
1535 static int get_atomic_mode(struct mlx5_ib_dev *dev,
1536 			   enum ib_qp_type qp_type)
1537 {
1538 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
1539 	u8 atomic = MLX5_CAP_GEN(dev->mdev, atomic);
1540 	int atomic_mode = -EOPNOTSUPP;
1541 	int atomic_size_mask;
1542 
1543 	if (!atomic)
1544 		return -EOPNOTSUPP;
1545 
1546 	if (qp_type == MLX5_IB_QPT_DCT)
1547 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
1548 	else
1549 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
1550 
1551 	if ((atomic_operations & MLX5_ATOMIC_OPS_MASKED_CMP_SWAP) ||
1552 	    (atomic_operations & MLX5_ATOMIC_OPS_MASKED_FETCH_ADD))
1553 		atomic_mode = atomic_size_to_mode(atomic_size_mask);
1554 
1555 	if (atomic_mode <= 0 &&
1556 	    (atomic_operations & MLX5_ATOMIC_OPS_CMP_SWAP &&
1557 	     atomic_operations & MLX5_ATOMIC_OPS_FETCH_ADD))
1558 		atomic_mode = MLX5_ATOMIC_MODE_IB_COMP;
1559 
1560 	return atomic_mode;
1561 }
1562 
1563 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1564 			    struct ib_qp_init_attr *init_attr,
1565 			    struct ib_udata *udata, struct mlx5_ib_qp *qp)
1566 {
1567 	struct mlx5_ib_resources *devr = &dev->devr;
1568 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1569 	struct mlx5_core_dev *mdev = dev->mdev;
1570 	struct mlx5_ib_create_qp_resp resp;
1571 	struct mlx5_ib_cq *send_cq;
1572 	struct mlx5_ib_cq *recv_cq;
1573 	unsigned long flags;
1574 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
1575 	struct mlx5_ib_create_qp ucmd;
1576 	struct mlx5_ib_qp_base *base;
1577 	void *qpc;
1578 	u32 *in;
1579 	int err;
1580 
1581 	base = init_attr->qp_type == IB_QPT_RAW_PACKET ?
1582 	       &qp->raw_packet_qp.rq.base :
1583 	       &qp->trans_qp.base;
1584 
1585 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1586 		mlx5_ib_odp_create_qp(qp);
1587 
1588 	mutex_init(&qp->mutex);
1589 	spin_lock_init(&qp->sq.lock);
1590 	spin_lock_init(&qp->rq.lock);
1591 
1592 	if (init_attr->rwq_ind_tbl) {
1593 		if (!udata)
1594 			return -ENOSYS;
1595 
1596 		err = create_rss_raw_qp_tir(dev, qp, pd, init_attr, udata);
1597 		return err;
1598 	}
1599 
1600 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
1601 		if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
1602 			mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
1603 			return -EINVAL;
1604 		} else {
1605 			qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1606 		}
1607 	}
1608 
1609 	if (init_attr->create_flags &
1610 			(IB_QP_CREATE_CROSS_CHANNEL |
1611 			 IB_QP_CREATE_MANAGED_SEND |
1612 			 IB_QP_CREATE_MANAGED_RECV)) {
1613 		if (!MLX5_CAP_GEN(mdev, cd)) {
1614 			mlx5_ib_dbg(dev, "cross-channel isn't supported\n");
1615 			return -EINVAL;
1616 		}
1617 		if (init_attr->create_flags & IB_QP_CREATE_CROSS_CHANNEL)
1618 			qp->flags |= MLX5_IB_QP_CROSS_CHANNEL;
1619 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_SEND)
1620 			qp->flags |= MLX5_IB_QP_MANAGED_SEND;
1621 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_RECV)
1622 			qp->flags |= MLX5_IB_QP_MANAGED_RECV;
1623 	}
1624 
1625 	if (init_attr->qp_type == IB_QPT_UD &&
1626 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO))
1627 		if (!MLX5_CAP_GEN(mdev, ipoib_ipoib_offloads)) {
1628 			mlx5_ib_dbg(dev, "ipoib UD lso qp isn't supported\n");
1629 			return -EOPNOTSUPP;
1630 		}
1631 
1632 	if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1633 		if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
1634 			mlx5_ib_dbg(dev, "Scatter FCS is supported only for Raw Packet QPs");
1635 			return -EOPNOTSUPP;
1636 		}
1637 		if (!MLX5_CAP_GEN(dev->mdev, eth_net_offloads) ||
1638 		    !MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
1639 			mlx5_ib_dbg(dev, "Scatter FCS isn't supported\n");
1640 			return -EOPNOTSUPP;
1641 		}
1642 		qp->flags |= MLX5_IB_QP_CAP_SCATTER_FCS;
1643 	}
1644 
1645 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1646 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
1647 
1648 	if (pd && pd->uobject) {
1649 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
1650 			mlx5_ib_dbg(dev, "copy failed\n");
1651 			return -EFAULT;
1652 		}
1653 
1654 		err = get_qp_user_index(to_mucontext(pd->uobject->context),
1655 					&ucmd, udata->inlen, &uidx);
1656 		if (err)
1657 			return err;
1658 
1659 		qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
1660 		qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
1661 	} else {
1662 		qp->wq_sig = !!wq_signature;
1663 	}
1664 
1665 	qp->has_rq = qp_has_rq(init_attr);
1666 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
1667 			  qp, (pd && pd->uobject) ? &ucmd : NULL);
1668 	if (err) {
1669 		mlx5_ib_dbg(dev, "err %d\n", err);
1670 		return err;
1671 	}
1672 
1673 	if (pd) {
1674 		if (pd->uobject) {
1675 			__u32 max_wqes =
1676 				1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1677 			mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
1678 			if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
1679 			    ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
1680 				mlx5_ib_dbg(dev, "invalid rq params\n");
1681 				return -EINVAL;
1682 			}
1683 			if (ucmd.sq_wqe_count > max_wqes) {
1684 				mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
1685 					    ucmd.sq_wqe_count, max_wqes);
1686 				return -EINVAL;
1687 			}
1688 			if (init_attr->create_flags &
1689 			    MLX5_IB_QP_CREATE_SQPN_QP1) {
1690 				mlx5_ib_dbg(dev, "user-space is not allowed to create UD QPs spoofing as QP1\n");
1691 				return -EINVAL;
1692 			}
1693 			err = create_user_qp(dev, pd, qp, udata, init_attr, &in,
1694 					     &resp, &inlen, base);
1695 			if (err)
1696 				mlx5_ib_dbg(dev, "err %d\n", err);
1697 		} else {
1698 			err = create_kernel_qp(dev, init_attr, qp, &in, &inlen,
1699 					       base);
1700 			if (err)
1701 				mlx5_ib_dbg(dev, "err %d\n", err);
1702 		}
1703 
1704 		if (err)
1705 			return err;
1706 	} else {
1707 		in = mlx5_vzalloc(inlen);
1708 		if (!in)
1709 			return -ENOMEM;
1710 
1711 		qp->create_type = MLX5_QP_EMPTY;
1712 	}
1713 
1714 	if (is_sqp(init_attr->qp_type))
1715 		qp->port = init_attr->port_num;
1716 
1717 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1718 
1719 	MLX5_SET(qpc, qpc, st, to_mlx5_st(init_attr->qp_type));
1720 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
1721 
1722 	if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
1723 		MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
1724 	else
1725 		MLX5_SET(qpc, qpc, latency_sensitive, 1);
1726 
1727 
1728 	if (qp->wq_sig)
1729 		MLX5_SET(qpc, qpc, wq_signature, 1);
1730 
1731 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
1732 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
1733 
1734 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
1735 		MLX5_SET(qpc, qpc, cd_master, 1);
1736 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
1737 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
1738 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
1739 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
1740 
1741 	if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
1742 		int rcqe_sz;
1743 		int scqe_sz;
1744 
1745 		rcqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->recv_cq);
1746 		scqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->send_cq);
1747 
1748 		if (rcqe_sz == 128)
1749 			MLX5_SET(qpc, qpc, cs_res, MLX5_RES_SCAT_DATA64_CQE);
1750 		else
1751 			MLX5_SET(qpc, qpc, cs_res, MLX5_RES_SCAT_DATA32_CQE);
1752 
1753 		if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) {
1754 			if (scqe_sz == 128)
1755 				MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA64_CQE);
1756 			else
1757 				MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA32_CQE);
1758 		}
1759 	}
1760 
1761 	if (qp->rq.wqe_cnt) {
1762 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
1763 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
1764 	}
1765 
1766 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1767 		MLX5_SET(qpc, qpc, ts_format, mlx5_get_qp_default_ts(dev->mdev));
1768 
1769 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
1770 
1771 	if (qp->sq.wqe_cnt)
1772 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
1773 	else
1774 		MLX5_SET(qpc, qpc, no_sq, 1);
1775 
1776 	/* Set default resources */
1777 	switch (init_attr->qp_type) {
1778 	case IB_QPT_XRC_TGT:
1779 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1780 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(devr->c0)->mcq.cqn);
1781 		MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(devr->s0)->msrq.srqn);
1782 		MLX5_SET(qpc, qpc, xrcd, to_mxrcd(init_attr->xrcd)->xrcdn);
1783 		break;
1784 	case IB_QPT_XRC_INI:
1785 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1786 		MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
1787 		MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(devr->s0)->msrq.srqn);
1788 		break;
1789 	default:
1790 		if (init_attr->srq) {
1791 			MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x0)->xrcdn);
1792 			MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(init_attr->srq)->msrq.srqn);
1793 		} else {
1794 			MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
1795 			MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(devr->s1)->msrq.srqn);
1796 		}
1797 	}
1798 
1799 	if (init_attr->send_cq)
1800 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(init_attr->send_cq)->mcq.cqn);
1801 
1802 	if (init_attr->recv_cq)
1803 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(init_attr->recv_cq)->mcq.cqn);
1804 
1805 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
1806 
1807 	/* 0xffffff means we ask to work with cqe version 0 */
1808 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
1809 		MLX5_SET(qpc, qpc, user_index, uidx);
1810 
1811 	/* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
1812 	if (init_attr->qp_type == IB_QPT_UD &&
1813 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)) {
1814 		MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
1815 		qp->flags |= MLX5_IB_QP_LSO;
1816 	}
1817 
1818 	if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
1819 		qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd.sq_buf_addr;
1820 		raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
1821 		err = create_raw_packet_qp(dev, qp, in, pd);
1822 	} else {
1823 		err = mlx5_core_create_qp(dev->mdev, &base->mqp, in, inlen);
1824 	}
1825 
1826 	if (err) {
1827 		mlx5_ib_dbg(dev, "create qp failed\n");
1828 		goto err_create;
1829 	}
1830 
1831 	kvfree(in);
1832 
1833 	base->container_mibqp = qp;
1834 	base->mqp.event = mlx5_ib_qp_event;
1835 
1836 	get_cqs(init_attr->qp_type, init_attr->send_cq, init_attr->recv_cq,
1837 		&send_cq, &recv_cq);
1838 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1839 	mlx5_ib_lock_cqs(send_cq, recv_cq);
1840 	/* Maintain device to QPs access, needed for further handling via reset
1841 	 * flow
1842 	 */
1843 	list_add_tail(&qp->qps_list, &dev->qp_list);
1844 	/* Maintain CQ to QPs access, needed for further handling via reset flow
1845 	 */
1846 	if (send_cq)
1847 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
1848 	if (recv_cq)
1849 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
1850 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
1851 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1852 
1853 	return 0;
1854 
1855 err_create:
1856 	if (qp->create_type == MLX5_QP_USER)
1857 		destroy_qp_user(dev, pd, qp, base);
1858 	else if (qp->create_type == MLX5_QP_KERNEL)
1859 		destroy_qp_kernel(dev, qp);
1860 
1861 	kvfree(in);
1862 	return err;
1863 }
1864 
1865 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1866 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1867 {
1868 	if (send_cq) {
1869 		if (recv_cq) {
1870 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1871 				spin_lock(&send_cq->lock);
1872 				spin_lock_nested(&recv_cq->lock,
1873 						 SINGLE_DEPTH_NESTING);
1874 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1875 				spin_lock(&send_cq->lock);
1876 				__acquire(&recv_cq->lock);
1877 			} else {
1878 				spin_lock(&recv_cq->lock);
1879 				spin_lock_nested(&send_cq->lock,
1880 						 SINGLE_DEPTH_NESTING);
1881 			}
1882 		} else {
1883 			spin_lock(&send_cq->lock);
1884 			__acquire(&recv_cq->lock);
1885 		}
1886 	} else if (recv_cq) {
1887 		spin_lock(&recv_cq->lock);
1888 		__acquire(&send_cq->lock);
1889 	} else {
1890 		__acquire(&send_cq->lock);
1891 		__acquire(&recv_cq->lock);
1892 	}
1893 }
1894 
1895 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1896 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1897 {
1898 	if (send_cq) {
1899 		if (recv_cq) {
1900 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1901 				spin_unlock(&recv_cq->lock);
1902 				spin_unlock(&send_cq->lock);
1903 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1904 				__release(&recv_cq->lock);
1905 				spin_unlock(&send_cq->lock);
1906 			} else {
1907 				spin_unlock(&send_cq->lock);
1908 				spin_unlock(&recv_cq->lock);
1909 			}
1910 		} else {
1911 			__release(&recv_cq->lock);
1912 			spin_unlock(&send_cq->lock);
1913 		}
1914 	} else if (recv_cq) {
1915 		__release(&send_cq->lock);
1916 		spin_unlock(&recv_cq->lock);
1917 	} else {
1918 		__release(&recv_cq->lock);
1919 		__release(&send_cq->lock);
1920 	}
1921 }
1922 
1923 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
1924 {
1925 	return to_mpd(qp->ibqp.pd);
1926 }
1927 
1928 static void get_cqs(enum ib_qp_type qp_type,
1929 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
1930 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
1931 {
1932 	switch (qp_type) {
1933 	case IB_QPT_XRC_TGT:
1934 		*send_cq = NULL;
1935 		*recv_cq = NULL;
1936 		break;
1937 	case MLX5_IB_QPT_REG_UMR:
1938 	case IB_QPT_XRC_INI:
1939 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
1940 		*recv_cq = NULL;
1941 		break;
1942 
1943 	case IB_QPT_SMI:
1944 	case MLX5_IB_QPT_HW_GSI:
1945 	case IB_QPT_RC:
1946 	case IB_QPT_UC:
1947 	case IB_QPT_UD:
1948 	case IB_QPT_RAW_IPV6:
1949 	case IB_QPT_RAW_ETHERTYPE:
1950 	case IB_QPT_RAW_PACKET:
1951 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
1952 		*recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
1953 		break;
1954 
1955 	case IB_QPT_MAX:
1956 	default:
1957 		*send_cq = NULL;
1958 		*recv_cq = NULL;
1959 		break;
1960 	}
1961 }
1962 
1963 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1964 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
1965 				u8 lag_tx_affinity);
1966 
1967 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1968 {
1969 	struct mlx5_ib_cq *send_cq, *recv_cq;
1970 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
1971 	unsigned long flags;
1972 	int err;
1973 
1974 	if (qp->ibqp.rwq_ind_tbl) {
1975 		destroy_rss_raw_qp_tir(dev, qp);
1976 		return;
1977 	}
1978 
1979 	base = qp->ibqp.qp_type == IB_QPT_RAW_PACKET ?
