xref: /openbmc/linux/drivers/infiniband/hw/mlx5/qp.c (revision de2bdb3d)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/module.h>
34 #include <rdma/ib_umem.h>
35 #include <rdma/ib_cache.h>
36 #include <rdma/ib_user_verbs.h>
37 #include "mlx5_ib.h"
38 
39 /* not supported currently */
40 static int wq_signature;
41 
42 enum {
43 	MLX5_IB_ACK_REQ_FREQ	= 8,
44 };
45 
46 enum {
47 	MLX5_IB_DEFAULT_SCHED_QUEUE	= 0x83,
48 	MLX5_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
49 	MLX5_IB_LINK_TYPE_IB		= 0,
50 	MLX5_IB_LINK_TYPE_ETH		= 1
51 };
52 
53 enum {
54 	MLX5_IB_SQ_STRIDE	= 6,
55 };
56 
57 static const u32 mlx5_ib_opcode[] = {
58 	[IB_WR_SEND]				= MLX5_OPCODE_SEND,
59 	[IB_WR_LSO]				= MLX5_OPCODE_LSO,
60 	[IB_WR_SEND_WITH_IMM]			= MLX5_OPCODE_SEND_IMM,
61 	[IB_WR_RDMA_WRITE]			= MLX5_OPCODE_RDMA_WRITE,
62 	[IB_WR_RDMA_WRITE_WITH_IMM]		= MLX5_OPCODE_RDMA_WRITE_IMM,
63 	[IB_WR_RDMA_READ]			= MLX5_OPCODE_RDMA_READ,
64 	[IB_WR_ATOMIC_CMP_AND_SWP]		= MLX5_OPCODE_ATOMIC_CS,
65 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= MLX5_OPCODE_ATOMIC_FA,
66 	[IB_WR_SEND_WITH_INV]			= MLX5_OPCODE_SEND_INVAL,
67 	[IB_WR_LOCAL_INV]			= MLX5_OPCODE_UMR,
68 	[IB_WR_REG_MR]				= MLX5_OPCODE_UMR,
69 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= MLX5_OPCODE_ATOMIC_MASKED_CS,
70 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= MLX5_OPCODE_ATOMIC_MASKED_FA,
71 	[MLX5_IB_WR_UMR]			= MLX5_OPCODE_UMR,
72 };
73 
74 struct mlx5_wqe_eth_pad {
75 	u8 rsvd0[16];
76 };
77 
78 enum raw_qp_set_mask_map {
79 	MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID		= 1UL << 0,
80 };
81 
82 struct mlx5_modify_raw_qp_param {
83 	u16 operation;
84 
85 	u32 set_mask; /* raw_qp_set_mask_map */
86 	u8 rq_q_ctr_id;
87 };
88 
89 static void get_cqs(enum ib_qp_type qp_type,
90 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
91 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq);
92 
93 static int is_qp0(enum ib_qp_type qp_type)
94 {
95 	return qp_type == IB_QPT_SMI;
96 }
97 
98 static int is_sqp(enum ib_qp_type qp_type)
99 {
100 	return is_qp0(qp_type) || is_qp1(qp_type);
101 }
102 
103 static void *get_wqe(struct mlx5_ib_qp *qp, int offset)
104 {
105 	return mlx5_buf_offset(&qp->buf, offset);
106 }
107 
108 static void *get_recv_wqe(struct mlx5_ib_qp *qp, int n)
109 {
110 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
111 }
112 
113 void *mlx5_get_send_wqe(struct mlx5_ib_qp *qp, int n)
114 {
115 	return get_wqe(qp, qp->sq.offset + (n << MLX5_IB_SQ_STRIDE));
116 }
117 
118 /**
119  * mlx5_ib_read_user_wqe() - Copy a user-space WQE to kernel space.
120  *
121  * @qp: QP to copy from.
122  * @send: copy from the send queue when non-zero, use the receive queue
123  *	  otherwise.
124  * @wqe_index:  index to start copying from. For send work queues, the
125  *		wqe_index is in units of MLX5_SEND_WQE_BB.
126  *		For receive work queue, it is the number of work queue
127  *		element in the queue.
128  * @buffer: destination buffer.
129  * @length: maximum number of bytes to copy.
130  *
131  * Copies at least a single WQE, but may copy more data.
132  *
133  * Return: the number of bytes copied, or an error code.
134  */
135 int mlx5_ib_read_user_wqe(struct mlx5_ib_qp *qp, int send, int wqe_index,
136 			  void *buffer, u32 length,
137 			  struct mlx5_ib_qp_base *base)
138 {
139 	struct ib_device *ibdev = qp->ibqp.device;
140 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
141 	struct mlx5_ib_wq *wq = send ? &qp->sq : &qp->rq;
142 	size_t offset;
143 	size_t wq_end;
144 	struct ib_umem *umem = base->ubuffer.umem;
145 	u32 first_copy_length;
146 	int wqe_length;
147 	int ret;
148 
149 	if (wq->wqe_cnt == 0) {
150 		mlx5_ib_dbg(dev, "mlx5_ib_read_user_wqe for a QP with wqe_cnt == 0. qp_type: 0x%x\n",
151 			    qp->ibqp.qp_type);
152 		return -EINVAL;
153 	}
154 
155 	offset = wq->offset + ((wqe_index % wq->wqe_cnt) << wq->wqe_shift);
156 	wq_end = wq->offset + (wq->wqe_cnt << wq->wqe_shift);
157 
158 	if (send && length < sizeof(struct mlx5_wqe_ctrl_seg))
159 		return -EINVAL;
160 
161 	if (offset > umem->length ||
162 	    (send && offset + sizeof(struct mlx5_wqe_ctrl_seg) > umem->length))
163 		return -EINVAL;
164 
165 	first_copy_length = min_t(u32, offset + length, wq_end) - offset;
166 	ret = ib_umem_copy_from(buffer, umem, offset, first_copy_length);
167 	if (ret)
168 		return ret;
169 
170 	if (send) {
171 		struct mlx5_wqe_ctrl_seg *ctrl = buffer;
172 		int ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
173 
174 		wqe_length = ds * MLX5_WQE_DS_UNITS;
175 	} else {
176 		wqe_length = 1 << wq->wqe_shift;
177 	}
178 
179 	if (wqe_length <= first_copy_length)
180 		return first_copy_length;
181 
182 	ret = ib_umem_copy_from(buffer + first_copy_length, umem, wq->offset,
183 				wqe_length - first_copy_length);
184 	if (ret)
185 		return ret;
186 
187 	return wqe_length;
188 }
189 
190 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
191 {
192 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
193 	struct ib_event event;
194 
195 	if (type == MLX5_EVENT_TYPE_PATH_MIG) {
196 		/* This event is only valid for trans_qps */
197 		to_mibqp(qp)->port = to_mibqp(qp)->trans_qp.alt_port;
198 	}
199 
200 	if (ibqp->event_handler) {
201 		event.device     = ibqp->device;
202 		event.element.qp = ibqp;
203 		switch (type) {
204 		case MLX5_EVENT_TYPE_PATH_MIG:
205 			event.event = IB_EVENT_PATH_MIG;
206 			break;
207 		case MLX5_EVENT_TYPE_COMM_EST:
208 			event.event = IB_EVENT_COMM_EST;
209 			break;
210 		case MLX5_EVENT_TYPE_SQ_DRAINED:
211 			event.event = IB_EVENT_SQ_DRAINED;
212 			break;
213 		case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
214 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
215 			break;
216 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
217 			event.event = IB_EVENT_QP_FATAL;
218 			break;
219 		case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
220 			event.event = IB_EVENT_PATH_MIG_ERR;
221 			break;
222 		case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
223 			event.event = IB_EVENT_QP_REQ_ERR;
224 			break;
225 		case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
226 			event.event = IB_EVENT_QP_ACCESS_ERR;
227 			break;
228 		default:
229 			pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
230 			return;
231 		}
232 
233 		ibqp->event_handler(&event, ibqp->qp_context);
234 	}
235 }
236 
237 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
238 		       int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
239 {
240 	int wqe_size;
241 	int wq_size;
242 
243 	/* Sanity check RQ size before proceeding */
244 	if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
245 		return -EINVAL;
246 
247 	if (!has_rq) {
248 		qp->rq.max_gs = 0;
249 		qp->rq.wqe_cnt = 0;
250 		qp->rq.wqe_shift = 0;
251 		cap->max_recv_wr = 0;
252 		cap->max_recv_sge = 0;
253 	} else {
254 		if (ucmd) {
255 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
256 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
257 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
258 			qp->rq.max_post = qp->rq.wqe_cnt;
259 		} else {
260 			wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
261 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
262 			wqe_size = roundup_pow_of_two(wqe_size);
263 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
264 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
265 			qp->rq.wqe_cnt = wq_size / wqe_size;
266 			if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
267 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
268 					    wqe_size,
269 					    MLX5_CAP_GEN(dev->mdev,
270 							 max_wqe_sz_rq));
271 				return -EINVAL;
272 			}
273 			qp->rq.wqe_shift = ilog2(wqe_size);
274 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
275 			qp->rq.max_post = qp->rq.wqe_cnt;
276 		}
277 	}
278 
279 	return 0;
280 }
281 
282 static int sq_overhead(struct ib_qp_init_attr *attr)
283 {
284 	int size = 0;
285 
286 	switch (attr->qp_type) {
287 	case IB_QPT_XRC_INI:
288 		size += sizeof(struct mlx5_wqe_xrc_seg);
289 		/* fall through */
290 	case IB_QPT_RC:
291 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
292 			max(sizeof(struct mlx5_wqe_atomic_seg) +
293 			    sizeof(struct mlx5_wqe_raddr_seg),
294 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
295 			    sizeof(struct mlx5_mkey_seg));
296 		break;
297 
298 	case IB_QPT_XRC_TGT:
299 		return 0;
300 
301 	case IB_QPT_UC:
302 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
303 			max(sizeof(struct mlx5_wqe_raddr_seg),
304 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
305 			    sizeof(struct mlx5_mkey_seg));
306 		break;
307 
308 	case IB_QPT_UD:
309 		if (attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
310 			size += sizeof(struct mlx5_wqe_eth_pad) +
311 				sizeof(struct mlx5_wqe_eth_seg);
312 		/* fall through */
313 	case IB_QPT_SMI:
314 	case MLX5_IB_QPT_HW_GSI:
315 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
316 			sizeof(struct mlx5_wqe_datagram_seg);
317 		break;
318 
319 	case MLX5_IB_QPT_REG_UMR:
320 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
321 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
322 			sizeof(struct mlx5_mkey_seg);
323 		break;
324 
325 	default:
326 		return -EINVAL;
327 	}
328 
329 	return size;
330 }
331 
332 static int calc_send_wqe(struct ib_qp_init_attr *attr)
333 {
334 	int inl_size = 0;
335 	int size;
336 
337 	size = sq_overhead(attr);
338 	if (size < 0)
339 		return size;
340 
341 	if (attr->cap.max_inline_data) {
342 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
343 			attr->cap.max_inline_data;
344 	}
345 
346 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
347 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN &&
348 	    ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
349 			return MLX5_SIG_WQE_SIZE;
350 	else
351 		return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
352 }
353 
354 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
355 			struct mlx5_ib_qp *qp)
356 {
357 	int wqe_size;
358 	int wq_size;
359 
360 	if (!attr->cap.max_send_wr)
361 		return 0;
362 
363 	wqe_size = calc_send_wqe(attr);
364 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
365 	if (wqe_size < 0)
366 		return wqe_size;
367 
368 	if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
369 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
370 			    wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
371 		return -EINVAL;
372 	}
373 
374 	qp->max_inline_data = wqe_size - sq_overhead(attr) -
375 			      sizeof(struct mlx5_wqe_inline_seg);
376 	attr->cap.max_inline_data = qp->max_inline_data;
377 
378 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN)
379 		qp->signature_en = true;
380 
381 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
382 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
383 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
384 		mlx5_ib_dbg(dev, "wqe count(%d) exceeds limits(%d)\n",
385 			    qp->sq.wqe_cnt,
386 			    1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
387 		return -ENOMEM;
388 	}
389 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
390 	qp->sq.max_gs = attr->cap.max_send_sge;
391 	qp->sq.max_post = wq_size / wqe_size;
392 	attr->cap.max_send_wr = qp->sq.max_post;
393 
394 	return wq_size;
395 }
396 
397 static int set_user_buf_size(struct mlx5_ib_dev *dev,
398 			    struct mlx5_ib_qp *qp,
399 			    struct mlx5_ib_create_qp *ucmd,
400 			    struct mlx5_ib_qp_base *base,
401 			    struct ib_qp_init_attr *attr)
402 {
403 	int desc_sz = 1 << qp->sq.wqe_shift;
404 
405 	if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
406 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
407 			     desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
408 		return -EINVAL;
409 	}
410 
411 	if (ucmd->sq_wqe_count && ((1 << ilog2(ucmd->sq_wqe_count)) != ucmd->sq_wqe_count)) {
412 		mlx5_ib_warn(dev, "sq_wqe_count %d, sq_wqe_count %d\n",
413 			     ucmd->sq_wqe_count, ucmd->sq_wqe_count);
414 		return -EINVAL;
415 	}
416 
417 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
418 
419 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
420 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
421 			     qp->sq.wqe_cnt,
422 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
423 		return -EINVAL;
424 	}
425 
426 	if (attr->qp_type == IB_QPT_RAW_PACKET) {
427 		base->ubuffer.buf_size = qp->rq.wqe_cnt << qp->rq.wqe_shift;
428 		qp->raw_packet_qp.sq.ubuffer.buf_size = qp->sq.wqe_cnt << 6;
429 	} else {
430 		base->ubuffer.buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
431 					 (qp->sq.wqe_cnt << 6);
432 	}
433 
434 	return 0;
435 }
436 
437 static int qp_has_rq(struct ib_qp_init_attr *attr)
438 {
439 	if (attr->qp_type == IB_QPT_XRC_INI ||
440 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
441 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
442 	    !attr->cap.max_recv_wr)
443 		return 0;
444 
445 	return 1;
446 }
447 
448 static int first_med_uuar(void)
449 {
450 	return 1;
451 }
452 
453 static int next_uuar(int n)
454 {
455 	n++;
456 
457 	while (((n % 4) & 2))
458 		n++;
459 
460 	return n;
461 }
462 
463 static int num_med_uuar(struct mlx5_uuar_info *uuari)
464 {
465 	int n;
466 
467 	n = uuari->num_uars * MLX5_NON_FP_BF_REGS_PER_PAGE -
468 		uuari->num_low_latency_uuars - 1;
469 
470 	return n >= 0 ? n : 0;
471 }
472 
473 static int max_uuari(struct mlx5_uuar_info *uuari)
474 {
475 	return uuari->num_uars * 4;
476 }
477 
478 static int first_hi_uuar(struct mlx5_uuar_info *uuari)
479 {
480 	int med;
481 	int i;
482 	int t;
483 
484 	med = num_med_uuar(uuari);
485 	for (t = 0, i = first_med_uuar();; i = next_uuar(i)) {
486 		t++;
487 		if (t == med)
488 			return next_uuar(i);
489 	}
490 
491 	return 0;
492 }
493 
494 static int alloc_high_class_uuar(struct mlx5_uuar_info *uuari)
495 {
496 	int i;
497 
498 	for (i = first_hi_uuar(uuari); i < max_uuari(uuari); i = next_uuar(i)) {
499 		if (!test_bit(i, uuari->bitmap)) {
500 			set_bit(i, uuari->bitmap);
501 			uuari->count[i]++;
502 			return i;
503 		}
504 	}
505 
506 	return -ENOMEM;
507 }
508 
509 static int alloc_med_class_uuar(struct mlx5_uuar_info *uuari)
510 {
511 	int minidx = first_med_uuar();
512 	int i;
513 
514 	for (i = first_med_uuar(); i < first_hi_uuar(uuari); i = next_uuar(i)) {
515 		if (uuari->count[i] < uuari->count[minidx])
516 			minidx = i;
517 	}
518 
519 	uuari->count[minidx]++;
520 	return minidx;
521 }
522 
523 static int alloc_uuar(struct mlx5_uuar_info *uuari,
524 		      enum mlx5_ib_latency_class lat)
525 {
526 	int uuarn = -EINVAL;
527 
528 	mutex_lock(&uuari->lock);
529 	switch (lat) {
530 	case MLX5_IB_LATENCY_CLASS_LOW:
531 		uuarn = 0;
532 		uuari->count[uuarn]++;
533 		break;
534 
535 	case MLX5_IB_LATENCY_CLASS_MEDIUM:
536 		if (uuari->ver < 2)
537 			uuarn = -ENOMEM;
538 		else
539 			uuarn = alloc_med_class_uuar(uuari);
540 		break;
541 
542 	case MLX5_IB_LATENCY_CLASS_HIGH:
543 		if (uuari->ver < 2)
544 			uuarn = -ENOMEM;
545 		else
546 			uuarn = alloc_high_class_uuar(uuari);
547 		break;
548 
549 	case MLX5_IB_LATENCY_CLASS_FAST_PATH:
550 		uuarn = 2;
551 		break;
552 	}
553 	mutex_unlock(&uuari->lock);
554 
555 	return uuarn;
556 }
557 
558 static void free_med_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
559 {
560 	clear_bit(uuarn, uuari->bitmap);
561 	--uuari->count[uuarn];
562 }
563 
564 static void free_high_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
565 {
566 	clear_bit(uuarn, uuari->bitmap);
567 	--uuari->count[uuarn];
568 }
569 
570 static void free_uuar(struct mlx5_uuar_info *uuari, int uuarn)
571 {
572 	int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
573 	int high_uuar = nuuars - uuari->num_low_latency_uuars;
574 
575 	mutex_lock(&uuari->lock);
576 	if (uuarn == 0) {
577 		--uuari->count[uuarn];
578 		goto out;
579 	}
580 
581 	if (uuarn < high_uuar) {
582 		free_med_class_uuar(uuari, uuarn);
583 		goto out;
584 	}
585 
586 	free_high_class_uuar(uuari, uuarn);
587 
588 out:
589 	mutex_unlock(&uuari->lock);
590 }
591 
592 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
593 {
594 	switch (state) {
595 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
596 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
597 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
598 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
599 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
600 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
601 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
602 	default:		return -1;
603 	}
604 }
605 
606 static int to_mlx5_st(enum ib_qp_type type)
607 {
608 	switch (type) {
609 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
610 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
611 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
612 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
613 	case IB_QPT_XRC_INI:
614 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
615 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
616 	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
617 	case IB_QPT_RAW_IPV6:		return MLX5_QP_ST_RAW_IPV6;
618 	case IB_QPT_RAW_PACKET:
619 	case IB_QPT_RAW_ETHERTYPE:	return MLX5_QP_ST_RAW_ETHERTYPE;
620 	case IB_QPT_MAX:
621 	default:		return -EINVAL;
622 	}
623 }
624 
625 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
626 			     struct mlx5_ib_cq *recv_cq);
627 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
628 			       struct mlx5_ib_cq *recv_cq);
629 
630 static int uuarn_to_uar_index(struct mlx5_uuar_info *uuari, int uuarn)
631 {
632 	return uuari->uars[uuarn / MLX5_BF_REGS_PER_PAGE].index;
633 }
634 
635 static int mlx5_ib_umem_get(struct mlx5_ib_dev *dev,
636 			    struct ib_pd *pd,
637 			    unsigned long addr, size_t size,
638 			    struct ib_umem **umem,
639 			    int *npages, int *page_shift, int *ncont,
640 			    u32 *offset)
641 {
642 	int err;
643 
644 	*umem = ib_umem_get(pd->uobject->context, addr, size, 0, 0);
645 	if (IS_ERR(*umem)) {
646 		mlx5_ib_dbg(dev, "umem_get failed\n");
647 		return PTR_ERR(*umem);
648 	}
649 
650 	mlx5_ib_cont_pages(*umem, addr, npages, page_shift, ncont, NULL);
651 
652 	err = mlx5_ib_get_buf_offset(addr, *page_shift, offset);
653 	if (err) {
654 		mlx5_ib_warn(dev, "bad offset\n");
655 		goto err_umem;
656 	}
657 
658 	mlx5_ib_dbg(dev, "addr 0x%lx, size %zu, npages %d, page_shift %d, ncont %d, offset %d\n",
659 		    addr, size, *npages, *page_shift, *ncont, *offset);
660 
661 	return 0;
662 
663 err_umem:
664 	ib_umem_release(*umem);
665 	*umem = NULL;
666 
667 	return err;
668 }
669 
670 static void destroy_user_rq(struct ib_pd *pd, struct mlx5_ib_rwq *rwq)
671 {
672 	struct mlx5_ib_ucontext *context;
673 
674 	context = to_mucontext(pd->uobject->context);
675 	mlx5_ib_db_unmap_user(context, &rwq->db);
676 	if (rwq->umem)
677 		ib_umem_release(rwq->umem);
678 }
679 
680 static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
681 			  struct mlx5_ib_rwq *rwq,
682 			  struct mlx5_ib_create_wq *ucmd)
683 {
684 	struct mlx5_ib_ucontext *context;
685 	int page_shift = 0;
686 	int npages;
687 	u32 offset = 0;
688 	int ncont = 0;
689 	int err;
690 
691 	if (!ucmd->buf_addr)
692 		return -EINVAL;
693 
694 	context = to_mucontext(pd->uobject->context);
695 	rwq->umem = ib_umem_get(pd->uobject->context, ucmd->buf_addr,
696 			       rwq->buf_size, 0, 0);
697 	if (IS_ERR(rwq->umem)) {
698 		mlx5_ib_dbg(dev, "umem_get failed\n");
699 		err = PTR_ERR(rwq->umem);
700 		return err;
701 	}
702 
703 	mlx5_ib_cont_pages(rwq->umem, ucmd->buf_addr, &npages, &page_shift,
704 			   &ncont, NULL);
705 	err = mlx5_ib_get_buf_offset(ucmd->buf_addr, page_shift,
706 				     &rwq->rq_page_offset);
707 	if (err) {
708 		mlx5_ib_warn(dev, "bad offset\n");
709 		goto err_umem;
710 	}
711 
712 	rwq->rq_num_pas = ncont;
713 	rwq->page_shift = page_shift;
714 	rwq->log_page_size =  page_shift - MLX5_ADAPTER_PAGE_SHIFT;
715 	rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
716 
717 	mlx5_ib_dbg(dev, "addr 0x%llx, size %zd, npages %d, page_shift %d, ncont %d, offset %d\n",
718 		    (unsigned long long)ucmd->buf_addr, rwq->buf_size,
719 		    npages, page_shift, ncont, offset);
720 
721 	err = mlx5_ib_db_map_user(context, ucmd->db_addr, &rwq->db);
722 	if (err) {
723 		mlx5_ib_dbg(dev, "map failed\n");
724 		goto err_umem;
725 	}
726 
727 	rwq->create_type = MLX5_WQ_USER;
728 	return 0;
729 
730 err_umem:
731 	ib_umem_release(rwq->umem);
732 	return err;
733 }
734 
735 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
736 			  struct mlx5_ib_qp *qp, struct ib_udata *udata,
737 			  struct ib_qp_init_attr *attr,
738 			  u32 **in,
739 			  struct mlx5_ib_create_qp_resp *resp, int *inlen,
740 			  struct mlx5_ib_qp_base *base)
741 {
742 	struct mlx5_ib_ucontext *context;
743 	struct mlx5_ib_create_qp ucmd;
744 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
745 	int page_shift = 0;
746 	int uar_index;
747 	int npages;
748 	u32 offset = 0;
749 	int uuarn;
750 	int ncont = 0;
751 	__be64 *pas;
752 	void *qpc;
753 	int err;
754 
755 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
756 	if (err) {
757 		mlx5_ib_dbg(dev, "copy failed\n");
758 		return err;
759 	}
760 
761 	context = to_mucontext(pd->uobject->context);
762 	/*
763 	 * TBD: should come from the verbs when we have the API
764 	 */
765 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
766 		/* In CROSS_CHANNEL CQ and QP must use the same UAR */
767 		uuarn = MLX5_CROSS_CHANNEL_UUAR;
768 	else {
769 		uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_HIGH);
770 		if (uuarn < 0) {
771 			mlx5_ib_dbg(dev, "failed to allocate low latency UUAR\n");
772 			mlx5_ib_dbg(dev, "reverting to medium latency\n");
773 			uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_MEDIUM);
774 			if (uuarn < 0) {
775 				mlx5_ib_dbg(dev, "failed to allocate medium latency UUAR\n");
776 				mlx5_ib_dbg(dev, "reverting to high latency\n");
777 				uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_LOW);
778 				if (uuarn < 0) {
779 					mlx5_ib_warn(dev, "uuar allocation failed\n");
780 					return uuarn;
781 				}
782 			}
783 		}
784 	}
785 
786 	uar_index = uuarn_to_uar_index(&context->uuari, uuarn);
787 	mlx5_ib_dbg(dev, "uuarn 0x%x, uar_index 0x%x\n", uuarn, uar_index);
