xref: /openbmc/linux/drivers/infiniband/hw/mlx5/qp.c (revision 0883c2c0)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/module.h>
34 #include <rdma/ib_umem.h>
35 #include <rdma/ib_cache.h>
36 #include <rdma/ib_user_verbs.h>
37 #include "mlx5_ib.h"
38 #include "user.h"
39 
40 /* not supported currently */
41 static int wq_signature;
42 
43 enum {
44 	MLX5_IB_ACK_REQ_FREQ	= 8,
45 };
46 
47 enum {
48 	MLX5_IB_DEFAULT_SCHED_QUEUE	= 0x83,
49 	MLX5_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
50 	MLX5_IB_LINK_TYPE_IB		= 0,
51 	MLX5_IB_LINK_TYPE_ETH		= 1
52 };
53 
54 enum {
55 	MLX5_IB_SQ_STRIDE	= 6,
56 	MLX5_IB_CACHE_LINE_SIZE	= 64,
57 };
58 
59 static const u32 mlx5_ib_opcode[] = {
60 	[IB_WR_SEND]				= MLX5_OPCODE_SEND,
61 	[IB_WR_LSO]				= MLX5_OPCODE_LSO,
62 	[IB_WR_SEND_WITH_IMM]			= MLX5_OPCODE_SEND_IMM,
63 	[IB_WR_RDMA_WRITE]			= MLX5_OPCODE_RDMA_WRITE,
64 	[IB_WR_RDMA_WRITE_WITH_IMM]		= MLX5_OPCODE_RDMA_WRITE_IMM,
65 	[IB_WR_RDMA_READ]			= MLX5_OPCODE_RDMA_READ,
66 	[IB_WR_ATOMIC_CMP_AND_SWP]		= MLX5_OPCODE_ATOMIC_CS,
67 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= MLX5_OPCODE_ATOMIC_FA,
68 	[IB_WR_SEND_WITH_INV]			= MLX5_OPCODE_SEND_INVAL,
69 	[IB_WR_LOCAL_INV]			= MLX5_OPCODE_UMR,
70 	[IB_WR_REG_MR]				= MLX5_OPCODE_UMR,
71 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= MLX5_OPCODE_ATOMIC_MASKED_CS,
72 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= MLX5_OPCODE_ATOMIC_MASKED_FA,
73 	[MLX5_IB_WR_UMR]			= MLX5_OPCODE_UMR,
74 };
75 
76 struct mlx5_wqe_eth_pad {
77 	u8 rsvd0[16];
78 };
79 
80 static int is_qp0(enum ib_qp_type qp_type)
81 {
82 	return qp_type == IB_QPT_SMI;
83 }
84 
85 static int is_sqp(enum ib_qp_type qp_type)
86 {
87 	return is_qp0(qp_type) || is_qp1(qp_type);
88 }
89 
90 static void *get_wqe(struct mlx5_ib_qp *qp, int offset)
91 {
92 	return mlx5_buf_offset(&qp->buf, offset);
93 }
94 
95 static void *get_recv_wqe(struct mlx5_ib_qp *qp, int n)
96 {
97 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
98 }
99 
100 void *mlx5_get_send_wqe(struct mlx5_ib_qp *qp, int n)
101 {
102 	return get_wqe(qp, qp->sq.offset + (n << MLX5_IB_SQ_STRIDE));
103 }
104 
105 /**
106  * mlx5_ib_read_user_wqe() - Copy a user-space WQE to kernel space.
107  *
108  * @qp: QP to copy from.
109  * @send: copy from the send queue when non-zero, use the receive queue
110  *	  otherwise.
111  * @wqe_index:  index to start copying from. For send work queues, the
112  *		wqe_index is in units of MLX5_SEND_WQE_BB.
113  *		For receive work queue, it is the number of work queue
114  *		element in the queue.
115  * @buffer: destination buffer.
116  * @length: maximum number of bytes to copy.
117  *
118  * Copies at least a single WQE, but may copy more data.
119  *
120  * Return: the number of bytes copied, or an error code.
121  */
122 int mlx5_ib_read_user_wqe(struct mlx5_ib_qp *qp, int send, int wqe_index,
123 			  void *buffer, u32 length,
124 			  struct mlx5_ib_qp_base *base)
125 {
126 	struct ib_device *ibdev = qp->ibqp.device;
127 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
128 	struct mlx5_ib_wq *wq = send ? &qp->sq : &qp->rq;
129 	size_t offset;
130 	size_t wq_end;
131 	struct ib_umem *umem = base->ubuffer.umem;
132 	u32 first_copy_length;
133 	int wqe_length;
134 	int ret;
135 
136 	if (wq->wqe_cnt == 0) {
137 		mlx5_ib_dbg(dev, "mlx5_ib_read_user_wqe for a QP with wqe_cnt == 0. qp_type: 0x%x\n",
138 			    qp->ibqp.qp_type);
139 		return -EINVAL;
140 	}
141 
142 	offset = wq->offset + ((wqe_index % wq->wqe_cnt) << wq->wqe_shift);
143 	wq_end = wq->offset + (wq->wqe_cnt << wq->wqe_shift);
144 
145 	if (send && length < sizeof(struct mlx5_wqe_ctrl_seg))
146 		return -EINVAL;
147 
148 	if (offset > umem->length ||
149 	    (send && offset + sizeof(struct mlx5_wqe_ctrl_seg) > umem->length))
150 		return -EINVAL;
151 
152 	first_copy_length = min_t(u32, offset + length, wq_end) - offset;
153 	ret = ib_umem_copy_from(buffer, umem, offset, first_copy_length);
154 	if (ret)
155 		return ret;
156 
157 	if (send) {
158 		struct mlx5_wqe_ctrl_seg *ctrl = buffer;
159 		int ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
160 
161 		wqe_length = ds * MLX5_WQE_DS_UNITS;
162 	} else {
163 		wqe_length = 1 << wq->wqe_shift;
164 	}
165 
166 	if (wqe_length <= first_copy_length)
167 		return first_copy_length;
168 
169 	ret = ib_umem_copy_from(buffer + first_copy_length, umem, wq->offset,
170 				wqe_length - first_copy_length);
171 	if (ret)
172 		return ret;
173 
174 	return wqe_length;
175 }
176 
177 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
178 {
179 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
180 	struct ib_event event;
181 
182 	if (type == MLX5_EVENT_TYPE_PATH_MIG) {
183 		/* This event is only valid for trans_qps */
184 		to_mibqp(qp)->port = to_mibqp(qp)->trans_qp.alt_port;
185 	}
186 
187 	if (ibqp->event_handler) {
188 		event.device     = ibqp->device;
189 		event.element.qp = ibqp;
190 		switch (type) {
191 		case MLX5_EVENT_TYPE_PATH_MIG:
192 			event.event = IB_EVENT_PATH_MIG;
193 			break;
194 		case MLX5_EVENT_TYPE_COMM_EST:
195 			event.event = IB_EVENT_COMM_EST;
196 			break;
197 		case MLX5_EVENT_TYPE_SQ_DRAINED:
198 			event.event = IB_EVENT_SQ_DRAINED;
199 			break;
200 		case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
201 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
202 			break;
203 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
204 			event.event = IB_EVENT_QP_FATAL;
205 			break;
206 		case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
207 			event.event = IB_EVENT_PATH_MIG_ERR;
208 			break;
209 		case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
210 			event.event = IB_EVENT_QP_REQ_ERR;
211 			break;
212 		case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
213 			event.event = IB_EVENT_QP_ACCESS_ERR;
214 			break;
215 		default:
216 			pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
217 			return;
218 		}
219 
220 		ibqp->event_handler(&event, ibqp->qp_context);
221 	}
222 }
223 
224 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
225 		       int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
226 {
227 	int wqe_size;
228 	int wq_size;
229 
230 	/* Sanity check RQ size before proceeding */
231 	if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
232 		return -EINVAL;
233 
234 	if (!has_rq) {
235 		qp->rq.max_gs = 0;
236 		qp->rq.wqe_cnt = 0;
237 		qp->rq.wqe_shift = 0;
238 		cap->max_recv_wr = 0;
239 		cap->max_recv_sge = 0;
240 	} else {
241 		if (ucmd) {
242 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
243 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
244 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
245 			qp->rq.max_post = qp->rq.wqe_cnt;
246 		} else {
247 			wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
248 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
249 			wqe_size = roundup_pow_of_two(wqe_size);
250 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
251 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
252 			qp->rq.wqe_cnt = wq_size / wqe_size;
253 			if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
254 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
255 					    wqe_size,
256 					    MLX5_CAP_GEN(dev->mdev,
257 							 max_wqe_sz_rq));
258 				return -EINVAL;
259 			}
260 			qp->rq.wqe_shift = ilog2(wqe_size);
261 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
262 			qp->rq.max_post = qp->rq.wqe_cnt;
263 		}
264 	}
265 
266 	return 0;
267 }
268 
269 static int sq_overhead(struct ib_qp_init_attr *attr)
270 {
271 	int size = 0;
272 
273 	switch (attr->qp_type) {
274 	case IB_QPT_XRC_INI:
275 		size += sizeof(struct mlx5_wqe_xrc_seg);
276 		/* fall through */
277 	case IB_QPT_RC:
278 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
279 			max(sizeof(struct mlx5_wqe_atomic_seg) +
280 			    sizeof(struct mlx5_wqe_raddr_seg),
281 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
282 			    sizeof(struct mlx5_mkey_seg));
283 		break;
284 
285 	case IB_QPT_XRC_TGT:
286 		return 0;
287 
288 	case IB_QPT_UC:
289 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
290 			max(sizeof(struct mlx5_wqe_raddr_seg),
291 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
292 			    sizeof(struct mlx5_mkey_seg));
293 		break;
294 
295 	case IB_QPT_UD:
296 		if (attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
297 			size += sizeof(struct mlx5_wqe_eth_pad) +
298 				sizeof(struct mlx5_wqe_eth_seg);
299 		/* fall through */
300 	case IB_QPT_SMI:
301 	case MLX5_IB_QPT_HW_GSI:
302 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
303 			sizeof(struct mlx5_wqe_datagram_seg);
304 		break;
305 
306 	case MLX5_IB_QPT_REG_UMR:
307 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
308 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
309 			sizeof(struct mlx5_mkey_seg);
310 		break;
311 
312 	default:
313 		return -EINVAL;
314 	}
315 
316 	return size;
317 }
318 
319 static int calc_send_wqe(struct ib_qp_init_attr *attr)
320 {
321 	int inl_size = 0;
322 	int size;
323 
324 	size = sq_overhead(attr);
325 	if (size < 0)
326 		return size;
327 
328 	if (attr->cap.max_inline_data) {
329 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
330 			attr->cap.max_inline_data;
331 	}
332 
333 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
334 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN &&
335 	    ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
336 			return MLX5_SIG_WQE_SIZE;
337 	else
338 		return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
339 }
340 
341 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
342 			struct mlx5_ib_qp *qp)
343 {
344 	int wqe_size;
345 	int wq_size;
346 
347 	if (!attr->cap.max_send_wr)
348 		return 0;
349 
350 	wqe_size = calc_send_wqe(attr);
351 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
352 	if (wqe_size < 0)
353 		return wqe_size;
354 
355 	if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
356 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
357 			    wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
358 		return -EINVAL;
359 	}
360 
361 	qp->max_inline_data = wqe_size - sq_overhead(attr) -
362 			      sizeof(struct mlx5_wqe_inline_seg);
363 	attr->cap.max_inline_data = qp->max_inline_data;
364 
365 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN)
366 		qp->signature_en = true;
367 
368 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
369 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
370 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
371 		mlx5_ib_dbg(dev, "wqe count(%d) exceeds limits(%d)\n",
372 			    qp->sq.wqe_cnt,
373 			    1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
374 		return -ENOMEM;
375 	}
376 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
377 	qp->sq.max_gs = attr->cap.max_send_sge;
378 	qp->sq.max_post = wq_size / wqe_size;
379 	attr->cap.max_send_wr = qp->sq.max_post;
380 
381 	return wq_size;
382 }
383 
384 static int set_user_buf_size(struct mlx5_ib_dev *dev,
385 			    struct mlx5_ib_qp *qp,
386 			    struct mlx5_ib_create_qp *ucmd,
387 			    struct mlx5_ib_qp_base *base,
388 			    struct ib_qp_init_attr *attr)
389 {
390 	int desc_sz = 1 << qp->sq.wqe_shift;
391 
392 	if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
393 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
394 			     desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
395 		return -EINVAL;
396 	}
397 
398 	if (ucmd->sq_wqe_count && ((1 << ilog2(ucmd->sq_wqe_count)) != ucmd->sq_wqe_count)) {
399 		mlx5_ib_warn(dev, "sq_wqe_count %d, sq_wqe_count %d\n",
400 			     ucmd->sq_wqe_count, ucmd->sq_wqe_count);
401 		return -EINVAL;
402 	}
403 
404 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
405 
406 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
407 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
408 			     qp->sq.wqe_cnt,
409 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
410 		return -EINVAL;
411 	}
412 
413 	if (attr->qp_type == IB_QPT_RAW_PACKET) {
414 		base->ubuffer.buf_size = qp->rq.wqe_cnt << qp->rq.wqe_shift;
415 		qp->raw_packet_qp.sq.ubuffer.buf_size = qp->sq.wqe_cnt << 6;
416 	} else {
417 		base->ubuffer.buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
418 					 (qp->sq.wqe_cnt << 6);
419 	}
420 
421 	return 0;
422 }
423 
424 static int qp_has_rq(struct ib_qp_init_attr *attr)
425 {
426 	if (attr->qp_type == IB_QPT_XRC_INI ||
427 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
428 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
429 	    !attr->cap.max_recv_wr)
430 		return 0;
431 
432 	return 1;
433 }
434 
435 static int first_med_uuar(void)
436 {
437 	return 1;
438 }
439 
440 static int next_uuar(int n)
441 {
442 	n++;
443 
444 	while (((n % 4) & 2))
445 		n++;
446 
447 	return n;
448 }
449 
450 static int num_med_uuar(struct mlx5_uuar_info *uuari)
451 {
452 	int n;
453 
454 	n = uuari->num_uars * MLX5_NON_FP_BF_REGS_PER_PAGE -
455 		uuari->num_low_latency_uuars - 1;
456 
457 	return n >= 0 ? n : 0;
458 }
459 
460 static int max_uuari(struct mlx5_uuar_info *uuari)
461 {
462 	return uuari->num_uars * 4;
463 }
464 
465 static int first_hi_uuar(struct mlx5_uuar_info *uuari)
466 {
467 	int med;
468 	int i;
469 	int t;
470 
471 	med = num_med_uuar(uuari);
472 	for (t = 0, i = first_med_uuar();; i = next_uuar(i)) {
473 		t++;
474 		if (t == med)
475 			return next_uuar(i);
476 	}
477 
478 	return 0;
479 }
480 
481 static int alloc_high_class_uuar(struct mlx5_uuar_info *uuari)
482 {
483 	int i;
484 
485 	for (i = first_hi_uuar(uuari); i < max_uuari(uuari); i = next_uuar(i)) {
486 		if (!test_bit(i, uuari->bitmap)) {
487 			set_bit(i, uuari->bitmap);
488 			uuari->count[i]++;
489 			return i;
490 		}
491 	}
492 
493 	return -ENOMEM;
494 }
495 
496 static int alloc_med_class_uuar(struct mlx5_uuar_info *uuari)
497 {
498 	int minidx = first_med_uuar();
499 	int i;
500 
501 	for (i = first_med_uuar(); i < first_hi_uuar(uuari); i = next_uuar(i)) {
502 		if (uuari->count[i] < uuari->count[minidx])
503 			minidx = i;
504 	}
505 
506 	uuari->count[minidx]++;
507 	return minidx;
508 }
509 
510 static int alloc_uuar(struct mlx5_uuar_info *uuari,
511 		      enum mlx5_ib_latency_class lat)
512 {
513 	int uuarn = -EINVAL;
514 
515 	mutex_lock(&uuari->lock);
516 	switch (lat) {
517 	case MLX5_IB_LATENCY_CLASS_LOW:
518 		uuarn = 0;
519 		uuari->count[uuarn]++;
520 		break;
521 
522 	case MLX5_IB_LATENCY_CLASS_MEDIUM:
523 		if (uuari->ver < 2)
524 			uuarn = -ENOMEM;
525 		else
526 			uuarn = alloc_med_class_uuar(uuari);
527 		break;
528 
529 	case MLX5_IB_LATENCY_CLASS_HIGH:
530 		if (uuari->ver < 2)
531 			uuarn = -ENOMEM;
532 		else
533 			uuarn = alloc_high_class_uuar(uuari);
534 		break;
535 
536 	case MLX5_IB_LATENCY_CLASS_FAST_PATH:
537 		uuarn = 2;
538 		break;
539 	}
540 	mutex_unlock(&uuari->lock);
541 
542 	return uuarn;
543 }
544 
545 static void free_med_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
546 {
547 	clear_bit(uuarn, uuari->bitmap);
548 	--uuari->count[uuarn];
549 }
550 
551 static void free_high_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
552 {
553 	clear_bit(uuarn, uuari->bitmap);
554 	--uuari->count[uuarn];
555 }
556 
557 static void free_uuar(struct mlx5_uuar_info *uuari, int uuarn)
558 {
559 	int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
560 	int high_uuar = nuuars - uuari->num_low_latency_uuars;
561 
562 	mutex_lock(&uuari->lock);
563 	if (uuarn == 0) {
564 		--uuari->count[uuarn];
565 		goto out;
566 	}
567 
568 	if (uuarn < high_uuar) {
569 		free_med_class_uuar(uuari, uuarn);
570 		goto out;
571 	}
572 
573 	free_high_class_uuar(uuari, uuarn);
574 
575 out:
576 	mutex_unlock(&uuari->lock);
577 }
578 
579 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
580 {
581 	switch (state) {
582 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
583 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
584 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
585 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
586 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
587 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
588 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
589 	default:		return -1;
590 	}
591 }
592 
593 static int to_mlx5_st(enum ib_qp_type type)
594 {
595 	switch (type) {
596 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
597 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
598 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
599 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
600 	case IB_QPT_XRC_INI:
601 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
602 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
603 	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
604 	case IB_QPT_RAW_IPV6:		return MLX5_QP_ST_RAW_IPV6;
605 	case IB_QPT_RAW_PACKET:
606 	case IB_QPT_RAW_ETHERTYPE:	return MLX5_QP_ST_RAW_ETHERTYPE;
607 	case IB_QPT_MAX:
608 	default:		return -EINVAL;
609 	}
610 }
611 
612 static int uuarn_to_uar_index(struct mlx5_uuar_info *uuari, int uuarn)
613 {
614 	return uuari->uars[uuarn / MLX5_BF_REGS_PER_PAGE].index;
615 }
616 
617 static int mlx5_ib_umem_get(struct mlx5_ib_dev *dev,
618 			    struct ib_pd *pd,
619 			    unsigned long addr, size_t size,
620 			    struct ib_umem **umem,
621 			    int *npages, int *page_shift, int *ncont,
622 			    u32 *offset)
623 {
624 	int err;
625 
626 	*umem = ib_umem_get(pd->uobject->context, addr, size, 0, 0);
627 	if (IS_ERR(*umem)) {
628 		mlx5_ib_dbg(dev, "umem_get failed\n");
629 		return PTR_ERR(*umem);
630 	}
631 
632 	mlx5_ib_cont_pages(*umem, addr, npages, page_shift, ncont, NULL);
633 
634 	err = mlx5_ib_get_buf_offset(addr, *page_shift, offset);
635 	if (err) {
636 		mlx5_ib_warn(dev, "bad offset\n");
637 		goto err_umem;
638 	}
639 
640 	mlx5_ib_dbg(dev, "addr 0x%lx, size %zu, npages %d, page_shift %d, ncont %d, offset %d\n",