1980 	       &qp->raw_packet_qp.rq.base :
1981 	       &qp->trans_qp.base;
1982 
1983 	if (qp->state != IB_QPS_RESET) {
1984 		if (qp->ibqp.qp_type != IB_QPT_RAW_PACKET) {
1985 			mlx5_ib_qp_disable_pagefaults(qp);
1986 			err = mlx5_core_qp_modify(dev->mdev,
1987 						  MLX5_CMD_OP_2RST_QP, 0,
1988 						  NULL, &base->mqp);
1989 		} else {
1990 			struct mlx5_modify_raw_qp_param raw_qp_param = {
1991 				.operation = MLX5_CMD_OP_2RST_QP
1992 			};
1993 
1994 			err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
1995 		}
1996 		if (err)
1997 			mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
1998 				     base->mqp.qpn);
1999 	}
2000 
2001 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
2002 		&send_cq, &recv_cq);
2003 
2004 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2005 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2006 	/* del from lists under both locks above to protect reset flow paths */
2007 	list_del(&qp->qps_list);
2008 	if (send_cq)
2009 		list_del(&qp->cq_send_list);
2010 
2011 	if (recv_cq)
2012 		list_del(&qp->cq_recv_list);
2013 
2014 	if (qp->create_type == MLX5_QP_KERNEL) {
2015 		__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2016 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
2017 		if (send_cq != recv_cq)
2018 			__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
2019 					   NULL);
2020 	}
2021 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2022 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2023 
2024 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
2025 		destroy_raw_packet_qp(dev, qp);
2026 	} else {
2027 		err = mlx5_core_destroy_qp(dev->mdev, &base->mqp);
2028 		if (err)
2029 			mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
2030 				     base->mqp.qpn);
2031 	}
2032 
2033 	if (qp->create_type == MLX5_QP_KERNEL)
2034 		destroy_qp_kernel(dev, qp);
2035 	else if (qp->create_type == MLX5_QP_USER)
2036 		destroy_qp_user(dev, &get_pd(qp)->ibpd, qp, base);
2037 }
2038 
2039 static const char *ib_qp_type_str(enum ib_qp_type type)
2040 {
2041 	switch (type) {
2042 	case IB_QPT_SMI:
2043 		return "IB_QPT_SMI";
2044 	case IB_QPT_GSI:
2045 		return "IB_QPT_GSI";
2046 	case IB_QPT_RC:
2047 		return "IB_QPT_RC";
2048 	case IB_QPT_UC:
2049 		return "IB_QPT_UC";
2050 	case IB_QPT_UD:
2051 		return "IB_QPT_UD";
2052 	case IB_QPT_RAW_IPV6:
2053 		return "IB_QPT_RAW_IPV6";
2054 	case IB_QPT_RAW_ETHERTYPE:
2055 		return "IB_QPT_RAW_ETHERTYPE";
2056 	case IB_QPT_XRC_INI:
2057 		return "IB_QPT_XRC_INI";
2058 	case IB_QPT_XRC_TGT:
2059 		return "IB_QPT_XRC_TGT";
2060 	case IB_QPT_RAW_PACKET:
2061 		return "IB_QPT_RAW_PACKET";
2062 	case MLX5_IB_QPT_REG_UMR:
2063 		return "MLX5_IB_QPT_REG_UMR";
2064 	case IB_QPT_MAX:
2065 	default:
2066 		return "Invalid QP type";
2067 	}
2068 }
2069 
2070 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
2071 				struct ib_qp_init_attr *init_attr,
2072 				struct ib_udata *udata)
2073 {
2074 	struct mlx5_ib_dev *dev;
2075 	struct mlx5_ib_qp *qp;
2076 	u16 xrcdn = 0;
2077 	int err;
2078 
2079 	if (pd) {
2080 		dev = to_mdev(pd->device);
2081 
2082 		if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
2083 			if (!pd->uobject) {
2084 				mlx5_ib_dbg(dev, "Raw Packet QP is not supported for kernel consumers\n");
2085 				return ERR_PTR(-EINVAL);
2086 			} else if (!to_mucontext(pd->uobject->context)->cqe_version) {
2087 				mlx5_ib_dbg(dev, "Raw Packet QP is only supported for CQE version > 0\n");
2088 				return ERR_PTR(-EINVAL);
2089 			}
2090 		}
2091 	} else {
2092 		/* being cautious here */
2093 		if (init_attr->qp_type != IB_QPT_XRC_TGT &&
2094 		    init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
2095 			pr_warn("%s: no PD for transport %s\n", __func__,
2096 				ib_qp_type_str(init_attr->qp_type));
2097 			return ERR_PTR(-EINVAL);
2098 		}
2099 		dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
2100 	}
2101 
2102 	switch (init_attr->qp_type) {
2103 	case IB_QPT_XRC_TGT:
2104 	case IB_QPT_XRC_INI:
2105 		if (!MLX5_CAP_GEN(dev->mdev, xrc)) {
2106 			mlx5_ib_dbg(dev, "XRC not supported\n");
2107 			return ERR_PTR(-ENOSYS);
2108 		}
2109 		init_attr->recv_cq = NULL;
2110 		if (init_attr->qp_type == IB_QPT_XRC_TGT) {
2111 			xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
2112 			init_attr->send_cq = NULL;
2113 		}
2114 
2115 		/* fall through */
2116 	case IB_QPT_RAW_PACKET:
2117 	case IB_QPT_RC:
2118 	case IB_QPT_UC:
2119 	case IB_QPT_UD:
2120 	case IB_QPT_SMI:
2121 	case MLX5_IB_QPT_HW_GSI:
2122 	case MLX5_IB_QPT_REG_UMR:
2123 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2124 		if (!qp)
2125 			return ERR_PTR(-ENOMEM);
2126 
2127 		err = create_qp_common(dev, pd, init_attr, udata, qp);
2128 		if (err) {
2129 			mlx5_ib_dbg(dev, "create_qp_common failed\n");
2130 			kfree(qp);
2131 			return ERR_PTR(err);
2132 		}
2133 
2134 		if (is_qp0(init_attr->qp_type))
2135 			qp->ibqp.qp_num = 0;
2136 		else if (is_qp1(init_attr->qp_type))
2137 			qp->ibqp.qp_num = 1;
2138 		else
2139 			qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
2140 
2141 		mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
2142 			    qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
2143 			    init_attr->recv_cq ? to_mcq(init_attr->recv_cq)->mcq.cqn : -1,
2144 			    init_attr->send_cq ? to_mcq(init_attr->send_cq)->mcq.cqn : -1);
2145 
2146 		qp->trans_qp.xrcdn = xrcdn;
2147 
2148 		break;
2149 
2150 	case IB_QPT_GSI:
2151 		return mlx5_ib_gsi_create_qp(pd, init_attr);
2152 
2153 	case IB_QPT_RAW_IPV6:
2154 	case IB_QPT_RAW_ETHERTYPE:
2155 	case IB_QPT_MAX:
2156 	default:
2157 		mlx5_ib_dbg(dev, "unsupported qp type %d\n",
2158 			    init_attr->qp_type);
2159 		/* Don't support raw QPs */
2160 		return ERR_PTR(-EINVAL);
2161 	}
2162 
2163 	return &qp->ibqp;
2164 }
2165 
2166 int mlx5_ib_destroy_qp(struct ib_qp *qp)
2167 {
2168 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2169 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2170 
2171 	if (unlikely(qp->qp_type == IB_QPT_GSI))
2172 		return mlx5_ib_gsi_destroy_qp(qp);
2173 
2174 	destroy_qp_common(dev, mqp);
2175 
2176 	kfree(mqp);
2177 
2178 	return 0;
2179 }
2180 
2181 static int to_mlx5_access_flags(struct mlx5_ib_qp *qp,
2182 				const struct ib_qp_attr *attr,
2183 				int attr_mask, __be32 *hw_access_flags_be)
2184 {
2185 	u8 dest_rd_atomic;
2186 	u32 access_flags, hw_access_flags = 0;
2187 
2188 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
2189 
2190 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2191 		dest_rd_atomic = attr->max_dest_rd_atomic;
2192 	else
2193 		dest_rd_atomic = qp->trans_qp.resp_depth;
2194 
2195 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2196 		access_flags = attr->qp_access_flags;
2197 	else
2198 		access_flags = qp->trans_qp.atomic_rd_en;
2199 
2200 	if (!dest_rd_atomic)
2201 		access_flags &= IB_ACCESS_REMOTE_WRITE;
2202 
2203 	if (access_flags & IB_ACCESS_REMOTE_READ)
2204 		hw_access_flags |= MLX5_QP_BIT_RRE;
2205 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC) {
2206 		int atomic_mode;
2207 
2208 		atomic_mode = get_atomic_mode(dev, qp->ibqp.qp_type);
2209 		if (atomic_mode < 0)
2210 			return -EOPNOTSUPP;
2211 
2212 		hw_access_flags |= MLX5_QP_BIT_RAE;
2213 		hw_access_flags |= atomic_mode << MLX5_ATOMIC_MODE_OFF;
2214 	}
2215 
2216 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
2217 		hw_access_flags |= MLX5_QP_BIT_RWE;
2218 
2219 	*hw_access_flags_be = cpu_to_be32(hw_access_flags);
2220 
2221 	return 0;
2222 }
2223 
2224 enum {
2225 	MLX5_PATH_FLAG_FL	= 1 << 0,
2226 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
2227 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
2228 };
2229 
2230 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
2231 {
2232 	if (rate == IB_RATE_PORT_CURRENT) {
2233 		return 0;
2234 	} else if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_600_GBPS) {
2235 		return -EINVAL;
2236 	} else {
2237 		while (rate != IB_RATE_2_5_GBPS &&
2238 		       !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
2239 			 MLX5_CAP_GEN(dev->mdev, stat_rate_support)))
2240 			--rate;
2241 	}
2242 
2243 	return rate + MLX5_STAT_RATE_OFFSET;
2244 }
2245 
2246 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
2247 				      struct mlx5_ib_sq *sq, u8 sl)
2248 {
2249 	void *in;
2250 	void *tisc;
2251 	int inlen;
2252 	int err;
2253 
2254 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2255 	in = mlx5_vzalloc(inlen);
2256 	if (!in)
2257 		return -ENOMEM;
2258 
2259 	MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
2260 
2261 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2262 	MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
2263 
2264 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2265 
2266 	kvfree(in);
2267 
2268 	return err;
2269 }
2270 
2271 static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,
2272 					 struct mlx5_ib_sq *sq, u8 tx_affinity)
2273 {
2274 	void *in;
2275 	void *tisc;
2276 	int inlen;
2277 	int err;
2278 
2279 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2280 	in = mlx5_vzalloc(inlen);
2281 	if (!in)
2282 		return -ENOMEM;
2283 
2284 	MLX5_SET(modify_tis_in, in, bitmask.lag_tx_port_affinity, 1);
2285 
2286 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2287 	MLX5_SET(tisc, tisc, lag_tx_port_affinity, tx_affinity);
2288 
2289 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2290 
2291 	kvfree(in);
2292 
2293 	return err;
2294 }
2295 
2296 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2297 			 const struct ib_ah_attr *ah,
2298 			 struct mlx5_qp_path *path, u8 port, int attr_mask,
2299 			 u32 path_flags, const struct ib_qp_attr *attr,
2300 			 bool alt)
2301 {
2302 	enum rdma_link_layer ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2303 	int err;
2304 	enum ib_gid_type gid_type;
2305 
2306 	if (attr_mask & IB_QP_PKEY_INDEX)
2307 		path->pkey_index = cpu_to_be16(alt ? attr->alt_pkey_index :
2308 						     attr->pkey_index);
2309 
2310 	if (ah->ah_flags & IB_AH_GRH) {
2311 		if (ah->grh.sgid_index >=
2312 		    dev->mdev->port_caps[port - 1].gid_table_len) {
2313 			pr_err("sgid_index (%u) too large. max is %d\n",
2314 			       ah->grh.sgid_index,
2315 			       dev->mdev->port_caps[port - 1].gid_table_len);
2316 			return -EINVAL;
2317 		}
2318 	}
2319 
2320 	if (ll == IB_LINK_LAYER_ETHERNET) {
2321 		if (!(ah->ah_flags & IB_AH_GRH))
2322 			return -EINVAL;
2323 		err = mlx5_get_roce_gid_type(dev, port, ah->grh.sgid_index,
2324 					     &gid_type);
2325 		if (err)
2326 			return err;
2327 		memcpy(path->rmac, ah->dmac, sizeof(ah->dmac));
2328 		path->udp_sport = mlx5_get_roce_udp_sport(dev, port,
2329 							  ah->grh.sgid_index);
2330 		path->dci_cfi_prio_sl = (ah->sl & 0x7) << 4;
2331 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
2332 			path->ecn_dscp = (ah->grh.traffic_class >> 2) & 0x3f;
2333 	} else {
2334 		path->fl_free_ar = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
2335 		path->fl_free_ar |=
2336 			(path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x40 : 0;
2337 		path->rlid = cpu_to_be16(ah->dlid);
2338 		path->grh_mlid = ah->src_path_bits & 0x7f;
2339 		if (ah->ah_flags & IB_AH_GRH)
2340 			path->grh_mlid	|= 1 << 7;
2341 		path->dci_cfi_prio_sl = ah->sl & 0xf;
2342 	}
2343 
2344 	if (ah->ah_flags & IB_AH_GRH) {
2345 		path->mgid_index = ah->grh.sgid_index;
2346 		path->hop_limit  = ah->grh.hop_limit;
2347 		path->tclass_flowlabel =
2348 			cpu_to_be32((ah->grh.traffic_class << 20) |
2349 				    (ah->grh.flow_label));
2350 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
2351 	}
2352 
2353 	err = ib_rate_to_mlx5(dev, ah->static_rate);
2354 	if (err < 0)
2355 		return err;
2356 	path->static_rate = err;
2357 	path->port = port;
2358 
2359 	if (attr_mask & IB_QP_TIMEOUT)
2360 		path->ackto_lt = (alt ? attr->alt_timeout : attr->timeout) << 3;
2361 
2362 	if ((qp->ibqp.qp_type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
2363 		return modify_raw_packet_eth_prio(dev->mdev,
2364 						  &qp->raw_packet_qp.sq,
2365 						  ah->sl & 0xf);
2366 
2367 	return 0;
2368 }
2369 
2370 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
2371 	[MLX5_QP_STATE_INIT] = {
2372 		[MLX5_QP_STATE_INIT] = {
2373 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
2374 					  MLX5_QP_OPTPAR_RAE		|
2375 					  MLX5_QP_OPTPAR_RWE		|
2376 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
2377 					  MLX5_QP_OPTPAR_PRI_PORT,
2378 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
2379 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
2380 					  MLX5_QP_OPTPAR_PRI_PORT,
2381 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
2382 					  MLX5_QP_OPTPAR_Q_KEY		|
2383 					  MLX5_QP_OPTPAR_PRI_PORT,
2384 		},
2385 		[MLX5_QP_STATE_RTR] = {
2386 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
2387 					  MLX5_QP_OPTPAR_RRE            |
2388 					  MLX5_QP_OPTPAR_RAE            |
2389 					  MLX5_QP_OPTPAR_RWE            |
2390 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2391 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
2392 					  MLX5_QP_OPTPAR_RWE            |
2393 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2394 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
2395 					  MLX5_QP_OPTPAR_Q_KEY,
2396 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
2397 					   MLX5_QP_OPTPAR_Q_KEY,
2398 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2399 					  MLX5_QP_OPTPAR_RRE            |
2400 					  MLX5_QP_OPTPAR_RAE            |
2401 					  MLX5_QP_OPTPAR_RWE            |
2402 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2403 		},
2404 	},
2405 	[MLX5_QP_STATE_RTR] = {
2406 		[MLX5_QP_STATE_RTS] = {
2407 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
2408 					  MLX5_QP_OPTPAR_RRE		|
2409 					  MLX5_QP_OPTPAR_RAE		|
2410 					  MLX5_QP_OPTPAR_RWE		|
2411 					  MLX5_QP_OPTPAR_PM_STATE	|
2412 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
2413 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
2414 					  MLX5_QP_OPTPAR_RWE		|
2415 					  MLX5_QP_OPTPAR_PM_STATE,
2416 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
2417 		},
2418 	},
2419 	[MLX5_QP_STATE_RTS] = {
2420 		[MLX5_QP_STATE_RTS] = {
2421 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
2422 					  MLX5_QP_OPTPAR_RAE		|
2423 					  MLX5_QP_OPTPAR_RWE		|
2424 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
2425 					  MLX5_QP_OPTPAR_PM_STATE	|
2426 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
2427 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
2428 					  MLX5_QP_OPTPAR_PM_STATE	|
2429 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
2430 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
2431 					  MLX5_QP_OPTPAR_SRQN		|
2432 					  MLX5_QP_OPTPAR_CQN_RCV,
2433 		},
2434 	},
2435 	[MLX5_QP_STATE_SQER] = {
2436 		[MLX5_QP_STATE_RTS] = {
2437 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
2438 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
2439 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
2440 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
2441 					   MLX5_QP_OPTPAR_RWE		|
2442 					   MLX5_QP_OPTPAR_RAE		|
2443 					   MLX5_QP_OPTPAR_RRE,