788 
789 	qp->rq.offset = 0;
790 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
791 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
792 
793 	err = set_user_buf_size(dev, qp, &ucmd, base, attr);
794 	if (err)
795 		goto err_uuar;
796 
797 	if (ucmd.buf_addr && ubuffer->buf_size) {
798 		ubuffer->buf_addr = ucmd.buf_addr;
799 		err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr,
800 				       ubuffer->buf_size,
801 				       &ubuffer->umem, &npages, &page_shift,
802 				       &ncont, &offset);
803 		if (err)
804 			goto err_uuar;
805 	} else {
806 		ubuffer->umem = NULL;
807 	}
808 
809 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
810 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * ncont;
811 	*in = mlx5_vzalloc(*inlen);
812 	if (!*in) {
813 		err = -ENOMEM;
814 		goto err_umem;
815 	}
816 
817 	pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas);
818 	if (ubuffer->umem)
819 		mlx5_ib_populate_pas(dev, ubuffer->umem, page_shift, pas, 0);
820 
821 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
822 
823 	MLX5_SET(qpc, qpc, log_page_size, page_shift - MLX5_ADAPTER_PAGE_SHIFT);
824 	MLX5_SET(qpc, qpc, page_offset, offset);
825 
826 	MLX5_SET(qpc, qpc, uar_page, uar_index);
827 	resp->uuar_index = uuarn;
828 	qp->uuarn = uuarn;
829 
830 	err = mlx5_ib_db_map_user(context, ucmd.db_addr, &qp->db);
831 	if (err) {
832 		mlx5_ib_dbg(dev, "map failed\n");
833 		goto err_free;
834 	}
835 
836 	err = ib_copy_to_udata(udata, resp, sizeof(*resp));
837 	if (err) {
838 		mlx5_ib_dbg(dev, "copy failed\n");
839 		goto err_unmap;
840 	}
841 	qp->create_type = MLX5_QP_USER;
842 
843 	return 0;
844 
845 err_unmap:
846 	mlx5_ib_db_unmap_user(context, &qp->db);
847 
848 err_free:
849 	kvfree(*in);
850 
851 err_umem:
852 	if (ubuffer->umem)
853 		ib_umem_release(ubuffer->umem);
854 
855 err_uuar:
856 	free_uuar(&context->uuari, uuarn);
857 	return err;
858 }
859 
860 static void destroy_qp_user(struct ib_pd *pd, struct mlx5_ib_qp *qp,
861 			    struct mlx5_ib_qp_base *base)
862 {
863 	struct mlx5_ib_ucontext *context;
864 
865 	context = to_mucontext(pd->uobject->context);
866 	mlx5_ib_db_unmap_user(context, &qp->db);
867 	if (base->ubuffer.umem)
868 		ib_umem_release(base->ubuffer.umem);
869 	free_uuar(&context->uuari, qp->uuarn);
870 }
871 
872 static int create_kernel_qp(struct mlx5_ib_dev *dev,
873 			    struct ib_qp_init_attr *init_attr,
874 			    struct mlx5_ib_qp *qp,
875 			    u32 **in, int *inlen,
876 			    struct mlx5_ib_qp_base *base)
877 {
878 	enum mlx5_ib_latency_class lc = MLX5_IB_LATENCY_CLASS_LOW;
879 	struct mlx5_uuar_info *uuari;
880 	int uar_index;
881 	void *qpc;
882 	int uuarn;
883 	int err;
884 
885 	uuari = &dev->mdev->priv.uuari;
886 	if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN |
887 					IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
888 					IB_QP_CREATE_IPOIB_UD_LSO |
889 					mlx5_ib_create_qp_sqpn_qp1()))
890 		return -EINVAL;
891 
892 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
893 		lc = MLX5_IB_LATENCY_CLASS_FAST_PATH;
894 
895 	uuarn = alloc_uuar(uuari, lc);
896 	if (uuarn < 0) {
897 		mlx5_ib_dbg(dev, "\n");
898 		return -ENOMEM;
899 	}
900 
901 	qp->bf = &uuari->bfs[uuarn];
902 	uar_index = qp->bf->uar->index;
903 
904 	err = calc_sq_size(dev, init_attr, qp);
905 	if (err < 0) {
906 		mlx5_ib_dbg(dev, "err %d\n", err);
907 		goto err_uuar;
908 	}
909 
910 	qp->rq.offset = 0;
911 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
912 	base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
913 
914 	err = mlx5_buf_alloc(dev->mdev, base->ubuffer.buf_size, &qp->buf);
915 	if (err) {
916 		mlx5_ib_dbg(dev, "err %d\n", err);
917 		goto err_uuar;
918 	}
919 
920 	qp->sq.qend = mlx5_get_send_wqe(qp, qp->sq.wqe_cnt);
921 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
922 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * qp->buf.npages;
923 	*in = mlx5_vzalloc(*inlen);
924 	if (!*in) {
925 		err = -ENOMEM;
926 		goto err_buf;
927 	}
928 
929 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
930 	MLX5_SET(qpc, qpc, uar_page, uar_index);
931 	MLX5_SET(qpc, qpc, log_page_size, qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
932 
933 	/* Set "fast registration enabled" for all kernel QPs */
934 	MLX5_SET(qpc, qpc, fre, 1);
935 	MLX5_SET(qpc, qpc, rlky, 1);
936 
937 	if (init_attr->create_flags & mlx5_ib_create_qp_sqpn_qp1()) {
938 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
939 		qp->flags |= MLX5_IB_QP_SQPN_QP1;
940 	}
941 
942 	mlx5_fill_page_array(&qp->buf,
943 			     (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas));
944 
945 	err = mlx5_db_alloc(dev->mdev, &qp->db);
946 	if (err) {
947 		mlx5_ib_dbg(dev, "err %d\n", err);
948 		goto err_free;
949 	}
950 
951 	qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
952 	qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
953 	qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
954 	qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
955 	qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
956 
957 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
958 	    !qp->sq.w_list || !qp->sq.wqe_head) {
959 		err = -ENOMEM;
960 		goto err_wrid;
961 	}
962 	qp->create_type = MLX5_QP_KERNEL;
963 
964 	return 0;
965 
966 err_wrid:
967 	mlx5_db_free(dev->mdev, &qp->db);
968 	kfree(qp->sq.wqe_head);
969 	kfree(qp->sq.w_list);
970 	kfree(qp->sq.wrid);
971 	kfree(qp->sq.wr_data);
972 	kfree(qp->rq.wrid);
973 
974 err_free:
975 	kvfree(*in);
976 
977 err_buf:
978 	mlx5_buf_free(dev->mdev, &qp->buf);
979 
980 err_uuar:
981 	free_uuar(&dev->mdev->priv.uuari, uuarn);
982 	return err;
983 }
984 
985 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
986 {
987 	mlx5_db_free(dev->mdev, &qp->db);
988 	kfree(qp->sq.wqe_head);
989 	kfree(qp->sq.w_list);
990 	kfree(qp->sq.wrid);
991 	kfree(qp->sq.wr_data);
992 	kfree(qp->rq.wrid);
993 	mlx5_buf_free(dev->mdev, &qp->buf);
994 	free_uuar(&dev->mdev->priv.uuari, qp->bf->uuarn);
995 }
996 
997 static u32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
998 {
999 	if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
1000 	    (attr->qp_type == IB_QPT_XRC_INI))
1001 		return MLX5_SRQ_RQ;
1002 	else if (!qp->has_rq)
1003 		return MLX5_ZERO_LEN_RQ;
1004 	else
1005 		return MLX5_NON_ZERO_RQ;
1006 }
1007 
1008 static int is_connected(enum ib_qp_type qp_type)
1009 {
1010 	if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC)
1011 		return 1;
1012 
1013 	return 0;
1014 }
1015 
1016 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1017 				    struct mlx5_ib_sq *sq, u32 tdn)
1018 {
1019 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {0};
1020 	void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1021 
1022 	MLX5_SET(tisc, tisc, transport_domain, tdn);
1023 	return mlx5_core_create_tis(dev->mdev, in, sizeof(in), &sq->tisn);
1024 }
1025 
1026 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1027 				      struct mlx5_ib_sq *sq)
1028 {
1029 	mlx5_core_destroy_tis(dev->mdev, sq->tisn);
1030 }
1031 
1032 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1033 				   struct mlx5_ib_sq *sq, void *qpin,
1034 				   struct ib_pd *pd)
1035 {
1036 	struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1037 	__be64 *pas;
1038 	void *in;
1039 	void *sqc;
1040 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1041 	void *wq;
1042 	int inlen;
1043 	int err;
1044 	int page_shift = 0;
1045 	int npages;
1046 	int ncont = 0;
1047 	u32 offset = 0;
1048 
1049 	err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr, ubuffer->buf_size,
1050 			       &sq->ubuffer.umem, &npages, &page_shift,
1051 			       &ncont, &offset);
1052 	if (err)
1053 		return err;
1054 
1055 	inlen = MLX5_ST_SZ_BYTES(create_sq_in) + sizeof(u64) * ncont;
1056 	in = mlx5_vzalloc(inlen);
1057 	if (!in) {
1058 		err = -ENOMEM;
1059 		goto err_umem;
1060 	}
1061 
1062 	sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1063 	MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1064 	MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1065 	MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1066 	MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1067 	MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1068 	MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1069 
1070 	wq = MLX5_ADDR_OF(sqc, sqc, wq);
1071 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1072 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1073 	MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1074 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1075 	MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1076 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1077 	MLX5_SET(wq, wq, log_wq_pg_sz,  page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1078 	MLX5_SET(wq, wq, page_offset, offset);
1079 
1080 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1081 	mlx5_ib_populate_pas(dev, sq->ubuffer.umem, page_shift, pas, 0);
1082 
1083 	err = mlx5_core_create_sq_tracked(dev->mdev, in, inlen, &sq->base.mqp);
1084 
1085 	kvfree(in);
1086 
1087 	if (err)
1088 		goto err_umem;
1089 
1090 	return 0;
1091 
1092 err_umem:
1093 	ib_umem_release(sq->ubuffer.umem);
1094 	sq->ubuffer.umem = NULL;
1095 
1096 	return err;
1097 }
1098 
1099 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1100 				     struct mlx5_ib_sq *sq)
1101 {
1102 	mlx5_core_destroy_sq_tracked(dev->mdev, &sq->base.mqp);
1103 	ib_umem_release(sq->ubuffer.umem);
1104 }
1105 
1106 static int get_rq_pas_size(void *qpc)
1107 {
1108 	u32 log_page_size = MLX5_GET(qpc, qpc, log_page_size) + 12;
1109 	u32 log_rq_stride = MLX5_GET(qpc, qpc, log_rq_stride);
1110 	u32 log_rq_size   = MLX5_GET(qpc, qpc, log_rq_size);
1111 	u32 page_offset   = MLX5_GET(qpc, qpc, page_offset);
1112 	u32 po_quanta	  = 1 << (log_page_size - 6);
1113 	u32 rq_sz	  = 1 << (log_rq_size + 4 + log_rq_stride);
1114 	u32 page_size	  = 1 << log_page_size;
1115 	u32 rq_sz_po      = rq_sz + (page_offset * po_quanta);
1116 	u32 rq_num_pas	  = (rq_sz_po + page_size - 1) / page_size;
1117 
1118 	return rq_num_pas * sizeof(u64);
1119 }
1120 
1121 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1122 				   struct mlx5_ib_rq *rq, void *qpin)
1123 {
1124 	struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1125 	__be64 *pas;
1126 	__be64 *qp_pas;
1127 	void *in;
1128 	void *rqc;
1129 	void *wq;
1130 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1131 	int inlen;
1132 	int err;
1133 	u32 rq_pas_size = get_rq_pas_size(qpc);
1134 
1135 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + rq_pas_size;
1136 	in = mlx5_vzalloc(inlen);
1137 	if (!in)
1138 		return -ENOMEM;
1139 
1140 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1141 	MLX5_SET(rqc, rqc, vsd, 1);
1142 	MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
1143 	MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1144 	MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1145 	MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1146 	MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1147 
1148 	if (mqp->flags & MLX5_IB_QP_CAP_SCATTER_FCS)
1149 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
1150 
1151 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
1152 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1153 	MLX5_SET(wq, wq, end_padding_mode,
1154 		 MLX5_GET(qpc, qpc, end_padding_mode));
1155 	MLX5_SET(wq, wq, page_offset, MLX5_GET(qpc, qpc, page_offset));
1156 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1157 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1158 	MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1159 	MLX5_SET(wq, wq, log_wq_pg_sz, MLX5_GET(qpc, qpc, log_page_size));
1160 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1161 
1162 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1163 	qp_pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, qpin, pas);
1164 	memcpy(pas, qp_pas, rq_pas_size);
1165 
1166 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rq->base.mqp);
1167 
1168 	kvfree(in);
1169 
1170 	return err;
1171 }
1172 
1173 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1174 				     struct mlx5_ib_rq *rq)
1175 {
1176 	mlx5_core_destroy_rq_tracked(dev->mdev, &rq->base.mqp);
1177 }
1178 
1179 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1180 				    struct mlx5_ib_rq *rq, u32 tdn)
1181 {
1182 	u32 *in;
1183 	void *tirc;
1184 	int inlen;
1185 	int err;
1186 
1187 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1188 	in = mlx5_vzalloc(inlen);
1189 	if (!in)
1190 		return -ENOMEM;
1191 
1192 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1193 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1194 	MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1195 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1196 
1197 	err = mlx5_core_create_tir(dev->mdev, in, inlen, &rq->tirn);
1198 
1199 	kvfree(in);
1200 
1201 	return err;
1202 }
1203 
1204 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1205 				      struct mlx5_ib_rq *rq)
1206 {
1207 	mlx5_core_destroy_tir(dev->mdev, rq->tirn);
1208 }
1209 
1210 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1211 				u32 *in,
1212 				struct ib_pd *pd)
1213 {
1214 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1215 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1216 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1217 	struct ib_uobject *uobj = pd->uobject;
1218 	struct ib_ucontext *ucontext = uobj->context;
1219 	struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1220 	int err;
1221 	u32 tdn = mucontext->tdn;
1222 
1223 	if (qp->sq.wqe_cnt) {
1224 		err = create_raw_packet_qp_tis(dev, sq, tdn);
1225 		if (err)
1226 			return err;
1227 
1228 		err = create_raw_packet_qp_sq(dev, sq, in, pd);
1229 		if (err)
1230 			goto err_destroy_tis;
1231 
1232 		sq->base.container_mibqp = qp;
1233 	}
1234 
1235 	if (qp->rq.wqe_cnt) {
1236 		rq->base.container_mibqp = qp;
1237 
1238 		err = create_raw_packet_qp_rq(dev, rq, in);
1239 		if (err)
1240 			goto err_destroy_sq;
1241 
1242 
1243 		err = create_raw_packet_qp_tir(dev, rq, tdn);
1244 		if (err)
1245 			goto err_destroy_rq;
1246 	}
1247 
1248 	qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1249 						     rq->base.mqp.qpn;
1250 
1251 	return 0;
1252 
1253 err_destroy_rq:
1254 	destroy_raw_packet_qp_rq(dev, rq);
1255 err_destroy_sq:
1256 	if (!qp->sq.wqe_cnt)
1257 		return err;
1258 	destroy_raw_packet_qp_sq(dev, sq);
1259 err_destroy_tis:
1260 	destroy_raw_packet_qp_tis(dev, sq);
1261 
1262 	return err;
1263 }
1264 
1265 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1266 				  struct mlx5_ib_qp *qp)
1267 {
1268 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1269 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1270 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1271 
1272 	if (qp->rq.wqe_cnt) {
1273 		destroy_raw_packet_qp_tir(dev, rq);
1274 		destroy_raw_packet_qp_rq(dev, rq);
1275 	}
1276 
1277 	if (qp->sq.wqe_cnt) {
1278 		destroy_raw_packet_qp_sq(dev, sq);
1279 		destroy_raw_packet_qp_tis(dev, sq);
1280 	}
1281 }
1282 
1283 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1284 				    struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1285 {
1286 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1287 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1288 
1289 	sq->sq = &qp->sq;
1290 	rq->rq = &qp->rq;
1291 	sq->doorbell = &qp->db;
1292 	rq->doorbell = &qp->db;
1293 }
1294 
1295 static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1296 {
1297 	mlx5_core_destroy_tir(dev->mdev, qp->rss_qp.tirn);
1298 }
1299 
1300 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1301 				 struct ib_pd *pd,
1302 				 struct ib_qp_init_attr *init_attr,
1303 				 struct ib_udata *udata)
1304 {
1305 	struct ib_uobject *uobj = pd->uobject;
1306 	struct ib_ucontext *ucontext = uobj->context;
1307 	struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1308 	struct mlx5_ib_create_qp_resp resp = {};
1309 	int inlen;
1310 	int err;
1311 	u32 *in;
1312 	void *tirc;
1313 	void *hfso;
1314 	u32 selected_fields = 0;
1315 	size_t min_resp_len;
1316 	u32 tdn = mucontext->tdn;
1317 	struct mlx5_ib_create_qp_rss ucmd = {};
1318 	size_t required_cmd_sz;
1319 
1320 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1321 		return -EOPNOTSUPP;
1322 
1323 	if (init_attr->create_flags || init_attr->send_cq)
1324 		return -EINVAL;
1325 
1326 	min_resp_len = offsetof(typeof(resp), uuar_index) + sizeof(resp.uuar_index);
1327 	if (udata->outlen < min_resp_len)
1328 		return -EINVAL;
1329 
1330 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) + sizeof(ucmd.reserved1);
1331 	if (udata->inlen < required_cmd_sz) {
1332 		mlx5_ib_dbg(dev, "invalid inlen\n");
1333 		return -EINVAL;
1334 	}
1335 
1336 	if (udata->inlen > sizeof(ucmd) &&
1337 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
1338 				 udata->inlen - sizeof(ucmd))) {
1339 		mlx5_ib_dbg(dev, "inlen is not supported\n");
1340 		return -EOPNOTSUPP;
1341 	}
1342 
1343 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
1344 		mlx5_ib_dbg(dev, "copy failed\n");
1345 		return -EFAULT;
1346 	}
1347 
1348 	if (ucmd.comp_mask) {
1349 		mlx5_ib_dbg(dev, "invalid comp mask\n");
1350 		return -EOPNOTSUPP;
1351 	}
1352 
1353 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)) || ucmd.reserved1) {
1354 		mlx5_ib_dbg(dev, "invalid reserved\n");
1355 		return -EOPNOTSUPP;
1356 	}
1357 
1358 	err = ib_copy_to_udata(udata, &resp, min_resp_len);
1359 	if (err) {
1360 		mlx5_ib_dbg(dev, "copy failed\n");
1361 		return -EINVAL;
1362 	}
1363 
1364 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1365 	in = mlx5_vzalloc(inlen);
1366 	if (!in)
1367 		return -ENOMEM;
1368 
1369 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1370 	MLX5_SET(tirc, tirc, disp_type,
1371 		 MLX5_TIRC_DISP_TYPE_INDIRECT);
1372 	MLX5_SET(tirc, tirc, indirect_table,
1373 		 init_attr->rwq_ind_tbl->ind_tbl_num);
1374 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1375 
1376 	hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1377 	switch (ucmd.rx_hash_function) {
1378 	case MLX5_RX_HASH_FUNC_TOEPLITZ:
1379 	{
1380 		void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1381 		size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1382 
1383 		if (len != ucmd.rx_key_len) {
1384 			err = -EINVAL;
1385 			goto err;
1386 		}
1387 
1388 		MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_TOEPLITZ);
1389 		MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
1390 		memcpy(rss_key, ucmd.rx_hash_key, len);
1391 		break;
1392 	}
1393 	default:
1394 		err = -EOPNOTSUPP;
1395 		goto err;
1396 	}
1397 
1398 	if (!ucmd.rx_hash_fields_mask) {
1399 		/* special case when this TIR serves as steering entry without hashing */
1400 		if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1401 			goto create_tir;
1402 		err = -EINVAL;
1403 		goto err;
1404 	}
1405 
1406 	if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1407 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1408 	     ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1409 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1410 		err = -EINVAL;
1411 		goto err;
1412 	}
1413 
1414 	/* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1415 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1416 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1417 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1418 			 MLX5_L3_PROT_TYPE_IPV4);
1419 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1420 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1421 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1422 			 MLX5_L3_PROT_TYPE_IPV6);
1423 
1424 	if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1425 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP)) &&
1426 	     ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1427 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))) {
1428 		err = -EINVAL;
1429 		goto err;
1430 	}
1431 
1432 	/* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1433 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1434 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1435 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1436 			 MLX5_L4_PROT_TYPE_TCP);
1437 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1438 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1439 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1440 			 MLX5_L4_PROT_TYPE_UDP);
1441 
1442 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1443 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1444 		selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1445 
1446 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1447 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1448 		selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1449 
1450 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1451 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1452 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1453 
1454 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1455 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1456 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1457 
1458 	MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1459 
1460 create_tir:
1461 	err = mlx5_core_create_tir(dev->mdev, in, inlen, &qp->rss_qp.tirn);
1462 
1463 	if (err)
1464 		goto err;
1465 
1466 	kvfree(in);
1467 	/* qpn is reserved for that QP */
1468 	qp->trans_qp.base.mqp.qpn = 0;
1469 	qp->flags |= MLX5_IB_QP_RSS;
1470 	return 0;
1471 
1472 err:
1473 	kvfree(in);
1474 	return err;
1475 }
1476 
1477 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1478 			    struct ib_qp_init_attr *init_attr,
1479 			    struct ib_udata *udata, struct mlx5_ib_qp *qp)
1480 {
1481 	struct mlx5_ib_resources *devr = &dev->devr;
1482 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1483 	struct mlx5_core_dev *mdev = dev->mdev;
1484 	struct mlx5_ib_create_qp_resp resp;
1485 	struct mlx5_ib_cq *send_cq;
1486 	struct mlx5_ib_cq *recv_cq;
1487 	unsigned long flags;
1488 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
1489 	struct mlx5_ib_create_qp ucmd;
1490 	struct mlx5_ib_qp_base *base;
1491 	void *qpc;
1492 	u32 *in;
1493 	int err;
1494 
1495 	base = init_attr->qp_type == IB_QPT_RAW_PACKET ?