641 		    addr, size, *npages, *page_shift, *ncont, *offset);
642 
643 	return 0;
644 
645 err_umem:
646 	ib_umem_release(*umem);
647 	*umem = NULL;
648 
649 	return err;
650 }
651 
652 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
653 			  struct mlx5_ib_qp *qp, struct ib_udata *udata,
654 			  struct ib_qp_init_attr *attr,
655 			  struct mlx5_create_qp_mbox_in **in,
656 			  struct mlx5_ib_create_qp_resp *resp, int *inlen,
657 			  struct mlx5_ib_qp_base *base)
658 {
659 	struct mlx5_ib_ucontext *context;
660 	struct mlx5_ib_create_qp ucmd;
661 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
662 	int page_shift = 0;
663 	int uar_index;
664 	int npages;
665 	u32 offset = 0;
666 	int uuarn;
667 	int ncont = 0;
668 	int err;
669 
670 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
671 	if (err) {
672 		mlx5_ib_dbg(dev, "copy failed\n");
673 		return err;
674 	}
675 
676 	context = to_mucontext(pd->uobject->context);
677 	/*
678 	 * TBD: should come from the verbs when we have the API
679 	 */
680 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
681 		/* In CROSS_CHANNEL CQ and QP must use the same UAR */
682 		uuarn = MLX5_CROSS_CHANNEL_UUAR;
683 	else {
684 		uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_HIGH);
685 		if (uuarn < 0) {
686 			mlx5_ib_dbg(dev, "failed to allocate low latency UUAR\n");
687 			mlx5_ib_dbg(dev, "reverting to medium latency\n");
688 			uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_MEDIUM);
689 			if (uuarn < 0) {
690 				mlx5_ib_dbg(dev, "failed to allocate medium latency UUAR\n");
691 				mlx5_ib_dbg(dev, "reverting to high latency\n");
692 				uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_LOW);
693 				if (uuarn < 0) {
694 					mlx5_ib_warn(dev, "uuar allocation failed\n");
695 					return uuarn;
696 				}
697 			}
698 		}
699 	}
700 
701 	uar_index = uuarn_to_uar_index(&context->uuari, uuarn);
702 	mlx5_ib_dbg(dev, "uuarn 0x%x, uar_index 0x%x\n", uuarn, uar_index);
703 
704 	qp->rq.offset = 0;
705 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
706 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
707 
708 	err = set_user_buf_size(dev, qp, &ucmd, base, attr);
709 	if (err)
710 		goto err_uuar;
711 
712 	if (ucmd.buf_addr && ubuffer->buf_size) {
713 		ubuffer->buf_addr = ucmd.buf_addr;
714 		err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr,
715 				       ubuffer->buf_size,
716 				       &ubuffer->umem, &npages, &page_shift,
717 				       &ncont, &offset);
718 		if (err)
719 			goto err_uuar;
720 	} else {
721 		ubuffer->umem = NULL;
722 	}
723 
724 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * ncont;
725 	*in = mlx5_vzalloc(*inlen);
726 	if (!*in) {
727 		err = -ENOMEM;
728 		goto err_umem;
729 	}
730 	if (ubuffer->umem)
731 		mlx5_ib_populate_pas(dev, ubuffer->umem, page_shift,
732 				     (*in)->pas, 0);
733 	(*in)->ctx.log_pg_sz_remote_qpn =
734 		cpu_to_be32((page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
735 	(*in)->ctx.params2 = cpu_to_be32(offset << 6);
736 
737 	(*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
738 	resp->uuar_index = uuarn;
739 	qp->uuarn = uuarn;
740 
741 	err = mlx5_ib_db_map_user(context, ucmd.db_addr, &qp->db);
742 	if (err) {
743 		mlx5_ib_dbg(dev, "map failed\n");
744 		goto err_free;
745 	}
746 
747 	err = ib_copy_to_udata(udata, resp, sizeof(*resp));
748 	if (err) {
749 		mlx5_ib_dbg(dev, "copy failed\n");
750 		goto err_unmap;
751 	}
752 	qp->create_type = MLX5_QP_USER;
753 
754 	return 0;
755 
756 err_unmap:
757 	mlx5_ib_db_unmap_user(context, &qp->db);
758 
759 err_free:
760 	kvfree(*in);
761 
762 err_umem:
763 	if (ubuffer->umem)
764 		ib_umem_release(ubuffer->umem);
765 
766 err_uuar:
767 	free_uuar(&context->uuari, uuarn);
768 	return err;
769 }
770 
771 static void destroy_qp_user(struct ib_pd *pd, struct mlx5_ib_qp *qp,
772 			    struct mlx5_ib_qp_base *base)
773 {
774 	struct mlx5_ib_ucontext *context;
775 
776 	context = to_mucontext(pd->uobject->context);
777 	mlx5_ib_db_unmap_user(context, &qp->db);
778 	if (base->ubuffer.umem)
779 		ib_umem_release(base->ubuffer.umem);
780 	free_uuar(&context->uuari, qp->uuarn);
781 }
782 
783 static int create_kernel_qp(struct mlx5_ib_dev *dev,
784 			    struct ib_qp_init_attr *init_attr,
785 			    struct mlx5_ib_qp *qp,
786 			    struct mlx5_create_qp_mbox_in **in, int *inlen,
787 			    struct mlx5_ib_qp_base *base)
788 {
789 	enum mlx5_ib_latency_class lc = MLX5_IB_LATENCY_CLASS_LOW;
790 	struct mlx5_uuar_info *uuari;
791 	int uar_index;
792 	int uuarn;
793 	int err;
794 
795 	uuari = &dev->mdev->priv.uuari;
796 	if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN |
797 					IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
798 					IB_QP_CREATE_IPOIB_UD_LSO |
799 					mlx5_ib_create_qp_sqpn_qp1()))
800 		return -EINVAL;
801 
802 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
803 		lc = MLX5_IB_LATENCY_CLASS_FAST_PATH;
804 
805 	uuarn = alloc_uuar(uuari, lc);
806 	if (uuarn < 0) {
807 		mlx5_ib_dbg(dev, "\n");
808 		return -ENOMEM;
809 	}
810 
811 	qp->bf = &uuari->bfs[uuarn];
812 	uar_index = qp->bf->uar->index;
813 
814 	err = calc_sq_size(dev, init_attr, qp);
815 	if (err < 0) {
816 		mlx5_ib_dbg(dev, "err %d\n", err);
817 		goto err_uuar;
818 	}
819 
820 	qp->rq.offset = 0;
821 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
822 	base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
823 
824 	err = mlx5_buf_alloc(dev->mdev, base->ubuffer.buf_size, &qp->buf);
825 	if (err) {
826 		mlx5_ib_dbg(dev, "err %d\n", err);
827 		goto err_uuar;
828 	}
829 
830 	qp->sq.qend = mlx5_get_send_wqe(qp, qp->sq.wqe_cnt);
831 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * qp->buf.npages;
832 	*in = mlx5_vzalloc(*inlen);
833 	if (!*in) {
834 		err = -ENOMEM;
835 		goto err_buf;
836 	}
837 	(*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
838 	(*in)->ctx.log_pg_sz_remote_qpn =
839 		cpu_to_be32((qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
840 	/* Set "fast registration enabled" for all kernel QPs */
841 	(*in)->ctx.params1 |= cpu_to_be32(1 << 11);
842 	(*in)->ctx.sq_crq_size |= cpu_to_be16(1 << 4);
843 
844 	if (init_attr->create_flags & mlx5_ib_create_qp_sqpn_qp1()) {
845 		(*in)->ctx.deth_sqpn = cpu_to_be32(1);
846 		qp->flags |= MLX5_IB_QP_SQPN_QP1;
847 	}
848 
849 	mlx5_fill_page_array(&qp->buf, (*in)->pas);
850 
851 	err = mlx5_db_alloc(dev->mdev, &qp->db);
852 	if (err) {
853 		mlx5_ib_dbg(dev, "err %d\n", err);
854 		goto err_free;
855 	}
856 
857 	qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
858 	qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
859 	qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
860 	qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
861 	qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
862 
863 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
864 	    !qp->sq.w_list || !qp->sq.wqe_head) {
865 		err = -ENOMEM;
866 		goto err_wrid;
867 	}
868 	qp->create_type = MLX5_QP_KERNEL;
869 
870 	return 0;
871 
872 err_wrid:
873 	mlx5_db_free(dev->mdev, &qp->db);
874 	kfree(qp->sq.wqe_head);
875 	kfree(qp->sq.w_list);
876 	kfree(qp->sq.wrid);
877 	kfree(qp->sq.wr_data);
878 	kfree(qp->rq.wrid);
879 
880 err_free:
881 	kvfree(*in);
882 
883 err_buf:
884 	mlx5_buf_free(dev->mdev, &qp->buf);
885 
886 err_uuar:
887 	free_uuar(&dev->mdev->priv.uuari, uuarn);
888 	return err;
889 }
890 
891 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
892 {
893 	mlx5_db_free(dev->mdev, &qp->db);
894 	kfree(qp->sq.wqe_head);
895 	kfree(qp->sq.w_list);
896 	kfree(qp->sq.wrid);
897 	kfree(qp->sq.wr_data);
898 	kfree(qp->rq.wrid);
899 	mlx5_buf_free(dev->mdev, &qp->buf);
900 	free_uuar(&dev->mdev->priv.uuari, qp->bf->uuarn);
901 }
902 
903 static __be32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
904 {
905 	if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
906 	    (attr->qp_type == IB_QPT_XRC_INI))
907 		return cpu_to_be32(MLX5_SRQ_RQ);
908 	else if (!qp->has_rq)
909 		return cpu_to_be32(MLX5_ZERO_LEN_RQ);
910 	else
911 		return cpu_to_be32(MLX5_NON_ZERO_RQ);
912 }
913 
914 static int is_connected(enum ib_qp_type qp_type)
915 {
916 	if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC)
917 		return 1;
918 
919 	return 0;
920 }
921 
922 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
923 				    struct mlx5_ib_sq *sq, u32 tdn)
924 {
925 	u32 in[MLX5_ST_SZ_DW(create_tis_in)];
926 	void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
927 
928 	memset(in, 0, sizeof(in));
929 
930 	MLX5_SET(tisc, tisc, transport_domain, tdn);
931 
932 	return mlx5_core_create_tis(dev->mdev, in, sizeof(in), &sq->tisn);
933 }
934 
935 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
936 				      struct mlx5_ib_sq *sq)
937 {
938 	mlx5_core_destroy_tis(dev->mdev, sq->tisn);
939 }
940 
941 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
942 				   struct mlx5_ib_sq *sq, void *qpin,
943 				   struct ib_pd *pd)
944 {
945 	struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
946 	__be64 *pas;
947 	void *in;
948 	void *sqc;
949 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
950 	void *wq;
951 	int inlen;
952 	int err;
953 	int page_shift = 0;
954 	int npages;
955 	int ncont = 0;
956 	u32 offset = 0;
957 
958 	err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr, ubuffer->buf_size,
959 			       &sq->ubuffer.umem, &npages, &page_shift,
960 			       &ncont, &offset);
961 	if (err)
962 		return err;
963 
964 	inlen = MLX5_ST_SZ_BYTES(create_sq_in) + sizeof(u64) * ncont;
965 	in = mlx5_vzalloc(inlen);
966 	if (!in) {
967 		err = -ENOMEM;
968 		goto err_umem;
969 	}
970 
971 	sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
972 	MLX5_SET(sqc, sqc, flush_in_error_en, 1);
973 	MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
974 	MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
975 	MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
976 	MLX5_SET(sqc, sqc, tis_lst_sz, 1);
977 	MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
978 
979 	wq = MLX5_ADDR_OF(sqc, sqc, wq);
980 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
981 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
982 	MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
983 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
984 	MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
985 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
986 	MLX5_SET(wq, wq, log_wq_pg_sz,  page_shift - MLX5_ADAPTER_PAGE_SHIFT);
987 	MLX5_SET(wq, wq, page_offset, offset);
988 
989 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
990 	mlx5_ib_populate_pas(dev, sq->ubuffer.umem, page_shift, pas, 0);
991 
992 	err = mlx5_core_create_sq_tracked(dev->mdev, in, inlen, &sq->base.mqp);
993 
994 	kvfree(in);
995 
996 	if (err)
997 		goto err_umem;
998 
999 	return 0;
1000 
1001 err_umem:
1002 	ib_umem_release(sq->ubuffer.umem);
1003 	sq->ubuffer.umem = NULL;
1004 
1005 	return err;
1006 }
1007 
1008 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1009 				     struct mlx5_ib_sq *sq)
1010 {
1011 	mlx5_core_destroy_sq_tracked(dev->mdev, &sq->base.mqp);
1012 	ib_umem_release(sq->ubuffer.umem);
1013 }
1014 
1015 static int get_rq_pas_size(void *qpc)
1016 {
1017 	u32 log_page_size = MLX5_GET(qpc, qpc, log_page_size) + 12;
1018 	u32 log_rq_stride = MLX5_GET(qpc, qpc, log_rq_stride);
1019 	u32 log_rq_size   = MLX5_GET(qpc, qpc, log_rq_size);
1020 	u32 page_offset   = MLX5_GET(qpc, qpc, page_offset);
1021 	u32 po_quanta	  = 1 << (log_page_size - 6);
1022 	u32 rq_sz	  = 1 << (log_rq_size + 4 + log_rq_stride);
1023 	u32 page_size	  = 1 << log_page_size;
1024 	u32 rq_sz_po      = rq_sz + (page_offset * po_quanta);
1025 	u32 rq_num_pas	  = (rq_sz_po + page_size - 1) / page_size;
1026 
1027 	return rq_num_pas * sizeof(u64);
1028 }
1029 
1030 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1031 				   struct mlx5_ib_rq *rq, void *qpin)
1032 {
1033 	struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1034 	__be64 *pas;
1035 	__be64 *qp_pas;
1036 	void *in;
1037 	void *rqc;
1038 	void *wq;
1039 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1040 	int inlen;
1041 	int err;
1042 	u32 rq_pas_size = get_rq_pas_size(qpc);
1043 
1044 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + rq_pas_size;
1045 	in = mlx5_vzalloc(inlen);
1046 	if (!in)
1047 		return -ENOMEM;
1048 
1049 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1050 	MLX5_SET(rqc, rqc, vsd, 1);
1051 	MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
1052 	MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1053 	MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1054 	MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1055 	MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1056 
1057 	if (mqp->flags & MLX5_IB_QP_CAP_SCATTER_FCS)
1058 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
1059 
1060 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
1061 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1062 	MLX5_SET(wq, wq, end_padding_mode,
1063 		 MLX5_GET(qpc, qpc, end_padding_mode));
1064 	MLX5_SET(wq, wq, page_offset, MLX5_GET(qpc, qpc, page_offset));
1065 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1066 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1067 	MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1068 	MLX5_SET(wq, wq, log_wq_pg_sz, MLX5_GET(qpc, qpc, log_page_size));
1069 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1070 
1071 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1072 	qp_pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, qpin, pas);
1073 	memcpy(pas, qp_pas, rq_pas_size);
1074 
1075 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rq->base.mqp);
1076 
1077 	kvfree(in);
1078 
1079 	return err;
1080 }
1081 
1082 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1083 				     struct mlx5_ib_rq *rq)
1084 {
1085 	mlx5_core_destroy_rq_tracked(dev->mdev, &rq->base.mqp);
1086 }
1087 
1088 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1089 				    struct mlx5_ib_rq *rq, u32 tdn)
1090 {
1091 	u32 *in;
1092 	void *tirc;
1093 	int inlen;
1094 	int err;
1095 
1096 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1097 	in = mlx5_vzalloc(inlen);
1098 	if (!in)
1099 		return -ENOMEM;
1100 
1101 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1102 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1103 	MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1104 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1105 
1106 	err = mlx5_core_create_tir(dev->mdev, in, inlen, &rq->tirn);
1107 
1108 	kvfree(in);
1109 
1110 	return err;
1111 }
1112 
1113 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1114 				      struct mlx5_ib_rq *rq)
1115 {
1116 	mlx5_core_destroy_tir(dev->mdev, rq->tirn);
1117 }
1118 
1119 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1120 				struct mlx5_create_qp_mbox_in *in,
1121 				struct ib_pd *pd)
1122 {
1123 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1124 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1125 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1126 	struct ib_uobject *uobj = pd->uobject;
1127 	struct ib_ucontext *ucontext = uobj->context;
1128 	struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1129 	int err;
1130 	u32 tdn = mucontext->tdn;
1131 
1132 	if (qp->sq.wqe_cnt) {
1133 		err = create_raw_packet_qp_tis(dev, sq, tdn);
1134 		if (err)
1135 			return err;
1136 
1137 		err = create_raw_packet_qp_sq(dev, sq, in, pd);
1138 		if (err)
1139 			goto err_destroy_tis;
1140 
1141 		sq->base.container_mibqp = qp;
1142 	}
1143 
1144 	if (qp->rq.wqe_cnt) {
1145 		rq->base.container_mibqp = qp;
1146 
1147 		err = create_raw_packet_qp_rq(dev, rq, in);
1148 		if (err)
1149 			goto err_destroy_sq;
1150 
1151 
1152 		err = create_raw_packet_qp_tir(dev, rq, tdn);
1153 		if (err)
1154 			goto err_destroy_rq;
1155 	}
1156 
1157 	qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1158 						     rq->base.mqp.qpn;
1159 
1160 	return 0;
1161 
1162 err_destroy_rq:
1163 	destroy_raw_packet_qp_rq(dev, rq);
1164 err_destroy_sq:
1165 	if (!qp->sq.wqe_cnt)
1166 		return err;
1167 	destroy_raw_packet_qp_sq(dev, sq);
1168 err_destroy_tis:
1169 	destroy_raw_packet_qp_tis(dev, sq);
1170 
1171 	return err;
1172 }
1173 
1174 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1175 				  struct mlx5_ib_qp *qp)
1176 {
1177 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1178 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1179 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1180 
1181 	if (qp->rq.wqe_cnt) {
1182 		destroy_raw_packet_qp_tir(dev, rq);
1183 		destroy_raw_packet_qp_rq(dev, rq);
1184 	}
1185 
1186 	if (qp->sq.wqe_cnt) {
1187 		destroy_raw_packet_qp_sq(dev, sq);
1188 		destroy_raw_packet_qp_tis(dev, sq);
1189 	}
1190 }
1191 
1192 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1193 				    struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1194 {
1195 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1196 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1197 
1198 	sq->sq = &qp->sq;
1199 	rq->rq = &qp->rq;
1200 	sq->doorbell = &qp->db;
1201 	rq->doorbell = &qp->db;
1202 }
1203 
1204 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1205 			    struct ib_qp_init_attr *init_attr,
1206 			    struct ib_udata *udata, struct mlx5_ib_qp *qp)
1207 {
1208 	struct mlx5_ib_resources *devr = &dev->devr;
1209 	struct mlx5_core_dev *mdev = dev->mdev;
1210 	struct mlx5_ib_qp_base *base;
1211 	struct mlx5_ib_create_qp_resp resp;
1212 	struct mlx5_create_qp_mbox_in *in;
1213 	struct mlx5_ib_create_qp ucmd;
1214 	int inlen = sizeof(*in);
1215 	int err;
1216 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
1217 	void *qpc;
1218 
1219 	base = init_attr->qp_type == IB_QPT_RAW_PACKET ?