2444 		},
2445 	},
2446 };
2447 
2448 static int ib_nr_to_mlx5_nr(int ib_mask)
2449 {
2450 	switch (ib_mask) {
2451 	case IB_QP_STATE:
2452 		return 0;
2453 	case IB_QP_CUR_STATE:
2454 		return 0;
2455 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
2456 		return 0;
2457 	case IB_QP_ACCESS_FLAGS:
2458 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
2459 			MLX5_QP_OPTPAR_RAE;
2460 	case IB_QP_PKEY_INDEX:
2461 		return MLX5_QP_OPTPAR_PKEY_INDEX;
2462 	case IB_QP_PORT:
2463 		return MLX5_QP_OPTPAR_PRI_PORT;
2464 	case IB_QP_QKEY:
2465 		return MLX5_QP_OPTPAR_Q_KEY;
2466 	case IB_QP_AV:
2467 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
2468 			MLX5_QP_OPTPAR_PRI_PORT;
2469 	case IB_QP_PATH_MTU:
2470 		return 0;
2471 	case IB_QP_TIMEOUT:
2472 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
2473 	case IB_QP_RETRY_CNT:
2474 		return MLX5_QP_OPTPAR_RETRY_COUNT;
2475 	case IB_QP_RNR_RETRY:
2476 		return MLX5_QP_OPTPAR_RNR_RETRY;
2477 	case IB_QP_RQ_PSN:
2478 		return 0;
2479 	case IB_QP_MAX_QP_RD_ATOMIC:
2480 		return MLX5_QP_OPTPAR_SRA_MAX;
2481 	case IB_QP_ALT_PATH:
2482 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
2483 	case IB_QP_MIN_RNR_TIMER:
2484 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
2485 	case IB_QP_SQ_PSN:
2486 		return 0;
2487 	case IB_QP_MAX_DEST_RD_ATOMIC:
2488 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
2489 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
2490 	case IB_QP_PATH_MIG_STATE:
2491 		return MLX5_QP_OPTPAR_PM_STATE;
2492 	case IB_QP_CAP:
2493 		return 0;
2494 	case IB_QP_DEST_QPN:
2495 		return 0;
2496 	}
2497 	return 0;
2498 }
2499 
2500 static int ib_mask_to_mlx5_opt(int ib_mask)
2501 {
2502 	int result = 0;
2503 	int i;
2504 
2505 	for (i = 0; i < 8 * sizeof(int); i++) {
2506 		if ((1 << i) & ib_mask)
2507 			result |= ib_nr_to_mlx5_nr(1 << i);
2508 	}
2509 
2510 	return result;
2511 }
2512 
2513 static int modify_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
2514 				   struct mlx5_ib_rq *rq, int new_state,
2515 				   const struct mlx5_modify_raw_qp_param *raw_qp_param)
2516 {
2517 	void *in;
2518 	void *rqc;
2519 	int inlen;
2520 	int err;
2521 
2522 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
2523 	in = mlx5_vzalloc(inlen);
2524 	if (!in)
2525 		return -ENOMEM;
2526 
2527 	MLX5_SET(modify_rq_in, in, rqn, rq->base.mqp.qpn);
2528 	MLX5_SET(modify_rq_in, in, rq_state, rq->state);
2529 
2530 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
2531 	MLX5_SET(rqc, rqc, state, new_state);
2532 
2533 	if (raw_qp_param->set_mask & MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID) {
2534 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counters_set_id)) {
2535 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
2536 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_MODIFY_RQ_COUNTER_SET_ID);
2537 			MLX5_SET(rqc, rqc, counter_set_id, raw_qp_param->rq_q_ctr_id);
2538 		} else
2539 			pr_info_once("%s: RAW PACKET QP counters are not supported on current FW\n",
2540 				     dev->ib_dev.name);
2541 	}
2542 
2543 	err = mlx5_core_modify_rq(dev->mdev, in, inlen);
2544 	if (err)
2545 		goto out;
2546 
2547 	rq->state = new_state;
2548 
2549 out:
2550 	kvfree(in);
2551 	return err;
2552 }
2553 
2554 static int modify_raw_packet_qp_sq(struct mlx5_core_dev *dev,
2555 				   struct mlx5_ib_sq *sq, int new_state)
2556 {
2557 	void *in;
2558 	void *sqc;
2559 	int inlen;
2560 	int err;
2561 
2562 	inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
2563 	in = mlx5_vzalloc(inlen);
2564 	if (!in)
2565 		return -ENOMEM;
2566 
2567 	MLX5_SET(modify_sq_in, in, sqn, sq->base.mqp.qpn);
2568 	MLX5_SET(modify_sq_in, in, sq_state, sq->state);
2569 
2570 	sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
2571 	MLX5_SET(sqc, sqc, state, new_state);
2572 
2573 	err = mlx5_core_modify_sq(dev, in, inlen);
2574 	if (err)
2575 		goto out;
2576 
2577 	sq->state = new_state;
2578 
2579 out:
2580 	kvfree(in);
2581 	return err;
2582 }
2583 
2584 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2585 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
2586 				u8 tx_affinity)
2587 {
2588 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
2589 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
2590 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
2591 	int rq_state;
2592 	int sq_state;
2593 	int err;
2594 
2595 	switch (raw_qp_param->operation) {
2596 	case MLX5_CMD_OP_RST2INIT_QP:
2597 		rq_state = MLX5_RQC_STATE_RDY;
2598 		sq_state = MLX5_SQC_STATE_RDY;
2599 		break;
2600 	case MLX5_CMD_OP_2ERR_QP:
2601 		rq_state = MLX5_RQC_STATE_ERR;
2602 		sq_state = MLX5_SQC_STATE_ERR;
2603 		break;
2604 	case MLX5_CMD_OP_2RST_QP:
2605 		rq_state = MLX5_RQC_STATE_RST;
2606 		sq_state = MLX5_SQC_STATE_RST;
2607 		break;
2608 	case MLX5_CMD_OP_INIT2INIT_QP:
2609 	case MLX5_CMD_OP_INIT2RTR_QP:
2610 	case MLX5_CMD_OP_RTR2RTS_QP:
2611 	case MLX5_CMD_OP_RTS2RTS_QP:
2612 		if (raw_qp_param->set_mask)
2613 			return -EINVAL;
2614 		else
2615 			return 0;
2616 	default:
2617 		WARN_ON(1);
2618 		return -EINVAL;
2619 	}
2620 
2621 	if (qp->rq.wqe_cnt) {
2622 		err = modify_raw_packet_qp_rq(dev, rq, rq_state, raw_qp_param);
2623 		if (err)
2624 			return err;
2625 	}
2626 
2627 	if (qp->sq.wqe_cnt) {
2628 		if (tx_affinity) {
2629 			err = modify_raw_packet_tx_affinity(dev->mdev, sq,
2630 							    tx_affinity);
2631 			if (err)
2632 				return err;
2633 		}
2634 
2635 		return modify_raw_packet_qp_sq(dev->mdev, sq, sq_state);
2636 	}
2637 
2638 	return 0;
2639 }
2640 
2641 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
2642 			       const struct ib_qp_attr *attr, int attr_mask,
2643 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
2644 {
2645 	static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
2646 		[MLX5_QP_STATE_RST] = {
2647 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2648 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2649 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_RST2INIT_QP,
2650 		},
2651 		[MLX5_QP_STATE_INIT]  = {
2652 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2653 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2654 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_INIT2INIT_QP,
2655 			[MLX5_QP_STATE_RTR]	= MLX5_CMD_OP_INIT2RTR_QP,
2656 		},
2657 		[MLX5_QP_STATE_RTR]   = {
2658 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2659 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2660 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTR2RTS_QP,
2661 		},
2662 		[MLX5_QP_STATE_RTS]   = {
2663 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2664 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2665 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTS2RTS_QP,
2666 		},
2667 		[MLX5_QP_STATE_SQD] = {
2668 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2669 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2670 		},
2671 		[MLX5_QP_STATE_SQER] = {
2672 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2673 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2674 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQERR2RTS_QP,
2675 		},
2676 		[MLX5_QP_STATE_ERR] = {
2677 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2678 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2679 		}
2680 	};
2681 
2682 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2683 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2684 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
2685 	struct mlx5_ib_cq *send_cq, *recv_cq;
2686 	struct mlx5_qp_context *context;
2687 	struct mlx5_ib_pd *pd;
2688 	struct mlx5_ib_port *mibport = NULL;
2689 	enum mlx5_qp_state mlx5_cur, mlx5_new;
2690 	enum mlx5_qp_optpar optpar;
2691 	int sqd_event;
2692 	int mlx5_st;
2693 	int err;
2694 	u16 op;
2695 
2696 	context = kzalloc(sizeof(*context), GFP_KERNEL);
2697 	if (!context)
2698 		return -ENOMEM;
2699 
2700 	err = to_mlx5_st(ibqp->qp_type);
2701 	if (err < 0) {
2702 		mlx5_ib_dbg(dev, "unsupported qp type %d\n", ibqp->qp_type);
2703 		goto out;
2704 	}
2705 
2706 	context->flags = cpu_to_be32(err << 16);
2707 
2708 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
2709 		context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2710 	} else {
2711 		switch (attr->path_mig_state) {
2712 		case IB_MIG_MIGRATED:
2713 			context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2714 			break;
2715 		case IB_MIG_REARM:
2716 			context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
2717 			break;
2718 		case IB_MIG_ARMED:
2719 			context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
2720 			break;
2721 		}
2722 	}
2723 
2724 	if (is_sqp(ibqp->qp_type)) {
2725 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
2726 	} else if (ibqp->qp_type == IB_QPT_UD ||
2727 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
2728 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
2729 	} else if (attr_mask & IB_QP_PATH_MTU) {
2730 		if (attr->path_mtu < IB_MTU_256 ||
2731 		    attr->path_mtu > IB_MTU_4096) {
2732 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
2733 			err = -EINVAL;
2734 			goto out;
2735 		}
2736 		context->mtu_msgmax = (attr->path_mtu << 5) |
2737 				      (u8)MLX5_CAP_GEN(dev->mdev, log_max_msg);
2738 	}
2739 
2740 	if (attr_mask & IB_QP_DEST_QPN)
2741 		context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
2742 
2743 	if (attr_mask & IB_QP_PKEY_INDEX)
2744 		context->pri_path.pkey_index = cpu_to_be16(attr->pkey_index);
2745 
2746 	/* todo implement counter_index functionality */
2747 
2748 	if (is_sqp(ibqp->qp_type))
2749 		context->pri_path.port = qp->port;
2750 
2751 	if (attr_mask & IB_QP_PORT)
2752 		context->pri_path.port = attr->port_num;
2753 
2754 	if (attr_mask & IB_QP_AV) {
2755 		err = mlx5_set_path(dev, qp, &attr->ah_attr, &context->pri_path,
2756 				    attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
2757 				    attr_mask, 0, attr, false);
2758 		if (err)
2759 			goto out;
2760 	}
2761 
2762 	if (attr_mask & IB_QP_TIMEOUT)
2763 		context->pri_path.ackto_lt |= attr->timeout << 3;
2764 
2765 	if (attr_mask & IB_QP_ALT_PATH) {
2766 		err = mlx5_set_path(dev, qp, &attr->alt_ah_attr,
2767 				    &context->alt_path,
2768 				    attr->alt_port_num,
2769 				    attr_mask | IB_QP_PKEY_INDEX | IB_QP_TIMEOUT,
2770 				    0, attr, true);
2771 		if (err)
2772 			goto out;
2773 	}
2774 
2775 	pd = get_pd(qp);
2776 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
2777 		&send_cq, &recv_cq);
2778 
2779 	context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
2780 	context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
2781 	context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
2782 	context->params1  = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
2783 
2784 	if (attr_mask & IB_QP_RNR_RETRY)
2785 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2786 
2787 	if (attr_mask & IB_QP_RETRY_CNT)
2788 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2789 
2790 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2791 		if (attr->max_rd_atomic)
2792 			context->params1 |=
2793 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2794 	}
2795 
2796 	if (attr_mask & IB_QP_SQ_PSN)
2797 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2798 
2799 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2800 		if (attr->max_dest_rd_atomic)
2801 			context->params2 |=
2802 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2803 	}
2804 
2805 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
2806 		__be32 access_flags;
2807 
2808 		err = to_mlx5_access_flags(qp, attr, attr_mask, &access_flags);
2809 		if (err)
2810 			goto out;
2811 
2812 		context->params2 |= access_flags;
2813 	}
2814 
2815 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
2816 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2817 
2818 	if (attr_mask & IB_QP_RQ_PSN)
2819 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2820 
2821 	if (attr_mask & IB_QP_QKEY)
2822 		context->qkey = cpu_to_be32(attr->qkey);
2823 
2824 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2825 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2826 
2827 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2828 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2829 		sqd_event = 1;
2830 	else
2831 		sqd_event = 0;
2832 
2833 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2834 		u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
2835 			       qp->port) - 1;
2836 		mibport = &dev->port[port_num];
2837 		context->qp_counter_set_usr_page |=
2838 			cpu_to_be32((u32)(mibport->q_cnt_id) << 24);
2839 	}
2840 
2841 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2842 		context->sq_crq_size |= cpu_to_be16(1 << 4);
2843 
2844 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
2845 		context->deth_sqpn = cpu_to_be32(1);
2846 
2847 	mlx5_cur = to_mlx5_state(cur_state);
2848 	mlx5_new = to_mlx5_state(new_state);
2849 	mlx5_st = to_mlx5_st(ibqp->qp_type);
2850 	if (mlx5_st < 0)
2851 		goto out;
2852 
2853 	/* If moving to a reset or error state, we must disable page faults on
2854 	 * this QP and flush all current page faults. Otherwise a stale page
2855 	 * fault may attempt to work on this QP after it is reset and moved
2856 	 * again to RTS, and may cause the driver and the device to get out of
2857 	 * sync. */
2858 	if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2859 	    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR) &&
2860 	    (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
2861 		mlx5_ib_qp_disable_pagefaults(qp);
2862 
2863 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
2864 	    !optab[mlx5_cur][mlx5_new])
2865 		goto out;
2866 
2867 	op = optab[mlx5_cur][mlx5_new];
2868 	optpar = ib_mask_to_mlx5_opt(attr_mask);
2869 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
2870 
2871 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
2872 		struct mlx5_modify_raw_qp_param raw_qp_param = {};
2873 
2874 		raw_qp_param.operation = op;
2875 		if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2876 			raw_qp_param.rq_q_ctr_id = mibport->q_cnt_id;
2877 			raw_qp_param.set_mask |= MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID;
2878 		}
2879 		err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
2880 	} else {
2881 		err = mlx5_core_qp_modify(dev->mdev, op, optpar, context,
2882 					  &base->mqp);
2883 	}
2884 
2885 	if (err)
2886 		goto out;
2887 
2888 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT &&
2889 	    (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
2890 		mlx5_ib_qp_enable_pagefaults(qp);
2891 
2892 	qp->state = new_state;
2893 
2894 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2895 		qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
2896 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2897 		qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
2898 	if (attr_mask & IB_QP_PORT)
2899 		qp->port = attr->port_num;
2900 	if (attr_mask & IB_QP_ALT_PATH)
2901 		qp->trans_qp.alt_port = attr->alt_port_num;
2902 
2903 	/*
2904 	 * If we moved a kernel QP to RESET, clean up all old CQ
2905 	 * entries and reinitialize the QP.