1496 	       &qp->raw_packet_qp.rq.base :
1497 	       &qp->trans_qp.base;
1498 
1499 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1500 		mlx5_ib_odp_create_qp(qp);
1501 
1502 	mutex_init(&qp->mutex);
1503 	spin_lock_init(&qp->sq.lock);
1504 	spin_lock_init(&qp->rq.lock);
1505 
1506 	if (init_attr->rwq_ind_tbl) {
1507 		if (!udata)
1508 			return -ENOSYS;
1509 
1510 		err = create_rss_raw_qp_tir(dev, qp, pd, init_attr, udata);
1511 		return err;
1512 	}
1513 
1514 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
1515 		if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
1516 			mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
1517 			return -EINVAL;
1518 		} else {
1519 			qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1520 		}
1521 	}
1522 
1523 	if (init_attr->create_flags &
1524 			(IB_QP_CREATE_CROSS_CHANNEL |
1525 			 IB_QP_CREATE_MANAGED_SEND |
1526 			 IB_QP_CREATE_MANAGED_RECV)) {
1527 		if (!MLX5_CAP_GEN(mdev, cd)) {
1528 			mlx5_ib_dbg(dev, "cross-channel isn't supported\n");
1529 			return -EINVAL;
1530 		}
1531 		if (init_attr->create_flags & IB_QP_CREATE_CROSS_CHANNEL)
1532 			qp->flags |= MLX5_IB_QP_CROSS_CHANNEL;
1533 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_SEND)
1534 			qp->flags |= MLX5_IB_QP_MANAGED_SEND;
1535 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_RECV)
1536 			qp->flags |= MLX5_IB_QP_MANAGED_RECV;
1537 	}
1538 
1539 	if (init_attr->qp_type == IB_QPT_UD &&
1540 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO))
1541 		if (!MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
1542 			mlx5_ib_dbg(dev, "ipoib UD lso qp isn't supported\n");
1543 			return -EOPNOTSUPP;
1544 		}
1545 
1546 	if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1547 		if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
1548 			mlx5_ib_dbg(dev, "Scatter FCS is supported only for Raw Packet QPs");
1549 			return -EOPNOTSUPP;
1550 		}
1551 		if (!MLX5_CAP_GEN(dev->mdev, eth_net_offloads) ||
1552 		    !MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
1553 			mlx5_ib_dbg(dev, "Scatter FCS isn't supported\n");
1554 			return -EOPNOTSUPP;
1555 		}
1556 		qp->flags |= MLX5_IB_QP_CAP_SCATTER_FCS;
1557 	}
1558 
1559 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1560 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
1561 
1562 	if (pd && pd->uobject) {
1563 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
1564 			mlx5_ib_dbg(dev, "copy failed\n");
1565 			return -EFAULT;
1566 		}
1567 
1568 		err = get_qp_user_index(to_mucontext(pd->uobject->context),
1569 					&ucmd, udata->inlen, &uidx);
1570 		if (err)
1571 			return err;
1572 
1573 		qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
1574 		qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
1575 	} else {
1576 		qp->wq_sig = !!wq_signature;
1577 	}
1578 
1579 	qp->has_rq = qp_has_rq(init_attr);
1580 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
1581 			  qp, (pd && pd->uobject) ? &ucmd : NULL);
1582 	if (err) {
1583 		mlx5_ib_dbg(dev, "err %d\n", err);
1584 		return err;
1585 	}
1586 
1587 	if (pd) {
1588 		if (pd->uobject) {
1589 			__u32 max_wqes =
1590 				1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1591 			mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
1592 			if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
1593 			    ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
1594 				mlx5_ib_dbg(dev, "invalid rq params\n");
1595 				return -EINVAL;
1596 			}
1597 			if (ucmd.sq_wqe_count > max_wqes) {
1598 				mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
1599 					    ucmd.sq_wqe_count, max_wqes);
1600 				return -EINVAL;
1601 			}
1602 			if (init_attr->create_flags &
1603 			    mlx5_ib_create_qp_sqpn_qp1()) {
1604 				mlx5_ib_dbg(dev, "user-space is not allowed to create UD QPs spoofing as QP1\n");
1605 				return -EINVAL;
1606 			}
1607 			err = create_user_qp(dev, pd, qp, udata, init_attr, &in,
1608 					     &resp, &inlen, base);
1609 			if (err)
1610 				mlx5_ib_dbg(dev, "err %d\n", err);
1611 		} else {
1612 			err = create_kernel_qp(dev, init_attr, qp, &in, &inlen,
1613 					       base);
1614 			if (err)
1615 				mlx5_ib_dbg(dev, "err %d\n", err);
1616 		}
1617 
1618 		if (err)
1619 			return err;
1620 	} else {
1621 		in = mlx5_vzalloc(inlen);
1622 		if (!in)
1623 			return -ENOMEM;
1624 
1625 		qp->create_type = MLX5_QP_EMPTY;
1626 	}
1627 
1628 	if (is_sqp(init_attr->qp_type))
1629 		qp->port = init_attr->port_num;
1630 
1631 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1632 
1633 	MLX5_SET(qpc, qpc, st, to_mlx5_st(init_attr->qp_type));
1634 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
1635 
1636 	if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
1637 		MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
1638 	else
1639 		MLX5_SET(qpc, qpc, latency_sensitive, 1);
1640 
1641 
1642 	if (qp->wq_sig)
1643 		MLX5_SET(qpc, qpc, wq_signature, 1);
1644 
1645 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
1646 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
1647 
1648 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
1649 		MLX5_SET(qpc, qpc, cd_master, 1);
1650 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
1651 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
1652 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
1653 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
1654 
1655 	if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
1656 		int rcqe_sz;
1657 		int scqe_sz;
1658 
1659 		rcqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->recv_cq);
1660 		scqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->send_cq);
1661 
1662 		if (rcqe_sz == 128)
1663 			MLX5_SET(qpc, qpc, cs_res, MLX5_RES_SCAT_DATA64_CQE);
1664 		else
1665 			MLX5_SET(qpc, qpc, cs_res, MLX5_RES_SCAT_DATA32_CQE);
1666 
1667 		if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) {
1668 			if (scqe_sz == 128)
1669 				MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA64_CQE);
1670 			else
1671 				MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA32_CQE);
1672 		}
1673 	}
1674 
1675 	if (qp->rq.wqe_cnt) {
1676 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
1677 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
1678 	}
1679 
1680 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
1681 
1682 	if (qp->sq.wqe_cnt)
1683 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
1684 	else
1685 		MLX5_SET(qpc, qpc, no_sq, 1);
1686 
1687 	/* Set default resources */
1688 	switch (init_attr->qp_type) {
1689 	case IB_QPT_XRC_TGT:
1690 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1691 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(devr->c0)->mcq.cqn);
1692 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
1693 		MLX5_SET(qpc, qpc, xrcd, to_mxrcd(init_attr->xrcd)->xrcdn);
1694 		break;
1695 	case IB_QPT_XRC_INI:
1696 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1697 		MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
1698 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
1699 		break;
1700 	default:
1701 		if (init_attr->srq) {
1702 			MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x0)->xrcdn);
1703 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(init_attr->srq)->msrq.srqn);
1704 		} else {
1705 			MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
1706 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s1)->msrq.srqn);
1707 		}
1708 	}
1709 
1710 	if (init_attr->send_cq)
1711 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(init_attr->send_cq)->mcq.cqn);
1712 
1713 	if (init_attr->recv_cq)
1714 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(init_attr->recv_cq)->mcq.cqn);
1715 
1716 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
1717 
1718 	/* 0xffffff means we ask to work with cqe version 0 */
1719 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
1720 		MLX5_SET(qpc, qpc, user_index, uidx);
1721 
1722 	/* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
1723 	if (init_attr->qp_type == IB_QPT_UD &&
1724 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)) {
1725 		MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
1726 		qp->flags |= MLX5_IB_QP_LSO;
1727 	}
1728 
1729 	if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
1730 		qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd.sq_buf_addr;
1731 		raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
1732 		err = create_raw_packet_qp(dev, qp, in, pd);
1733 	} else {
1734 		err = mlx5_core_create_qp(dev->mdev, &base->mqp, in, inlen);
1735 	}
1736 
1737 	if (err) {
1738 		mlx5_ib_dbg(dev, "create qp failed\n");
1739 		goto err_create;
1740 	}
1741 
1742 	kvfree(in);
1743 
1744 	base->container_mibqp = qp;
1745 	base->mqp.event = mlx5_ib_qp_event;
1746 
1747 	get_cqs(init_attr->qp_type, init_attr->send_cq, init_attr->recv_cq,
1748 		&send_cq, &recv_cq);
1749 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1750 	mlx5_ib_lock_cqs(send_cq, recv_cq);
1751 	/* Maintain device to QPs access, needed for further handling via reset
1752 	 * flow
1753 	 */
1754 	list_add_tail(&qp->qps_list, &dev->qp_list);
1755 	/* Maintain CQ to QPs access, needed for further handling via reset flow
1756 	 */
1757 	if (send_cq)
1758 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
1759 	if (recv_cq)
1760 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
1761 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
1762 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1763 
1764 	return 0;
1765 
1766 err_create:
1767 	if (qp->create_type == MLX5_QP_USER)
1768 		destroy_qp_user(pd, qp, base);
1769 	else if (qp->create_type == MLX5_QP_KERNEL)
1770 		destroy_qp_kernel(dev, qp);
1771 
1772 	kvfree(in);
1773 	return err;
1774 }
1775 
1776 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1777 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1778 {
1779 	if (send_cq) {
1780 		if (recv_cq) {
1781 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1782 				spin_lock(&send_cq->lock);
1783 				spin_lock_nested(&recv_cq->lock,
1784 						 SINGLE_DEPTH_NESTING);
1785 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1786 				spin_lock(&send_cq->lock);
1787 				__acquire(&recv_cq->lock);
1788 			} else {
1789 				spin_lock(&recv_cq->lock);
1790 				spin_lock_nested(&send_cq->lock,
1791 						 SINGLE_DEPTH_NESTING);
1792 			}
1793 		} else {
1794 			spin_lock(&send_cq->lock);
1795 			__acquire(&recv_cq->lock);
1796 		}
1797 	} else if (recv_cq) {
1798 		spin_lock(&recv_cq->lock);
1799 		__acquire(&send_cq->lock);
1800 	} else {
1801 		__acquire(&send_cq->lock);
1802 		__acquire(&recv_cq->lock);
1803 	}
1804 }
1805 
1806 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1807 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1808 {
1809 	if (send_cq) {
1810 		if (recv_cq) {
1811 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1812 				spin_unlock(&recv_cq->lock);
1813 				spin_unlock(&send_cq->lock);
1814 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1815 				__release(&recv_cq->lock);
1816 				spin_unlock(&send_cq->lock);
1817 			} else {
1818 				spin_unlock(&send_cq->lock);
1819 				spin_unlock(&recv_cq->lock);
1820 			}
1821 		} else {
1822 			__release(&recv_cq->lock);
1823 			spin_unlock(&send_cq->lock);
1824 		}
1825 	} else if (recv_cq) {
1826 		__release(&send_cq->lock);
1827 		spin_unlock(&recv_cq->lock);
1828 	} else {
1829 		__release(&recv_cq->lock);
1830 		__release(&send_cq->lock);
1831 	}
1832 }
1833 
1834 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
1835 {
1836 	return to_mpd(qp->ibqp.pd);
1837 }
1838 
1839 static void get_cqs(enum ib_qp_type qp_type,
1840 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
1841 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
1842 {
1843 	switch (qp_type) {
1844 	case IB_QPT_XRC_TGT:
1845 		*send_cq = NULL;
1846 		*recv_cq = NULL;
1847 		break;
1848 	case MLX5_IB_QPT_REG_UMR:
1849 	case IB_QPT_XRC_INI:
1850 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
1851 		*recv_cq = NULL;
1852 		break;
1853 
1854 	case IB_QPT_SMI:
1855 	case MLX5_IB_QPT_HW_GSI:
1856 	case IB_QPT_RC:
1857 	case IB_QPT_UC:
1858 	case IB_QPT_UD:
1859 	case IB_QPT_RAW_IPV6:
1860 	case IB_QPT_RAW_ETHERTYPE:
1861 	case IB_QPT_RAW_PACKET:
1862 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
1863 		*recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
1864 		break;
1865 
1866 	case IB_QPT_MAX:
1867 	default:
1868 		*send_cq = NULL;
1869 		*recv_cq = NULL;
1870 		break;
1871 	}
1872 }
1873 
1874 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1875 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
1876 				u8 lag_tx_affinity);
1877 
1878 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1879 {
1880 	struct mlx5_ib_cq *send_cq, *recv_cq;
1881 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
1882 	unsigned long flags;
1883 	int err;
1884 
1885 	if (qp->ibqp.rwq_ind_tbl) {
1886 		destroy_rss_raw_qp_tir(dev, qp);
1887 		return;
1888 	}
1889 
1890 	base = qp->ibqp.qp_type == IB_QPT_RAW_PACKET ?