1220 	       &qp->raw_packet_qp.rq.base :
1221 	       &qp->trans_qp.base;
1222 
1223 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1224 		mlx5_ib_odp_create_qp(qp);
1225 
1226 	mutex_init(&qp->mutex);
1227 	spin_lock_init(&qp->sq.lock);
1228 	spin_lock_init(&qp->rq.lock);
1229 
1230 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
1231 		if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
1232 			mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
1233 			return -EINVAL;
1234 		} else {
1235 			qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1236 		}
1237 	}
1238 
1239 	if (init_attr->create_flags &
1240 			(IB_QP_CREATE_CROSS_CHANNEL |
1241 			 IB_QP_CREATE_MANAGED_SEND |
1242 			 IB_QP_CREATE_MANAGED_RECV)) {
1243 		if (!MLX5_CAP_GEN(mdev, cd)) {
1244 			mlx5_ib_dbg(dev, "cross-channel isn't supported\n");
1245 			return -EINVAL;
1246 		}
1247 		if (init_attr->create_flags & IB_QP_CREATE_CROSS_CHANNEL)
1248 			qp->flags |= MLX5_IB_QP_CROSS_CHANNEL;
1249 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_SEND)
1250 			qp->flags |= MLX5_IB_QP_MANAGED_SEND;
1251 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_RECV)
1252 			qp->flags |= MLX5_IB_QP_MANAGED_RECV;
1253 	}
1254 
1255 	if (init_attr->qp_type == IB_QPT_UD &&
1256 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO))
1257 		if (!MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
1258 			mlx5_ib_dbg(dev, "ipoib UD lso qp isn't supported\n");
1259 			return -EOPNOTSUPP;
1260 		}
1261 
1262 	if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1263 		if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
1264 			mlx5_ib_dbg(dev, "Scatter FCS is supported only for Raw Packet QPs");
1265 			return -EOPNOTSUPP;
1266 		}
1267 		if (!MLX5_CAP_GEN(dev->mdev, eth_net_offloads) ||
1268 		    !MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
1269 			mlx5_ib_dbg(dev, "Scatter FCS isn't supported\n");
1270 			return -EOPNOTSUPP;
1271 		}
1272 		qp->flags |= MLX5_IB_QP_CAP_SCATTER_FCS;
1273 	}
1274 
1275 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1276 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
1277 
1278 	if (pd && pd->uobject) {
1279 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
1280 			mlx5_ib_dbg(dev, "copy failed\n");
1281 			return -EFAULT;
1282 		}
1283 
1284 		err = get_qp_user_index(to_mucontext(pd->uobject->context),
1285 					&ucmd, udata->inlen, &uidx);
1286 		if (err)
1287 			return err;
1288 
1289 		qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
1290 		qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
1291 	} else {
1292 		qp->wq_sig = !!wq_signature;
1293 	}
1294 
1295 	qp->has_rq = qp_has_rq(init_attr);
1296 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
1297 			  qp, (pd && pd->uobject) ? &ucmd : NULL);
1298 	if (err) {
1299 		mlx5_ib_dbg(dev, "err %d\n", err);
1300 		return err;
1301 	}
1302 
1303 	if (pd) {
1304 		if (pd->uobject) {
1305 			__u32 max_wqes =
1306 				1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1307 			mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
1308 			if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
1309 			    ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
1310 				mlx5_ib_dbg(dev, "invalid rq params\n");
1311 				return -EINVAL;
1312 			}
1313 			if (ucmd.sq_wqe_count > max_wqes) {
1314 				mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
1315 					    ucmd.sq_wqe_count, max_wqes);
1316 				return -EINVAL;
1317 			}
1318 			if (init_attr->create_flags &
1319 			    mlx5_ib_create_qp_sqpn_qp1()) {
1320 				mlx5_ib_dbg(dev, "user-space is not allowed to create UD QPs spoofing as QP1\n");
1321 				return -EINVAL;
1322 			}
1323 			err = create_user_qp(dev, pd, qp, udata, init_attr, &in,
1324 					     &resp, &inlen, base);
1325 			if (err)
1326 				mlx5_ib_dbg(dev, "err %d\n", err);
1327 		} else {
1328 			err = create_kernel_qp(dev, init_attr, qp, &in, &inlen,
1329 					       base);
1330 			if (err)
1331 				mlx5_ib_dbg(dev, "err %d\n", err);
1332 		}
1333 
1334 		if (err)
1335 			return err;
1336 	} else {
1337 		in = mlx5_vzalloc(sizeof(*in));
1338 		if (!in)
1339 			return -ENOMEM;
1340 
1341 		qp->create_type = MLX5_QP_EMPTY;
1342 	}
1343 
1344 	if (is_sqp(init_attr->qp_type))
1345 		qp->port = init_attr->port_num;
1346 
1347 	in->ctx.flags = cpu_to_be32(to_mlx5_st(init_attr->qp_type) << 16 |
1348 				    MLX5_QP_PM_MIGRATED << 11);
1349 
1350 	if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
1351 		in->ctx.flags_pd = cpu_to_be32(to_mpd(pd ? pd : devr->p0)->pdn);
1352 	else
1353 		in->ctx.flags_pd = cpu_to_be32(MLX5_QP_LAT_SENSITIVE);
1354 
1355 	if (qp->wq_sig)
1356 		in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_ENABLE_SIG);
1357 
1358 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
1359 		in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_BLOCK_MCAST);
1360 
1361 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
1362 		in->ctx.params2 |= cpu_to_be32(MLX5_QP_BIT_CC_MASTER);
1363 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
1364 		in->ctx.params2 |= cpu_to_be32(MLX5_QP_BIT_CC_SLAVE_SEND);
1365 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
1366 		in->ctx.params2 |= cpu_to_be32(MLX5_QP_BIT_CC_SLAVE_RECV);
1367 
1368 	if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
1369 		int rcqe_sz;
1370 		int scqe_sz;
1371 
1372 		rcqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->recv_cq);
1373 		scqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->send_cq);
1374 
1375 		if (rcqe_sz == 128)
1376 			in->ctx.cs_res = MLX5_RES_SCAT_DATA64_CQE;
1377 		else
1378 			in->ctx.cs_res = MLX5_RES_SCAT_DATA32_CQE;
1379 
1380 		if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) {
1381 			if (scqe_sz == 128)
1382 				in->ctx.cs_req = MLX5_REQ_SCAT_DATA64_CQE;
1383 			else
1384 				in->ctx.cs_req = MLX5_REQ_SCAT_DATA32_CQE;
1385 		}
1386 	}
1387 
1388 	if (qp->rq.wqe_cnt) {
1389 		in->ctx.rq_size_stride = (qp->rq.wqe_shift - 4);
1390 		in->ctx.rq_size_stride |= ilog2(qp->rq.wqe_cnt) << 3;
1391 	}
1392 
1393 	in->ctx.rq_type_srqn = get_rx_type(qp, init_attr);
1394 
1395 	if (qp->sq.wqe_cnt)
1396 		in->ctx.sq_crq_size |= cpu_to_be16(ilog2(qp->sq.wqe_cnt) << 11);
1397 	else
1398 		in->ctx.sq_crq_size |= cpu_to_be16(0x8000);
1399 
1400 	/* Set default resources */
1401 	switch (init_attr->qp_type) {
1402 	case IB_QPT_XRC_TGT:
1403 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
1404 		in->ctx.cqn_send = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
1405 		in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
1406 		in->ctx.xrcd = cpu_to_be32(to_mxrcd(init_attr->xrcd)->xrcdn);
1407 		break;
1408 	case IB_QPT_XRC_INI:
1409 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
1410 		in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
1411 		in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
1412 		break;
1413 	default:
1414 		if (init_attr->srq) {
1415 			in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x0)->xrcdn);
1416 			in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(init_attr->srq)->msrq.srqn);
1417 		} else {
1418 			in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
1419 			in->ctx.rq_type_srqn |=
1420 				cpu_to_be32(to_msrq(devr->s1)->msrq.srqn);
1421 		}
1422 	}
1423 
1424 	if (init_attr->send_cq)
1425 		in->ctx.cqn_send = cpu_to_be32(to_mcq(init_attr->send_cq)->mcq.cqn);
1426 
1427 	if (init_attr->recv_cq)
1428 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(init_attr->recv_cq)->mcq.cqn);
1429 
1430 	in->ctx.db_rec_addr = cpu_to_be64(qp->db.dma);
1431 
1432 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1) {
1433 		qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1434 		/* 0xffffff means we ask to work with cqe version 0 */
1435 		MLX5_SET(qpc, qpc, user_index, uidx);
1436 	}
1437 	/* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
1438 	if (init_attr->qp_type == IB_QPT_UD &&
1439 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)) {
1440 		qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1441 		MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
1442 		qp->flags |= MLX5_IB_QP_LSO;
1443 	}
1444 
1445 	if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
1446 		qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd.sq_buf_addr;
1447 		raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
1448 		err = create_raw_packet_qp(dev, qp, in, pd);
1449 	} else {
1450 		err = mlx5_core_create_qp(dev->mdev, &base->mqp, in, inlen);
1451 	}
1452 
1453 	if (err) {
1454 		mlx5_ib_dbg(dev, "create qp failed\n");
1455 		goto err_create;
1456 	}
1457 
1458 	kvfree(in);
1459 
1460 	base->container_mibqp = qp;
1461 	base->mqp.event = mlx5_ib_qp_event;
1462 
1463 	return 0;
1464 
1465 err_create:
1466 	if (qp->create_type == MLX5_QP_USER)
1467 		destroy_qp_user(pd, qp, base);
1468 	else if (qp->create_type == MLX5_QP_KERNEL)
1469 		destroy_qp_kernel(dev, qp);
1470 
1471 	kvfree(in);
1472 	return err;
1473 }
1474 
1475 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1476 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1477 {
1478 	if (send_cq) {
1479 		if (recv_cq) {
1480 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1481 				spin_lock_irq(&send_cq->lock);
1482 				spin_lock_nested(&recv_cq->lock,
1483 						 SINGLE_DEPTH_NESTING);
1484 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1485 				spin_lock_irq(&send_cq->lock);
1486 				__acquire(&recv_cq->lock);
1487 			} else {
1488 				spin_lock_irq(&recv_cq->lock);
1489 				spin_lock_nested(&send_cq->lock,
1490 						 SINGLE_DEPTH_NESTING);
1491 			}
1492 		} else {
1493 			spin_lock_irq(&send_cq->lock);
1494 			__acquire(&recv_cq->lock);
1495 		}
1496 	} else if (recv_cq) {
1497 		spin_lock_irq(&recv_cq->lock);
1498 		__acquire(&send_cq->lock);
1499 	} else {
1500 		__acquire(&send_cq->lock);
1501 		__acquire(&recv_cq->lock);
1502 	}
1503 }
1504 
1505 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1506 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1507 {
1508 	if (send_cq) {
1509 		if (recv_cq) {
1510 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1511 				spin_unlock(&recv_cq->lock);
1512 				spin_unlock_irq(&send_cq->lock);
1513 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1514 				__release(&recv_cq->lock);
1515 				spin_unlock_irq(&send_cq->lock);
1516 			} else {
1517 				spin_unlock(&send_cq->lock);
1518 				spin_unlock_irq(&recv_cq->lock);
1519 			}
1520 		} else {
1521 			__release(&recv_cq->lock);
1522 			spin_unlock_irq(&send_cq->lock);
1523 		}
1524 	} else if (recv_cq) {
1525 		__release(&send_cq->lock);
1526 		spin_unlock_irq(&recv_cq->lock);
1527 	} else {
1528 		__release(&recv_cq->lock);
1529 		__release(&send_cq->lock);
1530 	}
1531 }
1532 
1533 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
1534 {
1535 	return to_mpd(qp->ibqp.pd);
1536 }
1537 
1538 static void get_cqs(struct mlx5_ib_qp *qp,
1539 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
1540 {
1541 	switch (qp->ibqp.qp_type) {
1542 	case IB_QPT_XRC_TGT:
1543 		*send_cq = NULL;
1544 		*recv_cq = NULL;
1545 		break;
1546 	case MLX5_IB_QPT_REG_UMR:
1547 	case IB_QPT_XRC_INI:
1548 		*send_cq = to_mcq(qp->ibqp.send_cq);
1549 		*recv_cq = NULL;
1550 		break;
1551 
1552 	case IB_QPT_SMI:
1553 	case MLX5_IB_QPT_HW_GSI:
1554 	case IB_QPT_RC:
1555 	case IB_QPT_UC:
1556 	case IB_QPT_UD:
1557 	case IB_QPT_RAW_IPV6:
1558 	case IB_QPT_RAW_ETHERTYPE:
1559 	case IB_QPT_RAW_PACKET:
1560 		*send_cq = to_mcq(qp->ibqp.send_cq);
1561 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
1562 		break;
1563 
1564 	case IB_QPT_MAX:
1565 	default:
1566 		*send_cq = NULL;
1567 		*recv_cq = NULL;
1568 		break;
1569 	}
1570 }
1571 
1572 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1573 				u16 operation);
1574 
1575 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1576 {
1577 	struct mlx5_ib_cq *send_cq, *recv_cq;
1578 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
1579 	struct mlx5_modify_qp_mbox_in *in;
1580 	int err;
1581 
1582 	base = qp->ibqp.qp_type == IB_QPT_RAW_PACKET ?