2906 	 */
2907 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
2908 		mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2909 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
2910 		if (send_cq != recv_cq)
2911 			mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
2912 
2913 		qp->rq.head = 0;
2914 		qp->rq.tail = 0;
2915 		qp->sq.head = 0;
2916 		qp->sq.tail = 0;
2917 		qp->sq.cur_post = 0;
2918 		qp->sq.last_poll = 0;
2919 		qp->db.db[MLX5_RCV_DBR] = 0;
2920 		qp->db.db[MLX5_SND_DBR] = 0;
2921 	}
2922 
2923 out:
2924 	kfree(context);
2925 	return err;
2926 }
2927 
2928 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2929 		      int attr_mask, struct ib_udata *udata)
2930 {
2931 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2932 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2933 	enum ib_qp_type qp_type;
2934 	enum ib_qp_state cur_state, new_state;
2935 	int err = -EINVAL;
2936 	int port;
2937 
2938 	if (ibqp->rwq_ind_tbl)
2939 		return -ENOSYS;
2940 
2941 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
2942 		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
2943 
2944 	qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ?
2945 		IB_QPT_GSI : ibqp->qp_type;
2946 
2947 	mutex_lock(&qp->mutex);
2948 
2949 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2950 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2951 
2952 	if (qp_type != MLX5_IB_QPT_REG_UMR &&
2953 	    !ib_modify_qp_is_ok(cur_state, new_state, qp_type, attr_mask)) {
2954 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
2955 			    cur_state, new_state, ibqp->qp_type, attr_mask);
2956 		goto out;
2957 	}
2958 
2959 	if ((attr_mask & IB_QP_PORT) &&
2960 	    (attr->port_num == 0 ||
2961 	     attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports))) {
2962 		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
2963 			    attr->port_num, dev->num_ports);
2964 		goto out;
2965 	}
2966 
2967 	if (attr_mask & IB_QP_PKEY_INDEX) {
2968 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2969 		if (attr->pkey_index >=
2970 		    dev->mdev->port_caps[port - 1].pkey_table_len) {
2971 			mlx5_ib_dbg(dev, "invalid pkey index %d\n",
2972 				    attr->pkey_index);
2973 			goto out;
2974 		}
2975 	}
2976 
2977 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2978 	    attr->max_rd_atomic >
2979 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
2980 		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
2981 			    attr->max_rd_atomic);
2982 		goto out;
2983 	}
2984 
2985 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2986 	    attr->max_dest_rd_atomic >
2987 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
2988 		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
2989 			    attr->max_dest_rd_atomic);
2990 		goto out;
2991 	}
2992 
2993 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2994 		err = 0;
2995 		goto out;
2996 	}
2997 
2998 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2999 
3000 out:
3001 	mutex_unlock(&qp->mutex);
3002 	return err;
3003 }
3004 
3005 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3006 {
3007 	struct mlx5_ib_cq *cq;
3008 	unsigned cur;
3009 
3010 	cur = wq->head - wq->tail;
3011 	if (likely(cur + nreq < wq->max_post))
3012 		return 0;
3013 
3014 	cq = to_mcq(ib_cq);
3015 	spin_lock(&cq->lock);
3016 	cur = wq->head - wq->tail;
3017 	spin_unlock(&cq->lock);
3018 
3019 	return cur + nreq >= wq->max_post;
3020 }
3021 
3022 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
3023 					  u64 remote_addr, u32 rkey)
3024 {
3025 	rseg->raddr    = cpu_to_be64(remote_addr);
3026 	rseg->rkey     = cpu_to_be32(rkey);
3027 	rseg->reserved = 0;
3028 }
3029 
3030 static void *set_eth_seg(struct mlx5_wqe_eth_seg *eseg,
3031 			 const struct ib_send_wr *wr, void *qend,
3032 			 struct mlx5_ib_qp *qp, int *size)
3033 {
3034 	void *seg = eseg;
3035 
3036 	memset(eseg, 0, sizeof(struct mlx5_wqe_eth_seg));
3037 
3038 	if (wr->send_flags & IB_SEND_IP_CSUM)
3039 		eseg->cs_flags = MLX5_ETH_WQE_L3_CSUM |
3040 				 MLX5_ETH_WQE_L4_CSUM;
3041 
3042 	seg += sizeof(struct mlx5_wqe_eth_seg);
3043 	*size += sizeof(struct mlx5_wqe_eth_seg) / 16;
3044 
3045 	if (wr->opcode == IB_WR_LSO) {
3046 		struct ib_ud_wr *ud_wr = container_of(wr, struct ib_ud_wr, wr);
3047 		int size_of_inl_hdr_start = sizeof(eseg->inline_hdr_start);
3048 		u64 left, leftlen, copysz;
3049 		void *pdata = ud_wr->header;
3050 
3051 		left = ud_wr->hlen;
3052 		eseg->mss = cpu_to_be16(ud_wr->mss);
3053 		eseg->inline_hdr_sz = cpu_to_be16(left);
3054 
3055 		/*
3056 		 * check if there is space till the end of queue, if yes,
3057 		 * copy all in one shot, otherwise copy till the end of queue,
3058 		 * rollback and than the copy the left
3059 		 */
3060 		leftlen = qend - (void *)eseg->inline_hdr_start;
3061 		copysz = min_t(u64, leftlen, left);
3062 
3063 		memcpy(seg - size_of_inl_hdr_start, pdata, copysz);
3064 
3065 		if (likely(copysz > size_of_inl_hdr_start)) {
3066 			seg += ALIGN(copysz - size_of_inl_hdr_start, 16);
3067 			*size += ALIGN(copysz - size_of_inl_hdr_start, 16) / 16;
3068 		}
3069 
3070 		if (unlikely(copysz < left)) { /* the last wqe in the queue */
3071 			seg = mlx5_get_send_wqe(qp, 0);
3072 			left -= copysz;
3073 			pdata += copysz;
3074 			memcpy(seg, pdata, left);
3075 			seg += ALIGN(left, 16);
3076 			*size += ALIGN(left, 16) / 16;
3077 		}
3078 	}
3079 
3080 	return seg;
3081 }
3082 
3083 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
3084 			     const struct ib_send_wr *wr)
3085 {
3086 	memcpy(&dseg->av, &to_mah(ud_wr(wr)->ah)->av, sizeof(struct mlx5_av));
3087 	dseg->av.dqp_dct = cpu_to_be32(ud_wr(wr)->remote_qpn | MLX5_EXTENDED_UD_AV);
3088 	dseg->av.key.qkey.qkey = cpu_to_be32(ud_wr(wr)->remote_qkey);
3089 }
3090 
3091 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
3092 {
3093 	dseg->byte_count = cpu_to_be32(sg->length);
3094 	dseg->lkey       = cpu_to_be32(sg->lkey);
3095 	dseg->addr       = cpu_to_be64(sg->addr);
3096 }
3097 
3098 static __be16 get_klm_octo(int npages)
3099 {
3100 	return cpu_to_be16(ALIGN(npages, 8) / 2);
3101 }
3102 
3103 static __be64 frwr_mkey_mask(void)
3104 {
3105 	u64 result;
3106 
3107 	result = MLX5_MKEY_MASK_LEN		|
3108 		MLX5_MKEY_MASK_PAGE_SIZE	|
3109 		MLX5_MKEY_MASK_START_ADDR	|
3110 		MLX5_MKEY_MASK_EN_RINVAL	|
3111 		MLX5_MKEY_MASK_KEY		|
3112 		MLX5_MKEY_MASK_LR		|
3113 		MLX5_MKEY_MASK_LW		|
3114 		MLX5_MKEY_MASK_RR		|
3115 		MLX5_MKEY_MASK_RW		|
3116 		MLX5_MKEY_MASK_A		|
3117 		MLX5_MKEY_MASK_SMALL_FENCE	|
3118 		MLX5_MKEY_MASK_FREE;
3119 
3120 	return cpu_to_be64(result);
3121 }
3122 
3123 static __be64 sig_mkey_mask(void)
3124 {
3125 	u64 result;
3126 
3127 	result = MLX5_MKEY_MASK_LEN		|
3128 		MLX5_MKEY_MASK_PAGE_SIZE	|
3129 		MLX5_MKEY_MASK_START_ADDR	|
3130 		MLX5_MKEY_MASK_EN_SIGERR	|
3131 		MLX5_MKEY_MASK_EN_RINVAL	|
3132 		MLX5_MKEY_MASK_KEY		|
3133 		MLX5_MKEY_MASK_LR		|
3134 		MLX5_MKEY_MASK_LW		|
3135 		MLX5_MKEY_MASK_RR		|
3136 		MLX5_MKEY_MASK_RW		|
3137 		MLX5_MKEY_MASK_SMALL_FENCE	|
3138 		MLX5_MKEY_MASK_FREE		|
3139 		MLX5_MKEY_MASK_BSF_EN;
3140 
3141 	return cpu_to_be64(result);
3142 }
3143 
3144 static void set_reg_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr,
3145 				struct mlx5_ib_mr *mr)
3146 {
3147 	int ndescs = mr->ndescs;
3148 
3149 	memset(umr, 0, sizeof(*umr));
3150 
3151 	if (mr->access_mode == MLX5_ACCESS_MODE_KLM)
3152 		/* KLMs take twice the size of MTTs */
3153 		ndescs *= 2;
3154 
3155 	umr->flags = MLX5_UMR_CHECK_NOT_FREE;
3156 	umr->klm_octowords = get_klm_octo(ndescs);
3157 	umr->mkey_mask = frwr_mkey_mask();
3158 }
3159 
3160 static void set_linv_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr)
3161 {
3162 	memset(umr, 0, sizeof(*umr));
3163 	umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
3164 	umr->flags = 1 << 7;
3165 }
3166 
3167 static __be64 get_umr_reg_mr_mask(void)
3168 {
3169 	u64 result;
3170 
3171 	result = MLX5_MKEY_MASK_LEN		|
3172 		 MLX5_MKEY_MASK_PAGE_SIZE	|
3173 		 MLX5_MKEY_MASK_START_ADDR	|
3174 		 MLX5_MKEY_MASK_PD		|
3175 		 MLX5_MKEY_MASK_LR		|
3176 		 MLX5_MKEY_MASK_LW		|
3177 		 MLX5_MKEY_MASK_KEY		|
3178 		 MLX5_MKEY_MASK_RR		|
3179 		 MLX5_MKEY_MASK_RW		|
3180 		 MLX5_MKEY_MASK_A		|
3181 		 MLX5_MKEY_MASK_FREE;
3182 
3183 	return cpu_to_be64(result);
3184 }
3185 
3186 static __be64 get_umr_unreg_mr_mask(void)
3187 {
3188 	u64 result;
3189 
3190 	result = MLX5_MKEY_MASK_FREE;
3191 
3192 	return cpu_to_be64(result);
3193 }
3194 
3195 static __be64 get_umr_update_mtt_mask(void)
3196 {
3197 	u64 result;
3198 
3199 	result = MLX5_MKEY_MASK_FREE;
3200 
3201 	return cpu_to_be64(result);
3202 }
3203 
3204 static __be64 get_umr_update_translation_mask(void)
3205 {
3206 	u64 result;
3207 
3208 	result = MLX5_MKEY_MASK_LEN |
3209 		 MLX5_MKEY_MASK_PAGE_SIZE |
3210 		 MLX5_MKEY_MASK_START_ADDR |
3211 		 MLX5_MKEY_MASK_KEY |
3212 		 MLX5_MKEY_MASK_FREE;
3213 
3214 	return cpu_to_be64(result);
3215 }
3216 
3217 static __be64 get_umr_update_access_mask(void)
3218 {
3219 	u64 result;
3220 
3221 	result = MLX5_MKEY_MASK_LW |
3222 		 MLX5_MKEY_MASK_RR |
3223 		 MLX5_MKEY_MASK_RW |
3224 		 MLX5_MKEY_MASK_A |
3225 		 MLX5_MKEY_MASK_KEY |
3226 		 MLX5_MKEY_MASK_FREE;
3227 
3228 	return cpu_to_be64(result);
3229 }
3230 
3231 static __be64 get_umr_update_pd_mask(void)
3232 {
3233 	u64 result;
3234 
3235 	result = MLX5_MKEY_MASK_PD |
3236 		 MLX5_MKEY_MASK_KEY |
3237 		 MLX5_MKEY_MASK_FREE;
3238 
3239 	return cpu_to_be64(result);
3240 }
3241 
3242 static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
3243 				const struct ib_send_wr *wr)
3244 {
3245 	const struct mlx5_umr_wr *umrwr = umr_wr(wr);
3246 
3247 	memset(umr, 0, sizeof(*umr));
3248 
3249 	if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
3250 		umr->flags = MLX5_UMR_CHECK_FREE; /* fail if free */
3251 	else
3252 		umr->flags = MLX5_UMR_CHECK_NOT_FREE; /* fail if not free */
3253 
3254 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UNREG)) {
3255 		umr->klm_octowords = get_klm_octo(umrwr->npages);
3256 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT) {
3257 			umr->mkey_mask = get_umr_update_mtt_mask();
3258 			umr->bsf_octowords = get_klm_octo(umrwr->target.offset);
3259 			umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
3260 		}
3261 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION)
3262 			umr->mkey_mask |= get_umr_update_translation_mask();
3263 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_ACCESS)
3264 			umr->mkey_mask |= get_umr_update_access_mask();
3265 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_PD)
3266 			umr->mkey_mask |= get_umr_update_pd_mask();
3267 		if (!umr->mkey_mask)
3268 			umr->mkey_mask = get_umr_reg_mr_mask();
3269 	} else {
3270 		umr->mkey_mask = get_umr_unreg_mr_mask();
3271 	}
3272 
3273 	if (!wr->num_sge)
3274 		umr->flags |= MLX5_UMR_INLINE;
3275 }
3276 
3277 static u8 get_umr_flags(int acc)
3278 {
3279 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
3280 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
3281 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
3282 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
3283 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
3284 }
3285 
3286 static void set_reg_mkey_seg(struct mlx5_mkey_seg *seg,
3287 			     struct mlx5_ib_mr *mr,
3288 			     u32 key, int access)
3289 {
3290 	int ndescs = ALIGN(mr->ndescs, 8) >> 1;
3291 
3292 	memset(seg, 0, sizeof(*seg));
3293 
3294 	if (mr->access_mode == MLX5_ACCESS_MODE_MTT)
3295 		seg->log2_page_size = ilog2(mr->ibmr.page_size);
3296 	else if (mr->access_mode == MLX5_ACCESS_MODE_KLM)
3297 		/* KLMs take twice the size of MTTs */
3298 		ndescs *= 2;
3299 
3300 	seg->flags = get_umr_flags(access) | mr->access_mode;
3301 	seg->qpn_mkey7_0 = cpu_to_be32((key & 0xff) | 0xffffff00);
3302 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
3303 	seg->start_addr = cpu_to_be64(mr->ibmr.iova);
3304 	seg->len = cpu_to_be64(mr->ibmr.length);
3305 	seg->xlt_oct_size = cpu_to_be32(ndescs);
3306 }
3307 
3308 static void set_linv_mkey_seg(struct mlx5_mkey_seg *seg)
3309 {
3310 	memset(seg, 0, sizeof(*seg));
3311 	seg->status = MLX5_MKEY_STATUS_FREE;
3312 }
3313 
3314 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, const struct ib_send_wr *wr)
3315 {
3316 	const struct mlx5_umr_wr *umrwr = umr_wr(wr);
3317 
3318 	memset(seg, 0, sizeof(*seg));
3319 	if (wr->send_flags & MLX5_IB_SEND_UMR_UNREG) {
3320 		seg->status = MLX5_MKEY_STATUS_FREE;
3321 		return;
3322 	}
3323 
3324 	seg->flags = convert_access(umrwr->access_flags);
3325 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT)) {
3326 		if (umrwr->pd)
3327 			seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
3328 		seg->start_addr = cpu_to_be64(umrwr->target.virt_addr);
3329 	}
3330 	seg->len = cpu_to_be64(umrwr->length);
3331 	seg->log2_page_size = umrwr->page_shift;
3332 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
3333 				       mlx5_mkey_variant(umrwr->mkey));
3334 }
3335 
3336 static void set_reg_data_seg(struct mlx5_wqe_data_seg *dseg,
3337 			     struct mlx5_ib_mr *mr,
3338 			     struct mlx5_ib_pd *pd)
3339 {
3340 	int bcount = mr->desc_size * mr->ndescs;
3341 
3342 	dseg->addr = cpu_to_be64(mr->desc_map);