1891 	       &qp->raw_packet_qp.rq.base :
1892 	       &qp->trans_qp.base;
1893 
1894 	if (qp->state != IB_QPS_RESET) {
1895 		if (qp->ibqp.qp_type != IB_QPT_RAW_PACKET) {
1896 			mlx5_ib_qp_disable_pagefaults(qp);
1897 			err = mlx5_core_qp_modify(dev->mdev,
1898 						  MLX5_CMD_OP_2RST_QP, 0,
1899 						  NULL, &base->mqp);
1900 		} else {
1901 			struct mlx5_modify_raw_qp_param raw_qp_param = {
1902 				.operation = MLX5_CMD_OP_2RST_QP
1903 			};
1904 
1905 			err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
1906 		}
1907 		if (err)
1908 			mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
1909 				     base->mqp.qpn);
1910 	}
1911 
1912 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
1913 		&send_cq, &recv_cq);
1914 
1915 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1916 	mlx5_ib_lock_cqs(send_cq, recv_cq);
1917 	/* del from lists under both locks above to protect reset flow paths */
1918 	list_del(&qp->qps_list);
1919 	if (send_cq)
1920 		list_del(&qp->cq_send_list);
1921 
1922 	if (recv_cq)
1923 		list_del(&qp->cq_recv_list);
1924 
1925 	if (qp->create_type == MLX5_QP_KERNEL) {
1926 		__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
1927 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1928 		if (send_cq != recv_cq)
1929 			__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
1930 					   NULL);
1931 	}
1932 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
1933 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1934 
1935 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1936 		destroy_raw_packet_qp(dev, qp);
1937 	} else {
1938 		err = mlx5_core_destroy_qp(dev->mdev, &base->mqp);
1939 		if (err)
1940 			mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
1941 				     base->mqp.qpn);
1942 	}
1943 
1944 	if (qp->create_type == MLX5_QP_KERNEL)
1945 		destroy_qp_kernel(dev, qp);
1946 	else if (qp->create_type == MLX5_QP_USER)
1947 		destroy_qp_user(&get_pd(qp)->ibpd, qp, base);
1948 }
1949 
1950 static const char *ib_qp_type_str(enum ib_qp_type type)
1951 {
1952 	switch (type) {
1953 	case IB_QPT_SMI:
1954 		return "IB_QPT_SMI";
1955 	case IB_QPT_GSI:
1956 		return "IB_QPT_GSI";
1957 	case IB_QPT_RC:
1958 		return "IB_QPT_RC";
1959 	case IB_QPT_UC:
1960 		return "IB_QPT_UC";
1961 	case IB_QPT_UD:
1962 		return "IB_QPT_UD";
1963 	case IB_QPT_RAW_IPV6:
1964 		return "IB_QPT_RAW_IPV6";
1965 	case IB_QPT_RAW_ETHERTYPE:
1966 		return "IB_QPT_RAW_ETHERTYPE";
1967 	case IB_QPT_XRC_INI:
1968 		return "IB_QPT_XRC_INI";
1969 	case IB_QPT_XRC_TGT:
1970 		return "IB_QPT_XRC_TGT";
1971 	case IB_QPT_RAW_PACKET:
1972 		return "IB_QPT_RAW_PACKET";
1973 	case MLX5_IB_QPT_REG_UMR:
1974 		return "MLX5_IB_QPT_REG_UMR";
1975 	case IB_QPT_MAX:
1976 	default:
1977 		return "Invalid QP type";
1978 	}
1979 }
1980 
1981 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
1982 				struct ib_qp_init_attr *init_attr,
1983 				struct ib_udata *udata)
1984 {
1985 	struct mlx5_ib_dev *dev;
1986 	struct mlx5_ib_qp *qp;
1987 	u16 xrcdn = 0;
1988 	int err;
1989 
1990 	if (pd) {
1991 		dev = to_mdev(pd->device);
1992 
1993 		if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
1994 			if (!pd->uobject) {
1995 				mlx5_ib_dbg(dev, "Raw Packet QP is not supported for kernel consumers\n");
1996 				return ERR_PTR(-EINVAL);
1997 			} else if (!to_mucontext(pd->uobject->context)->cqe_version) {
1998 				mlx5_ib_dbg(dev, "Raw Packet QP is only supported for CQE version > 0\n");
1999 				return ERR_PTR(-EINVAL);
2000 			}
2001 		}
2002 	} else {
2003 		/* being cautious here */
2004 		if (init_attr->qp_type != IB_QPT_XRC_TGT &&
2005 		    init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
2006 			pr_warn("%s: no PD for transport %s\n", __func__,
2007 				ib_qp_type_str(init_attr->qp_type));
2008 			return ERR_PTR(-EINVAL);
2009 		}
2010 		dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
2011 	}
2012 
2013 	switch (init_attr->qp_type) {
2014 	case IB_QPT_XRC_TGT:
2015 	case IB_QPT_XRC_INI:
2016 		if (!MLX5_CAP_GEN(dev->mdev, xrc)) {
2017 			mlx5_ib_dbg(dev, "XRC not supported\n");
2018 			return ERR_PTR(-ENOSYS);
2019 		}
2020 		init_attr->recv_cq = NULL;
2021 		if (init_attr->qp_type == IB_QPT_XRC_TGT) {
2022 			xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
2023 			init_attr->send_cq = NULL;
2024 		}
2025 
2026 		/* fall through */
2027 	case IB_QPT_RAW_PACKET:
2028 	case IB_QPT_RC:
2029 	case IB_QPT_UC:
2030 	case IB_QPT_UD:
2031 	case IB_QPT_SMI:
2032 	case MLX5_IB_QPT_HW_GSI:
2033 	case MLX5_IB_QPT_REG_UMR:
2034 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2035 		if (!qp)
2036 			return ERR_PTR(-ENOMEM);
2037 
2038 		err = create_qp_common(dev, pd, init_attr, udata, qp);
2039 		if (err) {
2040 			mlx5_ib_dbg(dev, "create_qp_common failed\n");
2041 			kfree(qp);
2042 			return ERR_PTR(err);
2043 		}
2044 
2045 		if (is_qp0(init_attr->qp_type))
2046 			qp->ibqp.qp_num = 0;
2047 		else if (is_qp1(init_attr->qp_type))
2048 			qp->ibqp.qp_num = 1;
2049 		else
2050 			qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
2051 
2052 		mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
2053 			    qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
2054 			    to_mcq(init_attr->recv_cq)->mcq.cqn,
2055 			    to_mcq(init_attr->send_cq)->mcq.cqn);
2056 
2057 		qp->trans_qp.xrcdn = xrcdn;
2058 
2059 		break;
2060 
2061 	case IB_QPT_GSI:
2062 		return mlx5_ib_gsi_create_qp(pd, init_attr);
2063 
2064 	case IB_QPT_RAW_IPV6:
2065 	case IB_QPT_RAW_ETHERTYPE:
2066 	case IB_QPT_MAX:
2067 	default:
2068 		mlx5_ib_dbg(dev, "unsupported qp type %d\n",
2069 			    init_attr->qp_type);
2070 		/* Don't support raw QPs */
2071 		return ERR_PTR(-EINVAL);
2072 	}
2073 
2074 	return &qp->ibqp;
2075 }
2076 
2077 int mlx5_ib_destroy_qp(struct ib_qp *qp)
2078 {
2079 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2080 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2081 
2082 	if (unlikely(qp->qp_type == IB_QPT_GSI))
2083 		return mlx5_ib_gsi_destroy_qp(qp);
2084 
2085 	destroy_qp_common(dev, mqp);
2086 
2087 	kfree(mqp);
2088 
2089 	return 0;
2090 }
2091 
2092 static __be32 to_mlx5_access_flags(struct mlx5_ib_qp *qp, const struct ib_qp_attr *attr,
2093 				   int attr_mask)
2094 {
2095 	u32 hw_access_flags = 0;
2096 	u8 dest_rd_atomic;
2097 	u32 access_flags;
2098 
2099 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2100 		dest_rd_atomic = attr->max_dest_rd_atomic;
2101 	else
2102 		dest_rd_atomic = qp->trans_qp.resp_depth;
2103 
2104 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2105 		access_flags = attr->qp_access_flags;
2106 	else
2107 		access_flags = qp->trans_qp.atomic_rd_en;
2108 
2109 	if (!dest_rd_atomic)
2110 		access_flags &= IB_ACCESS_REMOTE_WRITE;
2111 
2112 	if (access_flags & IB_ACCESS_REMOTE_READ)
2113 		hw_access_flags |= MLX5_QP_BIT_RRE;
2114 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
2115 		hw_access_flags |= (MLX5_QP_BIT_RAE | MLX5_ATOMIC_MODE_CX);
2116 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
2117 		hw_access_flags |= MLX5_QP_BIT_RWE;
2118 
2119 	return cpu_to_be32(hw_access_flags);
2120 }
2121 
2122 enum {
2123 	MLX5_PATH_FLAG_FL	= 1 << 0,
2124 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
2125 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
2126 };
2127 
2128 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
2129 {
2130 	if (rate == IB_RATE_PORT_CURRENT) {
2131 		return 0;
2132 	} else if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_300_GBPS) {
2133 		return -EINVAL;
2134 	} else {
2135 		while (rate != IB_RATE_2_5_GBPS &&
2136 		       !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
2137 			 MLX5_CAP_GEN(dev->mdev, stat_rate_support)))
2138 			--rate;
2139 	}
2140 
2141 	return rate + MLX5_STAT_RATE_OFFSET;
2142 }
2143 
2144 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
2145 				      struct mlx5_ib_sq *sq, u8 sl)
2146 {
2147 	void *in;
2148 	void *tisc;
2149 	int inlen;
2150 	int err;
2151 
2152 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2153 	in = mlx5_vzalloc(inlen);
2154 	if (!in)
2155 		return -ENOMEM;
2156 
2157 	MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
2158 
2159 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2160 	MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
2161 
2162 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2163 
2164 	kvfree(in);
2165 
2166 	return err;
2167 }
2168 
2169 static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,
2170 					 struct mlx5_ib_sq *sq, u8 tx_affinity)
2171 {
2172 	void *in;
2173 	void *tisc;
2174 	int inlen;
2175 	int err;
2176 
2177 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2178 	in = mlx5_vzalloc(inlen);
2179 	if (!in)
2180 		return -ENOMEM;
2181 
2182 	MLX5_SET(modify_tis_in, in, bitmask.lag_tx_port_affinity, 1);
2183 
2184 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2185 	MLX5_SET(tisc, tisc, lag_tx_port_affinity, tx_affinity);
2186 
2187 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2188 
2189 	kvfree(in);
2190 
2191 	return err;
2192 }
2193 
2194 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2195 			 const struct ib_ah_attr *ah,
2196 			 struct mlx5_qp_path *path, u8 port, int attr_mask,
2197 			 u32 path_flags, const struct ib_qp_attr *attr,
2198 			 bool alt)
2199 {
2200 	enum rdma_link_layer ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2201 	int err;
2202 
2203 	if (attr_mask & IB_QP_PKEY_INDEX)
2204 		path->pkey_index = cpu_to_be16(alt ? attr->alt_pkey_index :
2205 						     attr->pkey_index);
2206 
2207 	if (ah->ah_flags & IB_AH_GRH) {
2208 		if (ah->grh.sgid_index >=
2209 		    dev->mdev->port_caps[port - 1].gid_table_len) {
2210 			pr_err("sgid_index (%u) too large. max is %d\n",
2211 			       ah->grh.sgid_index,
2212 			       dev->mdev->port_caps[port - 1].gid_table_len);
2213 			return -EINVAL;
2214 		}
2215 	}
2216 
2217 	if (ll == IB_LINK_LAYER_ETHERNET) {
2218 		if (!(ah->ah_flags & IB_AH_GRH))
2219 			return -EINVAL;
2220 		memcpy(path->rmac, ah->dmac, sizeof(ah->dmac));
2221 		path->udp_sport = mlx5_get_roce_udp_sport(dev, port,
2222 							  ah->grh.sgid_index);
2223 		path->dci_cfi_prio_sl = (ah->sl & 0x7) << 4;
2224 	} else {
2225 		path->fl_free_ar = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
2226 		path->fl_free_ar |=
2227 			(path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x40 : 0;
2228 		path->rlid = cpu_to_be16(ah->dlid);
2229 		path->grh_mlid = ah->src_path_bits & 0x7f;
2230 		if (ah->ah_flags & IB_AH_GRH)
2231 			path->grh_mlid	|= 1 << 7;
2232 		path->dci_cfi_prio_sl = ah->sl & 0xf;
2233 	}
2234 
2235 	if (ah->ah_flags & IB_AH_GRH) {
2236 		path->mgid_index = ah->grh.sgid_index;
2237 		path->hop_limit  = ah->grh.hop_limit;
2238 		path->tclass_flowlabel =
2239 			cpu_to_be32((ah->grh.traffic_class << 20) |
2240 				    (ah->grh.flow_label));
2241 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
2242 	}
2243 
2244 	err = ib_rate_to_mlx5(dev, ah->static_rate);
2245 	if (err < 0)
2246 		return err;
2247 	path->static_rate = err;
2248 	path->port = port;
2249 
2250 	if (attr_mask & IB_QP_TIMEOUT)
2251 		path->ackto_lt = (alt ? attr->alt_timeout : attr->timeout) << 3;
2252 
2253 	if ((qp->ibqp.qp_type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
2254 		return modify_raw_packet_eth_prio(dev->mdev,
2255 						  &qp->raw_packet_qp.sq,
2256 						  ah->sl & 0xf);
2257 
2258 	return 0;
2259 }
2260 
2261 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
2262 	[MLX5_QP_STATE_INIT] = {
2263 		[MLX5_QP_STATE_INIT] = {
2264 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
2265 					  MLX5_QP_OPTPAR_RAE		|
2266 					  MLX5_QP_OPTPAR_RWE		|
2267 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
2268 					  MLX5_QP_OPTPAR_PRI_PORT,
2269 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
2270 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
2271 					  MLX5_QP_OPTPAR_PRI_PORT,
2272 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
2273 					  MLX5_QP_OPTPAR_Q_KEY		|
2274 					  MLX5_QP_OPTPAR_PRI_PORT,
2275 		},
2276 		[MLX5_QP_STATE_RTR] = {
2277 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
2278 					  MLX5_QP_OPTPAR_RRE            |
2279 					  MLX5_QP_OPTPAR_RAE            |
2280 					  MLX5_QP_OPTPAR_RWE            |
2281 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2282 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
2283 					  MLX5_QP_OPTPAR_RWE            |
2284 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2285 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
2286 					  MLX5_QP_OPTPAR_Q_KEY,
2287 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
2288 					   MLX5_QP_OPTPAR_Q_KEY,
2289 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2290 					  MLX5_QP_OPTPAR_RRE            |
2291 					  MLX5_QP_OPTPAR_RAE            |
2292 					  MLX5_QP_OPTPAR_RWE            |
2293 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2294 		},
2295 	},
2296 	[MLX5_QP_STATE_RTR] = {
2297 		[MLX5_QP_STATE_RTS] = {
2298 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
2299 					  MLX5_QP_OPTPAR_RRE		|
2300 					  MLX5_QP_OPTPAR_RAE		|
2301 					  MLX5_QP_OPTPAR_RWE		|
2302 					  MLX5_QP_OPTPAR_PM_STATE	|
2303 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
2304 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
2305 					  MLX5_QP_OPTPAR_RWE		|
2306 					  MLX5_QP_OPTPAR_PM_STATE,
2307 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
2308 		},
2309 	},
2310 	[MLX5_QP_STATE_RTS] = {
2311 		[MLX5_QP_STATE_RTS] = {
2312 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
2313 					  MLX5_QP_OPTPAR_RAE		|
2314 					  MLX5_QP_OPTPAR_RWE		|
2315 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
2316 					  MLX5_QP_OPTPAR_PM_STATE	|
2317 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
2318 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
2319 					  MLX5_QP_OPTPAR_PM_STATE	|
2320 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
2321 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
2322 					  MLX5_QP_OPTPAR_SRQN		|
2323 					  MLX5_QP_OPTPAR_CQN_RCV,
2324 		},
2325 	},
2326 	[MLX5_QP_STATE_SQER] = {
2327 		[MLX5_QP_STATE_RTS] = {
2328 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
2329 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
2330 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
2331 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
2332 					   MLX5_QP_OPTPAR_RWE		|
2333 					   MLX5_QP_OPTPAR_RAE		|
2334 					   MLX5_QP_OPTPAR_RRE,
2335 		},
2336 	},
2337 };
2338 
2339 static int ib_nr_to_mlx5_nr(int ib_mask)
2340 {
2341 	switch (ib_mask) {
2342 	case IB_QP_STATE:
2343 		return 0;
2344 	case IB_QP_CUR_STATE:
2345 		return 0;
2346 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
2347 		return 0;
2348 	case IB_QP_ACCESS_FLAGS:
2349 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
2350 			MLX5_QP_OPTPAR_RAE;
2351 	case IB_QP_PKEY_INDEX:
2352 		return MLX5_QP_OPTPAR_PKEY_INDEX;
2353 	case IB_QP_PORT:
2354 		return MLX5_QP_OPTPAR_PRI_PORT;
2355 	case IB_QP_QKEY:
2356 		return MLX5_QP_OPTPAR_Q_KEY;
2357 	case IB_QP_AV:
2358 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
2359 			MLX5_QP_OPTPAR_PRI_PORT;
2360 	case IB_QP_PATH_MTU:
2361 		return 0;
2362 	case IB_QP_TIMEOUT:
2363 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
2364 	case IB_QP_RETRY_CNT:
2365 		return MLX5_QP_OPTPAR_RETRY_COUNT;
2366 	case IB_QP_RNR_RETRY:
2367 		return MLX5_QP_OPTPAR_RNR_RETRY;
2368 	case IB_QP_RQ_PSN:
2369 		return 0;
2370 	case IB_QP_MAX_QP_RD_ATOMIC:
2371 		return MLX5_QP_OPTPAR_SRA_MAX;
2372 	case IB_QP_ALT_PATH:
2373 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
2374 	case IB_QP_MIN_RNR_TIMER:
2375 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
2376 	case IB_QP_SQ_PSN:
2377 		return 0;
2378 	case IB_QP_MAX_DEST_RD_ATOMIC:
2379 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
2380 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
2381 	case IB_QP_PATH_MIG_STATE:
2382 		return MLX5_QP_OPTPAR_PM_STATE;
2383 	case IB_QP_CAP:
2384 		return 0;
2385 	case IB_QP_DEST_QPN:
2386 		return 0;
2387 	}
2388 	return 0;
2389 }
2390 
2391 static int ib_mask_to_mlx5_opt(int ib_mask)
2392 {
2393 	int result = 0;
2394 	int i;
2395 
2396 	for (i = 0; i < 8 * sizeof(int); i++) {
2397 		if ((1 << i) & ib_mask)
2398 			result |= ib_nr_to_mlx5_nr(1 << i);
2399 	}
2400 
2401 	return result;
2402 }
2403 
2404 static int modify_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
2405 				   struct mlx5_ib_rq *rq, int new_state,
2406 				   const struct mlx5_modify_raw_qp_param *raw_qp_param)
2407 {
2408 	void *in;
2409 	void *rqc;
2410 	int inlen;
2411 	int err;
2412 
2413 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
2414 	in = mlx5_vzalloc(inlen);
2415 	if (!in)
2416 		return -ENOMEM;
2417 
2418 	MLX5_SET(modify_rq_in, in, rq_state, rq->state);
2419 
2420 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
2421 	MLX5_SET(rqc, rqc, state, new_state);
2422 
2423 	if (raw_qp_param->set_mask & MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID) {
2424 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
2425 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
2426 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_MODIFY_RQ_COUNTER_SET_ID);
2427 			MLX5_SET(rqc, rqc, counter_set_id, raw_qp_param->rq_q_ctr_id);
2428 		} else
2429 			pr_info_once("%s: RAW PACKET QP counters are not supported on current FW\n",
2430 				     dev->ib_dev.name);
2431 	}
2432 
2433 	err = mlx5_core_modify_rq(dev->mdev, rq->base.mqp.qpn, in, inlen);
2434 	if (err)
2435 		goto out;
2436 
2437 	rq->state = new_state;
2438 
2439 out:
2440 	kvfree(in);
2441 	return err;
2442 }
2443 
2444 static int modify_raw_packet_qp_sq(struct mlx5_core_dev *dev,
2445 				   struct mlx5_ib_sq *sq, int new_state)
2446 {
2447 	void *in;
2448 	void *sqc;
2449 	int inlen;
2450 	int err;
2451 
2452 	inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
2453 	in = mlx5_vzalloc(inlen);
2454 	if (!in)
2455 		return -ENOMEM;
2456 
2457 	MLX5_SET(modify_sq_in, in, sq_state, sq->state);
2458 
2459 	sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
2460 	MLX5_SET(sqc, sqc, state, new_state);
2461 
2462 	err = mlx5_core_modify_sq(dev, sq->base.mqp.qpn, in, inlen);
2463 	if (err)
2464 		goto out;
2465 
2466 	sq->state = new_state;
2467 
2468 out:
2469 	kvfree(in);
2470 	return err;
2471 }
2472 
2473 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2474 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
2475 				u8 tx_affinity)
2476 {
2477 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
2478 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
2479 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
2480 	int rq_state;
2481 	int sq_state;
2482 	int err;
2483 
2484 	switch (raw_qp_param->operation) {
2485 	case MLX5_CMD_OP_RST2INIT_QP:
2486 		rq_state = MLX5_RQC_STATE_RDY;
2487 		sq_state = MLX5_SQC_STATE_RDY;
2488 		break;
2489 	case MLX5_CMD_OP_2ERR_QP:
2490 		rq_state = MLX5_RQC_STATE_ERR;
2491 		sq_state = MLX5_SQC_STATE_ERR;
2492 		break;
2493 	case MLX5_CMD_OP_2RST_QP:
2494 		rq_state = MLX5_RQC_STATE_RST;
2495 		sq_state = MLX5_SQC_STATE_RST;
2496 		break;
2497 	case MLX5_CMD_OP_INIT2INIT_QP:
2498 	case MLX5_CMD_OP_INIT2RTR_QP:
2499 	case MLX5_CMD_OP_RTR2RTS_QP:
2500 	case MLX5_CMD_OP_RTS2RTS_QP:
2501 		if (raw_qp_param->set_mask)
2502 			return -EINVAL;
2503 		else
2504 			return 0;
2505 	default:
2506 		WARN_ON(1);
2507 		return -EINVAL;
2508 	}
2509 
2510 	if (qp->rq.wqe_cnt) {
2511 		err = modify_raw_packet_qp_rq(dev, rq, rq_state, raw_qp_param);
2512 		if (err)
2513 			return err;
2514 	}
2515 
2516 	if (qp->sq.wqe_cnt) {
2517 		if (tx_affinity) {
2518 			err = modify_raw_packet_tx_affinity(dev->mdev, sq,
2519 							    tx_affinity);
2520 			if (err)
2521 				return err;
2522 		}
2523 
2524 		return modify_raw_packet_qp_sq(dev->mdev, sq, sq_state);
2525 	}
2526 
2527 	return 0;
2528 }
2529 
2530 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
2531 			       const struct ib_qp_attr *attr, int attr_mask,