1583 	       &qp->raw_packet_qp.rq.base :
1584 	       &qp->trans_qp.base;
1585 
1586 	in = kzalloc(sizeof(*in), GFP_KERNEL);
1587 	if (!in)
1588 		return;
1589 
1590 	if (qp->state != IB_QPS_RESET) {
1591 		if (qp->ibqp.qp_type != IB_QPT_RAW_PACKET) {
1592 			mlx5_ib_qp_disable_pagefaults(qp);
1593 			err = mlx5_core_qp_modify(dev->mdev,
1594 						  MLX5_CMD_OP_2RST_QP, in, 0,
1595 						  &base->mqp);
1596 		} else {
1597 			err = modify_raw_packet_qp(dev, qp,
1598 						   MLX5_CMD_OP_2RST_QP);
1599 		}
1600 		if (err)
1601 			mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
1602 				     base->mqp.qpn);
1603 	}
1604 
1605 	get_cqs(qp, &send_cq, &recv_cq);
1606 
1607 	if (qp->create_type == MLX5_QP_KERNEL) {
1608 		mlx5_ib_lock_cqs(send_cq, recv_cq);
1609 		__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
1610 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1611 		if (send_cq != recv_cq)
1612 			__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
1613 					   NULL);
1614 		mlx5_ib_unlock_cqs(send_cq, recv_cq);
1615 	}
1616 
1617 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1618 		destroy_raw_packet_qp(dev, qp);
1619 	} else {
1620 		err = mlx5_core_destroy_qp(dev->mdev, &base->mqp);
1621 		if (err)
1622 			mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
1623 				     base->mqp.qpn);
1624 	}
1625 
1626 	kfree(in);
1627 
1628 	if (qp->create_type == MLX5_QP_KERNEL)
1629 		destroy_qp_kernel(dev, qp);
1630 	else if (qp->create_type == MLX5_QP_USER)
1631 		destroy_qp_user(&get_pd(qp)->ibpd, qp, base);
1632 }
1633 
1634 static const char *ib_qp_type_str(enum ib_qp_type type)
1635 {
1636 	switch (type) {
1637 	case IB_QPT_SMI:
1638 		return "IB_QPT_SMI";
1639 	case IB_QPT_GSI:
1640 		return "IB_QPT_GSI";
1641 	case IB_QPT_RC:
1642 		return "IB_QPT_RC";
1643 	case IB_QPT_UC:
1644 		return "IB_QPT_UC";
1645 	case IB_QPT_UD:
1646 		return "IB_QPT_UD";
1647 	case IB_QPT_RAW_IPV6:
1648 		return "IB_QPT_RAW_IPV6";
1649 	case IB_QPT_RAW_ETHERTYPE:
1650 		return "IB_QPT_RAW_ETHERTYPE";
1651 	case IB_QPT_XRC_INI:
1652 		return "IB_QPT_XRC_INI";
1653 	case IB_QPT_XRC_TGT:
1654 		return "IB_QPT_XRC_TGT";
1655 	case IB_QPT_RAW_PACKET:
1656 		return "IB_QPT_RAW_PACKET";
1657 	case MLX5_IB_QPT_REG_UMR:
1658 		return "MLX5_IB_QPT_REG_UMR";
1659 	case IB_QPT_MAX:
1660 	default:
1661 		return "Invalid QP type";
1662 	}
1663 }
1664 
1665 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
1666 				struct ib_qp_init_attr *init_attr,
1667 				struct ib_udata *udata)
1668 {
1669 	struct mlx5_ib_dev *dev;
1670 	struct mlx5_ib_qp *qp;
1671 	u16 xrcdn = 0;
1672 	int err;
1673 
1674 	if (pd) {
1675 		dev = to_mdev(pd->device);
1676 
1677 		if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
1678 			if (!pd->uobject) {
1679 				mlx5_ib_dbg(dev, "Raw Packet QP is not supported for kernel consumers\n");
1680 				return ERR_PTR(-EINVAL);
1681 			} else if (!to_mucontext(pd->uobject->context)->cqe_version) {
1682 				mlx5_ib_dbg(dev, "Raw Packet QP is only supported for CQE version > 0\n");
1683 				return ERR_PTR(-EINVAL);
1684 			}
1685 		}
1686 	} else {
1687 		/* being cautious here */
1688 		if (init_attr->qp_type != IB_QPT_XRC_TGT &&
1689 		    init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
1690 			pr_warn("%s: no PD for transport %s\n", __func__,
1691 				ib_qp_type_str(init_attr->qp_type));
1692 			return ERR_PTR(-EINVAL);
1693 		}
1694 		dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
1695 	}
1696 
1697 	switch (init_attr->qp_type) {
1698 	case IB_QPT_XRC_TGT:
1699 	case IB_QPT_XRC_INI:
1700 		if (!MLX5_CAP_GEN(dev->mdev, xrc)) {
1701 			mlx5_ib_dbg(dev, "XRC not supported\n");
1702 			return ERR_PTR(-ENOSYS);
1703 		}
1704 		init_attr->recv_cq = NULL;
1705 		if (init_attr->qp_type == IB_QPT_XRC_TGT) {
1706 			xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1707 			init_attr->send_cq = NULL;
1708 		}
1709 
1710 		/* fall through */
1711 	case IB_QPT_RAW_PACKET:
1712 	case IB_QPT_RC:
1713 	case IB_QPT_UC:
1714 	case IB_QPT_UD:
1715 	case IB_QPT_SMI:
1716 	case MLX5_IB_QPT_HW_GSI:
1717 	case MLX5_IB_QPT_REG_UMR:
1718 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1719 		if (!qp)
1720 			return ERR_PTR(-ENOMEM);
1721 
1722 		err = create_qp_common(dev, pd, init_attr, udata, qp);
1723 		if (err) {
1724 			mlx5_ib_dbg(dev, "create_qp_common failed\n");
1725 			kfree(qp);
1726 			return ERR_PTR(err);
1727 		}
1728 
1729 		if (is_qp0(init_attr->qp_type))
1730 			qp->ibqp.qp_num = 0;
1731 		else if (is_qp1(init_attr->qp_type))
1732 			qp->ibqp.qp_num = 1;
1733 		else
1734 			qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
1735 
1736 		mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
1737 			    qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
1738 			    to_mcq(init_attr->recv_cq)->mcq.cqn,
1739 			    to_mcq(init_attr->send_cq)->mcq.cqn);
1740 
1741 		qp->trans_qp.xrcdn = xrcdn;
1742 
1743 		break;
1744 
1745 	case IB_QPT_GSI:
1746 		return mlx5_ib_gsi_create_qp(pd, init_attr);
1747 
1748 	case IB_QPT_RAW_IPV6:
1749 	case IB_QPT_RAW_ETHERTYPE:
1750 	case IB_QPT_MAX:
1751 	default:
1752 		mlx5_ib_dbg(dev, "unsupported qp type %d\n",
1753 			    init_attr->qp_type);
1754 		/* Don't support raw QPs */
1755 		return ERR_PTR(-EINVAL);
1756 	}
1757 
1758 	return &qp->ibqp;
1759 }
1760 
1761 int mlx5_ib_destroy_qp(struct ib_qp *qp)
1762 {
1763 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
1764 	struct mlx5_ib_qp *mqp = to_mqp(qp);
1765 
1766 	if (unlikely(qp->qp_type == IB_QPT_GSI))
1767 		return mlx5_ib_gsi_destroy_qp(qp);
1768 
1769 	destroy_qp_common(dev, mqp);
1770 
1771 	kfree(mqp);
1772 
1773 	return 0;
1774 }
1775 
1776 static __be32 to_mlx5_access_flags(struct mlx5_ib_qp *qp, const struct ib_qp_attr *attr,
1777 				   int attr_mask)
1778 {
1779 	u32 hw_access_flags = 0;
1780 	u8 dest_rd_atomic;
1781 	u32 access_flags;
1782 
1783 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1784 		dest_rd_atomic = attr->max_dest_rd_atomic;
1785 	else
1786 		dest_rd_atomic = qp->trans_qp.resp_depth;
1787 
1788 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1789 		access_flags = attr->qp_access_flags;
1790 	else
1791 		access_flags = qp->trans_qp.atomic_rd_en;
1792 
1793 	if (!dest_rd_atomic)
1794 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1795 
1796 	if (access_flags & IB_ACCESS_REMOTE_READ)
1797 		hw_access_flags |= MLX5_QP_BIT_RRE;
1798 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1799 		hw_access_flags |= (MLX5_QP_BIT_RAE | MLX5_ATOMIC_MODE_CX);
1800 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1801 		hw_access_flags |= MLX5_QP_BIT_RWE;
1802 
1803 	return cpu_to_be32(hw_access_flags);
1804 }
1805 
1806 enum {
1807 	MLX5_PATH_FLAG_FL	= 1 << 0,
1808 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
1809 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
1810 };
1811 
1812 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
1813 {
1814 	if (rate == IB_RATE_PORT_CURRENT) {
1815 		return 0;
1816 	} else if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_300_GBPS) {
1817 		return -EINVAL;
1818 	} else {
1819 		while (rate != IB_RATE_2_5_GBPS &&
1820 		       !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
1821 			 MLX5_CAP_GEN(dev->mdev, stat_rate_support)))
1822 			--rate;
1823 	}
1824 
1825 	return rate + MLX5_STAT_RATE_OFFSET;
1826 }
1827 
1828 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
1829 				      struct mlx5_ib_sq *sq, u8 sl)
1830 {
1831 	void *in;
1832 	void *tisc;
1833 	int inlen;
1834 	int err;
1835 
1836 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
1837 	in = mlx5_vzalloc(inlen);
1838 	if (!in)
1839 		return -ENOMEM;
1840 
1841 	MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
1842 
1843 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
1844 	MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
1845 
1846 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
1847 
1848 	kvfree(in);
1849 
1850 	return err;
1851 }
1852 
1853 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1854 			 const struct ib_ah_attr *ah,
1855 			 struct mlx5_qp_path *path, u8 port, int attr_mask,
1856 			 u32 path_flags, const struct ib_qp_attr *attr,
1857 			 bool alt)
1858 {
1859 	enum rdma_link_layer ll = rdma_port_get_link_layer(&dev->ib_dev, port);
1860 	int err;
1861 
1862 	if (attr_mask & IB_QP_PKEY_INDEX)
1863 		path->pkey_index = cpu_to_be16(alt ? attr->alt_pkey_index :
1864 						     attr->pkey_index);
1865 
1866 	if (ah->ah_flags & IB_AH_GRH) {
1867 		if (ah->grh.sgid_index >=
1868 		    dev->mdev->port_caps[port - 1].gid_table_len) {
1869 			pr_err("sgid_index (%u) too large. max is %d\n",
1870 			       ah->grh.sgid_index,
1871 			       dev->mdev->port_caps[port - 1].gid_table_len);
1872 			return -EINVAL;
1873 		}
1874 	}
1875 
1876 	if (ll == IB_LINK_LAYER_ETHERNET) {
1877 		if (!(ah->ah_flags & IB_AH_GRH))
1878 			return -EINVAL;
1879 		memcpy(path->rmac, ah->dmac, sizeof(ah->dmac));
1880 		path->udp_sport = mlx5_get_roce_udp_sport(dev, port,
1881 							  ah->grh.sgid_index);
1882 		path->dci_cfi_prio_sl = (ah->sl & 0x7) << 4;
1883 	} else {
1884 		path->fl_free_ar = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
1885 		path->fl_free_ar |=
1886 			(path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x40 : 0;
1887 		path->rlid = cpu_to_be16(ah->dlid);
1888 		path->grh_mlid = ah->src_path_bits & 0x7f;
1889 		if (ah->ah_flags & IB_AH_GRH)
1890 			path->grh_mlid	|= 1 << 7;
1891 		path->dci_cfi_prio_sl = ah->sl & 0xf;
1892 	}
1893 
1894 	if (ah->ah_flags & IB_AH_GRH) {
1895 		path->mgid_index = ah->grh.sgid_index;
1896 		path->hop_limit  = ah->grh.hop_limit;
1897 		path->tclass_flowlabel =
1898 			cpu_to_be32((ah->grh.traffic_class << 20) |
1899 				    (ah->grh.flow_label));
1900 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1901 	}
1902 
1903 	err = ib_rate_to_mlx5(dev, ah->static_rate);
1904 	if (err < 0)
1905 		return err;
1906 	path->static_rate = err;
1907 	path->port = port;
1908 
1909 	if (attr_mask & IB_QP_TIMEOUT)
1910 		path->ackto_lt = (alt ? attr->alt_timeout : attr->timeout) << 3;
1911 
1912 	if ((qp->ibqp.qp_type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
1913 		return modify_raw_packet_eth_prio(dev->mdev,
1914 						  &qp->raw_packet_qp.sq,
1915 						  ah->sl & 0xf);
1916 
1917 	return 0;
1918 }
1919 
1920 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
1921 	[MLX5_QP_STATE_INIT] = {
1922 		[MLX5_QP_STATE_INIT] = {
1923 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
1924 					  MLX5_QP_OPTPAR_RAE		|
1925 					  MLX5_QP_OPTPAR_RWE		|
1926 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
1927 					  MLX5_QP_OPTPAR_PRI_PORT,
1928 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
1929 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
1930 					  MLX5_QP_OPTPAR_PRI_PORT,
1931 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
1932 					  MLX5_QP_OPTPAR_Q_KEY		|
1933 					  MLX5_QP_OPTPAR_PRI_PORT,
1934 		},
1935 		[MLX5_QP_STATE_RTR] = {
1936 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
1937 					  MLX5_QP_OPTPAR_RRE            |
1938 					  MLX5_QP_OPTPAR_RAE            |
1939 					  MLX5_QP_OPTPAR_RWE            |
1940 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1941 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
1942 					  MLX5_QP_OPTPAR_RWE            |
1943 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1944 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
1945 					  MLX5_QP_OPTPAR_Q_KEY,
1946 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
1947 					   MLX5_QP_OPTPAR_Q_KEY,
1948 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
1949 					  MLX5_QP_OPTPAR_RRE            |
1950 					  MLX5_QP_OPTPAR_RAE            |
1951 					  MLX5_QP_OPTPAR_RWE            |
1952 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1953 		},
1954 	},
1955 	[MLX5_QP_STATE_RTR] = {
1956 		[MLX5_QP_STATE_RTS] = {
1957 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
1958 					  MLX5_QP_OPTPAR_RRE		|
1959 					  MLX5_QP_OPTPAR_RAE		|
1960 					  MLX5_QP_OPTPAR_RWE		|
1961 					  MLX5_QP_OPTPAR_PM_STATE	|
1962 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
1963 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
1964 					  MLX5_QP_OPTPAR_RWE		|
1965 					  MLX5_QP_OPTPAR_PM_STATE,
1966 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
1967 		},
1968 	},
1969 	[MLX5_QP_STATE_RTS] = {
1970 		[MLX5_QP_STATE_RTS] = {
1971 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
1972 					  MLX5_QP_OPTPAR_RAE		|
1973 					  MLX5_QP_OPTPAR_RWE		|
1974 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
1975 					  MLX5_QP_OPTPAR_PM_STATE	|
1976 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
1977 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
1978 					  MLX5_QP_OPTPAR_PM_STATE	|
1979 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
1980 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
1981 					  MLX5_QP_OPTPAR_SRQN		|
1982 					  MLX5_QP_OPTPAR_CQN_RCV,
1983 		},
1984 	},
1985 	[MLX5_QP_STATE_SQER] = {
1986 		[MLX5_QP_STATE_RTS] = {
1987 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
1988 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
1989 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
1990 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
1991 					   MLX5_QP_OPTPAR_RWE		|
1992 					   MLX5_QP_OPTPAR_RAE		|
1993 					   MLX5_QP_OPTPAR_RRE,
1994 		},
1995 	},
1996 };
1997 
1998 static int ib_nr_to_mlx5_nr(int ib_mask)
1999 {
2000 	switch (ib_mask) {
2001 	case IB_QP_STATE:
2002 		return 0;
2003 	case IB_QP_CUR_STATE:
2004 		return 0;
2005 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
2006 		return 0;
2007 	case IB_QP_ACCESS_FLAGS:
2008 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
2009 			MLX5_QP_OPTPAR_RAE;
2010 	case IB_QP_PKEY_INDEX:
2011 		return MLX5_QP_OPTPAR_PKEY_INDEX;
2012 	case IB_QP_PORT:
2013 		return MLX5_QP_OPTPAR_PRI_PORT;
2014 	case IB_QP_QKEY:
2015 		return MLX5_QP_OPTPAR_Q_KEY;
2016 	case IB_QP_AV:
2017 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
2018 			MLX5_QP_OPTPAR_PRI_PORT;
2019 	case IB_QP_PATH_MTU:
2020 		return 0;
2021 	case IB_QP_TIMEOUT:
2022 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
2023 	case IB_QP_RETRY_CNT:
2024 		return MLX5_QP_OPTPAR_RETRY_COUNT;
2025 	case IB_QP_RNR_RETRY:
2026 		return MLX5_QP_OPTPAR_RNR_RETRY;
2027 	case IB_QP_RQ_PSN:
2028 		return 0;
2029 	case IB_QP_MAX_QP_RD_ATOMIC:
2030 		return MLX5_QP_OPTPAR_SRA_MAX;
2031 	case IB_QP_ALT_PATH:
2032 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
2033 	case IB_QP_MIN_RNR_TIMER:
2034 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
2035 	case IB_QP_SQ_PSN:
2036 		return 0;
2037 	case IB_QP_MAX_DEST_RD_ATOMIC:
2038 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
2039 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
2040 	case IB_QP_PATH_MIG_STATE:
2041 		return MLX5_QP_OPTPAR_PM_STATE;
2042 	case IB_QP_CAP:
2043 		return 0;
2044 	case IB_QP_DEST_QPN:
2045 		return 0;
2046 	}
2047 	return 0;
2048 }
2049 
2050 static int ib_mask_to_mlx5_opt(int ib_mask)
2051 {
2052 	int result = 0;
2053 	int i;
2054 
2055 	for (i = 0; i < 8 * sizeof(int); i++) {
2056 		if ((1 << i) & ib_mask)
2057 			result |= ib_nr_to_mlx5_nr(1 << i);
2058 	}
2059 
2060 	return result;
2061 }
2062 
2063 static int modify_raw_packet_qp_rq(struct mlx5_core_dev *dev,