3343 	dseg->byte_count = cpu_to_be32(ALIGN(bcount, 64));
3344 	dseg->lkey = cpu_to_be32(pd->ibpd.local_dma_lkey);
3345 }
3346 
3347 static __be32 send_ieth(const struct ib_send_wr *wr)
3348 {
3349 	switch (wr->opcode) {
3350 	case IB_WR_SEND_WITH_IMM:
3351 	case IB_WR_RDMA_WRITE_WITH_IMM:
3352 		return wr->ex.imm_data;
3353 
3354 	case IB_WR_SEND_WITH_INV:
3355 		return cpu_to_be32(wr->ex.invalidate_rkey);
3356 
3357 	default:
3358 		return 0;
3359 	}
3360 }
3361 
3362 static u8 calc_sig(void *wqe, int size)
3363 {
3364 	u8 *p = wqe;
3365 	u8 res = 0;
3366 	int i;
3367 
3368 	for (i = 0; i < size; i++)
3369 		res ^= p[i];
3370 
3371 	return ~res;
3372 }
3373 
3374 static u8 wq_sig(void *wqe)
3375 {
3376 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
3377 }
3378 
3379 static int set_data_inl_seg(struct mlx5_ib_qp *qp, const struct ib_send_wr *wr,
3380 			    void *wqe, int *sz)
3381 {
3382 	struct mlx5_wqe_inline_seg *seg;
3383 	void *qend = qp->sq.qend;
3384 	void *addr;
3385 	int inl = 0;
3386 	int copy;
3387 	int len;
3388 	int i;
3389 
3390 	seg = wqe;
3391 	wqe += sizeof(*seg);
3392 	for (i = 0; i < wr->num_sge; i++) {
3393 		addr = (void *)(unsigned long)(wr->sg_list[i].addr);
3394 		len  = wr->sg_list[i].length;
3395 		inl += len;
3396 
3397 		if (unlikely(inl > qp->max_inline_data))
3398 			return -ENOMEM;
3399 
3400 		if (unlikely(wqe + len > qend)) {
3401 			copy = qend - wqe;
3402 			memcpy(wqe, addr, copy);
3403 			addr += copy;
3404 			len -= copy;
3405 			wqe = mlx5_get_send_wqe(qp, 0);
3406 		}
3407 		memcpy(wqe, addr, len);
3408 		wqe += len;
3409 	}
3410 
3411 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
3412 
3413 	*sz = ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
3414 
3415 	return 0;
3416 }
3417 
3418 static u16 prot_field_size(enum ib_signature_type type)
3419 {
3420 	switch (type) {
3421 	case IB_SIG_TYPE_T10_DIF:
3422 		return MLX5_DIF_SIZE;
3423 	default:
3424 		return 0;
3425 	}
3426 }
3427 
3428 static u8 bs_selector(int block_size)
3429 {
3430 	switch (block_size) {
3431 	case 512:	    return 0x1;
3432 	case 520:	    return 0x2;
3433 	case 4096:	    return 0x3;
3434 	case 4160:	    return 0x4;
3435 	case 1073741824:    return 0x5;
3436 	default:	    return 0;
3437 	}
3438 }
3439 
3440 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
3441 			      struct mlx5_bsf_inl *inl)
3442 {
3443 	/* Valid inline section and allow BSF refresh */
3444 	inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
3445 				       MLX5_BSF_REFRESH_DIF);
3446 	inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
3447 	inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
3448 	/* repeating block */
3449 	inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
3450 	inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
3451 			MLX5_DIF_CRC : MLX5_DIF_IPCS;
3452 
3453 	if (domain->sig.dif.ref_remap)
3454 		inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
3455 
3456 	if (domain->sig.dif.app_escape) {
3457 		if (domain->sig.dif.ref_escape)
3458 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
3459 		else
3460 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
3461 	}
3462 
3463 	inl->dif_app_bitmask_check =
3464 		cpu_to_be16(domain->sig.dif.apptag_check_mask);
3465 }
3466 
3467 static int mlx5_set_bsf(struct ib_mr *sig_mr,
3468 			struct ib_sig_attrs *sig_attrs,
3469 			struct mlx5_bsf *bsf, u32 data_size)
3470 {
3471 	struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
3472 	struct mlx5_bsf_basic *basic = &bsf->basic;
3473 	struct ib_sig_domain *mem = &sig_attrs->mem;
3474 	struct ib_sig_domain *wire = &sig_attrs->wire;
3475 
3476 	memset(bsf, 0, sizeof(*bsf));
3477 
3478 	/* Basic + Extended + Inline */
3479 	basic->bsf_size_sbs = 1 << 7;
3480 	/* Input domain check byte mask */
3481 	basic->check_byte_mask = sig_attrs->check_mask;
3482 	basic->raw_data_size = cpu_to_be32(data_size);
3483 
3484 	/* Memory domain */
3485 	switch (sig_attrs->mem.sig_type) {
3486 	case IB_SIG_TYPE_NONE:
3487 		break;
3488 	case IB_SIG_TYPE_T10_DIF:
3489 		basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
3490 		basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
3491 		mlx5_fill_inl_bsf(mem, &bsf->m_inl);
3492 		break;
3493 	default:
3494 		return -EINVAL;
3495 	}
3496 
3497 	/* Wire domain */
3498 	switch (sig_attrs->wire.sig_type) {
3499 	case IB_SIG_TYPE_NONE:
3500 		break;
3501 	case IB_SIG_TYPE_T10_DIF:
3502 		if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
3503 		    mem->sig_type == wire->sig_type) {
3504 			/* Same block structure */
3505 			basic->bsf_size_sbs |= 1 << 4;
3506 			if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
3507 				basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
3508 			if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
3509 				basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
3510 			if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
3511 				basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
3512 		} else
3513 			basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
3514 
3515 		basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
3516 		mlx5_fill_inl_bsf(wire, &bsf->w_inl);
3517 		break;
3518 	default:
3519 		return -EINVAL;
3520 	}
3521 
3522 	return 0;
3523 }
3524 
3525 static int set_sig_data_segment(const struct ib_sig_handover_wr *wr,
3526 				struct mlx5_ib_qp *qp, void **seg, int *size)
3527 {
3528 	struct ib_sig_attrs *sig_attrs = wr->sig_attrs;
3529 	struct ib_mr *sig_mr = wr->sig_mr;
3530 	struct mlx5_bsf *bsf;
3531 	u32 data_len = wr->wr.sg_list->length;
3532 	u32 data_key = wr->wr.sg_list->lkey;
3533 	u64 data_va = wr->wr.sg_list->addr;
3534 	int ret;
3535 	int wqe_size;
3536 
3537 	if (!wr->prot ||
3538 	    (data_key == wr->prot->lkey &&
3539 	     data_va == wr->prot->addr &&
3540 	     data_len == wr->prot->length)) {
3541 		/**
3542 		 * Source domain doesn't contain signature information
3543 		 * or data and protection are interleaved in memory.
3544 		 * So need construct:
3545 		 *                  ------------------
3546 		 *                 |     data_klm     |
3547 		 *                  ------------------
3548 		 *                 |       BSF        |
3549 		 *                  ------------------
3550 		 **/
3551 		struct mlx5_klm *data_klm = *seg;
3552 
3553 		data_klm->bcount = cpu_to_be32(data_len);
3554 		data_klm->key = cpu_to_be32(data_key);
3555 		data_klm->va = cpu_to_be64(data_va);
3556 		wqe_size = ALIGN(sizeof(*data_klm), 64);
3557 	} else {
3558 		/**
3559 		 * Source domain contains signature information
3560 		 * So need construct a strided block format:
3561 		 *               ---------------------------
3562 		 *              |     stride_block_ctrl     |
3563 		 *               ---------------------------
3564 		 *              |          data_klm         |
3565 		 *               ---------------------------
3566 		 *              |          prot_klm         |
3567 		 *               ---------------------------
3568 		 *              |             BSF           |
3569 		 *               ---------------------------
3570 		 **/
3571 		struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
3572 		struct mlx5_stride_block_entry *data_sentry;
3573 		struct mlx5_stride_block_entry *prot_sentry;
3574 		u32 prot_key = wr->prot->lkey;
3575 		u64 prot_va = wr->prot->addr;
3576 		u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
3577 		int prot_size;
3578 
3579 		sblock_ctrl = *seg;
3580 		data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
3581 		prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
3582 
3583 		prot_size = prot_field_size(sig_attrs->mem.sig_type);
3584 		if (!prot_size) {
3585 			pr_err("Bad block size given: %u\n", block_size);
3586 			return -EINVAL;
3587 		}
3588 		sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
3589 							    prot_size);
3590 		sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
3591 		sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
3592 		sblock_ctrl->num_entries = cpu_to_be16(2);
3593 
3594 		data_sentry->bcount = cpu_to_be16(block_size);
3595 		data_sentry->key = cpu_to_be32(data_key);
3596 		data_sentry->va = cpu_to_be64(data_va);
3597 		data_sentry->stride = cpu_to_be16(block_size);
3598 
3599 		prot_sentry->bcount = cpu_to_be16(prot_size);
3600 		prot_sentry->key = cpu_to_be32(prot_key);
3601 		prot_sentry->va = cpu_to_be64(prot_va);
3602 		prot_sentry->stride = cpu_to_be16(prot_size);
3603 
3604 		wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
3605 				 sizeof(*prot_sentry), 64);
3606 	}
3607 
3608 	*seg += wqe_size;
3609 	*size += wqe_size / 16;
3610 	if (unlikely((*seg == qp->sq.qend)))
3611 		*seg = mlx5_get_send_wqe(qp, 0);
3612 
3613 	bsf = *seg;
3614 	ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
3615 	if (ret)
3616 		return -EINVAL;
3617 
3618 	*seg += sizeof(*bsf);
3619 	*size += sizeof(*bsf) / 16;
3620 	if (unlikely((*seg == qp->sq.qend)))
3621 		*seg = mlx5_get_send_wqe(qp, 0);
3622 
3623 	return 0;
3624 }
3625 
3626 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
3627 				 const struct ib_sig_handover_wr *wr, u32 nelements,
3628 				 u32 length, u32 pdn)
3629 {
3630 	struct ib_mr *sig_mr = wr->sig_mr;
3631 	u32 sig_key = sig_mr->rkey;
3632 	u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
3633 
3634 	memset(seg, 0, sizeof(*seg));
3635 
3636 	seg->flags = get_umr_flags(wr->access_flags) |
3637 				   MLX5_ACCESS_MODE_KLM;
3638 	seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
3639 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
3640 				    MLX5_MKEY_BSF_EN | pdn);
3641 	seg->len = cpu_to_be64(length);
3642 	seg->xlt_oct_size = cpu_to_be32(be16_to_cpu(get_klm_octo(nelements)));
3643 	seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
3644 }
3645 
3646 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
3647 				u32 nelements)
3648 {
3649 	memset(umr, 0, sizeof(*umr));
3650 
3651 	umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
3652 	umr->klm_octowords = get_klm_octo(nelements);
3653 	umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
3654 	umr->mkey_mask = sig_mkey_mask();
3655 }
3656 
3657 
3658 static int set_sig_umr_wr(const struct ib_send_wr *send_wr, struct mlx5_ib_qp *qp,
3659 			  void **seg, int *size)
3660 {
3661 	const struct ib_sig_handover_wr *wr = sig_handover_wr(send_wr);
3662 	struct mlx5_ib_mr *sig_mr = to_mmr(wr->sig_mr);
3663 	u32 pdn = get_pd(qp)->pdn;
3664 	u32 klm_oct_size;
3665 	int region_len, ret;
3666 
3667 	if (unlikely(wr->wr.num_sge != 1) ||
3668 	    unlikely(wr->access_flags & IB_ACCESS_REMOTE_ATOMIC) ||
3669 	    unlikely(!sig_mr->sig) || unlikely(!qp->signature_en) ||
3670 	    unlikely(!sig_mr->sig->sig_status_checked))
3671 		return -EINVAL;
3672 
3673 	/* length of the protected region, data + protection */
3674 	region_len = wr->wr.sg_list->length;
3675 	if (wr->prot &&
3676 	    (wr->prot->lkey != wr->wr.sg_list->lkey  ||
3677 	     wr->prot->addr != wr->wr.sg_list->addr  ||
3678 	     wr->prot->length != wr->wr.sg_list->length))
3679 		region_len += wr->prot->length;
3680 
3681 	/**
3682 	 * KLM octoword size - if protection was provided
3683 	 * then we use strided block format (3 octowords),
3684 	 * else we use single KLM (1 octoword)
3685 	 **/
3686 	klm_oct_size = wr->prot ? 3 : 1;
3687 
3688 	set_sig_umr_segment(*seg, klm_oct_size);
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 
3694 	set_sig_mkey_segment(*seg, wr, klm_oct_size, region_len, pdn);
3695 	*seg += sizeof(struct mlx5_mkey_seg);
3696 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3697 	if (unlikely((*seg == qp->sq.qend)))
3698 		*seg = mlx5_get_send_wqe(qp, 0);
3699 
3700 	ret = set_sig_data_segment(wr, qp, seg, size);
3701 	if (ret)
3702 		return ret;
3703 
3704 	sig_mr->sig->sig_status_checked = false;
3705 	return 0;
3706 }
3707 
3708 static int set_psv_wr(struct ib_sig_domain *domain,
3709 		      u32 psv_idx, void **seg, int *size)
3710 {
3711 	struct mlx5_seg_set_psv *psv_seg = *seg;
3712 
3713 	memset(psv_seg, 0, sizeof(*psv_seg));
3714 	psv_seg->psv_num = cpu_to_be32(psv_idx);
3715 	switch (domain->sig_type) {
3716 	case IB_SIG_TYPE_NONE:
3717 		break;
3718 	case IB_SIG_TYPE_T10_DIF:
3719 		psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
3720 						     domain->sig.dif.app_tag);
3721 		psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
3722 		break;
3723 	default:
3724 		pr_err("Bad signature type given.\n");
3725 		return 1;
3726 	}
3727 
3728 	*seg += sizeof(*psv_seg);
3729 	*size += sizeof(*psv_seg) / 16;
3730 
3731 	return 0;
3732 }
3733 
3734 static int set_reg_wr(struct mlx5_ib_qp *qp,
3735 		      const struct ib_reg_wr *wr,
3736 		      void **seg, int *size)
3737 {
3738 	struct mlx5_ib_mr *mr = to_mmr(wr->mr);
3739 	struct mlx5_ib_pd *pd = to_mpd(qp->ibqp.pd);
3740 
3741 	if (unlikely(wr->wr.send_flags & IB_SEND_INLINE)) {
3742 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
3743 			     "Invalid IB_SEND_INLINE send flag\n");
3744 		return -EINVAL;
3745 	}
3746 
3747 	set_reg_umr_seg(*seg, mr);
3748 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3749 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3750 	if (unlikely((*seg == qp->sq.qend)))