2532 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
2533 {
2534 	static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
2535 		[MLX5_QP_STATE_RST] = {
2536 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2537 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2538 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_RST2INIT_QP,
2539 		},
2540 		[MLX5_QP_STATE_INIT]  = {
2541 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2542 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2543 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_INIT2INIT_QP,
2544 			[MLX5_QP_STATE_RTR]	= MLX5_CMD_OP_INIT2RTR_QP,
2545 		},
2546 		[MLX5_QP_STATE_RTR]   = {
2547 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2548 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2549 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTR2RTS_QP,
2550 		},
2551 		[MLX5_QP_STATE_RTS]   = {
2552 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2553 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2554 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTS2RTS_QP,
2555 		},
2556 		[MLX5_QP_STATE_SQD] = {
2557 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2558 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2559 		},
2560 		[MLX5_QP_STATE_SQER] = {
2561 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2562 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2563 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQERR2RTS_QP,
2564 		},
2565 		[MLX5_QP_STATE_ERR] = {
2566 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2567 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2568 		}
2569 	};
2570 
2571 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2572 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2573 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
2574 	struct mlx5_ib_cq *send_cq, *recv_cq;
2575 	struct mlx5_qp_context *context;
2576 	struct mlx5_ib_pd *pd;
2577 	struct mlx5_ib_port *mibport = NULL;
2578 	enum mlx5_qp_state mlx5_cur, mlx5_new;
2579 	enum mlx5_qp_optpar optpar;
2580 	int sqd_event;
2581 	int mlx5_st;
2582 	int err;
2583 	u16 op;
2584 	u8 tx_affinity = 0;
2585 
2586 	context = kzalloc(sizeof(*context), GFP_KERNEL);
2587 	if (!context)
2588 		return -ENOMEM;
2589 
2590 	err = to_mlx5_st(ibqp->qp_type);
2591 	if (err < 0) {
2592 		mlx5_ib_dbg(dev, "unsupported qp type %d\n", ibqp->qp_type);
2593 		goto out;
2594 	}
2595 
2596 	context->flags = cpu_to_be32(err << 16);
2597 
2598 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
2599 		context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2600 	} else {
2601 		switch (attr->path_mig_state) {
2602 		case IB_MIG_MIGRATED:
2603 			context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2604 			break;
2605 		case IB_MIG_REARM:
2606 			context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
2607 			break;
2608 		case IB_MIG_ARMED:
2609 			context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
2610 			break;
2611 		}
2612 	}
2613 
2614 	if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2615 		if ((ibqp->qp_type == IB_QPT_RC) ||
2616 		    (ibqp->qp_type == IB_QPT_UD &&
2617 		     !(qp->flags & MLX5_IB_QP_SQPN_QP1)) ||
2618 		    (ibqp->qp_type == IB_QPT_UC) ||
2619 		    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2620 		    (ibqp->qp_type == IB_QPT_XRC_INI) ||
2621 		    (ibqp->qp_type == IB_QPT_XRC_TGT)) {
2622 			if (mlx5_lag_is_active(dev->mdev)) {
2623 				tx_affinity = (unsigned int)atomic_add_return(1,
2624 						&dev->roce.next_port) %
2625 						MLX5_MAX_PORTS + 1;
2626 				context->flags |= cpu_to_be32(tx_affinity << 24);
2627 			}
2628 		}
2629 	}
2630 
2631 	if (is_sqp(ibqp->qp_type)) {
2632 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
2633 	} else if (ibqp->qp_type == IB_QPT_UD ||
2634 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
2635 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
2636 	} else if (attr_mask & IB_QP_PATH_MTU) {
2637 		if (attr->path_mtu < IB_MTU_256 ||
2638 		    attr->path_mtu > IB_MTU_4096) {
2639 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
2640 			err = -EINVAL;
2641 			goto out;
2642 		}
2643 		context->mtu_msgmax = (attr->path_mtu << 5) |
2644 				      (u8)MLX5_CAP_GEN(dev->mdev, log_max_msg);
2645 	}
2646 
2647 	if (attr_mask & IB_QP_DEST_QPN)
2648 		context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
2649 
2650 	if (attr_mask & IB_QP_PKEY_INDEX)
2651 		context->pri_path.pkey_index = cpu_to_be16(attr->pkey_index);
2652 
2653 	/* todo implement counter_index functionality */
2654 
2655 	if (is_sqp(ibqp->qp_type))
2656 		context->pri_path.port = qp->port;
2657 
2658 	if (attr_mask & IB_QP_PORT)
2659 		context->pri_path.port = attr->port_num;
2660 
2661 	if (attr_mask & IB_QP_AV) {
2662 		err = mlx5_set_path(dev, qp, &attr->ah_attr, &context->pri_path,
2663 				    attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
2664 				    attr_mask, 0, attr, false);
2665 		if (err)
2666 			goto out;
2667 	}
2668 
2669 	if (attr_mask & IB_QP_TIMEOUT)
2670 		context->pri_path.ackto_lt |= attr->timeout << 3;
2671 
2672 	if (attr_mask & IB_QP_ALT_PATH) {
2673 		err = mlx5_set_path(dev, qp, &attr->alt_ah_attr,
2674 				    &context->alt_path,
2675 				    attr->alt_port_num,
2676 				    attr_mask | IB_QP_PKEY_INDEX | IB_QP_TIMEOUT,
2677 				    0, attr, true);
2678 		if (err)
2679 			goto out;
2680 	}
2681 
2682 	pd = get_pd(qp);
2683 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
2684 		&send_cq, &recv_cq);
2685 
2686 	context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
2687 	context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
2688 	context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
2689 	context->params1  = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
2690 
2691 	if (attr_mask & IB_QP_RNR_RETRY)
2692 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2693 
2694 	if (attr_mask & IB_QP_RETRY_CNT)
2695 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2696 
2697 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2698 		if (attr->max_rd_atomic)
2699 			context->params1 |=
2700 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2701 	}
2702 
2703 	if (attr_mask & IB_QP_SQ_PSN)
2704 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2705 
2706 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2707 		if (attr->max_dest_rd_atomic)
2708 			context->params2 |=
2709 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2710 	}
2711 
2712 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
2713 		context->params2 |= to_mlx5_access_flags(qp, attr, attr_mask);
2714 
2715 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
2716 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2717 
2718 	if (attr_mask & IB_QP_RQ_PSN)
2719 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2720 
2721 	if (attr_mask & IB_QP_QKEY)
2722 		context->qkey = cpu_to_be32(attr->qkey);
2723 
2724 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2725 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2726 
2727 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2728 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2729 		sqd_event = 1;
2730 	else
2731 		sqd_event = 0;
2732 
2733 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2734 		u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
2735 			       qp->port) - 1;
2736 		mibport = &dev->port[port_num];
2737 		context->qp_counter_set_usr_page |=
2738 			cpu_to_be32((u32)(mibport->q_cnt_id) << 24);
2739 	}
2740 
2741 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2742 		context->sq_crq_size |= cpu_to_be16(1 << 4);
2743 
2744 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
2745 		context->deth_sqpn = cpu_to_be32(1);
2746 
2747 	mlx5_cur = to_mlx5_state(cur_state);
2748 	mlx5_new = to_mlx5_state(new_state);
2749 	mlx5_st = to_mlx5_st(ibqp->qp_type);
2750 	if (mlx5_st < 0)
2751 		goto out;
2752 
2753 	/* If moving to a reset or error state, we must disable page faults on
2754 	 * this QP and flush all current page faults. Otherwise a stale page
2755 	 * fault may attempt to work on this QP after it is reset and moved
2756 	 * again to RTS, and may cause the driver and the device to get out of
2757 	 * sync. */
2758 	if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2759 	    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR) &&
2760 	    (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
2761 		mlx5_ib_qp_disable_pagefaults(qp);
2762 
2763 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
2764 	    !optab[mlx5_cur][mlx5_new])
2765 		goto out;
2766 
2767 	op = optab[mlx5_cur][mlx5_new];
2768 	optpar = ib_mask_to_mlx5_opt(attr_mask);
2769 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
2770 
2771 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
2772 		struct mlx5_modify_raw_qp_param raw_qp_param = {};
2773 
2774 		raw_qp_param.operation = op;
2775 		if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2776 			raw_qp_param.rq_q_ctr_id = mibport->q_cnt_id;
2777 			raw_qp_param.set_mask |= MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID;
2778 		}
2779 		err = modify_raw_packet_qp(dev, qp, &raw_qp_param, tx_affinity);
2780 	} else {
2781 		err = mlx5_core_qp_modify(dev->mdev, op, optpar, context,
2782 					  &base->mqp);
2783 	}
2784 
2785 	if (err)
2786 		goto out;
2787 
2788 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT &&
2789 	    (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
2790 		mlx5_ib_qp_enable_pagefaults(qp);
2791 
2792 	qp->state = new_state;
2793 
2794 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2795 		qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
2796 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2797 		qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
2798 	if (attr_mask & IB_QP_PORT)
2799 		qp->port = attr->port_num;
2800 	if (attr_mask & IB_QP_ALT_PATH)
2801 		qp->trans_qp.alt_port = attr->alt_port_num;
2802 
2803 	/*
2804 	 * If we moved a kernel QP to RESET, clean up all old CQ
2805 	 * entries and reinitialize the QP.
2806 	 */
2807 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
2808 		mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2809 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
2810 		if (send_cq != recv_cq)
2811 			mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
2812 
2813 		qp->rq.head = 0;
2814 		qp->rq.tail = 0;
2815 		qp->sq.head = 0;
2816 		qp->sq.tail = 0;
2817 		qp->sq.cur_post = 0;
2818 		qp->sq.last_poll = 0;
2819 		qp->db.db[MLX5_RCV_DBR] = 0;
2820 		qp->db.db[MLX5_SND_DBR] = 0;
2821 	}
2822 
2823 out:
2824 	kfree(context);
2825 	return err;
2826 }
2827 
2828 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2829 		      int attr_mask, struct ib_udata *udata)
2830 {
2831 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2832 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2833 	enum ib_qp_type qp_type;
2834 	enum ib_qp_state cur_state, new_state;
2835 	int err = -EINVAL;
2836 	int port;
2837 	enum rdma_link_layer ll = IB_LINK_LAYER_UNSPECIFIED;
2838 
2839 	if (ibqp->rwq_ind_tbl)
2840 		return -ENOSYS;
2841 
2842 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
2843 		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
2844 
2845 	qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ?
2846 		IB_QPT_GSI : ibqp->qp_type;
2847 
2848 	mutex_lock(&qp->mutex);
2849 
2850 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2851 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2852 
2853 	if (!(cur_state == new_state && cur_state == IB_QPS_RESET)) {
2854 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2855 		ll = dev->ib_dev.get_link_layer(&dev->ib_dev, port);
2856 	}
2857 
2858 	if (qp_type != MLX5_IB_QPT_REG_UMR &&
2859 	    !ib_modify_qp_is_ok(cur_state, new_state, qp_type, attr_mask, ll)) {
2860 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
2861 			    cur_state, new_state, ibqp->qp_type, attr_mask);
2862 		goto out;
2863 	}
2864 
2865 	if ((attr_mask & IB_QP_PORT) &&
2866 	    (attr->port_num == 0 ||
2867 	     attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports))) {
2868 		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
2869 			    attr->port_num, dev->num_ports);
2870 		goto out;
2871 	}
2872 
2873 	if (attr_mask & IB_QP_PKEY_INDEX) {
2874 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2875 		if (attr->pkey_index >=
2876 		    dev->mdev->port_caps[port - 1].pkey_table_len) {
2877 			mlx5_ib_dbg(dev, "invalid pkey index %d\n",
2878 				    attr->pkey_index);
2879 			goto out;
2880 		}
2881 	}
2882 
2883 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2884 	    attr->max_rd_atomic >
2885 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
2886 		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
2887 			    attr->max_rd_atomic);
2888 		goto out;
2889 	}
2890 
2891 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2892 	    attr->max_dest_rd_atomic >
2893 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
2894 		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
2895 			    attr->max_dest_rd_atomic);
2896 		goto out;
2897 	}
2898 
2899 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2900 		err = 0;
2901 		goto out;
2902 	}
2903 
2904 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2905 
2906 out:
2907 	mutex_unlock(&qp->mutex);
2908 	return err;
2909 }
2910 
2911 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2912 {
2913 	struct mlx5_ib_cq *cq;
2914 	unsigned cur;
2915 
2916 	cur = wq->head - wq->tail;
2917 	if (likely(cur + nreq < wq->max_post))
2918 		return 0;
2919 
2920 	cq = to_mcq(ib_cq);
2921 	spin_lock(&cq->lock);
2922 	cur = wq->head - wq->tail;
2923 	spin_unlock(&cq->lock);
2924 
2925 	return cur + nreq >= wq->max_post;
2926 }
2927 
2928 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
2929 					  u64 remote_addr, u32 rkey)
2930 {
2931 	rseg->raddr    = cpu_to_be64(remote_addr);
2932 	rseg->rkey     = cpu_to_be32(rkey);
2933 	rseg->reserved = 0;
2934 }
2935 
2936 static void *set_eth_seg(struct mlx5_wqe_eth_seg *eseg,
2937 			 struct ib_send_wr *wr, void *qend,
2938 			 struct mlx5_ib_qp *qp, int *size)
2939 {
2940 	void *seg = eseg;
2941 
2942 	memset(eseg, 0, sizeof(struct mlx5_wqe_eth_seg));
2943 
2944 	if (wr->send_flags & IB_SEND_IP_CSUM)
2945 		eseg->cs_flags = MLX5_ETH_WQE_L3_CSUM |
2946 				 MLX5_ETH_WQE_L4_CSUM;
2947 
2948 	seg += sizeof(struct mlx5_wqe_eth_seg);
2949 	*size += sizeof(struct mlx5_wqe_eth_seg) / 16;
2950 
2951 	if (wr->opcode == IB_WR_LSO) {
2952 		struct ib_ud_wr *ud_wr = container_of(wr, struct ib_ud_wr, wr);
2953 		int size_of_inl_hdr_start = sizeof(eseg->inline_hdr_start);
2954 		u64 left, leftlen, copysz;
2955 		void *pdata = ud_wr->header;
2956 
2957 		left = ud_wr->hlen;
2958 		eseg->mss = cpu_to_be16(ud_wr->mss);
2959 		eseg->inline_hdr_sz = cpu_to_be16(left);
2960 
2961 		/*
2962 		 * check if there is space till the end of queue, if yes,
2963 		 * copy all in one shot, otherwise copy till the end of queue,
2964 		 * rollback and than the copy the left
2965 		 */
2966 		leftlen = qend - (void *)eseg->inline_hdr_start;
2967 		copysz = min_t(u64, leftlen, left);
2968 
2969 		memcpy(seg - size_of_inl_hdr_start, pdata, copysz);
2970 
2971 		if (likely(copysz > size_of_inl_hdr_start)) {
2972 			seg += ALIGN(copysz - size_of_inl_hdr_start, 16);
2973 			*size += ALIGN(copysz - size_of_inl_hdr_start, 16) / 16;
2974 		}
2975 
2976 		if (unlikely(copysz < left)) { /* the last wqe in the queue */
2977 			seg = mlx5_get_send_wqe(qp, 0);
2978 			left -= copysz;
2979 			pdata += copysz;
2980 			memcpy(seg, pdata, left);
2981 			seg += ALIGN(left, 16);
2982 			*size += ALIGN(left, 16) / 16;
2983 		}
2984 	}
2985 
2986 	return seg;
2987 }
2988 
2989 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
2990 			     struct ib_send_wr *wr)
2991 {
2992 	memcpy(&dseg->av, &to_mah(ud_wr(wr)->ah)->av, sizeof(struct mlx5_av));
2993 	dseg->av.dqp_dct = cpu_to_be32(ud_wr(wr)->remote_qpn | MLX5_EXTENDED_UD_AV);
2994 	dseg->av.key.qkey.qkey = cpu_to_be32(ud_wr(wr)->remote_qkey);
2995 }
2996 
2997 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
2998 {
2999 	dseg->byte_count = cpu_to_be32(sg->length);
3000 	dseg->lkey       = cpu_to_be32(sg->lkey);
3001 	dseg->addr       = cpu_to_be64(sg->addr);
3002 }
3003 
3004 static __be16 get_klm_octo(int npages)
3005 {
3006 	return cpu_to_be16(ALIGN(npages, 8) / 2);
3007 }
3008 
3009 static __be64 frwr_mkey_mask(void)
3010 {
3011 	u64 result;
3012 
3013 	result = MLX5_MKEY_MASK_LEN		|
3014 		MLX5_MKEY_MASK_PAGE_SIZE	|
3015 		MLX5_MKEY_MASK_START_ADDR	|
3016 		MLX5_MKEY_MASK_EN_RINVAL	|
3017 		MLX5_MKEY_MASK_KEY		|
3018 		MLX5_MKEY_MASK_LR		|
3019 		MLX5_MKEY_MASK_LW		|
3020 		MLX5_MKEY_MASK_RR		|
3021 		MLX5_MKEY_MASK_RW		|
3022 		MLX5_MKEY_MASK_A		|
3023 		MLX5_MKEY_MASK_SMALL_FENCE	|
3024 		MLX5_MKEY_MASK_FREE;
3025 
3026 	return cpu_to_be64(result);
3027 }
3028 
3029 static __be64 sig_mkey_mask(void)
3030 {
3031 	u64 result;
3032 
3033 	result = MLX5_MKEY_MASK_LEN		|
3034 		MLX5_MKEY_MASK_PAGE_SIZE	|
3035 		MLX5_MKEY_MASK_START_ADDR	|
3036 		MLX5_MKEY_MASK_EN_SIGERR	|
3037 		MLX5_MKEY_MASK_EN_RINVAL	|
3038 		MLX5_MKEY_MASK_KEY		|
3039 		MLX5_MKEY_MASK_LR		|
3040 		MLX5_MKEY_MASK_LW		|
3041 		MLX5_MKEY_MASK_RR		|
3042 		MLX5_MKEY_MASK_RW		|
3043 		MLX5_MKEY_MASK_SMALL_FENCE	|
3044 		MLX5_MKEY_MASK_FREE		|
3045 		MLX5_MKEY_MASK_BSF_EN;
3046 
3047 	return cpu_to_be64(result);
3048 }
3049 
3050 static void set_reg_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr,
3051 				struct mlx5_ib_mr *mr)
3052 {
3053 	int ndescs = mr->ndescs;
3054 
3055 	memset(umr, 0, sizeof(*umr));
3056 
3057 	if (mr->access_mode == MLX5_MKC_ACCESS_MODE_KLMS)
3058 		/* KLMs take twice the size of MTTs */
3059 		ndescs *= 2;
3060 
3061 	umr->flags = MLX5_UMR_CHECK_NOT_FREE;
3062 	umr->klm_octowords = get_klm_octo(ndescs);
3063 	umr->mkey_mask = frwr_mkey_mask();
3064 }
3065 
3066 static void set_linv_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr)
3067 {
3068 	memset(umr, 0, sizeof(*umr));
3069 	umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
3070 	umr->flags = 1 << 7;
3071 }
3072 
3073 static __be64 get_umr_reg_mr_mask(void)
3074 {
3075 	u64 result;
3076 
3077 	result = MLX5_MKEY_MASK_LEN		|
3078 		 MLX5_MKEY_MASK_PAGE_SIZE	|
3079 		 MLX5_MKEY_MASK_START_ADDR	|
3080 		 MLX5_MKEY_MASK_PD		|
3081 		 MLX5_MKEY_MASK_LR		|
3082 		 MLX5_MKEY_MASK_LW		|
3083 		 MLX5_MKEY_MASK_KEY		|
3084 		 MLX5_MKEY_MASK_RR		|
3085 		 MLX5_MKEY_MASK_RW		|
3086 		 MLX5_MKEY_MASK_A		|
3087 		 MLX5_MKEY_MASK_FREE;
3088 
3089 	return cpu_to_be64(result);
3090 }
3091 
3092 static __be64 get_umr_unreg_mr_mask(void)
3093 {
3094 	u64 result;
3095 
3096 	result = MLX5_MKEY_MASK_FREE;
3097 
3098 	return cpu_to_be64(result);
3099 }
3100 
3101 static __be64 get_umr_update_mtt_mask(void)
3102 {
3103 	u64 result;
3104 
3105 	result = MLX5_MKEY_MASK_FREE;
3106 
3107 	return cpu_to_be64(result);
3108 }
3109 
3110 static __be64 get_umr_update_translation_mask(void)
3111 {
3112 	u64 result;
3113 
3114 	result = MLX5_MKEY_MASK_LEN |
3115 		 MLX5_MKEY_MASK_PAGE_SIZE |
3116 		 MLX5_MKEY_MASK_START_ADDR |
3117 		 MLX5_MKEY_MASK_KEY |
3118 		 MLX5_MKEY_MASK_FREE;
3119 
3120 	return cpu_to_be64(result);
3121 }
3122 
3123 static __be64 get_umr_update_access_mask(void)
3124 {
3125 	u64 result;
3126 
3127 	result = MLX5_MKEY_MASK_LW |
3128 		 MLX5_MKEY_MASK_RR |
3129 		 MLX5_MKEY_MASK_RW |
3130 		 MLX5_MKEY_MASK_A |
3131 		 MLX5_MKEY_MASK_KEY |
3132 		 MLX5_MKEY_MASK_FREE;
3133 
3134 	return cpu_to_be64(result);
3135 }
3136 
3137 static __be64 get_umr_update_pd_mask(void)
3138 {
3139 	u64 result;
3140 
3141 	result = MLX5_MKEY_MASK_PD |
3142 		 MLX5_MKEY_MASK_KEY |
3143 		 MLX5_MKEY_MASK_FREE;
3144 
3145 	return cpu_to_be64(result);
3146 }
3147 
3148 static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
3149 				struct ib_send_wr *wr)
3150 {
3151 	struct mlx5_umr_wr *umrwr = umr_wr(wr);
3152 
3153 	memset(umr, 0, sizeof(*umr));
3154 
3155 	if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