2064 				   struct mlx5_ib_rq *rq, int new_state)
2065 {
2066 	void *in;
2067 	void *rqc;
2068 	int inlen;
2069 	int err;
2070 
2071 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
2072 	in = mlx5_vzalloc(inlen);
2073 	if (!in)
2074 		return -ENOMEM;
2075 
2076 	MLX5_SET(modify_rq_in, in, rq_state, rq->state);
2077 
2078 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
2079 	MLX5_SET(rqc, rqc, state, new_state);
2080 
2081 	err = mlx5_core_modify_rq(dev, rq->base.mqp.qpn, in, inlen);
2082 	if (err)
2083 		goto out;
2084 
2085 	rq->state = new_state;
2086 
2087 out:
2088 	kvfree(in);
2089 	return err;
2090 }
2091 
2092 static int modify_raw_packet_qp_sq(struct mlx5_core_dev *dev,
2093 				   struct mlx5_ib_sq *sq, int new_state)
2094 {
2095 	void *in;
2096 	void *sqc;
2097 	int inlen;
2098 	int err;
2099 
2100 	inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
2101 	in = mlx5_vzalloc(inlen);
2102 	if (!in)
2103 		return -ENOMEM;
2104 
2105 	MLX5_SET(modify_sq_in, in, sq_state, sq->state);
2106 
2107 	sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
2108 	MLX5_SET(sqc, sqc, state, new_state);
2109 
2110 	err = mlx5_core_modify_sq(dev, sq->base.mqp.qpn, in, inlen);
2111 	if (err)
2112 		goto out;
2113 
2114 	sq->state = new_state;
2115 
2116 out:
2117 	kvfree(in);
2118 	return err;
2119 }
2120 
2121 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2122 				u16 operation)
2123 {
2124 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
2125 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
2126 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
2127 	int rq_state;
2128 	int sq_state;
2129 	int err;
2130 
2131 	switch (operation) {
2132 	case MLX5_CMD_OP_RST2INIT_QP:
2133 		rq_state = MLX5_RQC_STATE_RDY;
2134 		sq_state = MLX5_SQC_STATE_RDY;
2135 		break;
2136 	case MLX5_CMD_OP_2ERR_QP:
2137 		rq_state = MLX5_RQC_STATE_ERR;
2138 		sq_state = MLX5_SQC_STATE_ERR;
2139 		break;
2140 	case MLX5_CMD_OP_2RST_QP:
2141 		rq_state = MLX5_RQC_STATE_RST;
2142 		sq_state = MLX5_SQC_STATE_RST;
2143 		break;
2144 	case MLX5_CMD_OP_INIT2INIT_QP:
2145 	case MLX5_CMD_OP_INIT2RTR_QP:
2146 	case MLX5_CMD_OP_RTR2RTS_QP:
2147 	case MLX5_CMD_OP_RTS2RTS_QP:
2148 		/* Nothing to do here... */
2149 		return 0;
2150 	default:
2151 		WARN_ON(1);
2152 		return -EINVAL;
2153 	}
2154 
2155 	if (qp->rq.wqe_cnt) {
2156 		err =  modify_raw_packet_qp_rq(dev->mdev, rq, rq_state);
2157 		if (err)
2158 			return err;
2159 	}
2160 
2161 	if (qp->sq.wqe_cnt)
2162 		return modify_raw_packet_qp_sq(dev->mdev, sq, sq_state);
2163 
2164 	return 0;
2165 }
2166 
2167 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
2168 			       const struct ib_qp_attr *attr, int attr_mask,
2169 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
2170 {
2171 	static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
2172 		[MLX5_QP_STATE_RST] = {
2173 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2174 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2175 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_RST2INIT_QP,
2176 		},
2177 		[MLX5_QP_STATE_INIT]  = {
2178 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2179 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2180 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_INIT2INIT_QP,
2181 			[MLX5_QP_STATE_RTR]	= MLX5_CMD_OP_INIT2RTR_QP,
2182 		},
2183 		[MLX5_QP_STATE_RTR]   = {
2184 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2185 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2186 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTR2RTS_QP,
2187 		},
2188 		[MLX5_QP_STATE_RTS]   = {
2189 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2190 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2191 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTS2RTS_QP,
2192 		},
2193 		[MLX5_QP_STATE_SQD] = {
2194 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2195 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2196 		},
2197 		[MLX5_QP_STATE_SQER] = {
2198 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2199 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2200 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQERR2RTS_QP,
2201 		},
2202 		[MLX5_QP_STATE_ERR] = {
2203 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2204 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2205 		}
2206 	};
2207 
2208 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2209 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2210 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
2211 	struct mlx5_ib_cq *send_cq, *recv_cq;
2212 	struct mlx5_qp_context *context;
2213 	struct mlx5_modify_qp_mbox_in *in;
2214 	struct mlx5_ib_pd *pd;
2215 	enum mlx5_qp_state mlx5_cur, mlx5_new;
2216 	enum mlx5_qp_optpar optpar;
2217 	int sqd_event;
2218 	int mlx5_st;
2219 	int err;
2220 	u16 op;
2221 
2222 	in = kzalloc(sizeof(*in), GFP_KERNEL);
2223 	if (!in)
2224 		return -ENOMEM;
2225 
2226 	context = &in->ctx;
2227 	err = to_mlx5_st(ibqp->qp_type);
2228 	if (err < 0) {
2229 		mlx5_ib_dbg(dev, "unsupported qp type %d\n", ibqp->qp_type);
2230 		goto out;
2231 	}
2232 
2233 	context->flags = cpu_to_be32(err << 16);
2234 
2235 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
2236 		context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2237 	} else {
2238 		switch (attr->path_mig_state) {
2239 		case IB_MIG_MIGRATED:
2240 			context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2241 			break;
2242 		case IB_MIG_REARM:
2243 			context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
2244 			break;
2245 		case IB_MIG_ARMED:
2246 			context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
2247 			break;
2248 		}
2249 	}
2250 
2251 	if (is_sqp(ibqp->qp_type)) {
2252 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
2253 	} else if (ibqp->qp_type == IB_QPT_UD ||
2254 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
2255 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
2256 	} else if (attr_mask & IB_QP_PATH_MTU) {
2257 		if (attr->path_mtu < IB_MTU_256 ||
2258 		    attr->path_mtu > IB_MTU_4096) {
2259 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
2260 			err = -EINVAL;
2261 			goto out;
2262 		}
2263 		context->mtu_msgmax = (attr->path_mtu << 5) |
2264 				      (u8)MLX5_CAP_GEN(dev->mdev, log_max_msg);
2265 	}
2266 
2267 	if (attr_mask & IB_QP_DEST_QPN)
2268 		context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
2269 
2270 	if (attr_mask & IB_QP_PKEY_INDEX)
2271 		context->pri_path.pkey_index = cpu_to_be16(attr->pkey_index);
2272 
2273 	/* todo implement counter_index functionality */
2274 
2275 	if (is_sqp(ibqp->qp_type))
2276 		context->pri_path.port = qp->port;
2277 
2278 	if (attr_mask & IB_QP_PORT)
2279 		context->pri_path.port = attr->port_num;
2280 
2281 	if (attr_mask & IB_QP_AV) {
2282 		err = mlx5_set_path(dev, qp, &attr->ah_attr, &context->pri_path,
2283 				    attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
2284 				    attr_mask, 0, attr, false);
2285 		if (err)
2286 			goto out;
2287 	}
2288 
2289 	if (attr_mask & IB_QP_TIMEOUT)
2290 		context->pri_path.ackto_lt |= attr->timeout << 3;
2291 
2292 	if (attr_mask & IB_QP_ALT_PATH) {
2293 		err = mlx5_set_path(dev, qp, &attr->alt_ah_attr,
2294 				    &context->alt_path,
2295 				    attr->alt_port_num,
2296 				    attr_mask | IB_QP_PKEY_INDEX | IB_QP_TIMEOUT,
2297 				    0, attr, true);
2298 		if (err)
2299 			goto out;
2300 	}
2301 
2302 	pd = get_pd(qp);
2303 	get_cqs(qp, &send_cq, &recv_cq);
2304 
2305 	context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
2306 	context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
2307 	context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
2308 	context->params1  = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
2309 
2310 	if (attr_mask & IB_QP_RNR_RETRY)
2311 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2312 
2313 	if (attr_mask & IB_QP_RETRY_CNT)
2314 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2315 
2316 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2317 		if (attr->max_rd_atomic)
2318 			context->params1 |=
2319 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2320 	}
2321 
2322 	if (attr_mask & IB_QP_SQ_PSN)
2323 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2324 
2325 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2326 		if (attr->max_dest_rd_atomic)
2327 			context->params2 |=
2328 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2329 	}
2330 
2331 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
2332 		context->params2 |= to_mlx5_access_flags(qp, attr, attr_mask);
2333 
2334 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
2335 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2336 
2337 	if (attr_mask & IB_QP_RQ_PSN)
2338 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2339 
2340 	if (attr_mask & IB_QP_QKEY)
2341 		context->qkey = cpu_to_be32(attr->qkey);
2342 
2343 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2344 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2345 
2346 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2347 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2348 		sqd_event = 1;
2349 	else
2350 		sqd_event = 0;
2351 
2352 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2353 		context->sq_crq_size |= cpu_to_be16(1 << 4);
2354 
2355 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
2356 		context->deth_sqpn = cpu_to_be32(1);
2357 
2358 	mlx5_cur = to_mlx5_state(cur_state);
2359 	mlx5_new = to_mlx5_state(new_state);
2360 	mlx5_st = to_mlx5_st(ibqp->qp_type);
2361 	if (mlx5_st < 0)
2362 		goto out;
2363 
2364 	/* If moving to a reset or error state, we must disable page faults on
2365 	 * this QP and flush all current page faults. Otherwise a stale page
2366 	 * fault may attempt to work on this QP after it is reset and moved
2367 	 * again to RTS, and may cause the driver and the device to get out of
2368 	 * sync. */
2369 	if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2370 	    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR) &&
2371 	    (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
2372 		mlx5_ib_qp_disable_pagefaults(qp);
2373 
2374 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
2375 	    !optab[mlx5_cur][mlx5_new])
2376 		goto out;
2377 
2378 	op = optab[mlx5_cur][mlx5_new];
2379 	optpar = ib_mask_to_mlx5_opt(attr_mask);
2380 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
2381 	in->optparam = cpu_to_be32(optpar);
2382 
2383 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET)
2384 		err = modify_raw_packet_qp(dev, qp, op);
2385 	else
2386 		err = mlx5_core_qp_modify(dev->mdev, op, in, sqd_event,
2387 					  &base->mqp);
2388 	if (err)
2389 		goto out;
2390 
2391 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT &&
2392 	    (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
2393 		mlx5_ib_qp_enable_pagefaults(qp);
2394 
2395 	qp->state = new_state;
2396 
2397 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2398 		qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
2399 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2400 		qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
2401 	if (attr_mask & IB_QP_PORT)
2402 		qp->port = attr->port_num;
2403 	if (attr_mask & IB_QP_ALT_PATH)
2404 		qp->trans_qp.alt_port = attr->alt_port_num;
2405 
2406 	/*
2407 	 * If we moved a kernel QP to RESET, clean up all old CQ
2408 	 * entries and reinitialize the QP.
2409 	 */
2410 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
2411 		mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2412 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
2413 		if (send_cq != recv_cq)
2414 			mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
2415 
2416 		qp->rq.head = 0;
2417 		qp->rq.tail = 0;
2418 		qp->sq.head = 0;
2419 		qp->sq.tail = 0;
2420 		qp->sq.cur_post = 0;
2421 		qp->sq.last_poll = 0;
2422 		qp->db.db[MLX5_RCV_DBR] = 0;
2423 		qp->db.db[MLX5_SND_DBR] = 0;
2424 	}
2425 
2426 out:
2427 	kfree(in);
2428 	return err;
2429 }
2430 
2431 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2432 		      int attr_mask, struct ib_udata *udata)
2433 {
2434 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2435 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2436 	enum ib_qp_type qp_type;
2437 	enum ib_qp_state cur_state, new_state;
2438 	int err = -EINVAL;
2439 	int port;
2440 	enum rdma_link_layer ll = IB_LINK_LAYER_UNSPECIFIED;
2441 
2442 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
2443 		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
2444 
2445 	qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ?
2446 		IB_QPT_GSI : ibqp->qp_type;
2447 
2448 	mutex_lock(&qp->mutex);
2449 
2450 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2451 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2452 
2453 	if (!(cur_state == new_state && cur_state == IB_QPS_RESET)) {
2454 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2455 		ll = dev->ib_dev.get_link_layer(&dev->ib_dev, port);
2456 	}
2457 
2458 	if (qp_type != MLX5_IB_QPT_REG_UMR &&
2459 	    !ib_modify_qp_is_ok(cur_state, new_state, qp_type, attr_mask, ll)) {
2460 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
2461 			    cur_state, new_state, ibqp->qp_type, attr_mask);
2462 		goto out;
2463 	}
2464 
2465 	if ((attr_mask & IB_QP_PORT) &&
2466 	    (attr->port_num == 0 ||
2467 	     attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports))) {
2468 		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
2469 			    attr->port_num, dev->num_ports);
2470 		goto out;
2471 	}
2472 
2473 	if (attr_mask & IB_QP_PKEY_INDEX) {
2474 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2475 		if (attr->pkey_index >=
2476 		    dev->mdev->port_caps[port - 1].pkey_table_len) {
2477 			mlx5_ib_dbg(dev, "invalid pkey index %d\n",
2478 				    attr->pkey_index);
2479 			goto out;
2480 		}
2481 	}
2482 
2483 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2484 	    attr->max_rd_atomic >
2485 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
2486 		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
2487 			    attr->max_rd_atomic);
2488 		goto out;
2489 	}
2490 
2491 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2492 	    attr->max_dest_rd_atomic >
2493 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
2494 		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
2495 			    attr->max_dest_rd_atomic);
2496 		goto out;
2497 	}
2498 
2499 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2500 		err = 0;
2501 		goto out;
2502 	}
2503 
2504 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2505 
2506 out:
2507 	mutex_unlock(&qp->mutex);
2508 	return err;
2509 }
2510 
2511 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2512 {
2513 	struct mlx5_ib_cq *cq;
2514 	unsigned cur;
2515 
2516 	cur = wq->head - wq->tail;
2517 	if (likely(cur + nreq < wq->max_post))
2518 		return 0;
2519 
2520 	cq = to_mcq(ib_cq);
2521 	spin_lock(&cq->lock);
2522 	cur = wq->head - wq->tail;
2523 	spin_unlock(&cq->lock);
2524 
2525 	return cur + nreq >= wq->max_post;
2526 }
2527 
2528 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
2529 					  u64 remote_addr, u32 rkey)
2530 {
2531 	rseg->raddr    = cpu_to_be64(remote_addr);
2532 	rseg->rkey     = cpu_to_be32(rkey);
2533 	rseg->reserved = 0;
2534 }
2535 
2536 static void *set_eth_seg(struct mlx5_wqe_eth_seg *eseg,