3751 		*seg = mlx5_get_send_wqe(qp, 0);
3752 
3753 	set_reg_mkey_seg(*seg, mr, wr->key, wr->access);
3754 	*seg += sizeof(struct mlx5_mkey_seg);
3755 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3756 	if (unlikely((*seg == qp->sq.qend)))
3757 		*seg = mlx5_get_send_wqe(qp, 0);
3758 
3759 	set_reg_data_seg(*seg, mr, pd);
3760 	*seg += sizeof(struct mlx5_wqe_data_seg);
3761 	*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
3762 
3763 	return 0;
3764 }
3765 
3766 static void set_linv_wr(struct mlx5_ib_qp *qp, void **seg, int *size)
3767 {
3768 	set_linv_umr_seg(*seg);
3769 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3770 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3771 	if (unlikely((*seg == qp->sq.qend)))
3772 		*seg = mlx5_get_send_wqe(qp, 0);
3773 	set_linv_mkey_seg(*seg);
3774 	*seg += sizeof(struct mlx5_mkey_seg);
3775 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3776 	if (unlikely((*seg == qp->sq.qend)))
3777 		*seg = mlx5_get_send_wqe(qp, 0);
3778 }
3779 
3780 static void dump_wqe(struct mlx5_ib_qp *qp, int idx, int size_16)
3781 {
3782 	__be32 *p = NULL;
3783 	int tidx = idx;
3784 	int i, j;
3785 
3786 	pr_debug("dump wqe at %p\n", mlx5_get_send_wqe(qp, tidx));
3787 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
3788 		if ((i & 0xf) == 0) {
3789 			void *buf = mlx5_get_send_wqe(qp, tidx);
3790 			tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
3791 			p = buf;
3792 			j = 0;
3793 		}
3794 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
3795 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
3796 			 be32_to_cpu(p[j + 3]));
3797 	}
3798 }
3799 
3800 static u8 get_fence(u8 fence, const struct ib_send_wr *wr)
3801 {
3802 	if (unlikely(wr->opcode == IB_WR_LOCAL_INV &&
3803 		     wr->send_flags & IB_SEND_FENCE))
3804 		return MLX5_FENCE_MODE_STRONG_ORDERING;
3805 
3806 	if (unlikely(fence)) {
3807 		if (wr->send_flags & IB_SEND_FENCE)
3808 			return MLX5_FENCE_MODE_SMALL_AND_FENCE;
3809 		else
3810 			return fence;
3811 	} else if (unlikely(wr->send_flags & IB_SEND_FENCE)) {
3812 		return MLX5_FENCE_MODE_FENCE;
3813 	}
3814 
3815 	return 0;
3816 }
3817 
3818 static int __begin_wqe(struct mlx5_ib_qp *qp, void **seg,
3819 		       struct mlx5_wqe_ctrl_seg **ctrl,
3820 		       const struct ib_send_wr *wr, unsigned *idx,
3821 		       int *size, int nreq, bool send_signaled, bool solicited)
3822 {
3823 	if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)))
3824 		return -ENOMEM;
3825 
3826 	*idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
3827 	*seg = mlx5_get_send_wqe(qp, *idx);
3828 	*ctrl = *seg;
3829 	*(uint32_t *)(*seg + 8) = 0;
3830 	(*ctrl)->imm = send_ieth(wr);
3831 	(*ctrl)->fm_ce_se = qp->sq_signal_bits |
3832 		(send_signaled ? MLX5_WQE_CTRL_CQ_UPDATE : 0) |
3833 		(solicited ? MLX5_WQE_CTRL_SOLICITED : 0);
3834 
3835 	*seg += sizeof(**ctrl);
3836 	*size = sizeof(**ctrl) / 16;
3837 
3838 	return 0;
3839 }
3840 
3841 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
3842 		     struct mlx5_wqe_ctrl_seg **ctrl,
3843 		     const struct ib_send_wr *wr, unsigned *idx,
3844 		     int *size, int nreq)
3845 {
3846 	return __begin_wqe(qp, seg, ctrl, wr, idx, size, nreq,
3847 			   wr->send_flags & IB_SEND_SIGNALED,
3848 			   wr->send_flags & IB_SEND_SOLICITED);
3849 }
3850 
3851 static void finish_wqe(struct mlx5_ib_qp *qp,
3852 		       struct mlx5_wqe_ctrl_seg *ctrl,
3853 		       u8 size, unsigned idx, u64 wr_id,
3854 		       int nreq, u8 fence, u8 next_fence,
3855 		       u32 mlx5_opcode)
3856 {
3857 	u8 opmod = 0;
3858 
3859 	ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
3860 					     mlx5_opcode | ((u32)opmod << 24));
3861 	ctrl->qpn_ds = cpu_to_be32(size | (qp->trans_qp.base.mqp.qpn << 8));
3862 	ctrl->fm_ce_se |= fence;
3863 	qp->fm_cache = next_fence;
3864 	if (unlikely(qp->wq_sig))
3865 		ctrl->signature = wq_sig(ctrl);
3866 
3867 	qp->sq.wrid[idx] = wr_id;
3868 	qp->sq.w_list[idx].opcode = mlx5_opcode;
3869 	qp->sq.wqe_head[idx] = qp->sq.head + nreq;
3870 	qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
3871 	qp->sq.w_list[idx].next = qp->sq.cur_post;
3872 }
3873 
3874 
3875 int mlx5_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3876 		      const struct ib_send_wr **bad_wr)
3877 {
3878 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
3879 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3880 	struct mlx5_core_dev *mdev = dev->mdev;
3881 	struct mlx5_ib_qp *qp;
3882 	struct mlx5_ib_mr *mr;
3883 	struct mlx5_wqe_data_seg *dpseg;
3884 	struct mlx5_wqe_xrc_seg *xrc;
3885 	struct mlx5_bf *bf;
3886 	int uninitialized_var(size);
3887 	void *qend;
3888 	unsigned long flags;
3889 	unsigned idx;
3890 	int err = 0;
3891 	int inl = 0;
3892 	int num_sge;
3893 	void *seg;
3894 	int nreq;
3895 	int i;
3896 	u8 next_fence = 0;
3897 	u8 fence;
3898 
3899 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
3900 		return mlx5_ib_gsi_post_send(ibqp, wr, bad_wr);
3901 
3902 	qp = to_mqp(ibqp);
3903 	bf = &qp->bf;
3904 	qend = qp->sq.qend;
3905 
3906 	spin_lock_irqsave(&qp->sq.lock, flags);
3907 
3908 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
3909 		err = -EIO;
3910 		*bad_wr = wr;
3911 		nreq = 0;
3912 		goto out;
3913 	}
3914 
3915 	for (nreq = 0; wr; nreq++, wr = wr->next) {
3916 		if (unlikely(wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
3917 			mlx5_ib_warn(dev, "\n");
3918 			err = -EINVAL;
3919 			*bad_wr = wr;
3920 			goto out;
3921 		}
3922 
3923 		fence = qp->fm_cache;
3924 		num_sge = wr->num_sge;
3925 		if (unlikely(num_sge > qp->sq.max_gs)) {
3926 			mlx5_ib_warn(dev, "\n");
3927 			err = -EINVAL;
3928 			*bad_wr = wr;
3929 			goto out;
3930 		}
3931 
3932 		err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, nreq);
3933 		if (err) {
3934 			mlx5_ib_warn(dev, "\n");
3935 			err = -ENOMEM;
3936 			*bad_wr = wr;
3937 			goto out;
3938 		}
3939 
3940 		switch (ibqp->qp_type) {
3941 		case IB_QPT_XRC_INI:
3942 			xrc = seg;
3943 			seg += sizeof(*xrc);
3944 			size += sizeof(*xrc) / 16;
3945 			/* fall through */
3946 		case IB_QPT_RC:
3947 			switch (wr->opcode) {
3948 			case IB_WR_RDMA_READ:
3949 			case IB_WR_RDMA_WRITE:
3950 			case IB_WR_RDMA_WRITE_WITH_IMM:
3951 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
3952 					      rdma_wr(wr)->rkey);
3953 				seg += sizeof(struct mlx5_wqe_raddr_seg);
3954 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
3955 				break;
3956 
3957 			case IB_WR_ATOMIC_CMP_AND_SWP:
3958 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3959 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3960 				mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
3961 				err = -ENOSYS;
3962 				*bad_wr = wr;
3963 				goto out;
3964 
3965 			case IB_WR_LOCAL_INV:
3966 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3967 				qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
3968 				ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
3969 				set_linv_wr(qp, &seg, &size);
3970 				num_sge = 0;
3971 				break;
3972 
3973 			case IB_WR_REG_MR:
3974 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3975 				qp->sq.wr_data[idx] = IB_WR_REG_MR;
3976 				ctrl->imm = cpu_to_be32(reg_wr(wr)->key);
3977 				err = set_reg_wr(qp, reg_wr(wr), &seg, &size);
3978 				if (err) {
3979 					*bad_wr = wr;
3980 					goto out;
3981 				}
3982 				num_sge = 0;
3983 				break;
3984 
3985 			case IB_WR_REG_SIG_MR:
3986 				qp->sq.wr_data[idx] = IB_WR_REG_SIG_MR;
3987 				mr = to_mmr(sig_handover_wr(wr)->sig_mr);
3988 
3989 				ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
3990 				err = set_sig_umr_wr(wr, qp, &seg, &size);
3991 				if (err) {
3992 					mlx5_ib_warn(dev, "\n");
3993 					*bad_wr = wr;
3994 					goto out;
3995 				}
3996 
3997 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3998 					   nreq, get_fence(fence, wr),
3999 					   next_fence, MLX5_OPCODE_UMR);
4000 				/*
4001 				 * SET_PSV WQEs are not signaled and solicited
4002 				 * on error
4003 				 */
4004 				err = __begin_wqe(qp, &seg, &ctrl, wr,
4005 						  &idx, &size, nreq, false, true);
4006 				if (err) {
4007 					mlx5_ib_warn(dev, "\n");
4008 					err = -ENOMEM;
4009 					*bad_wr = wr;
4010 					goto out;
4011 				}
4012 
4013 				err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->mem,
4014 						 mr->sig->psv_memory.psv_idx, &seg,
4015 						 &size);
4016 				if (err) {
4017 					mlx5_ib_warn(dev, "\n");
4018 					*bad_wr = wr;
4019 					goto out;
4020 				}
4021 
4022 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
4023 					   nreq, get_fence(fence, wr),
4024 					   next_fence, MLX5_OPCODE_SET_PSV);
4025 				err = __begin_wqe(qp, &seg, &ctrl, wr,
4026 						  &idx, &size, nreq, false, true);
4027 				if (err) {
4028 					mlx5_ib_warn(dev, "\n");
4029 					err = -ENOMEM;
4030 					*bad_wr = wr;
4031 					goto out;
4032 				}
4033 
4034 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
4035 				err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->wire,
4036 						 mr->sig->psv_wire.psv_idx, &seg,
4037 						 &size);
4038 				if (err) {
4039 					mlx5_ib_warn(dev, "\n");
4040 					*bad_wr = wr;
4041 					goto out;
4042 				}
4043 
4044 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
4045 					   nreq, get_fence(fence, wr),
4046 					   next_fence, MLX5_OPCODE_SET_PSV);
4047 				num_sge = 0;
4048 				goto skip_psv;
4049 
4050 			default:
4051 				break;
4052 			}
4053 			break;
4054 
4055 		case IB_QPT_UC:
4056 			switch (wr->opcode) {
4057 			case IB_WR_RDMA_WRITE:
4058 			case IB_WR_RDMA_WRITE_WITH_IMM:
4059 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
4060 					      rdma_wr(wr)->rkey);
4061 				seg  += sizeof(struct mlx5_wqe_raddr_seg);
4062 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
4063 				break;
4064 
4065 			default:
4066 				break;
4067 			}
4068 			break;
4069 
4070 		case IB_QPT_SMI:
4071 		case MLX5_IB_QPT_HW_GSI:
4072 			set_datagram_seg(seg, wr);
4073 			seg += sizeof(struct mlx5_wqe_datagram_seg);
4074 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
4075 			if (unlikely((seg == qend)))
4076 				seg = mlx5_get_send_wqe(qp, 0);
4077 			break;
4078 		case IB_QPT_UD:
4079 			set_datagram_seg(seg, wr);
4080 			seg += sizeof(struct mlx5_wqe_datagram_seg);
4081 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
4082 
4083 			if (unlikely((seg == qend)))
4084 				seg = mlx5_get_send_wqe(qp, 0);
4085 
4086 			/* handle qp that supports ud offload */
4087 			if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO) {
4088 				struct mlx5_wqe_eth_pad *pad;
4089 
4090 				pad = seg;
4091 				memset(pad, 0, sizeof(struct mlx5_wqe_eth_pad));
4092 				seg += sizeof(struct mlx5_wqe_eth_pad);
4093 				size += sizeof(struct mlx5_wqe_eth_pad) / 16;
4094 
4095 				seg = set_eth_seg(seg, wr, qend, qp, &size);
4096 
4097 				if (unlikely((seg == qend)))
4098 					seg = mlx5_get_send_wqe(qp, 0);
4099 			}
4100 			break;
4101 		case MLX5_IB_QPT_REG_UMR:
4102 			if (wr->opcode != MLX5_IB_WR_UMR) {
4103 				err = -EINVAL;
4104 				mlx5_ib_warn(dev, "bad opcode\n");
4105 				goto out;
4106 			}
4107 			qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
4108 			ctrl->imm = cpu_to_be32(umr_wr(wr)->mkey);
4109 			set_reg_umr_segment(seg, wr);
4110 			seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4111 			size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4112 			if (unlikely((seg == qend)))
4113 				seg = mlx5_get_send_wqe(qp, 0);
4114 			set_reg_mkey_segment(seg, wr);
4115 			seg += sizeof(struct mlx5_mkey_seg);
4116 			size += sizeof(struct mlx5_mkey_seg) / 16;
4117 			if (unlikely((seg == qend)))
4118 				seg = mlx5_get_send_wqe(qp, 0);
4119 			break;
4120 
4121 		default:
4122 			break;
4123 		}
4124 
4125 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
4126 			int uninitialized_var(sz);
4127 
4128 			err = set_data_inl_seg(qp, wr, seg, &sz);
4129 			if (unlikely(err)) {
4130 				mlx5_ib_warn(dev, "\n");
4131 				*bad_wr = wr;
4132 				goto out;
4133 			}
4134 			inl = 1;
4135 			size += sz;
4136 		} else {
4137 			dpseg = seg;
4138 			for (i = 0; i < num_sge; i++) {
4139 				if (unlikely(dpseg == qend)) {
4140 					seg = mlx5_get_send_wqe(qp, 0);
4141 					dpseg = seg;
4142 				}
4143 				if (likely(wr->sg_list[i].length)) {
4144 					set_data_ptr_seg(dpseg, wr->sg_list + i);
4145 					size += sizeof(struct mlx5_wqe_data_seg) / 16;
4146 					dpseg++;
4147 				}
4148 			}
4149 		}
4150 
4151 		finish_wqe(qp, ctrl, size, idx, wr->wr_id, nreq,
4152 			   get_fence(fence, wr), next_fence,
4153 			   mlx5_ib_opcode[wr->opcode]);
4154 skip_psv:
4155 		if (0)
4156 			dump_wqe(qp, idx, size);
4157 	}
4158 
4159 out:
4160 	if (likely(nreq)) {
4161 		qp->sq.head += nreq;
4162 
4163 		/* Make sure that descriptors are written before
4164 		 * updating doorbell record and ringing the doorbell
4165 		 */
4166 		wmb();
4167 
4168 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
4169 
4170 		/* Make sure doorbell record is visible to the HCA before
4171 		 * we hit doorbell */
4172 		wmb();
4173 
4174 		mlx5_write64((__be32 *)ctrl, bf->bfreg->map + bf->offset,
4175 			     MLX5_GET_DOORBELL_LOCK(&bf->lock32));
4176 		/* Make sure doorbells don't leak out of SQ spinlock
4177 		 * and reach the HCA out of order.