3156 		umr->flags = MLX5_UMR_CHECK_FREE; /* fail if free */
3157 	else
3158 		umr->flags = MLX5_UMR_CHECK_NOT_FREE; /* fail if not free */
3159 
3160 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UNREG)) {
3161 		umr->klm_octowords = get_klm_octo(umrwr->npages);
3162 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT) {
3163 			umr->mkey_mask = get_umr_update_mtt_mask();
3164 			umr->bsf_octowords = get_klm_octo(umrwr->target.offset);
3165 			umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
3166 		}
3167 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION)
3168 			umr->mkey_mask |= get_umr_update_translation_mask();
3169 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_ACCESS)
3170 			umr->mkey_mask |= get_umr_update_access_mask();
3171 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_PD)
3172 			umr->mkey_mask |= get_umr_update_pd_mask();
3173 		if (!umr->mkey_mask)
3174 			umr->mkey_mask = get_umr_reg_mr_mask();
3175 	} else {
3176 		umr->mkey_mask = get_umr_unreg_mr_mask();
3177 	}
3178 
3179 	if (!wr->num_sge)
3180 		umr->flags |= MLX5_UMR_INLINE;
3181 }
3182 
3183 static u8 get_umr_flags(int acc)
3184 {
3185 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
3186 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
3187 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
3188 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
3189 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
3190 }
3191 
3192 static void set_reg_mkey_seg(struct mlx5_mkey_seg *seg,
3193 			     struct mlx5_ib_mr *mr,
3194 			     u32 key, int access)
3195 {
3196 	int ndescs = ALIGN(mr->ndescs, 8) >> 1;
3197 
3198 	memset(seg, 0, sizeof(*seg));
3199 
3200 	if (mr->access_mode == MLX5_MKC_ACCESS_MODE_MTT)
3201 		seg->log2_page_size = ilog2(mr->ibmr.page_size);
3202 	else if (mr->access_mode == MLX5_MKC_ACCESS_MODE_KLMS)
3203 		/* KLMs take twice the size of MTTs */
3204 		ndescs *= 2;
3205 
3206 	seg->flags = get_umr_flags(access) | mr->access_mode;
3207 	seg->qpn_mkey7_0 = cpu_to_be32((key & 0xff) | 0xffffff00);
3208 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
3209 	seg->start_addr = cpu_to_be64(mr->ibmr.iova);
3210 	seg->len = cpu_to_be64(mr->ibmr.length);
3211 	seg->xlt_oct_size = cpu_to_be32(ndescs);
3212 }
3213 
3214 static void set_linv_mkey_seg(struct mlx5_mkey_seg *seg)
3215 {
3216 	memset(seg, 0, sizeof(*seg));
3217 	seg->status = MLX5_MKEY_STATUS_FREE;
3218 }
3219 
3220 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr)
3221 {
3222 	struct mlx5_umr_wr *umrwr = umr_wr(wr);
3223 
3224 	memset(seg, 0, sizeof(*seg));
3225 	if (wr->send_flags & MLX5_IB_SEND_UMR_UNREG) {
3226 		seg->status = MLX5_MKEY_STATUS_FREE;
3227 		return;
3228 	}
3229 
3230 	seg->flags = convert_access(umrwr->access_flags);
3231 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT)) {
3232 		if (umrwr->pd)
3233 			seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
3234 		seg->start_addr = cpu_to_be64(umrwr->target.virt_addr);
3235 	}
3236 	seg->len = cpu_to_be64(umrwr->length);
3237 	seg->log2_page_size = umrwr->page_shift;
3238 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
3239 				       mlx5_mkey_variant(umrwr->mkey));
3240 }
3241 
3242 static void set_reg_data_seg(struct mlx5_wqe_data_seg *dseg,
3243 			     struct mlx5_ib_mr *mr,
3244 			     struct mlx5_ib_pd *pd)
3245 {
3246 	int bcount = mr->desc_size * mr->ndescs;
3247 
3248 	dseg->addr = cpu_to_be64(mr->desc_map);
3249 	dseg->byte_count = cpu_to_be32(ALIGN(bcount, 64));
3250 	dseg->lkey = cpu_to_be32(pd->ibpd.local_dma_lkey);
3251 }
3252 
3253 static __be32 send_ieth(struct ib_send_wr *wr)
3254 {
3255 	switch (wr->opcode) {
3256 	case IB_WR_SEND_WITH_IMM:
3257 	case IB_WR_RDMA_WRITE_WITH_IMM:
3258 		return wr->ex.imm_data;
3259 
3260 	case IB_WR_SEND_WITH_INV:
3261 		return cpu_to_be32(wr->ex.invalidate_rkey);
3262 
3263 	default:
3264 		return 0;
3265 	}
3266 }
3267 
3268 static u8 calc_sig(void *wqe, int size)
3269 {
3270 	u8 *p = wqe;
3271 	u8 res = 0;
3272 	int i;
3273 
3274 	for (i = 0; i < size; i++)
3275 		res ^= p[i];
3276 
3277 	return ~res;
3278 }
3279 
3280 static u8 wq_sig(void *wqe)
3281 {
3282 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
3283 }
3284 
3285 static int set_data_inl_seg(struct mlx5_ib_qp *qp, struct ib_send_wr *wr,
3286 			    void *wqe, int *sz)
3287 {
3288 	struct mlx5_wqe_inline_seg *seg;
3289 	void *qend = qp->sq.qend;
3290 	void *addr;
3291 	int inl = 0;
3292 	int copy;
3293 	int len;
3294 	int i;
3295 
3296 	seg = wqe;
3297 	wqe += sizeof(*seg);
3298 	for (i = 0; i < wr->num_sge; i++) {
3299 		addr = (void *)(unsigned long)(wr->sg_list[i].addr);
3300 		len  = wr->sg_list[i].length;
3301 		inl += len;
3302 
3303 		if (unlikely(inl > qp->max_inline_data))
3304 			return -ENOMEM;
3305 
3306 		if (unlikely(wqe + len > qend)) {
3307 			copy = qend - wqe;
3308 			memcpy(wqe, addr, copy);
3309 			addr += copy;
3310 			len -= copy;
3311 			wqe = mlx5_get_send_wqe(qp, 0);
3312 		}
3313 		memcpy(wqe, addr, len);
3314 		wqe += len;
3315 	}
3316 
3317 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
3318 
3319 	*sz = ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
3320 
3321 	return 0;
3322 }
3323 
3324 static u16 prot_field_size(enum ib_signature_type type)
3325 {
3326 	switch (type) {
3327 	case IB_SIG_TYPE_T10_DIF:
3328 		return MLX5_DIF_SIZE;
3329 	default:
3330 		return 0;
3331 	}
3332 }
3333 
3334 static u8 bs_selector(int block_size)
3335 {
3336 	switch (block_size) {
3337 	case 512:	    return 0x1;
3338 	case 520:	    return 0x2;
3339 	case 4096:	    return 0x3;
3340 	case 4160:	    return 0x4;
3341 	case 1073741824:    return 0x5;
3342 	default:	    return 0;
3343 	}
3344 }
3345 
3346 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
3347 			      struct mlx5_bsf_inl *inl)
3348 {
3349 	/* Valid inline section and allow BSF refresh */
3350 	inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
3351 				       MLX5_BSF_REFRESH_DIF);
3352 	inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
3353 	inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
3354 	/* repeating block */
3355 	inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
3356 	inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
3357 			MLX5_DIF_CRC : MLX5_DIF_IPCS;
3358 
3359 	if (domain->sig.dif.ref_remap)
3360 		inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
3361 
3362 	if (domain->sig.dif.app_escape) {
3363 		if (domain->sig.dif.ref_escape)
3364 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
3365 		else
3366 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
3367 	}
3368 
3369 	inl->dif_app_bitmask_check =
3370 		cpu_to_be16(domain->sig.dif.apptag_check_mask);
3371 }
3372 
3373 static int mlx5_set_bsf(struct ib_mr *sig_mr,
3374 			struct ib_sig_attrs *sig_attrs,
3375 			struct mlx5_bsf *bsf, u32 data_size)
3376 {
3377 	struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
3378 	struct mlx5_bsf_basic *basic = &bsf->basic;
3379 	struct ib_sig_domain *mem = &sig_attrs->mem;
3380 	struct ib_sig_domain *wire = &sig_attrs->wire;
3381 
3382 	memset(bsf, 0, sizeof(*bsf));
3383 
3384 	/* Basic + Extended + Inline */
3385 	basic->bsf_size_sbs = 1 << 7;
3386 	/* Input domain check byte mask */
3387 	basic->check_byte_mask = sig_attrs->check_mask;
3388 	basic->raw_data_size = cpu_to_be32(data_size);
3389 
3390 	/* Memory domain */
3391 	switch (sig_attrs->mem.sig_type) {
3392 	case IB_SIG_TYPE_NONE:
3393 		break;
3394 	case IB_SIG_TYPE_T10_DIF:
3395 		basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
3396 		basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
3397 		mlx5_fill_inl_bsf(mem, &bsf->m_inl);
3398 		break;
3399 	default:
3400 		return -EINVAL;
3401 	}
3402 
3403 	/* Wire domain */
3404 	switch (sig_attrs->wire.sig_type) {
3405 	case IB_SIG_TYPE_NONE:
3406 		break;
3407 	case IB_SIG_TYPE_T10_DIF:
3408 		if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
3409 		    mem->sig_type == wire->sig_type) {
3410 			/* Same block structure */
3411 			basic->bsf_size_sbs |= 1 << 4;
3412 			if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
3413 				basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
3414 			if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
3415 				basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
3416 			if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
3417 				basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
3418 		} else
3419 			basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
3420 
3421 		basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
3422 		mlx5_fill_inl_bsf(wire, &bsf->w_inl);
3423 		break;
3424 	default:
3425 		return -EINVAL;
3426 	}
3427 
3428 	return 0;
3429 }
3430 
3431 static int set_sig_data_segment(struct ib_sig_handover_wr *wr,
3432 				struct mlx5_ib_qp *qp, void **seg, int *size)
3433 {
3434 	struct ib_sig_attrs *sig_attrs = wr->sig_attrs;
3435 	struct ib_mr *sig_mr = wr->sig_mr;
3436 	struct mlx5_bsf *bsf;
3437 	u32 data_len = wr->wr.sg_list->length;
3438 	u32 data_key = wr->wr.sg_list->lkey;
3439 	u64 data_va = wr->wr.sg_list->addr;
3440 	int ret;
3441 	int wqe_size;
3442 
3443 	if (!wr->prot ||
3444 	    (data_key == wr->prot->lkey &&
3445 	     data_va == wr->prot->addr &&
3446 	     data_len == wr->prot->length)) {
3447 		/**
3448 		 * Source domain doesn't contain signature information
3449 		 * or data and protection are interleaved in memory.
3450 		 * So need construct:
3451 		 *                  ------------------
3452 		 *                 |     data_klm     |
3453 		 *                  ------------------
3454 		 *                 |       BSF        |
3455 		 *                  ------------------
3456 		 **/
3457 		struct mlx5_klm *data_klm = *seg;
3458 
3459 		data_klm->bcount = cpu_to_be32(data_len);
3460 		data_klm->key = cpu_to_be32(data_key);
3461 		data_klm->va = cpu_to_be64(data_va);
3462 		wqe_size = ALIGN(sizeof(*data_klm), 64);
3463 	} else {
3464 		/**
3465 		 * Source domain contains signature information
3466 		 * So need construct a strided block format:
3467 		 *               ---------------------------
3468 		 *              |     stride_block_ctrl     |
3469 		 *               ---------------------------
3470 		 *              |          data_klm         |
3471 		 *               ---------------------------
3472 		 *              |          prot_klm         |
3473 		 *               ---------------------------
3474 		 *              |             BSF           |
3475 		 *               ---------------------------
3476 		 **/
3477 		struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
3478 		struct mlx5_stride_block_entry *data_sentry;
3479 		struct mlx5_stride_block_entry *prot_sentry;
3480 		u32 prot_key = wr->prot->lkey;
3481 		u64 prot_va = wr->prot->addr;
3482 		u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
3483 		int prot_size;
3484 
3485 		sblock_ctrl = *seg;
3486 		data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
3487 		prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
3488 
3489 		prot_size = prot_field_size(sig_attrs->mem.sig_type);
3490 		if (!prot_size) {
3491 			pr_err("Bad block size given: %u\n", block_size);
3492 			return -EINVAL;
3493 		}
3494 		sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
3495 							    prot_size);
3496 		sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
3497 		sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
3498 		sblock_ctrl->num_entries = cpu_to_be16(2);
3499 
3500 		data_sentry->bcount = cpu_to_be16(block_size);
3501 		data_sentry->key = cpu_to_be32(data_key);
3502 		data_sentry->va = cpu_to_be64(data_va);
3503 		data_sentry->stride = cpu_to_be16(block_size);
3504 
3505 		prot_sentry->bcount = cpu_to_be16(prot_size);
3506 		prot_sentry->key = cpu_to_be32(prot_key);
3507 		prot_sentry->va = cpu_to_be64(prot_va);
3508 		prot_sentry->stride = cpu_to_be16(prot_size);
3509 
3510 		wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
3511 				 sizeof(*prot_sentry), 64);
3512 	}
3513 
3514 	*seg += wqe_size;
3515 	*size += wqe_size / 16;
3516 	if (unlikely((*seg == qp->sq.qend)))
3517 		*seg = mlx5_get_send_wqe(qp, 0);
3518 
3519 	bsf = *seg;
3520 	ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
3521 	if (ret)
3522 		return -EINVAL;
3523 
3524 	*seg += sizeof(*bsf);
3525 	*size += sizeof(*bsf) / 16;
3526 	if (unlikely((*seg == qp->sq.qend)))
3527 		*seg = mlx5_get_send_wqe(qp, 0);
3528 
3529 	return 0;
3530 }
3531 
3532 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
3533 				 struct ib_sig_handover_wr *wr, u32 nelements,
3534 				 u32 length, u32 pdn)
3535 {
3536 	struct ib_mr *sig_mr = wr->sig_mr;
3537 	u32 sig_key = sig_mr->rkey;
3538 	u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
3539 
3540 	memset(seg, 0, sizeof(*seg));
3541 
3542 	seg->flags = get_umr_flags(wr->access_flags) |
3543 				   MLX5_MKC_ACCESS_MODE_KLMS;
3544 	seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
3545 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
3546 				    MLX5_MKEY_BSF_EN | pdn);
3547 	seg->len = cpu_to_be64(length);
3548 	seg->xlt_oct_size = cpu_to_be32(be16_to_cpu(get_klm_octo(nelements)));
3549 	seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
3550 }
3551 
3552 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
3553 				u32 nelements)
3554 {
3555 	memset(umr, 0, sizeof(*umr));
3556 
3557 	umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
3558 	umr->klm_octowords = get_klm_octo(nelements);
3559 	umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
3560 	umr->mkey_mask = sig_mkey_mask();
3561 }
3562 
3563 
3564 static int set_sig_umr_wr(struct ib_send_wr *send_wr, struct mlx5_ib_qp *qp,
3565 			  void **seg, int *size)
3566 {
3567 	struct ib_sig_handover_wr *wr = sig_handover_wr(send_wr);
3568 	struct mlx5_ib_mr *sig_mr = to_mmr(wr->sig_mr);
3569 	u32 pdn = get_pd(qp)->pdn;
3570 	u32 klm_oct_size;
3571 	int region_len, ret;
3572 
3573 	if (unlikely(wr->wr.num_sge != 1) ||
3574 	    unlikely(wr->access_flags & IB_ACCESS_REMOTE_ATOMIC) ||
3575 	    unlikely(!sig_mr->sig) || unlikely(!qp->signature_en) ||
3576 	    unlikely(!sig_mr->sig->sig_status_checked))
3577 		return -EINVAL;
3578 
3579 	/* length of the protected region, data + protection */
3580 	region_len = wr->wr.sg_list->length;
3581 	if (wr->prot &&
3582 	    (wr->prot->lkey != wr->wr.sg_list->lkey  ||
3583 	     wr->prot->addr != wr->wr.sg_list->addr  ||
3584 	     wr->prot->length != wr->wr.sg_list->length))
3585 		region_len += wr->prot->length;
3586 
3587 	/**
3588 	 * KLM octoword size - if protection was provided
3589 	 * then we use strided block format (3 octowords),
3590 	 * else we use single KLM (1 octoword)
3591 	 **/
3592 	klm_oct_size = wr->prot ? 3 : 1;
3593 
3594 	set_sig_umr_segment(*seg, klm_oct_size);
3595 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3596 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3597 	if (unlikely((*seg == qp->sq.qend)))
3598 		*seg = mlx5_get_send_wqe(qp, 0);
3599 
3600 	set_sig_mkey_segment(*seg, wr, klm_oct_size, region_len, pdn);
3601 	*seg += sizeof(struct mlx5_mkey_seg);
3602 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3603 	if (unlikely((*seg == qp->sq.qend)))
3604 		*seg = mlx5_get_send_wqe(qp, 0);
3605 
3606 	ret = set_sig_data_segment(wr, qp, seg, size);
3607 	if (ret)
3608 		return ret;
3609 
3610 	sig_mr->sig->sig_status_checked = false;
3611 	return 0;
3612 }
3613 
3614 static int set_psv_wr(struct ib_sig_domain *domain,
3615 		      u32 psv_idx, void **seg, int *size)
3616 {
3617 	struct mlx5_seg_set_psv *psv_seg = *seg;
3618 
3619 	memset(psv_seg, 0, sizeof(*psv_seg));
3620 	psv_seg->psv_num = cpu_to_be32(psv_idx);
3621 	switch (domain->sig_type) {
3622 	case IB_SIG_TYPE_NONE:
3623 		break;
3624 	case IB_SIG_TYPE_T10_DIF:
3625 		psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
3626 						     domain->sig.dif.app_tag);
3627 		psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
3628 		break;
3629 	default:
3630 		pr_err("Bad signature type given.\n");
3631 		return 1;
3632 	}
3633 
3634 	*seg += sizeof(*psv_seg);
3635 	*size += sizeof(*psv_seg) / 16;
3636 
3637 	return 0;
3638 }
3639 
3640 static int set_reg_wr(struct mlx5_ib_qp *qp,
3641 		      struct ib_reg_wr *wr,
3642 		      void **seg, int *size)
3643 {
3644 	struct mlx5_ib_mr *mr = to_mmr(wr->mr);
3645 	struct mlx5_ib_pd *pd = to_mpd(qp->ibqp.pd);
3646 
3647 	if (unlikely(wr->wr.send_flags & IB_SEND_INLINE)) {
3648 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
3649 			     "Invalid IB_SEND_INLINE send flag\n");
3650 		return -EINVAL;
3651 	}
3652 
3653 	set_reg_umr_seg(*seg, mr);
3654 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3655 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3656 	if (unlikely((*seg == qp->sq.qend)))
3657 		*seg = mlx5_get_send_wqe(qp, 0);
3658 
3659 	set_reg_mkey_seg(*seg, mr, wr->key, wr->access);
3660 	*seg += sizeof(struct mlx5_mkey_seg);
3661 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3662 	if (unlikely((*seg == qp->sq.qend)))
3663 		*seg = mlx5_get_send_wqe(qp, 0);
3664 
3665 	set_reg_data_seg(*seg, mr, pd);
3666 	*seg += sizeof(struct mlx5_wqe_data_seg);
3667 	*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
3668 
3669 	return 0;
3670 }
3671 
3672 static void set_linv_wr(struct mlx5_ib_qp *qp, void **seg, int *size)
3673 {
3674 	set_linv_umr_seg(*seg);
3675 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3676 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3677 	if (unlikely((*seg == qp->sq.qend)))
3678 		*seg = mlx5_get_send_wqe(qp, 0);
3679 	set_linv_mkey_seg(*seg);
3680 	*seg += sizeof(struct mlx5_mkey_seg);
3681 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3682 	if (unlikely((*seg == qp->sq.qend)))
3683 		*seg = mlx5_get_send_wqe(qp, 0);
3684 }
3685 
3686 static void dump_wqe(struct mlx5_ib_qp *qp, int idx, int size_16)
3687 {
3688 	__be32 *p = NULL;
3689 	int tidx = idx;
3690 	int i, j;
3691 
3692 	pr_debug("dump wqe at %p\n", mlx5_get_send_wqe(qp, tidx));
3693 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
3694 		if ((i & 0xf) == 0) {
3695 			void *buf = mlx5_get_send_wqe(qp, tidx);
3696 			tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
3697 			p = buf;
3698 			j = 0;
3699 		}
3700 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
3701 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
3702 			 be32_to_cpu(p[j + 3]));
3703 	}
3704 }
3705 
3706 static void mlx5_bf_copy(u64 __iomem *dst, u64 *src,
3707 			 unsigned bytecnt, struct mlx5_ib_qp *qp)
3708 {
3709 	while (bytecnt > 0) {
3710 		__iowrite64_copy(dst++, src++, 8);
3711 		__iowrite64_copy(dst++, src++, 8);
3712 		__iowrite64_copy(dst++, src++, 8);
3713 		__iowrite64_copy(dst++, src++, 8);
3714 		__iowrite64_copy(dst++, src++, 8);
3715 		__iowrite64_copy(dst++, src++, 8);
3716 		__iowrite64_copy(dst++, src++, 8);
3717 		__iowrite64_copy(dst++, src++, 8);
3718 		bytecnt -= 64;
3719 		if (unlikely(src == qp->sq.qend))
3720 			src = mlx5_get_send_wqe(qp, 0);
3721 	}
3722 }
3723 
3724 static u8 get_fence(u8 fence, struct ib_send_wr *wr)
3725 {
3726 	if (unlikely(wr->opcode == IB_WR_LOCAL_INV &&
3727 		     wr->send_flags & IB_SEND_FENCE))
3728 		return MLX5_FENCE_MODE_STRONG_ORDERING;
3729 
3730 	if (unlikely(fence)) {
3731 		if (wr->send_flags & IB_SEND_FENCE)
3732 			return MLX5_FENCE_MODE_SMALL_AND_FENCE;
3733 		else
3734 			return fence;
3735 	} else if (unlikely(wr->send_flags & IB_SEND_FENCE)) {
3736 		return MLX5_FENCE_MODE_FENCE;
3737 	}
3738 
3739 	return 0;
3740 }
3741 
3742 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
3743 		     struct mlx5_wqe_ctrl_seg **ctrl,
3744 		     struct ib_send_wr *wr, unsigned *idx,
3745 		     int *size, int nreq)
3746 {
3747 	if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)))
3748 		return -ENOMEM;
3749 
3750 	*idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
3751 	*seg = mlx5_get_send_wqe(qp, *idx);
3752 	*ctrl = *seg;
3753 	*(uint32_t *)(*seg + 8) = 0;
3754 	(*ctrl)->imm = send_ieth(wr);
3755 	(*ctrl)->fm_ce_se = qp->sq_signal_bits |
3756 		(wr->send_flags & IB_SEND_SIGNALED ?