2537 			 struct ib_send_wr *wr, void *qend,
2538 			 struct mlx5_ib_qp *qp, int *size)
2539 {
2540 	void *seg = eseg;
2541 
2542 	memset(eseg, 0, sizeof(struct mlx5_wqe_eth_seg));
2543 
2544 	if (wr->send_flags & IB_SEND_IP_CSUM)
2545 		eseg->cs_flags = MLX5_ETH_WQE_L3_CSUM |
2546 				 MLX5_ETH_WQE_L4_CSUM;
2547 
2548 	seg += sizeof(struct mlx5_wqe_eth_seg);
2549 	*size += sizeof(struct mlx5_wqe_eth_seg) / 16;
2550 
2551 	if (wr->opcode == IB_WR_LSO) {
2552 		struct ib_ud_wr *ud_wr = container_of(wr, struct ib_ud_wr, wr);
2553 		int size_of_inl_hdr_start = sizeof(eseg->inline_hdr_start);
2554 		u64 left, leftlen, copysz;
2555 		void *pdata = ud_wr->header;
2556 
2557 		left = ud_wr->hlen;
2558 		eseg->mss = cpu_to_be16(ud_wr->mss);
2559 		eseg->inline_hdr_sz = cpu_to_be16(left);
2560 
2561 		/*
2562 		 * check if there is space till the end of queue, if yes,
2563 		 * copy all in one shot, otherwise copy till the end of queue,
2564 		 * rollback and than the copy the left
2565 		 */
2566 		leftlen = qend - (void *)eseg->inline_hdr_start;
2567 		copysz = min_t(u64, leftlen, left);
2568 
2569 		memcpy(seg - size_of_inl_hdr_start, pdata, copysz);
2570 
2571 		if (likely(copysz > size_of_inl_hdr_start)) {
2572 			seg += ALIGN(copysz - size_of_inl_hdr_start, 16);
2573 			*size += ALIGN(copysz - size_of_inl_hdr_start, 16) / 16;
2574 		}
2575 
2576 		if (unlikely(copysz < left)) { /* the last wqe in the queue */
2577 			seg = mlx5_get_send_wqe(qp, 0);
2578 			left -= copysz;
2579 			pdata += copysz;
2580 			memcpy(seg, pdata, left);
2581 			seg += ALIGN(left, 16);
2582 			*size += ALIGN(left, 16) / 16;
2583 		}
2584 	}
2585 
2586 	return seg;
2587 }
2588 
2589 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
2590 			     struct ib_send_wr *wr)
2591 {
2592 	memcpy(&dseg->av, &to_mah(ud_wr(wr)->ah)->av, sizeof(struct mlx5_av));
2593 	dseg->av.dqp_dct = cpu_to_be32(ud_wr(wr)->remote_qpn | MLX5_EXTENDED_UD_AV);
2594 	dseg->av.key.qkey.qkey = cpu_to_be32(ud_wr(wr)->remote_qkey);
2595 }
2596 
2597 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
2598 {
2599 	dseg->byte_count = cpu_to_be32(sg->length);
2600 	dseg->lkey       = cpu_to_be32(sg->lkey);
2601 	dseg->addr       = cpu_to_be64(sg->addr);
2602 }
2603 
2604 static __be16 get_klm_octo(int npages)
2605 {
2606 	return cpu_to_be16(ALIGN(npages, 8) / 2);
2607 }
2608 
2609 static __be64 frwr_mkey_mask(void)
2610 {
2611 	u64 result;
2612 
2613 	result = MLX5_MKEY_MASK_LEN		|
2614 		MLX5_MKEY_MASK_PAGE_SIZE	|
2615 		MLX5_MKEY_MASK_START_ADDR	|
2616 		MLX5_MKEY_MASK_EN_RINVAL	|
2617 		MLX5_MKEY_MASK_KEY		|
2618 		MLX5_MKEY_MASK_LR		|
2619 		MLX5_MKEY_MASK_LW		|
2620 		MLX5_MKEY_MASK_RR		|
2621 		MLX5_MKEY_MASK_RW		|
2622 		MLX5_MKEY_MASK_A		|
2623 		MLX5_MKEY_MASK_SMALL_FENCE	|
2624 		MLX5_MKEY_MASK_FREE;
2625 
2626 	return cpu_to_be64(result);
2627 }
2628 
2629 static __be64 sig_mkey_mask(void)
2630 {
2631 	u64 result;
2632 
2633 	result = MLX5_MKEY_MASK_LEN		|
2634 		MLX5_MKEY_MASK_PAGE_SIZE	|
2635 		MLX5_MKEY_MASK_START_ADDR	|
2636 		MLX5_MKEY_MASK_EN_SIGERR	|
2637 		MLX5_MKEY_MASK_EN_RINVAL	|
2638 		MLX5_MKEY_MASK_KEY		|
2639 		MLX5_MKEY_MASK_LR		|
2640 		MLX5_MKEY_MASK_LW		|
2641 		MLX5_MKEY_MASK_RR		|
2642 		MLX5_MKEY_MASK_RW		|
2643 		MLX5_MKEY_MASK_SMALL_FENCE	|
2644 		MLX5_MKEY_MASK_FREE		|
2645 		MLX5_MKEY_MASK_BSF_EN;
2646 
2647 	return cpu_to_be64(result);
2648 }
2649 
2650 static void set_reg_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr,
2651 				struct mlx5_ib_mr *mr)
2652 {
2653 	int ndescs = mr->ndescs;
2654 
2655 	memset(umr, 0, sizeof(*umr));
2656 
2657 	if (mr->access_mode == MLX5_ACCESS_MODE_KLM)
2658 		/* KLMs take twice the size of MTTs */
2659 		ndescs *= 2;
2660 
2661 	umr->flags = MLX5_UMR_CHECK_NOT_FREE;
2662 	umr->klm_octowords = get_klm_octo(ndescs);
2663 	umr->mkey_mask = frwr_mkey_mask();
2664 }
2665 
2666 static void set_linv_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr)
2667 {
2668 	memset(umr, 0, sizeof(*umr));
2669 	umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
2670 	umr->flags = 1 << 7;
2671 }
2672 
2673 static __be64 get_umr_reg_mr_mask(void)
2674 {
2675 	u64 result;
2676 
2677 	result = MLX5_MKEY_MASK_LEN		|
2678 		 MLX5_MKEY_MASK_PAGE_SIZE	|
2679 		 MLX5_MKEY_MASK_START_ADDR	|
2680 		 MLX5_MKEY_MASK_PD		|
2681 		 MLX5_MKEY_MASK_LR		|
2682 		 MLX5_MKEY_MASK_LW		|
2683 		 MLX5_MKEY_MASK_KEY		|
2684 		 MLX5_MKEY_MASK_RR		|
2685 		 MLX5_MKEY_MASK_RW		|
2686 		 MLX5_MKEY_MASK_A		|
2687 		 MLX5_MKEY_MASK_FREE;
2688 
2689 	return cpu_to_be64(result);
2690 }
2691 
2692 static __be64 get_umr_unreg_mr_mask(void)
2693 {
2694 	u64 result;
2695 
2696 	result = MLX5_MKEY_MASK_FREE;
2697 
2698 	return cpu_to_be64(result);
2699 }
2700 
2701 static __be64 get_umr_update_mtt_mask(void)
2702 {
2703 	u64 result;
2704 
2705 	result = MLX5_MKEY_MASK_FREE;
2706 
2707 	return cpu_to_be64(result);
2708 }
2709 
2710 static __be64 get_umr_update_translation_mask(void)
2711 {
2712 	u64 result;
2713 
2714 	result = MLX5_MKEY_MASK_LEN |
2715 		 MLX5_MKEY_MASK_PAGE_SIZE |
2716 		 MLX5_MKEY_MASK_START_ADDR |
2717 		 MLX5_MKEY_MASK_KEY |
2718 		 MLX5_MKEY_MASK_FREE;
2719 
2720 	return cpu_to_be64(result);
2721 }
2722 
2723 static __be64 get_umr_update_access_mask(void)
2724 {
2725 	u64 result;
2726 
2727 	result = MLX5_MKEY_MASK_LW |
2728 		 MLX5_MKEY_MASK_RR |
2729 		 MLX5_MKEY_MASK_RW |
2730 		 MLX5_MKEY_MASK_A |
2731 		 MLX5_MKEY_MASK_KEY |
2732 		 MLX5_MKEY_MASK_FREE;
2733 
2734 	return cpu_to_be64(result);
2735 }
2736 
2737 static __be64 get_umr_update_pd_mask(void)
2738 {
2739 	u64 result;
2740 
2741 	result = MLX5_MKEY_MASK_PD |
2742 		 MLX5_MKEY_MASK_KEY |
2743 		 MLX5_MKEY_MASK_FREE;
2744 
2745 	return cpu_to_be64(result);
2746 }
2747 
2748 static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
2749 				struct ib_send_wr *wr)
2750 {
2751 	struct mlx5_umr_wr *umrwr = umr_wr(wr);
2752 
2753 	memset(umr, 0, sizeof(*umr));
2754 
2755 	if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
2756 		umr->flags = MLX5_UMR_CHECK_FREE; /* fail if free */
2757 	else
2758 		umr->flags = MLX5_UMR_CHECK_NOT_FREE; /* fail if not free */
2759 
2760 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UNREG)) {
2761 		umr->klm_octowords = get_klm_octo(umrwr->npages);
2762 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT) {
2763 			umr->mkey_mask = get_umr_update_mtt_mask();
2764 			umr->bsf_octowords = get_klm_octo(umrwr->target.offset);
2765 			umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
2766 		}
2767 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION)
2768 			umr->mkey_mask |= get_umr_update_translation_mask();
2769 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_ACCESS)
2770 			umr->mkey_mask |= get_umr_update_access_mask();
2771 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_PD)
2772 			umr->mkey_mask |= get_umr_update_pd_mask();
2773 		if (!umr->mkey_mask)
2774 			umr->mkey_mask = get_umr_reg_mr_mask();
2775 	} else {
2776 		umr->mkey_mask = get_umr_unreg_mr_mask();
2777 	}
2778 
2779 	if (!wr->num_sge)
2780 		umr->flags |= MLX5_UMR_INLINE;
2781 }
2782 
2783 static u8 get_umr_flags(int acc)
2784 {
2785 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
2786 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
2787 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
2788 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
2789 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
2790 }
2791 
2792 static void set_reg_mkey_seg(struct mlx5_mkey_seg *seg,
2793 			     struct mlx5_ib_mr *mr,
2794 			     u32 key, int access)
2795 {
2796 	int ndescs = ALIGN(mr->ndescs, 8) >> 1;
2797 
2798 	memset(seg, 0, sizeof(*seg));
2799 
2800 	if (mr->access_mode == MLX5_ACCESS_MODE_MTT)
2801 		seg->log2_page_size = ilog2(mr->ibmr.page_size);
2802 	else if (mr->access_mode == MLX5_ACCESS_MODE_KLM)
2803 		/* KLMs take twice the size of MTTs */
2804 		ndescs *= 2;
2805 
2806 	seg->flags = get_umr_flags(access) | mr->access_mode;
2807 	seg->qpn_mkey7_0 = cpu_to_be32((key & 0xff) | 0xffffff00);
2808 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
2809 	seg->start_addr = cpu_to_be64(mr->ibmr.iova);
2810 	seg->len = cpu_to_be64(mr->ibmr.length);
2811 	seg->xlt_oct_size = cpu_to_be32(ndescs);
2812 }
2813 
2814 static void set_linv_mkey_seg(struct mlx5_mkey_seg *seg)
2815 {
2816 	memset(seg, 0, sizeof(*seg));
2817 	seg->status = MLX5_MKEY_STATUS_FREE;
2818 }
2819 
2820 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr)
2821 {
2822 	struct mlx5_umr_wr *umrwr = umr_wr(wr);
2823 
2824 	memset(seg, 0, sizeof(*seg));
2825 	if (wr->send_flags & MLX5_IB_SEND_UMR_UNREG) {
2826 		seg->status = MLX5_MKEY_STATUS_FREE;
2827 		return;
2828 	}
2829 
2830 	seg->flags = convert_access(umrwr->access_flags);
2831 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT)) {
2832 		if (umrwr->pd)
2833 			seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
2834 		seg->start_addr = cpu_to_be64(umrwr->target.virt_addr);
2835 	}
2836 	seg->len = cpu_to_be64(umrwr->length);
2837 	seg->log2_page_size = umrwr->page_shift;
2838 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
2839 				       mlx5_mkey_variant(umrwr->mkey));
2840 }
2841 
2842 static void set_reg_data_seg(struct mlx5_wqe_data_seg *dseg,
2843 			     struct mlx5_ib_mr *mr,
2844 			     struct mlx5_ib_pd *pd)
2845 {
2846 	int bcount = mr->desc_size * mr->ndescs;
2847 
2848 	dseg->addr = cpu_to_be64(mr->desc_map);
2849 	dseg->byte_count = cpu_to_be32(ALIGN(bcount, 64));
2850 	dseg->lkey = cpu_to_be32(pd->ibpd.local_dma_lkey);
2851 }
2852 
2853 static __be32 send_ieth(struct ib_send_wr *wr)
2854 {
2855 	switch (wr->opcode) {
2856 	case IB_WR_SEND_WITH_IMM:
2857 	case IB_WR_RDMA_WRITE_WITH_IMM:
2858 		return wr->ex.imm_data;
2859 
2860 	case IB_WR_SEND_WITH_INV:
2861 		return cpu_to_be32(wr->ex.invalidate_rkey);
2862 
2863 	default:
2864 		return 0;
2865 	}
2866 }
2867 
2868 static u8 calc_sig(void *wqe, int size)
2869 {
2870 	u8 *p = wqe;
2871 	u8 res = 0;
2872 	int i;
2873 
2874 	for (i = 0; i < size; i++)
2875 		res ^= p[i];
2876 
2877 	return ~res;
2878 }
2879 
2880 static u8 wq_sig(void *wqe)
2881 {
2882 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
2883 }
2884 
2885 static int set_data_inl_seg(struct mlx5_ib_qp *qp, struct ib_send_wr *wr,
2886 			    void *wqe, int *sz)
2887 {
2888 	struct mlx5_wqe_inline_seg *seg;
2889 	void *qend = qp->sq.qend;
2890 	void *addr;
2891 	int inl = 0;
2892 	int copy;
2893 	int len;
2894 	int i;
2895 
2896 	seg = wqe;
2897 	wqe += sizeof(*seg);
2898 	for (i = 0; i < wr->num_sge; i++) {
2899 		addr = (void *)(unsigned long)(wr->sg_list[i].addr);
2900 		len  = wr->sg_list[i].length;
2901 		inl += len;
2902 
2903 		if (unlikely(inl > qp->max_inline_data))
2904 			return -ENOMEM;
2905 
2906 		if (unlikely(wqe + len > qend)) {
2907 			copy = qend - wqe;
2908 			memcpy(wqe, addr, copy);
2909 			addr += copy;
2910 			len -= copy;
2911 			wqe = mlx5_get_send_wqe(qp, 0);
2912 		}
2913 		memcpy(wqe, addr, len);
2914 		wqe += len;
2915 	}
2916 
2917 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
2918 
2919 	*sz = ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
2920 
2921 	return 0;
2922 }
2923 
2924 static u16 prot_field_size(enum ib_signature_type type)
2925 {
2926 	switch (type) {
2927 	case IB_SIG_TYPE_T10_DIF:
2928 		return MLX5_DIF_SIZE;
2929 	default:
2930 		return 0;
2931 	}
2932 }
2933 
2934 static u8 bs_selector(int block_size)
2935 {
2936 	switch (block_size) {
2937 	case 512:	    return 0x1;
2938 	case 520:	    return 0x2;
2939 	case 4096:	    return 0x3;
2940 	case 4160:	    return 0x4;
2941 	case 1073741824:    return 0x5;
2942 	default:	    return 0;
2943 	}
2944 }
2945 
2946 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
2947 			      struct mlx5_bsf_inl *inl)
2948 {
2949 	/* Valid inline section and allow BSF refresh */
2950 	inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
2951 				       MLX5_BSF_REFRESH_DIF);
2952 	inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
2953 	inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
2954 	/* repeating block */
2955 	inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
2956 	inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
2957 			MLX5_DIF_CRC : MLX5_DIF_IPCS;
2958 
2959 	if (domain->sig.dif.ref_remap)
2960 		inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
2961 
2962 	if (domain->sig.dif.app_escape) {
2963 		if (domain->sig.dif.ref_escape)
2964 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
2965 		else
2966 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
2967 	}
2968 
2969 	inl->dif_app_bitmask_check =
2970 		cpu_to_be16(domain->sig.dif.apptag_check_mask);
2971 }
2972 
2973 static int mlx5_set_bsf(struct ib_mr *sig_mr,
2974 			struct ib_sig_attrs *sig_attrs,
2975 			struct mlx5_bsf *bsf, u32 data_size)
2976 {
2977 	struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
2978 	struct mlx5_bsf_basic *basic = &bsf->basic;
2979 	struct ib_sig_domain *mem = &sig_attrs->mem;
2980 	struct ib_sig_domain *wire = &sig_attrs->wire;
2981 
2982 	memset(bsf, 0, sizeof(*bsf));
2983 
2984 	/* Basic + Extended + Inline */
2985 	basic->bsf_size_sbs = 1 << 7;
2986 	/* Input domain check byte mask */
2987 	basic->check_byte_mask = sig_attrs->check_mask;
2988 	basic->raw_data_size = cpu_to_be32(data_size);
2989 
2990 	/* Memory domain */
2991 	switch (sig_attrs->mem.sig_type) {
2992 	case IB_SIG_TYPE_NONE:
2993 		break;
2994 	case IB_SIG_TYPE_T10_DIF:
2995 		basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
2996 		basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
2997 		mlx5_fill_inl_bsf(mem, &bsf->m_inl);
2998 		break;
2999 	default:
3000 		return -EINVAL;
3001 	}
3002 
3003 	/* Wire domain */
3004 	switch (sig_attrs->wire.sig_type) {
3005 	case IB_SIG_TYPE_NONE:
3006 		break;
3007 	case IB_SIG_TYPE_T10_DIF:
3008 		if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
3009 		    mem->sig_type == wire->sig_type) {
3010 			/* Same block structure */
3011 			basic->bsf_size_sbs |= 1 << 4;
3012 			if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
3013 				basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
3014 			if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
3015 				basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
3016 			if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
3017 				basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
3018 		} else
3019 			basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
3020 
3021 		basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
3022 		mlx5_fill_inl_bsf(wire, &bsf->w_inl);
3023 		break;
3024 	default:
3025 		return -EINVAL;
3026 	}
3027 
3028 	return 0;
3029 }
3030 
3031 static int set_sig_data_segment(struct ib_sig_handover_wr *wr,
3032 				struct mlx5_ib_qp *qp, void **seg, int *size)
3033 {
3034 	struct ib_sig_attrs *sig_attrs = wr->sig_attrs;
3035 	struct ib_mr *sig_mr = wr->sig_mr;
3036 	struct mlx5_bsf *bsf;
3037 	u32 data_len = wr->wr.sg_list->length;
3038 	u32 data_key = wr->wr.sg_list->lkey;
3039 	u64 data_va = wr->wr.sg_list->addr;
3040 	int ret;
3041 	int wqe_size;
3042 
3043 	if (!wr->prot ||
3044 	    (data_key == wr->prot->lkey &&
3045 	     data_va == wr->prot->addr &&
3046 	     data_len == wr->prot->length)) {
3047 		/**
3048 		 * Source domain doesn't contain signature information
3049 		 * or data and protection are interleaved in memory.