4178 		 */
4179 		bf->offset ^= bf->buf_size;
4180 	}
4181 
4182 	spin_unlock_irqrestore(&qp->sq.lock, flags);
4183 
4184 	return err;
4185 }
4186 
4187 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
4188 {
4189 	sig->signature = calc_sig(sig, size);
4190 }
4191 
4192 int mlx5_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
4193 		      const struct ib_recv_wr **bad_wr)
4194 {
4195 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4196 	struct mlx5_wqe_data_seg *scat;
4197 	struct mlx5_rwqe_sig *sig;
4198 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4199 	struct mlx5_core_dev *mdev = dev->mdev;
4200 	unsigned long flags;
4201 	int err = 0;
4202 	int nreq;
4203 	int ind;
4204 	int i;
4205 
4206 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4207 		return mlx5_ib_gsi_post_recv(ibqp, wr, bad_wr);
4208 
4209 	spin_lock_irqsave(&qp->rq.lock, flags);
4210 
4211 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
4212 		err = -EIO;
4213 		*bad_wr = wr;
4214 		nreq = 0;
4215 		goto out;
4216 	}
4217 
4218 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
4219 
4220 	for (nreq = 0; wr; nreq++, wr = wr->next) {
4221 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
4222 			err = -ENOMEM;
4223 			*bad_wr = wr;
4224 			goto out;
4225 		}
4226 
4227 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
4228 			err = -EINVAL;
4229 			*bad_wr = wr;
4230 			goto out;
4231 		}
4232 
4233 		scat = get_recv_wqe(qp, ind);
4234 		if (qp->wq_sig)
4235 			scat++;
4236 
4237 		for (i = 0; i < wr->num_sge; i++)
4238 			set_data_ptr_seg(scat + i, wr->sg_list + i);
4239 
4240 		if (i < qp->rq.max_gs) {
4241 			scat[i].byte_count = 0;
4242 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
4243 			scat[i].addr       = 0;
4244 		}
4245 
4246 		if (qp->wq_sig) {
4247 			sig = (struct mlx5_rwqe_sig *)scat;
4248 			set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
4249 		}
4250 
4251 		qp->rq.wrid[ind] = wr->wr_id;
4252 
4253 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
4254 	}
4255 
4256 out:
4257 	if (likely(nreq)) {
4258 		qp->rq.head += nreq;
4259 
4260 		/* Make sure that descriptors are written before
4261 		 * doorbell record.
4262 		 */
4263 		wmb();
4264 
4265 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
4266 	}
4267 
4268 	spin_unlock_irqrestore(&qp->rq.lock, flags);
4269 
4270 	return err;
4271 }
4272 
4273 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
4274 {
4275 	switch (mlx5_state) {
4276 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
4277 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
4278 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
4279 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
4280 	case MLX5_QP_STATE_SQ_DRAINING:
4281 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
4282 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
4283 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
4284 	default:		     return -1;
4285 	}
4286 }
4287 
4288 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
4289 {
4290 	switch (mlx5_mig_state) {
4291 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
4292 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
4293 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
4294 	default: return -1;
4295 	}
4296 }
4297 
4298 static int to_ib_qp_access_flags(int mlx5_flags)
4299 {
4300 	int ib_flags = 0;
4301 
4302 	if (mlx5_flags & MLX5_QP_BIT_RRE)
4303 		ib_flags |= IB_ACCESS_REMOTE_READ;
4304 	if (mlx5_flags & MLX5_QP_BIT_RWE)
4305 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
4306 	if (mlx5_flags & MLX5_QP_BIT_RAE)
4307 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
4308 
4309 	return ib_flags;
4310 }
4311 
4312 static void to_ib_ah_attr(struct mlx5_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
4313 				struct mlx5_qp_path *path)
4314 {
4315 	struct mlx5_core_dev *dev = ibdev->mdev;
4316 
4317 	memset(ib_ah_attr, 0, sizeof(*ib_ah_attr));
4318 	ib_ah_attr->port_num	  = path->port;
4319 
4320 	if (ib_ah_attr->port_num == 0 ||
4321 	    ib_ah_attr->port_num > MLX5_CAP_GEN(dev, num_ports))
4322 		return;
4323 
4324 	ib_ah_attr->sl = path->dci_cfi_prio_sl & 0xf;
4325 
4326 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
4327 	ib_ah_attr->src_path_bits = path->grh_mlid & 0x7f;
4328 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
4329 	ib_ah_attr->ah_flags      = (path->grh_mlid & (1 << 7)) ? IB_AH_GRH : 0;
4330 	if (ib_ah_attr->ah_flags) {
4331 		ib_ah_attr->grh.sgid_index = path->mgid_index;
4332 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
4333 		ib_ah_attr->grh.traffic_class =
4334 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
4335 		ib_ah_attr->grh.flow_label =
4336 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
4337 		memcpy(ib_ah_attr->grh.dgid.raw,
4338 		       path->rgid, sizeof(ib_ah_attr->grh.dgid.raw));
4339 	}
4340 }
4341 
4342 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
4343 					struct mlx5_ib_sq *sq,
4344 					u8 *sq_state)
4345 {
4346 	void *out;
4347 	void *sqc;
4348 	int inlen;
4349 	int err;
4350 
4351 	inlen = MLX5_ST_SZ_BYTES(query_sq_out);
4352 	out = mlx5_vzalloc(inlen);
4353 	if (!out)
4354 		return -ENOMEM;
4355 
4356 	err = mlx5_core_query_sq(dev->mdev, sq->base.mqp.qpn, out);
4357 	if (err)
4358 		goto out;
4359 
4360 	sqc = MLX5_ADDR_OF(query_sq_out, out, sq_context);
4361 	*sq_state = MLX5_GET(sqc, sqc, state);
4362 	sq->state = *sq_state;
4363 
4364 out:
4365 	kvfree(out);
4366 	return err;
4367 }
4368 
4369 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
4370 					struct mlx5_ib_rq *rq,
4371 					u8 *rq_state)
4372 {
4373 	void *out;
4374 	void *rqc;
4375 	int inlen;
4376 	int err;
4377 
4378 	inlen = MLX5_ST_SZ_BYTES(query_rq_out);
4379 	out = mlx5_vzalloc(inlen);
4380 	if (!out)
4381 		return -ENOMEM;
4382 
4383 	err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
4384 	if (err)
4385 		goto out;
4386 
4387 	rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
4388 	*rq_state = MLX5_GET(rqc, rqc, state);
4389 	rq->state = *rq_state;
4390 
4391 out:
4392 	kvfree(out);
4393 	return err;
4394 }
4395 
4396 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
4397 				  struct mlx5_ib_qp *qp, u8 *qp_state)
4398 {
4399 	static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
4400 		[MLX5_RQC_STATE_RST] = {
4401 			[MLX5_SQC_STATE_RST]	= IB_QPS_RESET,
4402 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4403 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE_BAD,
4404 			[MLX5_SQ_STATE_NA]	= IB_QPS_RESET,
4405 		},
4406 		[MLX5_RQC_STATE_RDY] = {
4407 			[MLX5_SQC_STATE_RST]	= MLX5_QP_STATE_BAD,
4408 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4409 			[MLX5_SQC_STATE_ERR]	= IB_QPS_SQE,
4410 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE,
4411 		},
4412 		[MLX5_RQC_STATE_ERR] = {
4413 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE_BAD,
4414 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4415 			[MLX5_SQC_STATE_ERR]	= IB_QPS_ERR,
4416 			[MLX5_SQ_STATE_NA]	= IB_QPS_ERR,
4417 		},
4418 		[MLX5_RQ_STATE_NA] = {
4419 			[MLX5_SQC_STATE_RST]    = IB_QPS_RESET,
4420 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4421 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE,
4422 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE_BAD,
4423 		},
4424 	};
4425 
4426 	*qp_state = sqrq_trans[rq_state][sq_state];
4427 
4428 	if (*qp_state == MLX5_QP_STATE_BAD) {
4429 		WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
4430 		     qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
4431 		     qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
4432 		return -EINVAL;
4433 	}
4434 
4435 	if (*qp_state == MLX5_QP_STATE)
4436 		*qp_state = qp->state;
4437 
4438 	return 0;
4439 }
4440 
4441 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
4442 				     struct mlx5_ib_qp *qp,
4443 				     u8 *raw_packet_qp_state)
4444 {
4445 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
4446 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
4447 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
4448 	int err;
4449 	u8 sq_state = MLX5_SQ_STATE_NA;
4450 	u8 rq_state = MLX5_RQ_STATE_NA;
4451 
4452 	if (qp->sq.wqe_cnt) {
4453 		err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
4454 		if (err)
4455 			return err;
4456 	}
4457 
4458 	if (qp->rq.wqe_cnt) {
4459 		err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
4460 		if (err)
4461 			return err;
4462 	}
4463 
4464 	return sqrq_state_to_qp_state(sq_state, rq_state, qp,
4465 				      raw_packet_qp_state);
4466 }
4467 
4468 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
4469 			 struct ib_qp_attr *qp_attr)
4470 {
4471 	int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
4472 	struct mlx5_qp_context *context;
4473 	int mlx5_state;
4474 	u32 *outb;
4475 	int err = 0;
4476 
4477 	outb = kzalloc(outlen, GFP_KERNEL);
4478 	if (!outb)
4479 		return -ENOMEM;
4480 
4481 	err = mlx5_core_qp_query(dev->mdev, &qp->trans_qp.base.mqp, outb,
4482 				 outlen);
4483 	if (err)
4484 		goto out;
4485 
4486 	/* FIXME: use MLX5_GET rather than mlx5_qp_context manual struct */
4487 	context = (struct mlx5_qp_context *)MLX5_ADDR_OF(query_qp_out, outb, qpc);
4488 
4489 	mlx5_state = be32_to_cpu(context->flags) >> 28;
4490 
4491 	qp->state		     = to_ib_qp_state(mlx5_state);
4492 	qp_attr->path_mtu	     = context->mtu_msgmax >> 5;
4493 	qp_attr->path_mig_state	     =
4494 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
4495 	qp_attr->qkey		     = be32_to_cpu(context->qkey);
4496 	qp_attr->rq_psn		     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
4497 	qp_attr->sq_psn		     = be32_to_cpu(context->next_send_psn) & 0xffffff;
4498 	qp_attr->dest_qp_num	     = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
4499 	qp_attr->qp_access_flags     =
4500 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
4501 
4502 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4503 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
4504 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
4505 		qp_attr->alt_pkey_index =
4506 			be16_to_cpu(context->alt_path.pkey_index);
4507 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
4508 	}
4509 
4510 	qp_attr->pkey_index = be16_to_cpu(context->pri_path.pkey_index);
4511 	qp_attr->port_num = context->pri_path.port;
4512 
4513 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4514 	qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
4515 
4516 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
4517 
4518 	qp_attr->max_dest_rd_atomic =
4519 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
4520 	qp_attr->min_rnr_timer	    =
4521 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
4522 	qp_attr->timeout	    = context->pri_path.ackto_lt >> 3;
4523 	qp_attr->retry_cnt	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
4524 	qp_attr->rnr_retry	    = (be32_to_cpu(context->params1) >> 13) & 0x7;
4525 	qp_attr->alt_timeout	    = context->alt_path.ackto_lt >> 3;
4526 
4527 out:
4528 	kfree(outb);
4529 	return err;
4530 }
4531 
4532 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
4533 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
4534 {
4535 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4536 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4537 	int err = 0;
4538 	u8 raw_packet_qp_state;
4539 
4540 	if (ibqp->rwq_ind_tbl)
4541 		return -ENOSYS;
4542 
4543 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4544 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
4545 					    qp_init_attr);
4546 
4547 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4548 	/*
4549 	 * Wait for any outstanding page faults, in case the user frees memory
4550 	 * based upon this query's result.