3757 		 MLX5_WQE_CTRL_CQ_UPDATE : 0) |
3758 		(wr->send_flags & IB_SEND_SOLICITED ?
3759 		 MLX5_WQE_CTRL_SOLICITED : 0);
3760 
3761 	*seg += sizeof(**ctrl);
3762 	*size = sizeof(**ctrl) / 16;
3763 
3764 	return 0;
3765 }
3766 
3767 static void finish_wqe(struct mlx5_ib_qp *qp,
3768 		       struct mlx5_wqe_ctrl_seg *ctrl,
3769 		       u8 size, unsigned idx, u64 wr_id,
3770 		       int nreq, u8 fence, u8 next_fence,
3771 		       u32 mlx5_opcode)
3772 {
3773 	u8 opmod = 0;
3774 
3775 	ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
3776 					     mlx5_opcode | ((u32)opmod << 24));
3777 	ctrl->qpn_ds = cpu_to_be32(size | (qp->trans_qp.base.mqp.qpn << 8));
3778 	ctrl->fm_ce_se |= fence;
3779 	qp->fm_cache = next_fence;
3780 	if (unlikely(qp->wq_sig))
3781 		ctrl->signature = wq_sig(ctrl);
3782 
3783 	qp->sq.wrid[idx] = wr_id;
3784 	qp->sq.w_list[idx].opcode = mlx5_opcode;
3785 	qp->sq.wqe_head[idx] = qp->sq.head + nreq;
3786 	qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
3787 	qp->sq.w_list[idx].next = qp->sq.cur_post;
3788 }
3789 
3790 
3791 int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
3792 		      struct ib_send_wr **bad_wr)
3793 {
3794 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
3795 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3796 	struct mlx5_core_dev *mdev = dev->mdev;
3797 	struct mlx5_ib_qp *qp;
3798 	struct mlx5_ib_mr *mr;
3799 	struct mlx5_wqe_data_seg *dpseg;
3800 	struct mlx5_wqe_xrc_seg *xrc;
3801 	struct mlx5_bf *bf;
3802 	int uninitialized_var(size);
3803 	void *qend;
3804 	unsigned long flags;
3805 	unsigned idx;
3806 	int err = 0;
3807 	int inl = 0;
3808 	int num_sge;
3809 	void *seg;
3810 	int nreq;
3811 	int i;
3812 	u8 next_fence = 0;
3813 	u8 fence;
3814 
3815 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
3816 		return mlx5_ib_gsi_post_send(ibqp, wr, bad_wr);
3817 
3818 	qp = to_mqp(ibqp);
3819 	bf = qp->bf;
3820 	qend = qp->sq.qend;
3821 
3822 	spin_lock_irqsave(&qp->sq.lock, flags);
3823 
3824 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
3825 		err = -EIO;
3826 		*bad_wr = wr;
3827 		nreq = 0;
3828 		goto out;
3829 	}
3830 
3831 	for (nreq = 0; wr; nreq++, wr = wr->next) {
3832 		if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
3833 			mlx5_ib_warn(dev, "\n");
3834 			err = -EINVAL;
3835 			*bad_wr = wr;
3836 			goto out;
3837 		}
3838 
3839 		fence = qp->fm_cache;
3840 		num_sge = wr->num_sge;
3841 		if (unlikely(num_sge > qp->sq.max_gs)) {
3842 			mlx5_ib_warn(dev, "\n");
3843 			err = -EINVAL;
3844 			*bad_wr = wr;
3845 			goto out;
3846 		}
3847 
3848 		err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, nreq);
3849 		if (err) {
3850 			mlx5_ib_warn(dev, "\n");
3851 			err = -ENOMEM;
3852 			*bad_wr = wr;
3853 			goto out;
3854 		}
3855 
3856 		switch (ibqp->qp_type) {
3857 		case IB_QPT_XRC_INI:
3858 			xrc = seg;
3859 			seg += sizeof(*xrc);
3860 			size += sizeof(*xrc) / 16;
3861 			/* fall through */
3862 		case IB_QPT_RC:
3863 			switch (wr->opcode) {
3864 			case IB_WR_RDMA_READ:
3865 			case IB_WR_RDMA_WRITE:
3866 			case IB_WR_RDMA_WRITE_WITH_IMM:
3867 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
3868 					      rdma_wr(wr)->rkey);
3869 				seg += sizeof(struct mlx5_wqe_raddr_seg);
3870 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
3871 				break;
3872 
3873 			case IB_WR_ATOMIC_CMP_AND_SWP:
3874 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3875 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3876 				mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
3877 				err = -ENOSYS;
3878 				*bad_wr = wr;
3879 				goto out;
3880 
3881 			case IB_WR_LOCAL_INV:
3882 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3883 				qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
3884 				ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
3885 				set_linv_wr(qp, &seg, &size);
3886 				num_sge = 0;
3887 				break;
3888 
3889 			case IB_WR_REG_MR:
3890 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3891 				qp->sq.wr_data[idx] = IB_WR_REG_MR;
3892 				ctrl->imm = cpu_to_be32(reg_wr(wr)->key);
3893 				err = set_reg_wr(qp, reg_wr(wr), &seg, &size);
3894 				if (err) {
3895 					*bad_wr = wr;
3896 					goto out;
3897 				}
3898 				num_sge = 0;
3899 				break;
3900 
3901 			case IB_WR_REG_SIG_MR:
3902 				qp->sq.wr_data[idx] = IB_WR_REG_SIG_MR;
3903 				mr = to_mmr(sig_handover_wr(wr)->sig_mr);
3904 
3905 				ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
3906 				err = set_sig_umr_wr(wr, qp, &seg, &size);
3907 				if (err) {
3908 					mlx5_ib_warn(dev, "\n");
3909 					*bad_wr = wr;
3910 					goto out;
3911 				}
3912 
3913 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3914 					   nreq, get_fence(fence, wr),
3915 					   next_fence, MLX5_OPCODE_UMR);
3916 				/*
3917 				 * SET_PSV WQEs are not signaled and solicited
3918 				 * on error
3919 				 */
3920 				wr->send_flags &= ~IB_SEND_SIGNALED;
3921 				wr->send_flags |= IB_SEND_SOLICITED;
3922 				err = begin_wqe(qp, &seg, &ctrl, wr,
3923 						&idx, &size, nreq);
3924 				if (err) {
3925 					mlx5_ib_warn(dev, "\n");
3926 					err = -ENOMEM;
3927 					*bad_wr = wr;
3928 					goto out;
3929 				}
3930 
3931 				err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->mem,
3932 						 mr->sig->psv_memory.psv_idx, &seg,
3933 						 &size);
3934 				if (err) {
3935 					mlx5_ib_warn(dev, "\n");
3936 					*bad_wr = wr;
3937 					goto out;
3938 				}
3939 
3940 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3941 					   nreq, get_fence(fence, wr),
3942 					   next_fence, MLX5_OPCODE_SET_PSV);
3943 				err = begin_wqe(qp, &seg, &ctrl, wr,
3944 						&idx, &size, nreq);
3945 				if (err) {
3946 					mlx5_ib_warn(dev, "\n");
3947 					err = -ENOMEM;
3948 					*bad_wr = wr;
3949 					goto out;
3950 				}
3951 
3952 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3953 				err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->wire,
3954 						 mr->sig->psv_wire.psv_idx, &seg,
3955 						 &size);
3956 				if (err) {
3957 					mlx5_ib_warn(dev, "\n");
3958 					*bad_wr = wr;
3959 					goto out;
3960 				}
3961 
3962 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3963 					   nreq, get_fence(fence, wr),
3964 					   next_fence, MLX5_OPCODE_SET_PSV);
3965 				num_sge = 0;
3966 				goto skip_psv;
3967 
3968 			default:
3969 				break;
3970 			}
3971 			break;
3972 
3973 		case IB_QPT_UC:
3974 			switch (wr->opcode) {
3975 			case IB_WR_RDMA_WRITE:
3976 			case IB_WR_RDMA_WRITE_WITH_IMM:
3977 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
3978 					      rdma_wr(wr)->rkey);
3979 				seg  += sizeof(struct mlx5_wqe_raddr_seg);
3980 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
3981 				break;
3982 
3983 			default:
3984 				break;
3985 			}
3986 			break;
3987 
3988 		case IB_QPT_SMI:
3989 		case MLX5_IB_QPT_HW_GSI:
3990 			set_datagram_seg(seg, wr);
3991 			seg += sizeof(struct mlx5_wqe_datagram_seg);
3992 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
3993 			if (unlikely((seg == qend)))
3994 				seg = mlx5_get_send_wqe(qp, 0);
3995 			break;
3996 		case IB_QPT_UD:
3997 			set_datagram_seg(seg, wr);
3998 			seg += sizeof(struct mlx5_wqe_datagram_seg);
3999 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
4000 
4001 			if (unlikely((seg == qend)))
4002 				seg = mlx5_get_send_wqe(qp, 0);
4003 
4004 			/* handle qp that supports ud offload */
4005 			if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO) {
4006 				struct mlx5_wqe_eth_pad *pad;
4007 
4008 				pad = seg;
4009 				memset(pad, 0, sizeof(struct mlx5_wqe_eth_pad));
4010 				seg += sizeof(struct mlx5_wqe_eth_pad);
4011 				size += sizeof(struct mlx5_wqe_eth_pad) / 16;
4012 
4013 				seg = set_eth_seg(seg, wr, qend, qp, &size);
4014 
4015 				if (unlikely((seg == qend)))
4016 					seg = mlx5_get_send_wqe(qp, 0);
4017 			}
4018 			break;
4019 		case MLX5_IB_QPT_REG_UMR:
4020 			if (wr->opcode != MLX5_IB_WR_UMR) {
4021 				err = -EINVAL;
4022 				mlx5_ib_warn(dev, "bad opcode\n");
4023 				goto out;
4024 			}
4025 			qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
4026 			ctrl->imm = cpu_to_be32(umr_wr(wr)->mkey);
4027 			set_reg_umr_segment(seg, wr);
4028 			seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4029 			size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4030 			if (unlikely((seg == qend)))
4031 				seg = mlx5_get_send_wqe(qp, 0);
4032 			set_reg_mkey_segment(seg, wr);
4033 			seg += sizeof(struct mlx5_mkey_seg);
4034 			size += sizeof(struct mlx5_mkey_seg) / 16;
4035 			if (unlikely((seg == qend)))
4036 				seg = mlx5_get_send_wqe(qp, 0);
4037 			break;
4038 
4039 		default:
4040 			break;
4041 		}
4042 
4043 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
4044 			int uninitialized_var(sz);
4045 
4046 			err = set_data_inl_seg(qp, wr, seg, &sz);
4047 			if (unlikely(err)) {
4048 				mlx5_ib_warn(dev, "\n");
4049 				*bad_wr = wr;
4050 				goto out;
4051 			}
4052 			inl = 1;
4053 			size += sz;
4054 		} else {
4055 			dpseg = seg;
4056 			for (i = 0; i < num_sge; i++) {
4057 				if (unlikely(dpseg == qend)) {
4058 					seg = mlx5_get_send_wqe(qp, 0);
4059 					dpseg = seg;
4060 				}
4061 				if (likely(wr->sg_list[i].length)) {
4062 					set_data_ptr_seg(dpseg, wr->sg_list + i);
4063 					size += sizeof(struct mlx5_wqe_data_seg) / 16;
4064 					dpseg++;
4065 				}
4066 			}
4067 		}
4068 
4069 		finish_wqe(qp, ctrl, size, idx, wr->wr_id, nreq,
4070 			   get_fence(fence, wr), next_fence,
4071 			   mlx5_ib_opcode[wr->opcode]);
4072 skip_psv:
4073 		if (0)
4074 			dump_wqe(qp, idx, size);
4075 	}
4076 
4077 out:
4078 	if (likely(nreq)) {
4079 		qp->sq.head += nreq;
4080 
4081 		/* Make sure that descriptors are written before
4082 		 * updating doorbell record and ringing the doorbell
4083 		 */
4084 		wmb();
4085 
4086 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
4087 
4088 		/* Make sure doorbell record is visible to the HCA before
4089 		 * we hit doorbell */
4090 		wmb();
4091 
4092 		if (bf->need_lock)
4093 			spin_lock(&bf->lock);
4094 		else
4095 			__acquire(&bf->lock);
4096 
4097 		/* TBD enable WC */
4098 		if (0 && nreq == 1 && bf->uuarn && inl && size > 1 && size <= bf->buf_size / 16) {
4099 			mlx5_bf_copy(bf->reg + bf->offset, (u64 *)ctrl, ALIGN(size * 16, 64), qp);
4100 			/* wc_wmb(); */
4101 		} else {
4102 			mlx5_write64((__be32 *)ctrl, bf->regreg + bf->offset,
4103 				     MLX5_GET_DOORBELL_LOCK(&bf->lock32));
4104 			/* Make sure doorbells don't leak out of SQ spinlock
4105 			 * and reach the HCA out of order.
4106 			 */
4107 			mmiowb();
4108 		}
4109 		bf->offset ^= bf->buf_size;
4110 		if (bf->need_lock)
4111 			spin_unlock(&bf->lock);
4112 		else
4113 			__release(&bf->lock);
4114 	}
4115 
4116 	spin_unlock_irqrestore(&qp->sq.lock, flags);
4117 
4118 	return err;
4119 }
4120 
4121 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
4122 {
4123 	sig->signature = calc_sig(sig, size);
4124 }
4125 
4126 int mlx5_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
4127 		      struct ib_recv_wr **bad_wr)
4128 {
4129 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4130 	struct mlx5_wqe_data_seg *scat;
4131 	struct mlx5_rwqe_sig *sig;
4132 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4133 	struct mlx5_core_dev *mdev = dev->mdev;
4134 	unsigned long flags;
4135 	int err = 0;
4136 	int nreq;
4137 	int ind;
4138 	int i;
4139 
4140 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4141 		return mlx5_ib_gsi_post_recv(ibqp, wr, bad_wr);
4142 
4143 	spin_lock_irqsave(&qp->rq.lock, flags);
4144 
4145 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
4146 		err = -EIO;
4147 		*bad_wr = wr;
4148 		nreq = 0;
4149 		goto out;
4150 	}
4151 
4152 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
4153 
4154 	for (nreq = 0; wr; nreq++, wr = wr->next) {
4155 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
4156 			err = -ENOMEM;
4157 			*bad_wr = wr;
4158 			goto out;
4159 		}
4160 
4161 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
4162 			err = -EINVAL;
4163 			*bad_wr = wr;
4164 			goto out;
4165 		}
4166 
4167 		scat = get_recv_wqe(qp, ind);
4168 		if (qp->wq_sig)
4169 			scat++;
4170 
4171 		for (i = 0; i < wr->num_sge; i++)
4172 			set_data_ptr_seg(scat + i, wr->sg_list + i);
4173 
4174 		if (i < qp->rq.max_gs) {
4175 			scat[i].byte_count = 0;
4176 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
4177 			scat[i].addr       = 0;
4178 		}
4179 
4180 		if (qp->wq_sig) {
4181 			sig = (struct mlx5_rwqe_sig *)scat;
4182 			set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
4183 		}
4184 
4185 		qp->rq.wrid[ind] = wr->wr_id;
4186 
4187 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
4188 	}
4189 
4190 out:
4191 	if (likely(nreq)) {
4192 		qp->rq.head += nreq;
4193 
4194 		/* Make sure that descriptors are written before
4195 		 * doorbell record.