3050 		 * So need construct:
3051 		 *                  ------------------
3052 		 *                 |     data_klm     |
3053 		 *                  ------------------
3054 		 *                 |       BSF        |
3055 		 *                  ------------------
3056 		 **/
3057 		struct mlx5_klm *data_klm = *seg;
3058 
3059 		data_klm->bcount = cpu_to_be32(data_len);
3060 		data_klm->key = cpu_to_be32(data_key);
3061 		data_klm->va = cpu_to_be64(data_va);
3062 		wqe_size = ALIGN(sizeof(*data_klm), 64);
3063 	} else {
3064 		/**
3065 		 * Source domain contains signature information
3066 		 * So need construct a strided block format:
3067 		 *               ---------------------------
3068 		 *              |     stride_block_ctrl     |
3069 		 *               ---------------------------
3070 		 *              |          data_klm         |
3071 		 *               ---------------------------
3072 		 *              |          prot_klm         |
3073 		 *               ---------------------------
3074 		 *              |             BSF           |
3075 		 *               ---------------------------
3076 		 **/
3077 		struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
3078 		struct mlx5_stride_block_entry *data_sentry;
3079 		struct mlx5_stride_block_entry *prot_sentry;
3080 		u32 prot_key = wr->prot->lkey;
3081 		u64 prot_va = wr->prot->addr;
3082 		u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
3083 		int prot_size;
3084 
3085 		sblock_ctrl = *seg;
3086 		data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
3087 		prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
3088 
3089 		prot_size = prot_field_size(sig_attrs->mem.sig_type);
3090 		if (!prot_size) {
3091 			pr_err("Bad block size given: %u\n", block_size);
3092 			return -EINVAL;
3093 		}
3094 		sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
3095 							    prot_size);
3096 		sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
3097 		sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
3098 		sblock_ctrl->num_entries = cpu_to_be16(2);
3099 
3100 		data_sentry->bcount = cpu_to_be16(block_size);
3101 		data_sentry->key = cpu_to_be32(data_key);
3102 		data_sentry->va = cpu_to_be64(data_va);
3103 		data_sentry->stride = cpu_to_be16(block_size);
3104 
3105 		prot_sentry->bcount = cpu_to_be16(prot_size);
3106 		prot_sentry->key = cpu_to_be32(prot_key);
3107 		prot_sentry->va = cpu_to_be64(prot_va);
3108 		prot_sentry->stride = cpu_to_be16(prot_size);
3109 
3110 		wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
3111 				 sizeof(*prot_sentry), 64);
3112 	}
3113 
3114 	*seg += wqe_size;
3115 	*size += wqe_size / 16;
3116 	if (unlikely((*seg == qp->sq.qend)))
3117 		*seg = mlx5_get_send_wqe(qp, 0);
3118 
3119 	bsf = *seg;
3120 	ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
3121 	if (ret)
3122 		return -EINVAL;
3123 
3124 	*seg += sizeof(*bsf);
3125 	*size += sizeof(*bsf) / 16;
3126 	if (unlikely((*seg == qp->sq.qend)))
3127 		*seg = mlx5_get_send_wqe(qp, 0);
3128 
3129 	return 0;
3130 }
3131 
3132 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
3133 				 struct ib_sig_handover_wr *wr, u32 nelements,
3134 				 u32 length, u32 pdn)
3135 {
3136 	struct ib_mr *sig_mr = wr->sig_mr;
3137 	u32 sig_key = sig_mr->rkey;
3138 	u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
3139 
3140 	memset(seg, 0, sizeof(*seg));
3141 
3142 	seg->flags = get_umr_flags(wr->access_flags) |
3143 				   MLX5_ACCESS_MODE_KLM;
3144 	seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
3145 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
3146 				    MLX5_MKEY_BSF_EN | pdn);
3147 	seg->len = cpu_to_be64(length);
3148 	seg->xlt_oct_size = cpu_to_be32(be16_to_cpu(get_klm_octo(nelements)));
3149 	seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
3150 }
3151 
3152 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
3153 				u32 nelements)
3154 {
3155 	memset(umr, 0, sizeof(*umr));
3156 
3157 	umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
3158 	umr->klm_octowords = get_klm_octo(nelements);
3159 	umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
3160 	umr->mkey_mask = sig_mkey_mask();
3161 }
3162 
3163 
3164 static int set_sig_umr_wr(struct ib_send_wr *send_wr, struct mlx5_ib_qp *qp,
3165 			  void **seg, int *size)
3166 {
3167 	struct ib_sig_handover_wr *wr = sig_handover_wr(send_wr);
3168 	struct mlx5_ib_mr *sig_mr = to_mmr(wr->sig_mr);
3169 	u32 pdn = get_pd(qp)->pdn;
3170 	u32 klm_oct_size;
3171 	int region_len, ret;
3172 
3173 	if (unlikely(wr->wr.num_sge != 1) ||
3174 	    unlikely(wr->access_flags & IB_ACCESS_REMOTE_ATOMIC) ||
3175 	    unlikely(!sig_mr->sig) || unlikely(!qp->signature_en) ||
3176 	    unlikely(!sig_mr->sig->sig_status_checked))
3177 		return -EINVAL;
3178 
3179 	/* length of the protected region, data + protection */
3180 	region_len = wr->wr.sg_list->length;
3181 	if (wr->prot &&
3182 	    (wr->prot->lkey != wr->wr.sg_list->lkey  ||
3183 	     wr->prot->addr != wr->wr.sg_list->addr  ||
3184 	     wr->prot->length != wr->wr.sg_list->length))
3185 		region_len += wr->prot->length;
3186 
3187 	/**
3188 	 * KLM octoword size - if protection was provided
3189 	 * then we use strided block format (3 octowords),
3190 	 * else we use single KLM (1 octoword)
3191 	 **/
3192 	klm_oct_size = wr->prot ? 3 : 1;
3193 
3194 	set_sig_umr_segment(*seg, klm_oct_size);
3195 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3196 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3197 	if (unlikely((*seg == qp->sq.qend)))
3198 		*seg = mlx5_get_send_wqe(qp, 0);
3199 
3200 	set_sig_mkey_segment(*seg, wr, klm_oct_size, region_len, pdn);
3201 	*seg += sizeof(struct mlx5_mkey_seg);
3202 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3203 	if (unlikely((*seg == qp->sq.qend)))
3204 		*seg = mlx5_get_send_wqe(qp, 0);
3205 
3206 	ret = set_sig_data_segment(wr, qp, seg, size);
3207 	if (ret)
3208 		return ret;
3209 
3210 	sig_mr->sig->sig_status_checked = false;
3211 	return 0;
3212 }
3213 
3214 static int set_psv_wr(struct ib_sig_domain *domain,
3215 		      u32 psv_idx, void **seg, int *size)
3216 {
3217 	struct mlx5_seg_set_psv *psv_seg = *seg;
3218 
3219 	memset(psv_seg, 0, sizeof(*psv_seg));
3220 	psv_seg->psv_num = cpu_to_be32(psv_idx);
3221 	switch (domain->sig_type) {
3222 	case IB_SIG_TYPE_NONE:
3223 		break;
3224 	case IB_SIG_TYPE_T10_DIF:
3225 		psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
3226 						     domain->sig.dif.app_tag);
3227 		psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
3228 		break;
3229 	default:
3230 		pr_err("Bad signature type given.\n");
3231 		return 1;
3232 	}
3233 
3234 	*seg += sizeof(*psv_seg);
3235 	*size += sizeof(*psv_seg) / 16;
3236 
3237 	return 0;
3238 }
3239 
3240 static int set_reg_wr(struct mlx5_ib_qp *qp,
3241 		      struct ib_reg_wr *wr,
3242 		      void **seg, int *size)
3243 {
3244 	struct mlx5_ib_mr *mr = to_mmr(wr->mr);
3245 	struct mlx5_ib_pd *pd = to_mpd(qp->ibqp.pd);
3246 
3247 	if (unlikely(wr->wr.send_flags & IB_SEND_INLINE)) {
3248 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
3249 			     "Invalid IB_SEND_INLINE send flag\n");
3250 		return -EINVAL;
3251 	}
3252 
3253 	set_reg_umr_seg(*seg, mr);
3254 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3255 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3256 	if (unlikely((*seg == qp->sq.qend)))
3257 		*seg = mlx5_get_send_wqe(qp, 0);
3258 
3259 	set_reg_mkey_seg(*seg, mr, wr->key, wr->access);
3260 	*seg += sizeof(struct mlx5_mkey_seg);
3261 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3262 	if (unlikely((*seg == qp->sq.qend)))
3263 		*seg = mlx5_get_send_wqe(qp, 0);
3264 
3265 	set_reg_data_seg(*seg, mr, pd);
3266 	*seg += sizeof(struct mlx5_wqe_data_seg);
3267 	*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
3268 
3269 	return 0;
3270 }
3271 
3272 static void set_linv_wr(struct mlx5_ib_qp *qp, void **seg, int *size)
3273 {
3274 	set_linv_umr_seg(*seg);
3275 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3276 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3277 	if (unlikely((*seg == qp->sq.qend)))
3278 		*seg = mlx5_get_send_wqe(qp, 0);
3279 	set_linv_mkey_seg(*seg);
3280 	*seg += sizeof(struct mlx5_mkey_seg);
3281 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3282 	if (unlikely((*seg == qp->sq.qend)))
3283 		*seg = mlx5_get_send_wqe(qp, 0);
3284 }
3285 
3286 static void dump_wqe(struct mlx5_ib_qp *qp, int idx, int size_16)
3287 {
3288 	__be32 *p = NULL;
3289 	int tidx = idx;
3290 	int i, j;
3291 
3292 	pr_debug("dump wqe at %p\n", mlx5_get_send_wqe(qp, tidx));
3293 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
3294 		if ((i & 0xf) == 0) {
3295 			void *buf = mlx5_get_send_wqe(qp, tidx);
3296 			tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
3297 			p = buf;
3298 			j = 0;
3299 		}
3300 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
3301 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
3302 			 be32_to_cpu(p[j + 3]));
3303 	}
3304 }
3305 
3306 static void mlx5_bf_copy(u64 __iomem *dst, u64 *src,
3307 			 unsigned bytecnt, struct mlx5_ib_qp *qp)
3308 {
3309 	while (bytecnt > 0) {
3310 		__iowrite64_copy(dst++, src++, 8);
3311 		__iowrite64_copy(dst++, src++, 8);
3312 		__iowrite64_copy(dst++, src++, 8);
3313 		__iowrite64_copy(dst++, src++, 8);
3314 		__iowrite64_copy(dst++, src++, 8);
3315 		__iowrite64_copy(dst++, src++, 8);
3316 		__iowrite64_copy(dst++, src++, 8);
3317 		__iowrite64_copy(dst++, src++, 8);
3318 		bytecnt -= 64;
3319 		if (unlikely(src == qp->sq.qend))
3320 			src = mlx5_get_send_wqe(qp, 0);
3321 	}
3322 }
3323 
3324 static u8 get_fence(u8 fence, struct ib_send_wr *wr)
3325 {
3326 	if (unlikely(wr->opcode == IB_WR_LOCAL_INV &&
3327 		     wr->send_flags & IB_SEND_FENCE))
3328 		return MLX5_FENCE_MODE_STRONG_ORDERING;
3329 
3330 	if (unlikely(fence)) {
3331 		if (wr->send_flags & IB_SEND_FENCE)
3332 			return MLX5_FENCE_MODE_SMALL_AND_FENCE;
3333 		else
3334 			return fence;
3335 
3336 	} else {
3337 		return 0;
3338 	}
3339 }
3340 
3341 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
3342 		     struct mlx5_wqe_ctrl_seg **ctrl,
3343 		     struct ib_send_wr *wr, unsigned *idx,
3344 		     int *size, int nreq)
3345 {
3346 	int err = 0;
3347 
3348 	if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq))) {
3349 		err = -ENOMEM;
3350 		return err;
3351 	}
3352 
3353 	*idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
3354 	*seg = mlx5_get_send_wqe(qp, *idx);
3355 	*ctrl = *seg;
3356 	*(uint32_t *)(*seg + 8) = 0;
3357 	(*ctrl)->imm = send_ieth(wr);
3358 	(*ctrl)->fm_ce_se = qp->sq_signal_bits |
3359 		(wr->send_flags & IB_SEND_SIGNALED ?
3360 		 MLX5_WQE_CTRL_CQ_UPDATE : 0) |
3361 		(wr->send_flags & IB_SEND_SOLICITED ?
3362 		 MLX5_WQE_CTRL_SOLICITED : 0);
3363 
3364 	*seg += sizeof(**ctrl);
3365 	*size = sizeof(**ctrl) / 16;
3366 
3367 	return err;
3368 }
3369 
3370 static void finish_wqe(struct mlx5_ib_qp *qp,
3371 		       struct mlx5_wqe_ctrl_seg *ctrl,
3372 		       u8 size, unsigned idx, u64 wr_id,
3373 		       int nreq, u8 fence, u8 next_fence,
3374 		       u32 mlx5_opcode)
3375 {
3376 	u8 opmod = 0;
3377 
3378 	ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
3379 					     mlx5_opcode | ((u32)opmod << 24));
3380 	ctrl->qpn_ds = cpu_to_be32(size | (qp->trans_qp.base.mqp.qpn << 8));
3381 	ctrl->fm_ce_se |= fence;
3382 	qp->fm_cache = next_fence;
3383 	if (unlikely(qp->wq_sig))
3384 		ctrl->signature = wq_sig(ctrl);
3385 
3386 	qp->sq.wrid[idx] = wr_id;
3387 	qp->sq.w_list[idx].opcode = mlx5_opcode;
3388 	qp->sq.wqe_head[idx] = qp->sq.head + nreq;
3389 	qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
3390 	qp->sq.w_list[idx].next = qp->sq.cur_post;
3391 }
3392 
3393 
3394 int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
3395 		      struct ib_send_wr **bad_wr)
3396 {
3397 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
3398 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3399 	struct mlx5_ib_qp *qp;
3400 	struct mlx5_ib_mr *mr;
3401 	struct mlx5_wqe_data_seg *dpseg;
3402 	struct mlx5_wqe_xrc_seg *xrc;
3403 	struct mlx5_bf *bf;
3404 	int uninitialized_var(size);
3405 	void *qend;
3406 	unsigned long flags;
3407 	unsigned idx;
3408 	int err = 0;
3409 	int inl = 0;
3410 	int num_sge;
3411 	void *seg;
3412 	int nreq;
3413 	int i;
3414 	u8 next_fence = 0;
3415 	u8 fence;
3416 
3417 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
3418 		return mlx5_ib_gsi_post_send(ibqp, wr, bad_wr);
3419 
3420 	qp = to_mqp(ibqp);
3421 	bf = qp->bf;
3422 	qend = qp->sq.qend;
3423 
3424 	spin_lock_irqsave(&qp->sq.lock, flags);
3425 
3426 	for (nreq = 0; wr; nreq++, wr = wr->next) {
3427 		if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
3428 			mlx5_ib_warn(dev, "\n");
3429 			err = -EINVAL;
3430 			*bad_wr = wr;
3431 			goto out;
3432 		}
3433 
3434 		fence = qp->fm_cache;
3435 		num_sge = wr->num_sge;
3436 		if (unlikely(num_sge > qp->sq.max_gs)) {
3437 			mlx5_ib_warn(dev, "\n");
3438 			err = -ENOMEM;
3439 			*bad_wr = wr;
3440 			goto out;
3441 		}
3442 
3443 		err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, nreq);
3444 		if (err) {
3445 			mlx5_ib_warn(dev, "\n");
3446 			err = -ENOMEM;
3447 			*bad_wr = wr;
3448 			goto out;
3449 		}
3450 
3451 		switch (ibqp->qp_type) {
3452 		case IB_QPT_XRC_INI:
3453 			xrc = seg;
3454 			seg += sizeof(*xrc);
3455 			size += sizeof(*xrc) / 16;
3456 			/* fall through */
3457 		case IB_QPT_RC:
3458 			switch (wr->opcode) {
3459 			case IB_WR_RDMA_READ:
3460 			case IB_WR_RDMA_WRITE:
3461 			case IB_WR_RDMA_WRITE_WITH_IMM:
3462 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
3463 					      rdma_wr(wr)->rkey);
3464 				seg += sizeof(struct mlx5_wqe_raddr_seg);
3465 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
3466 				break;
3467 
3468 			case IB_WR_ATOMIC_CMP_AND_SWP:
3469 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3470 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3471 				mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
3472 				err = -ENOSYS;
3473 				*bad_wr = wr;
3474 				goto out;
3475 
3476 			case IB_WR_LOCAL_INV:
3477 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3478 				qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
3479 				ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
3480 				set_linv_wr(qp, &seg, &size);
3481 				num_sge = 0;
3482 				break;
3483 
3484 			case IB_WR_REG_MR:
3485 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3486 				qp->sq.wr_data[idx] = IB_WR_REG_MR;
3487 				ctrl->imm = cpu_to_be32(reg_wr(wr)->key);
3488 				err = set_reg_wr(qp, reg_wr(wr), &seg, &size);
3489 				if (err) {
3490 					*bad_wr = wr;
3491 					goto out;
3492 				}
3493 				num_sge = 0;
3494 				break;
3495 
3496 			case IB_WR_REG_SIG_MR:
3497 				qp->sq.wr_data[idx] = IB_WR_REG_SIG_MR;
3498 				mr = to_mmr(sig_handover_wr(wr)->sig_mr);
3499 
3500 				ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
3501 				err = set_sig_umr_wr(wr, qp, &seg, &size);
3502 				if (err) {
3503 					mlx5_ib_warn(dev, "\n");
3504 					*bad_wr = wr;
3505 					goto out;
3506 				}
3507 
3508 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3509 					   nreq, get_fence(fence, wr),
3510 					   next_fence, MLX5_OPCODE_UMR);
3511 				/*
3512 				 * SET_PSV WQEs are not signaled and solicited
3513 				 * on error
3514 				 */
3515 				wr->send_flags &= ~IB_SEND_SIGNALED;
3516 				wr->send_flags |= IB_SEND_SOLICITED;
3517 				err = begin_wqe(qp, &seg, &ctrl, wr,
3518 						&idx, &size, nreq);
3519 				if (err) {
3520 					mlx5_ib_warn(dev, "\n");
3521 					err = -ENOMEM;
3522 					*bad_wr = wr;
3523 					goto out;
3524 				}
3525 
3526 				err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->mem,
3527 						 mr->sig->psv_memory.psv_idx, &seg,
3528 						 &size);
3529 				if (err) {
3530 					mlx5_ib_warn(dev, "\n");
3531 					*bad_wr = wr;
3532 					goto out;
3533 				}
3534 
3535 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3536 					   nreq, get_fence(fence, wr),
3537 					   next_fence, MLX5_OPCODE_SET_PSV);
3538 				err = begin_wqe(qp, &seg, &ctrl, wr,
3539 						&idx, &size, nreq);
3540 				if (err) {
3541 					mlx5_ib_warn(dev, "\n");
3542 					err = -ENOMEM;
3543 					*bad_wr = wr;
3544 					goto out;
3545 				}
3546 
3547 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3548 				err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->wire,
3549 						 mr->sig->psv_wire.psv_idx, &seg,
3550 						 &size);
3551 				if (err) {
3552 					mlx5_ib_warn(dev, "\n");
3553 					*bad_wr = wr;
3554 					goto out;
3555 				}
3556 
3557 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3558 					   nreq, get_fence(fence, wr),
3559 					   next_fence, MLX5_OPCODE_SET_PSV);
3560 				num_sge = 0;
3561 				goto skip_psv;
3562 
3563 			default:
3564 				break;
3565 			}
3566 			break;
3567 
3568 		case IB_QPT_UC:
3569 			switch (wr->opcode) {
3570 			case IB_WR_RDMA_WRITE:
3571 			case IB_WR_RDMA_WRITE_WITH_IMM:
3572 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
3573 					      rdma_wr(wr)->rkey);
3574 				seg  += sizeof(struct mlx5_wqe_raddr_seg);
3575 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
3576 				break;
3577 
3578 			default:
3579 				break;
3580 			}
3581 			break;
3582 
3583 		case IB_QPT_SMI:
3584 		case MLX5_IB_QPT_HW_GSI:
3585 			set_datagram_seg(seg, wr);
3586 			seg += sizeof(struct mlx5_wqe_datagram_seg);
3587 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
3588 			if (unlikely((seg == qend)))
3589 				seg = mlx5_get_send_wqe(qp, 0);
3590 			break;
3591 		case IB_QPT_UD:
3592 			set_datagram_seg(seg, wr);
3593 			seg += sizeof(struct mlx5_wqe_datagram_seg);
3594 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
3595 
3596 			if (unlikely((seg == qend)))
3597 				seg = mlx5_get_send_wqe(qp, 0);
3598 
3599 			/* handle qp that supports ud offload */
3600 			if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO) {
3601 				struct mlx5_wqe_eth_pad *pad;
3602 
3603 				pad = seg;
3604 				memset(pad, 0, sizeof(struct mlx5_wqe_eth_pad));
3605 				seg += sizeof(struct mlx5_wqe_eth_pad);
3606 				size += sizeof(struct mlx5_wqe_eth_pad) / 16;
3607 
3608 				seg = set_eth_seg(seg, wr, qend, qp, &size);
3609 
3610 				if (unlikely((seg == qend)))
3611 					seg = mlx5_get_send_wqe(qp, 0);
3612 			}
3613 			break;
3614 		case MLX5_IB_QPT_REG_UMR:
3615 			if (wr->opcode != MLX5_IB_WR_UMR) {
3616 				err = -EINVAL;
3617 				mlx5_ib_warn(dev, "bad opcode\n");
3618 				goto out;
3619 			}
3620 			qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
3621 			ctrl->imm = cpu_to_be32(umr_wr(wr)->mkey);
3622 			set_reg_umr_segment(seg, wr);
3623 			seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3624 			size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3625 			if (unlikely((seg == qend)))
3626 				seg = mlx5_get_send_wqe(qp, 0);
3627 			set_reg_mkey_segment(seg, wr);
3628 			seg += sizeof(struct mlx5_mkey_seg);
3629 			size += sizeof(struct mlx5_mkey_seg) / 16;
3630 			if (unlikely((seg == qend)))
3631 				seg = mlx5_get_send_wqe(qp, 0);
3632 			break;
3633 
3634 		default:
3635 			break;
3636 		}
3637 
3638 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
3639 			int uninitialized_var(sz);
3640 
3641 			err = set_data_inl_seg(qp, wr, seg, &sz);
3642 			if (unlikely(err)) {
3643 				mlx5_ib_warn(dev, "\n");
3644 				*bad_wr = wr;
3645 				goto out;
3646 			}
3647 			inl = 1;
3648 			size += sz;
3649 		} else {
3650 			dpseg = seg;
3651 			for (i = 0; i < num_sge; i++) {
3652 				if (unlikely(dpseg == qend)) {
3653 					seg = mlx5_get_send_wqe(qp, 0);
3654 					dpseg = seg;
3655 				}
3656 				if (likely(wr->sg_list[i].length)) {
3657 					set_data_ptr_seg(dpseg, wr->sg_list + i);
3658 					size += sizeof(struct mlx5_wqe_data_seg) / 16;
3659 					dpseg++;
3660 				}
3661 			}
3662 		}
3663 
3664 		finish_wqe(qp, ctrl, size, idx, wr->wr_id, nreq,
3665 			   get_fence(fence, wr), next_fence,
3666 			   mlx5_ib_opcode[wr->opcode]);
3667 skip_psv:
3668 		if (0)
3669 			dump_wqe(qp, idx, size);
3670 	}
3671 
3672 out:
3673 	if (likely(nreq)) {
3674 		qp->sq.head += nreq;
3675 
3676 		/* Make sure that descriptors are written before
3677 		 * updating doorbell record and ringing the doorbell
3678 		 */
3679 		wmb();
3680 
3681 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
3682 
3683 		/* Make sure doorbell record is visible to the HCA before
3684 		 * we hit doorbell */
3685 		wmb();
3686 
3687 		if (bf->need_lock)
3688 			spin_lock(&bf->lock);
3689 		else
3690 			__acquire(&bf->lock);
3691 
3692 		/* TBD enable WC */
3693 		if (0 && nreq == 1 && bf->uuarn && inl && size > 1 && size <= bf->buf_size / 16) {
3694 			mlx5_bf_copy(bf->reg + bf->offset, (u64 *)ctrl, ALIGN(size * 16, 64), qp);
3695 			/* wc_wmb(); */
3696 		} else {
3697 			mlx5_write64((__be32 *)ctrl, bf->regreg + bf->offset,
3698 				     MLX5_GET_DOORBELL_LOCK(&bf->lock32));
3699 			/* Make sure doorbells don't leak out of SQ spinlock
3700 			 * and reach the HCA out of order.