4551 	 */
4552 	flush_workqueue(mlx5_ib_page_fault_wq);
4553 #endif
4554 
4555 	mutex_lock(&qp->mutex);
4556 
4557 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
4558 		err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
4559 		if (err)
4560 			goto out;
4561 		qp->state = raw_packet_qp_state;
4562 		qp_attr->port_num = 1;
4563 	} else {
4564 		err = query_qp_attr(dev, qp, qp_attr);
4565 		if (err)
4566 			goto out;
4567 	}
4568 
4569 	qp_attr->qp_state	     = qp->state;
4570 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4571 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4572 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4573 
4574 	if (!ibqp->uobject) {
4575 		qp_attr->cap.max_send_wr  = qp->sq.max_post;
4576 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4577 		qp_init_attr->qp_context = ibqp->qp_context;
4578 	} else {
4579 		qp_attr->cap.max_send_wr  = 0;
4580 		qp_attr->cap.max_send_sge = 0;
4581 	}
4582 
4583 	qp_init_attr->qp_type = ibqp->qp_type;
4584 	qp_init_attr->recv_cq = ibqp->recv_cq;
4585 	qp_init_attr->send_cq = ibqp->send_cq;
4586 	qp_init_attr->srq = ibqp->srq;
4587 	qp_attr->cap.max_inline_data = qp->max_inline_data;
4588 
4589 	qp_init_attr->cap	     = qp_attr->cap;
4590 
4591 	qp_init_attr->create_flags = 0;
4592 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4593 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4594 
4595 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
4596 		qp_init_attr->create_flags |= IB_QP_CREATE_CROSS_CHANNEL;
4597 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
4598 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_SEND;
4599 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
4600 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_RECV;
4601 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
4602 		qp_init_attr->create_flags |= MLX5_IB_QP_CREATE_SQPN_QP1;
4603 
4604 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
4605 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4606 
4607 out:
4608 	mutex_unlock(&qp->mutex);
4609 	return err;
4610 }
4611 
4612 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
4613 					  struct ib_ucontext *context,
4614 					  struct ib_udata *udata)
4615 {
4616 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4617 	struct mlx5_ib_xrcd *xrcd;
4618 	int err;
4619 
4620 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
4621 		return ERR_PTR(-ENOSYS);
4622 
4623 	xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
4624 	if (!xrcd)
4625 		return ERR_PTR(-ENOMEM);
4626 
4627 	err = mlx5_core_xrcd_alloc(dev->mdev, &xrcd->xrcdn);
4628 	if (err) {
4629 		kfree(xrcd);
4630 		return ERR_PTR(-ENOMEM);
4631 	}
4632 
4633 	return &xrcd->ibxrcd;
4634 }
4635 
4636 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
4637 {
4638 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
4639 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
4640 	int err;
4641 
4642 	err = mlx5_core_xrcd_dealloc(dev->mdev, xrcdn);
4643 	if (err) {
4644 		mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
4645 		return err;
4646 	}
4647 
4648 	kfree(xrcd);
4649 
4650 	return 0;
4651 }
4652 
4653 static void mlx5_ib_wq_event(struct mlx5_core_qp *core_qp, int type)
4654 {
4655 	struct mlx5_ib_rwq *rwq = to_mibrwq(core_qp);
4656 	struct mlx5_ib_dev *dev = to_mdev(rwq->ibwq.device);
4657 	struct ib_event event;
4658 
4659 	if (rwq->ibwq.event_handler) {
4660 		event.device     = rwq->ibwq.device;
4661 		event.element.wq = &rwq->ibwq;
4662 		switch (type) {
4663 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
4664 			event.event = IB_EVENT_WQ_FATAL;
4665 			break;
4666 		default:
4667 			mlx5_ib_warn(dev, "Unexpected event type %d on WQ %06x\n", type, core_qp->qpn);
4668 			return;
4669 		}
4670 
4671 		rwq->ibwq.event_handler(&event, rwq->ibwq.wq_context);
4672 	}
4673 }
4674 
4675 static int  create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
4676 		      struct ib_wq_init_attr *init_attr)
4677 {
4678 	struct mlx5_ib_dev *dev;
4679 	__be64 *rq_pas0;
4680 	void *in;
4681 	void *rqc;
4682 	void *wq;
4683 	int inlen;
4684 	int err;
4685 
4686 	dev = to_mdev(pd->device);
4687 
4688 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
4689 	in = mlx5_vzalloc(inlen);
4690 	if (!in)
4691 		return -ENOMEM;
4692 
4693 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
4694 	MLX5_SET(rqc,  rqc, mem_rq_type,
4695 		 MLX5_RQC_RQ_TYPE_MEMORY_RQ_INLINE);
4696 	MLX5_SET(rqc, rqc, user_index, rwq->user_index);
4697 	MLX5_SET(rqc,  rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
4698 	MLX5_SET(rqc,  rqc, state, MLX5_RQC_STATE_RST);
4699 	MLX5_SET(rqc,  rqc, flush_in_error_en, 1);
4700 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
4701 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
4702 	MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
4703 	MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
4704 	MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
4705 	MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
4706 	MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
4707 	MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
4708 	MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
4709 	MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
4710 	rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
4711 	mlx5_ib_populate_pas(dev, rwq->umem, rwq->page_shift, rq_pas0, 0);
4712 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rwq->core_qp);
4713 	kvfree(in);
4714 	return err;
4715 }
4716 
4717 static int set_user_rq_size(struct mlx5_ib_dev *dev,
4718 			    struct ib_wq_init_attr *wq_init_attr,
4719 			    struct mlx5_ib_create_wq *ucmd,
4720 			    struct mlx5_ib_rwq *rwq)
4721 {
4722 	/* Sanity check RQ size before proceeding */
4723 	if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
4724 		return -EINVAL;
4725 
4726 	if (!ucmd->rq_wqe_count)
4727 		return -EINVAL;
4728 
4729 	rwq->wqe_count = ucmd->rq_wqe_count;
4730 	rwq->wqe_shift = ucmd->rq_wqe_shift;
4731 	rwq->buf_size = (rwq->wqe_count << rwq->wqe_shift);
4732 	rwq->log_rq_stride = rwq->wqe_shift;
4733 	rwq->log_rq_size = ilog2(rwq->wqe_count);
4734 	return 0;
4735 }
4736 
4737 static int prepare_user_rq(struct ib_pd *pd,
4738 			   struct ib_wq_init_attr *init_attr,
4739 			   struct ib_udata *udata,
4740 			   struct mlx5_ib_rwq *rwq)
4741 {
4742 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
4743 	struct mlx5_ib_create_wq ucmd = {};
4744 	int err;
4745 	size_t required_cmd_sz;
4746 
4747 	required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
4748 	if (udata->inlen < required_cmd_sz) {
4749 		mlx5_ib_dbg(dev, "invalid inlen\n");
4750 		return -EINVAL;
4751 	}
4752 
4753 	if (udata->inlen > sizeof(ucmd) &&
4754 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4755 				 udata->inlen - sizeof(ucmd))) {
4756 		mlx5_ib_dbg(dev, "inlen is not supported\n");
4757 		return -EOPNOTSUPP;
4758 	}
4759 
4760 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
4761 		mlx5_ib_dbg(dev, "copy failed\n");
4762 		return -EFAULT;
4763 	}
4764 
4765 	if (ucmd.comp_mask) {
4766 		mlx5_ib_dbg(dev, "invalid comp mask\n");
4767 		return -EOPNOTSUPP;
4768 	}
4769 
4770 	if (ucmd.reserved) {
4771 		mlx5_ib_dbg(dev, "invalid reserved\n");
4772 		return -EOPNOTSUPP;
4773 	}
4774 
4775 	err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
4776 	if (err) {
4777 		mlx5_ib_dbg(dev, "err %d\n", err);
4778 		return err;
4779 	}
4780 
4781 	err = create_user_rq(dev, pd, rwq, &ucmd);
4782 	if (err) {
4783 		mlx5_ib_dbg(dev, "err %d\n", err);
4784 		if (err)
4785 			return err;
4786 	}
4787 
4788 	rwq->user_index = ucmd.user_index;
4789 	return 0;
4790 }
4791 
4792 struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
4793 				struct ib_wq_init_attr *init_attr,
4794 				struct ib_udata *udata)
4795 {
4796 	struct mlx5_ib_dev *dev;
4797 	struct mlx5_ib_rwq *rwq;
4798 	struct mlx5_ib_create_wq_resp resp = {};
4799 	size_t min_resp_len;
4800 	int err;
4801 
4802 	if (!udata)
4803 		return ERR_PTR(-ENOSYS);
4804 
4805 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4806 	if (udata->outlen && udata->outlen < min_resp_len)
4807 		return ERR_PTR(-EINVAL);
4808 
4809 	dev = to_mdev(pd->device);
4810 	switch (init_attr->wq_type) {
4811 	case IB_WQT_RQ:
4812 		rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
4813 		if (!rwq)
4814 			return ERR_PTR(-ENOMEM);
4815 		err = prepare_user_rq(pd, init_attr, udata, rwq);
4816 		if (err)
4817 			goto err;
4818 		err = create_rq(rwq, pd, init_attr);
4819 		if (err)
4820 			goto err_user_rq;
4821 		break;
4822 	default:
4823 		mlx5_ib_dbg(dev, "unsupported wq type %d\n",
4824 			    init_attr->wq_type);
4825 		return ERR_PTR(-EINVAL);
4826 	}
4827 
4828 	rwq->ibwq.wq_num = rwq->core_qp.qpn;
4829 	rwq->ibwq.state = IB_WQS_RESET;
4830 	if (udata->outlen) {
4831 		resp.response_length = offsetof(typeof(resp), response_length) +
4832 				sizeof(resp.response_length);
4833 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4834 		if (err)
4835 			goto err_copy;
4836 	}
4837 
4838 	rwq->core_qp.event = mlx5_ib_wq_event;
4839 	rwq->ibwq.event_handler = init_attr->event_handler;
4840 	return &rwq->ibwq;
4841 
4842 err_copy:
4843 	mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
4844 err_user_rq:
4845 	destroy_user_rq(pd, rwq);
4846 err:
4847 	kfree(rwq);
4848 	return ERR_PTR(err);
4849 }
4850 
4851 int mlx5_ib_destroy_wq(struct ib_wq *wq)
4852 {
4853 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
4854 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
4855 
4856 	mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
4857 	destroy_user_rq(wq->pd, rwq);
4858 	kfree(rwq);
4859 
4860 	return 0;
4861 }
4862 
4863 struct ib_rwq_ind_table *mlx5_ib_create_rwq_ind_table(struct ib_device *device,
4864 						      struct ib_rwq_ind_table_init_attr *init_attr,
4865 						      struct ib_udata *udata)
4866 {
4867 	struct mlx5_ib_dev *dev = to_mdev(device);
4868 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl;
4869 	int sz = 1 << init_attr->log_ind_tbl_size;
4870 	struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
4871 	size_t min_resp_len;
4872 	int inlen;
4873 	int err;
4874 	int i;
4875 	u32 *in;
4876 	void *rqtc;
4877 
4878 	if (udata->inlen > 0 &&
4879 	    !ib_is_udata_cleared(udata, 0,
4880 				 udata->inlen))
4881 		return ERR_PTR(-EOPNOTSUPP);
4882 
4883 	if (init_attr->log_ind_tbl_size >
4884 	    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size)) {
4885 		mlx5_ib_dbg(dev, "log_ind_tbl_size = %d is bigger than supported = %d\n",
4886 			    init_attr->log_ind_tbl_size,
4887 			    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size));
4888 		return ERR_PTR(-EINVAL);
4889 	}
4890 
4891 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4892 	if (udata->outlen && udata->outlen < min_resp_len)
4893 		return ERR_PTR(-EINVAL);
4894 
4895 	rwq_ind_tbl = kzalloc(sizeof(*rwq_ind_tbl), GFP_KERNEL);
4896 	if (!rwq_ind_tbl)
4897 		return ERR_PTR(-ENOMEM);
4898 
4899 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
4900 	in = mlx5_vzalloc(inlen);
4901 	if (!in) {
4902 		err = -ENOMEM;
4903 		goto err;
4904 	}
4905 
4906 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
4907 
4908 	MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
4909 	MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
4910 
4911 	for (i = 0; i < sz; i++)
4912 		MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
4913 
4914 	err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
4915 	kvfree(in);
4916 
4917 	if (err)
4918 		goto err;
4919 
4920 	rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
4921 	if (udata->outlen) {
4922 		resp.response_length = offsetof(typeof(resp), response_length) +
4923 					sizeof(resp.response_length);
4924 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4925 		if (err)
4926 			goto err_copy;
4927 	}
4928 
4929 	return &rwq_ind_tbl->ib_rwq_ind_tbl;
4930 
4931 err_copy:
4932 	mlx5_core_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn);
4933 err:
4934 	kfree(rwq_ind_tbl);
4935 	return ERR_PTR(err);
4936 }
4937 
4938 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4939 {
4940 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
4941 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
4942 
4943 	mlx5_core_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn);
4944 
4945 	kfree(rwq_ind_tbl);
4946 	return 0;
4947 }
4948 
4949 int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
4950 		      u32 wq_attr_mask, struct ib_udata *udata)
4951 {
4952 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
4953 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
4954 	struct mlx5_ib_modify_wq ucmd = {};
4955 	size_t required_cmd_sz;
4956 	int curr_wq_state;
4957 	int wq_state;
4958 	int inlen;
4959 	int err;
4960 	void *rqc;
4961 	void *in;
4962 
4963 	required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
4964 	if (udata->inlen < required_cmd_sz)
4965 		return -EINVAL;
4966 
4967 	if (udata->inlen > sizeof(ucmd) &&
4968 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4969 				 udata->inlen - sizeof(ucmd)))
4970 		return -EOPNOTSUPP;
4971 
4972 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4973 		return -EFAULT;
4974 
4975 	if (ucmd.comp_mask || ucmd.reserved)
4976 		return -EOPNOTSUPP;
4977 
4978 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
4979 	in = mlx5_vzalloc(inlen);
4980 	if (!in)
4981 		return -ENOMEM;
4982 
4983 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
4984 
4985 	MLX5_SET(modify_rq_in, in, rqn, rwq->core_qp.qpn);
4986 	curr_wq_state = (wq_attr_mask & IB_WQ_CUR_STATE) ?
4987 		wq_attr->curr_wq_state : wq->state;
4988 	wq_state = (wq_attr_mask & IB_WQ_STATE) ?
4989 		wq_attr->wq_state : curr_wq_state;
4990 	if (curr_wq_state == IB_WQS_ERR)
4991 		curr_wq_state = MLX5_RQC_STATE_ERR;
4992 	if (wq_state == IB_WQS_ERR)
4993 		wq_state = MLX5_RQC_STATE_ERR;
4994 	MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
4995 	MLX5_SET(rqc, rqc, state, wq_state);
4996 
4997 	err = mlx5_core_modify_rq(dev->mdev, in, inlen);
4998 	kvfree(in);
4999 	if (!err)
5000 		rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;
5001 
5002 	return err;
5003 }
5004