4196 		 */
4197 		wmb();
4198 
4199 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
4200 	}
4201 
4202 	spin_unlock_irqrestore(&qp->rq.lock, flags);
4203 
4204 	return err;
4205 }
4206 
4207 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
4208 {
4209 	switch (mlx5_state) {
4210 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
4211 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
4212 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
4213 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
4214 	case MLX5_QP_STATE_SQ_DRAINING:
4215 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
4216 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
4217 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
4218 	default:		     return -1;
4219 	}
4220 }
4221 
4222 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
4223 {
4224 	switch (mlx5_mig_state) {
4225 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
4226 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
4227 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
4228 	default: return -1;
4229 	}
4230 }
4231 
4232 static int to_ib_qp_access_flags(int mlx5_flags)
4233 {
4234 	int ib_flags = 0;
4235 
4236 	if (mlx5_flags & MLX5_QP_BIT_RRE)
4237 		ib_flags |= IB_ACCESS_REMOTE_READ;
4238 	if (mlx5_flags & MLX5_QP_BIT_RWE)
4239 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
4240 	if (mlx5_flags & MLX5_QP_BIT_RAE)
4241 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
4242 
4243 	return ib_flags;
4244 }
4245 
4246 static void to_ib_ah_attr(struct mlx5_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
4247 				struct mlx5_qp_path *path)
4248 {
4249 	struct mlx5_core_dev *dev = ibdev->mdev;
4250 
4251 	memset(ib_ah_attr, 0, sizeof(*ib_ah_attr));
4252 	ib_ah_attr->port_num	  = path->port;
4253 
4254 	if (ib_ah_attr->port_num == 0 ||
4255 	    ib_ah_attr->port_num > MLX5_CAP_GEN(dev, num_ports))
4256 		return;
4257 
4258 	ib_ah_attr->sl = path->dci_cfi_prio_sl & 0xf;
4259 
4260 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
4261 	ib_ah_attr->src_path_bits = path->grh_mlid & 0x7f;
4262 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
4263 	ib_ah_attr->ah_flags      = (path->grh_mlid & (1 << 7)) ? IB_AH_GRH : 0;
4264 	if (ib_ah_attr->ah_flags) {
4265 		ib_ah_attr->grh.sgid_index = path->mgid_index;
4266 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
4267 		ib_ah_attr->grh.traffic_class =
4268 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
4269 		ib_ah_attr->grh.flow_label =
4270 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
4271 		memcpy(ib_ah_attr->grh.dgid.raw,
4272 		       path->rgid, sizeof(ib_ah_attr->grh.dgid.raw));
4273 	}
4274 }
4275 
4276 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
4277 					struct mlx5_ib_sq *sq,
4278 					u8 *sq_state)
4279 {
4280 	void *out;
4281 	void *sqc;
4282 	int inlen;
4283 	int err;
4284 
4285 	inlen = MLX5_ST_SZ_BYTES(query_sq_out);
4286 	out = mlx5_vzalloc(inlen);
4287 	if (!out)
4288 		return -ENOMEM;
4289 
4290 	err = mlx5_core_query_sq(dev->mdev, sq->base.mqp.qpn, out);
4291 	if (err)
4292 		goto out;
4293 
4294 	sqc = MLX5_ADDR_OF(query_sq_out, out, sq_context);
4295 	*sq_state = MLX5_GET(sqc, sqc, state);
4296 	sq->state = *sq_state;
4297 
4298 out:
4299 	kvfree(out);
4300 	return err;
4301 }
4302 
4303 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
4304 					struct mlx5_ib_rq *rq,
4305 					u8 *rq_state)
4306 {
4307 	void *out;
4308 	void *rqc;
4309 	int inlen;
4310 	int err;
4311 
4312 	inlen = MLX5_ST_SZ_BYTES(query_rq_out);
4313 	out = mlx5_vzalloc(inlen);
4314 	if (!out)
4315 		return -ENOMEM;
4316 
4317 	err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
4318 	if (err)
4319 		goto out;
4320 
4321 	rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
4322 	*rq_state = MLX5_GET(rqc, rqc, state);
4323 	rq->state = *rq_state;
4324 
4325 out:
4326 	kvfree(out);
4327 	return err;
4328 }
4329 
4330 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
4331 				  struct mlx5_ib_qp *qp, u8 *qp_state)
4332 {
4333 	static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
4334 		[MLX5_RQC_STATE_RST] = {
4335 			[MLX5_SQC_STATE_RST]	= IB_QPS_RESET,
4336 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4337 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE_BAD,
4338 			[MLX5_SQ_STATE_NA]	= IB_QPS_RESET,
4339 		},
4340 		[MLX5_RQC_STATE_RDY] = {
4341 			[MLX5_SQC_STATE_RST]	= MLX5_QP_STATE_BAD,
4342 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4343 			[MLX5_SQC_STATE_ERR]	= IB_QPS_SQE,
4344 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE,
4345 		},
4346 		[MLX5_RQC_STATE_ERR] = {
4347 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE_BAD,
4348 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4349 			[MLX5_SQC_STATE_ERR]	= IB_QPS_ERR,
4350 			[MLX5_SQ_STATE_NA]	= IB_QPS_ERR,
4351 		},
4352 		[MLX5_RQ_STATE_NA] = {
4353 			[MLX5_SQC_STATE_RST]    = IB_QPS_RESET,
4354 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4355 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE,
4356 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE_BAD,
4357 		},
4358 	};
4359 
4360 	*qp_state = sqrq_trans[rq_state][sq_state];
4361 
4362 	if (*qp_state == MLX5_QP_STATE_BAD) {
4363 		WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
4364 		     qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
4365 		     qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
4366 		return -EINVAL;
4367 	}
4368 
4369 	if (*qp_state == MLX5_QP_STATE)
4370 		*qp_state = qp->state;
4371 
4372 	return 0;
4373 }
4374 
4375 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
4376 				     struct mlx5_ib_qp *qp,
4377 				     u8 *raw_packet_qp_state)
4378 {
4379 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
4380 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
4381 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
4382 	int err;
4383 	u8 sq_state = MLX5_SQ_STATE_NA;
4384 	u8 rq_state = MLX5_RQ_STATE_NA;
4385 
4386 	if (qp->sq.wqe_cnt) {
4387 		err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
4388 		if (err)
4389 			return err;
4390 	}
4391 
4392 	if (qp->rq.wqe_cnt) {
4393 		err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
4394 		if (err)
4395 			return err;
4396 	}
4397 
4398 	return sqrq_state_to_qp_state(sq_state, rq_state, qp,
4399 				      raw_packet_qp_state);
4400 }
4401 
4402 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
4403 			 struct ib_qp_attr *qp_attr)
4404 {
4405 	int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
4406 	struct mlx5_qp_context *context;
4407 	int mlx5_state;
4408 	u32 *outb;
4409 	int err = 0;
4410 
4411 	outb = kzalloc(outlen, GFP_KERNEL);
4412 	if (!outb)
4413 		return -ENOMEM;
4414 
4415 	err = mlx5_core_qp_query(dev->mdev, &qp->trans_qp.base.mqp, outb,
4416 				 outlen);
4417 	if (err)
4418 		goto out;
4419 
4420 	/* FIXME: use MLX5_GET rather than mlx5_qp_context manual struct */
4421 	context = (struct mlx5_qp_context *)MLX5_ADDR_OF(query_qp_out, outb, qpc);
4422 
4423 	mlx5_state = be32_to_cpu(context->flags) >> 28;
4424 
4425 	qp->state		     = to_ib_qp_state(mlx5_state);
4426 	qp_attr->path_mtu	     = context->mtu_msgmax >> 5;
4427 	qp_attr->path_mig_state	     =
4428 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
4429 	qp_attr->qkey		     = be32_to_cpu(context->qkey);
4430 	qp_attr->rq_psn		     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
4431 	qp_attr->sq_psn		     = be32_to_cpu(context->next_send_psn) & 0xffffff;
4432 	qp_attr->dest_qp_num	     = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
4433 	qp_attr->qp_access_flags     =
4434 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
4435 
4436 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4437 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
4438 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
4439 		qp_attr->alt_pkey_index =
4440 			be16_to_cpu(context->alt_path.pkey_index);
4441 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
4442 	}
4443 
4444 	qp_attr->pkey_index = be16_to_cpu(context->pri_path.pkey_index);
4445 	qp_attr->port_num = context->pri_path.port;
4446 
4447 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4448 	qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
4449 
4450 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
4451 
4452 	qp_attr->max_dest_rd_atomic =
4453 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
4454 	qp_attr->min_rnr_timer	    =
4455 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
4456 	qp_attr->timeout	    = context->pri_path.ackto_lt >> 3;
4457 	qp_attr->retry_cnt	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
4458 	qp_attr->rnr_retry	    = (be32_to_cpu(context->params1) >> 13) & 0x7;
4459 	qp_attr->alt_timeout	    = context->alt_path.ackto_lt >> 3;
4460 
4461 out:
4462 	kfree(outb);
4463 	return err;
4464 }
4465 
4466 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
4467 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
4468 {
4469 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4470 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4471 	int err = 0;
4472 	u8 raw_packet_qp_state;
4473 
4474 	if (ibqp->rwq_ind_tbl)
4475 		return -ENOSYS;
4476 
4477 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4478 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
4479 					    qp_init_attr);
4480 
4481 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4482 	/*
4483 	 * Wait for any outstanding page faults, in case the user frees memory
4484 	 * based upon this query's result.
4485 	 */
4486 	flush_workqueue(mlx5_ib_page_fault_wq);
4487 #endif
4488 
4489 	mutex_lock(&qp->mutex);
4490 
4491 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
4492 		err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
4493 		if (err)
4494 			goto out;
4495 		qp->state = raw_packet_qp_state;
4496 		qp_attr->port_num = 1;
4497 	} else {
4498 		err = query_qp_attr(dev, qp, qp_attr);
4499 		if (err)
4500 			goto out;
4501 	}
4502 
4503 	qp_attr->qp_state	     = qp->state;
4504 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4505 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4506 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4507 
4508 	if (!ibqp->uobject) {
4509 		qp_attr->cap.max_send_wr  = qp->sq.max_post;
4510 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4511 		qp_init_attr->qp_context = ibqp->qp_context;
4512 	} else {
4513 		qp_attr->cap.max_send_wr  = 0;
4514 		qp_attr->cap.max_send_sge = 0;
4515 	}
4516 
4517 	qp_init_attr->qp_type = ibqp->qp_type;
4518 	qp_init_attr->recv_cq = ibqp->recv_cq;
4519 	qp_init_attr->send_cq = ibqp->send_cq;
4520 	qp_init_attr->srq = ibqp->srq;
4521 	qp_attr->cap.max_inline_data = qp->max_inline_data;
4522 
4523 	qp_init_attr->cap	     = qp_attr->cap;
4524 
4525 	qp_init_attr->create_flags = 0;
4526 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4527 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4528 
4529 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
4530 		qp_init_attr->create_flags |= IB_QP_CREATE_CROSS_CHANNEL;
4531 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
4532 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_SEND;
4533 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
4534 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_RECV;
4535 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
4536 		qp_init_attr->create_flags |= mlx5_ib_create_qp_sqpn_qp1();
4537 
4538 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
4539 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4540 
4541 out:
4542 	mutex_unlock(&qp->mutex);
4543 	return err;
4544 }
4545 
4546 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
4547 					  struct ib_ucontext *context,
4548 					  struct ib_udata *udata)
4549 {
4550 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4551 	struct mlx5_ib_xrcd *xrcd;
4552 	int err;
4553 
4554 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
4555 		return ERR_PTR(-ENOSYS);
4556 
4557 	xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
4558 	if (!xrcd)
4559 		return ERR_PTR(-ENOMEM);
4560 
4561 	err = mlx5_core_xrcd_alloc(dev->mdev, &xrcd->xrcdn);
4562 	if (err) {
4563 		kfree(xrcd);
4564 		return ERR_PTR(-ENOMEM);
4565 	}
4566 
4567 	return &xrcd->ibxrcd;
4568 }
4569 
4570 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
4571 {
4572 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
4573 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
4574 	int err;
4575 
4576 	err = mlx5_core_xrcd_dealloc(dev->mdev, xrcdn);
4577 	if (err) {
4578 		mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
4579 		return err;
4580 	}
4581 
4582 	kfree(xrcd);
4583 
4584 	return 0;
4585 }
4586 
4587 static void mlx5_ib_wq_event(struct mlx5_core_qp *core_qp, int type)
4588 {
4589 	struct mlx5_ib_rwq *rwq = to_mibrwq(core_qp);
4590 	struct mlx5_ib_dev *dev = to_mdev(rwq->ibwq.device);
4591 	struct ib_event event;
4592 
4593 	if (rwq->ibwq.event_handler) {
4594 		event.device     = rwq->ibwq.device;
4595 		event.element.wq = &rwq->ibwq;
4596 		switch (type) {
4597 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
4598 			event.event = IB_EVENT_WQ_FATAL;
4599 			break;
4600 		default:
4601 			mlx5_ib_warn(dev, "Unexpected event type %d on WQ %06x\n", type, core_qp->qpn);
4602 			return;
4603 		}
4604 
4605 		rwq->ibwq.event_handler(&event, rwq->ibwq.wq_context);
4606 	}
4607 }
4608 
4609 static int  create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
4610 		      struct ib_wq_init_attr *init_attr)
4611 {
4612 	struct mlx5_ib_dev *dev;
4613 	__be64 *rq_pas0;
4614 	void *in;
4615 	void *rqc;
4616 	void *wq;
4617 	int inlen;
4618 	int err;
4619 
4620 	dev = to_mdev(pd->device);
4621 
4622 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
4623 	in = mlx5_vzalloc(inlen);
4624 	if (!in)
4625 		return -ENOMEM;
4626 
4627 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
4628 	MLX5_SET(rqc,  rqc, mem_rq_type,
4629 		 MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
4630 	MLX5_SET(rqc, rqc, user_index, rwq->user_index);
4631 	MLX5_SET(rqc,  rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
4632 	MLX5_SET(rqc,  rqc, state, MLX5_RQC_STATE_RST);
4633 	MLX5_SET(rqc,  rqc, flush_in_error_en, 1);
4634 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
4635 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
4636 	MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
4637 	MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
4638 	MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
4639 	MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
4640 	MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
4641 	MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
4642 	MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
4643 	MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
4644 	rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
4645 	mlx5_ib_populate_pas(dev, rwq->umem, rwq->page_shift, rq_pas0, 0);
4646 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rwq->core_qp);
4647 	kvfree(in);
4648 	return err;
4649 }
4650 
4651 static int set_user_rq_size(struct mlx5_ib_dev *dev,
4652 			    struct ib_wq_init_attr *wq_init_attr,
4653 			    struct mlx5_ib_create_wq *ucmd,
4654 			    struct mlx5_ib_rwq *rwq)
4655 {
4656 	/* Sanity check RQ size before proceeding */
4657 	if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
4658 		return -EINVAL;
4659 
4660 	if (!ucmd->rq_wqe_count)
4661 		return -EINVAL;
4662 
4663 	rwq->wqe_count = ucmd->rq_wqe_count;
4664 	rwq->wqe_shift = ucmd->rq_wqe_shift;
4665 	rwq->buf_size = (rwq->wqe_count << rwq->wqe_shift);
4666 	rwq->log_rq_stride = rwq->wqe_shift;
4667 	rwq->log_rq_size = ilog2(rwq->wqe_count);
4668 	return 0;
4669 }
4670 
4671 static int prepare_user_rq(struct ib_pd *pd,
4672 			   struct ib_wq_init_attr *init_attr,
4673 			   struct ib_udata *udata,
4674 			   struct mlx5_ib_rwq *rwq)
4675 {
4676 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
4677 	struct mlx5_ib_create_wq ucmd = {};
4678 	int err;
4679 	size_t required_cmd_sz;
4680 
4681 	required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
4682 	if (udata->inlen < required_cmd_sz) {
4683 		mlx5_ib_dbg(dev, "invalid inlen\n");
4684 		return -EINVAL;
4685 	}
4686 
4687 	if (udata->inlen > sizeof(ucmd) &&
4688 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4689 				 udata->inlen - sizeof(ucmd))) {
4690 		mlx5_ib_dbg(dev, "inlen is not supported\n");
4691 		return -EOPNOTSUPP;
4692 	}
4693 
4694 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
4695 		mlx5_ib_dbg(dev, "copy failed\n");
4696 		return -EFAULT;
4697 	}
4698 
4699 	if (ucmd.comp_mask) {
4700 		mlx5_ib_dbg(dev, "invalid comp mask\n");
4701 		return -EOPNOTSUPP;
4702 	}
4703 
4704 	if (ucmd.reserved) {
4705 		mlx5_ib_dbg(dev, "invalid reserved\n");
4706 		return -EOPNOTSUPP;
4707 	}
4708 
4709 	err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
4710 	if (err) {
4711 		mlx5_ib_dbg(dev, "err %d\n", err);
4712 		return err;
4713 	}
4714 
4715 	err = create_user_rq(dev, pd, rwq, &ucmd);
4716 	if (err) {
4717 		mlx5_ib_dbg(dev, "err %d\n", err);
4718 		if (err)
4719 			return err;
4720 	}
4721 
4722 	rwq->user_index = ucmd.user_index;
4723 	return 0;
4724 }
4725 
4726 struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
4727 				struct ib_wq_init_attr *init_attr,
4728 				struct ib_udata *udata)
4729 {
4730 	struct mlx5_ib_dev *dev;
4731 	struct mlx5_ib_rwq *rwq;
4732 	struct mlx5_ib_create_wq_resp resp = {};
4733 	size_t min_resp_len;
4734 	int err;
4735 
4736 	if (!udata)
4737 		return ERR_PTR(-ENOSYS);
4738 
4739 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4740 	if (udata->outlen && udata->outlen < min_resp_len)
4741 		return ERR_PTR(-EINVAL);
4742 
4743 	dev = to_mdev(pd->device);
4744 	switch (init_attr->wq_type) {
4745 	case IB_WQT_RQ:
4746 		rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
4747 		if (!rwq)
4748 			return ERR_PTR(-ENOMEM);
4749 		err = prepare_user_rq(pd, init_attr, udata, rwq);
4750 		if (err)
4751 			goto err;
4752 		err = create_rq(rwq, pd, init_attr);
4753 		if (err)
4754 			goto err_user_rq;
4755 		break;
4756 	default:
4757 		mlx5_ib_dbg(dev, "unsupported wq type %d\n",
4758 			    init_attr->wq_type);
4759 		return ERR_PTR(-EINVAL);
4760 	}
4761 
4762 	rwq->ibwq.wq_num = rwq->core_qp.qpn;
4763 	rwq->ibwq.state = IB_WQS_RESET;
4764 	if (udata->outlen) {
4765 		resp.response_length = offsetof(typeof(resp), response_length) +
4766 				sizeof(resp.response_length);
4767 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4768 		if (err)
4769 			goto err_copy;
4770 	}
4771 
4772 	rwq->core_qp.event = mlx5_ib_wq_event;
4773 	rwq->ibwq.event_handler = init_attr->event_handler;
4774 	return &rwq->ibwq;
4775 
4776 err_copy:
4777 	mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
4778 err_user_rq:
4779 	destroy_user_rq(pd, rwq);
4780 err:
4781 	kfree(rwq);
4782 	return ERR_PTR(err);
4783 }
4784 
4785 int mlx5_ib_destroy_wq(struct ib_wq *wq)
4786 {
4787 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
4788 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
4789 
4790 	mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
4791 	destroy_user_rq(wq->pd, rwq);
4792 	kfree(rwq);
4793 
4794 	return 0;
4795 }
4796 
4797 struct ib_rwq_ind_table *mlx5_ib_create_rwq_ind_table(struct ib_device *device,
4798 						      struct ib_rwq_ind_table_init_attr *init_attr,
4799 						      struct ib_udata *udata)
4800 {
4801 	struct mlx5_ib_dev *dev = to_mdev(device);
4802 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl;
4803 	int sz = 1 << init_attr->log_ind_tbl_size;
4804 	struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
4805 	size_t min_resp_len;
4806 	int inlen;
4807 	int err;
4808 	int i;
4809 	u32 *in;
4810 	void *rqtc;
4811 
4812 	if (udata->inlen > 0 &&
4813 	    !ib_is_udata_cleared(udata, 0,
4814 				 udata->inlen))
4815 		return ERR_PTR(-EOPNOTSUPP);
4816 
4817 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4818 	if (udata->outlen && udata->outlen < min_resp_len)
4819 		return ERR_PTR(-EINVAL);
4820 
4821 	rwq_ind_tbl = kzalloc(sizeof(*rwq_ind_tbl), GFP_KERNEL);
4822 	if (!rwq_ind_tbl)
4823 		return ERR_PTR(-ENOMEM);
4824 
4825 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
4826 	in = mlx5_vzalloc(inlen);
4827 	if (!in) {
4828 		err = -ENOMEM;
4829 		goto err;
4830 	}
4831 
4832 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
4833 
4834 	MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
4835 	MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
4836 
4837 	for (i = 0; i < sz; i++)
4838 		MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
4839 
4840 	err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
4841 	kvfree(in);
4842 
4843 	if (err)
4844 		goto err;
4845 
4846 	rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
4847 	if (udata->outlen) {
4848 		resp.response_length = offsetof(typeof(resp), response_length) +
4849 					sizeof(resp.response_length);
4850 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4851 		if (err)
4852 			goto err_copy;
4853 	}
4854 
4855 	return &rwq_ind_tbl->ib_rwq_ind_tbl;
4856 
4857 err_copy:
4858 	mlx5_core_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn);
4859 err:
4860 	kfree(rwq_ind_tbl);
4861 	return ERR_PTR(err);
4862 }
4863 
4864 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4865 {
4866 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
4867 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
4868 
4869 	mlx5_core_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn);
4870 
4871 	kfree(rwq_ind_tbl);
4872 	return 0;
4873 }
4874 
4875 int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
4876 		      u32 wq_attr_mask, struct ib_udata *udata)
4877 {
4878 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
4879 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
4880 	struct mlx5_ib_modify_wq ucmd = {};
4881 	size_t required_cmd_sz;
4882 	int curr_wq_state;
4883 	int wq_state;
4884 	int inlen;
4885 	int err;
4886 	void *rqc;
4887 	void *in;
4888 
4889 	required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
4890 	if (udata->inlen < required_cmd_sz)
4891 		return -EINVAL;
4892 
4893 	if (udata->inlen > sizeof(ucmd) &&
4894 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4895 				 udata->inlen - sizeof(ucmd)))
4896 		return -EOPNOTSUPP;
4897 
4898 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4899 		return -EFAULT;
4900 
4901 	if (ucmd.comp_mask || ucmd.reserved)
4902 		return -EOPNOTSUPP;
4903 
4904 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
4905 	in = mlx5_vzalloc(inlen);
4906 	if (!in)
4907 		return -ENOMEM;
4908 
4909 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
4910 
4911 	curr_wq_state = (wq_attr_mask & IB_WQ_CUR_STATE) ?
4912 		wq_attr->curr_wq_state : wq->state;
4913 	wq_state = (wq_attr_mask & IB_WQ_STATE) ?
4914 		wq_attr->wq_state : curr_wq_state;
4915 	if (curr_wq_state == IB_WQS_ERR)
4916 		curr_wq_state = MLX5_RQC_STATE_ERR;
4917 	if (wq_state == IB_WQS_ERR)
4918 		wq_state = MLX5_RQC_STATE_ERR;
4919 	MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
4920 	MLX5_SET(rqc, rqc, state, wq_state);
4921 
4922 	err = mlx5_core_modify_rq(dev->mdev, rwq->core_qp.qpn, in, inlen);
4923 	kvfree(in);
4924 	if (!err)
4925 		rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;
4926 
4927 	return err;
4928 }
4929