3701 			 */
3702 			mmiowb();
3703 		}
3704 		bf->offset ^= bf->buf_size;
3705 		if (bf->need_lock)
3706 			spin_unlock(&bf->lock);
3707 		else
3708 			__release(&bf->lock);
3709 	}
3710 
3711 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3712 
3713 	return err;
3714 }
3715 
3716 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
3717 {
3718 	sig->signature = calc_sig(sig, size);
3719 }
3720 
3721 int mlx5_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3722 		      struct ib_recv_wr **bad_wr)
3723 {
3724 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3725 	struct mlx5_wqe_data_seg *scat;
3726 	struct mlx5_rwqe_sig *sig;
3727 	unsigned long flags;
3728 	int err = 0;
3729 	int nreq;
3730 	int ind;
3731 	int i;
3732 
3733 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
3734 		return mlx5_ib_gsi_post_recv(ibqp, wr, bad_wr);
3735 
3736 	spin_lock_irqsave(&qp->rq.lock, flags);
3737 
3738 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3739 
3740 	for (nreq = 0; wr; nreq++, wr = wr->next) {
3741 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3742 			err = -ENOMEM;
3743 			*bad_wr = wr;
3744 			goto out;
3745 		}
3746 
3747 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3748 			err = -EINVAL;
3749 			*bad_wr = wr;
3750 			goto out;
3751 		}
3752 
3753 		scat = get_recv_wqe(qp, ind);
3754 		if (qp->wq_sig)
3755 			scat++;
3756 
3757 		for (i = 0; i < wr->num_sge; i++)
3758 			set_data_ptr_seg(scat + i, wr->sg_list + i);
3759 
3760 		if (i < qp->rq.max_gs) {
3761 			scat[i].byte_count = 0;
3762 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
3763 			scat[i].addr       = 0;
3764 		}
3765 
3766 		if (qp->wq_sig) {
3767 			sig = (struct mlx5_rwqe_sig *)scat;
3768 			set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
3769 		}
3770 
3771 		qp->rq.wrid[ind] = wr->wr_id;
3772 
3773 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3774 	}
3775 
3776 out:
3777 	if (likely(nreq)) {
3778 		qp->rq.head += nreq;
3779 
3780 		/* Make sure that descriptors are written before
3781 		 * doorbell record.
3782 		 */
3783 		wmb();
3784 
3785 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3786 	}
3787 
3788 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3789 
3790 	return err;
3791 }
3792 
3793 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
3794 {
3795 	switch (mlx5_state) {
3796 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
3797 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
3798 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
3799 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
3800 	case MLX5_QP_STATE_SQ_DRAINING:
3801 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
3802 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
3803 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
3804 	default:		     return -1;
3805 	}
3806 }
3807 
3808 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
3809 {
3810 	switch (mlx5_mig_state) {
3811 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
3812 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
3813 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3814 	default: return -1;
3815 	}
3816 }
3817 
3818 static int to_ib_qp_access_flags(int mlx5_flags)
3819 {
3820 	int ib_flags = 0;
3821 
3822 	if (mlx5_flags & MLX5_QP_BIT_RRE)
3823 		ib_flags |= IB_ACCESS_REMOTE_READ;
3824 	if (mlx5_flags & MLX5_QP_BIT_RWE)
3825 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3826 	if (mlx5_flags & MLX5_QP_BIT_RAE)
3827 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3828 
3829 	return ib_flags;
3830 }
3831 
3832 static void to_ib_ah_attr(struct mlx5_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
3833 				struct mlx5_qp_path *path)
3834 {
3835 	struct mlx5_core_dev *dev = ibdev->mdev;
3836 
3837 	memset(ib_ah_attr, 0, sizeof(*ib_ah_attr));
3838 	ib_ah_attr->port_num	  = path->port;
3839 
3840 	if (ib_ah_attr->port_num == 0 ||
3841 	    ib_ah_attr->port_num > MLX5_CAP_GEN(dev, num_ports))
3842 		return;
3843 
3844 	ib_ah_attr->sl = path->dci_cfi_prio_sl & 0xf;
3845 
3846 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
3847 	ib_ah_attr->src_path_bits = path->grh_mlid & 0x7f;
3848 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
3849 	ib_ah_attr->ah_flags      = (path->grh_mlid & (1 << 7)) ? IB_AH_GRH : 0;
3850 	if (ib_ah_attr->ah_flags) {
3851 		ib_ah_attr->grh.sgid_index = path->mgid_index;
3852 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
3853 		ib_ah_attr->grh.traffic_class =
3854 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3855 		ib_ah_attr->grh.flow_label =
3856 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3857 		memcpy(ib_ah_attr->grh.dgid.raw,
3858 		       path->rgid, sizeof(ib_ah_attr->grh.dgid.raw));
3859 	}
3860 }
3861 
3862 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
3863 					struct mlx5_ib_sq *sq,
3864 					u8 *sq_state)
3865 {
3866 	void *out;
3867 	void *sqc;
3868 	int inlen;
3869 	int err;
3870 
3871 	inlen = MLX5_ST_SZ_BYTES(query_sq_out);
3872 	out = mlx5_vzalloc(inlen);
3873 	if (!out)
3874 		return -ENOMEM;
3875 
3876 	err = mlx5_core_query_sq(dev->mdev, sq->base.mqp.qpn, out);
3877 	if (err)
3878 		goto out;
3879 
3880 	sqc = MLX5_ADDR_OF(query_sq_out, out, sq_context);
3881 	*sq_state = MLX5_GET(sqc, sqc, state);
3882 	sq->state = *sq_state;
3883 
3884 out:
3885 	kvfree(out);
3886 	return err;
3887 }
3888 
3889 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
3890 					struct mlx5_ib_rq *rq,
3891 					u8 *rq_state)
3892 {
3893 	void *out;
3894 	void *rqc;
3895 	int inlen;
3896 	int err;
3897 
3898 	inlen = MLX5_ST_SZ_BYTES(query_rq_out);
3899 	out = mlx5_vzalloc(inlen);
3900 	if (!out)
3901 		return -ENOMEM;
3902 
3903 	err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
3904 	if (err)
3905 		goto out;
3906 
3907 	rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
3908 	*rq_state = MLX5_GET(rqc, rqc, state);
3909 	rq->state = *rq_state;
3910 
3911 out:
3912 	kvfree(out);
3913 	return err;
3914 }
3915 
3916 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
3917 				  struct mlx5_ib_qp *qp, u8 *qp_state)
3918 {
3919 	static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
3920 		[MLX5_RQC_STATE_RST] = {
3921 			[MLX5_SQC_STATE_RST]	= IB_QPS_RESET,
3922 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
3923 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE_BAD,
3924 			[MLX5_SQ_STATE_NA]	= IB_QPS_RESET,
3925 		},
3926 		[MLX5_RQC_STATE_RDY] = {
3927 			[MLX5_SQC_STATE_RST]	= MLX5_QP_STATE_BAD,
3928 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
3929 			[MLX5_SQC_STATE_ERR]	= IB_QPS_SQE,
3930 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE,
3931 		},
3932 		[MLX5_RQC_STATE_ERR] = {
3933 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE_BAD,
3934 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
3935 			[MLX5_SQC_STATE_ERR]	= IB_QPS_ERR,
3936 			[MLX5_SQ_STATE_NA]	= IB_QPS_ERR,
3937 		},
3938 		[MLX5_RQ_STATE_NA] = {
3939 			[MLX5_SQC_STATE_RST]    = IB_QPS_RESET,
3940 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
3941 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE,
3942 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE_BAD,
3943 		},
3944 	};
3945 
3946 	*qp_state = sqrq_trans[rq_state][sq_state];
3947 
3948 	if (*qp_state == MLX5_QP_STATE_BAD) {
3949 		WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
3950 		     qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
3951 		     qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
3952 		return -EINVAL;
3953 	}
3954 
3955 	if (*qp_state == MLX5_QP_STATE)
3956 		*qp_state = qp->state;
3957 
3958 	return 0;
3959 }
3960 
3961 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
3962 				     struct mlx5_ib_qp *qp,
3963 				     u8 *raw_packet_qp_state)
3964 {
3965 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
3966 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
3967 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
3968 	int err;
3969 	u8 sq_state = MLX5_SQ_STATE_NA;
3970 	u8 rq_state = MLX5_RQ_STATE_NA;
3971 
3972 	if (qp->sq.wqe_cnt) {
3973 		err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
3974 		if (err)
3975 			return err;
3976 	}
3977 
3978 	if (qp->rq.wqe_cnt) {
3979 		err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
3980 		if (err)
3981 			return err;
3982 	}
3983 
3984 	return sqrq_state_to_qp_state(sq_state, rq_state, qp,
3985 				      raw_packet_qp_state);
3986 }
3987 
3988 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3989 			 struct ib_qp_attr *qp_attr)
3990 {
3991 	struct mlx5_query_qp_mbox_out *outb;
3992 	struct mlx5_qp_context *context;
3993 	int mlx5_state;
3994 	int err = 0;
3995 
3996 	outb = kzalloc(sizeof(*outb), GFP_KERNEL);
3997 	if (!outb)
3998 		return -ENOMEM;
3999 
4000 	context = &outb->ctx;
4001 	err = mlx5_core_qp_query(dev->mdev, &qp->trans_qp.base.mqp, outb,
4002 				 sizeof(*outb));
4003 	if (err)
4004 		goto out;
4005 
4006 	mlx5_state = be32_to_cpu(context->flags) >> 28;
4007 
4008 	qp->state		     = to_ib_qp_state(mlx5_state);
4009 	qp_attr->path_mtu	     = context->mtu_msgmax >> 5;
4010 	qp_attr->path_mig_state	     =
4011 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
4012 	qp_attr->qkey		     = be32_to_cpu(context->qkey);
4013 	qp_attr->rq_psn		     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
4014 	qp_attr->sq_psn		     = be32_to_cpu(context->next_send_psn) & 0xffffff;
4015 	qp_attr->dest_qp_num	     = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
4016 	qp_attr->qp_access_flags     =
4017 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
4018 
4019 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4020 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
4021 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
4022 		qp_attr->alt_pkey_index =
4023 			be16_to_cpu(context->alt_path.pkey_index);
4024 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
4025 	}
4026 
4027 	qp_attr->pkey_index = be16_to_cpu(context->pri_path.pkey_index);
4028 	qp_attr->port_num = context->pri_path.port;
4029 
4030 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4031 	qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
4032 
4033 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
4034 
4035 	qp_attr->max_dest_rd_atomic =
4036 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
4037 	qp_attr->min_rnr_timer	    =
4038 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
4039 	qp_attr->timeout	    = context->pri_path.ackto_lt >> 3;
4040 	qp_attr->retry_cnt	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
4041 	qp_attr->rnr_retry	    = (be32_to_cpu(context->params1) >> 13) & 0x7;
4042 	qp_attr->alt_timeout	    = context->alt_path.ackto_lt >> 3;
4043 
4044 out:
4045 	kfree(outb);
4046 	return err;
4047 }
4048 
4049 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
4050 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
4051 {
4052 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4053 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4054 	int err = 0;
4055 	u8 raw_packet_qp_state;
4056 
4057 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4058 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
4059 					    qp_init_attr);
4060 
4061 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4062 	/*
4063 	 * Wait for any outstanding page faults, in case the user frees memory
4064 	 * based upon this query's result.
4065 	 */
4066 	flush_workqueue(mlx5_ib_page_fault_wq);
4067 #endif
4068 
4069 	mutex_lock(&qp->mutex);
4070 
4071 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
4072 		err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
4073 		if (err)
4074 			goto out;
4075 		qp->state = raw_packet_qp_state;
4076 		qp_attr->port_num = 1;
4077 	} else {
4078 		err = query_qp_attr(dev, qp, qp_attr);
4079 		if (err)
4080 			goto out;
4081 	}
4082 
4083 	qp_attr->qp_state	     = qp->state;
4084 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4085 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4086 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4087 
4088 	if (!ibqp->uobject) {
4089 		qp_attr->cap.max_send_wr  = qp->sq.max_post;
4090 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4091 		qp_init_attr->qp_context = ibqp->qp_context;
4092 	} else {
4093 		qp_attr->cap.max_send_wr  = 0;
4094 		qp_attr->cap.max_send_sge = 0;
4095 	}
4096 
4097 	qp_init_attr->qp_type = ibqp->qp_type;
4098 	qp_init_attr->recv_cq = ibqp->recv_cq;
4099 	qp_init_attr->send_cq = ibqp->send_cq;
4100 	qp_init_attr->srq = ibqp->srq;
4101 	qp_attr->cap.max_inline_data = qp->max_inline_data;
4102 
4103 	qp_init_attr->cap	     = qp_attr->cap;
4104 
4105 	qp_init_attr->create_flags = 0;
4106 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4107 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4108 
4109 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
4110 		qp_init_attr->create_flags |= IB_QP_CREATE_CROSS_CHANNEL;
4111 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
4112 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_SEND;
4113 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
4114 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_RECV;
4115 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
4116 		qp_init_attr->create_flags |= mlx5_ib_create_qp_sqpn_qp1();
4117 
4118 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
4119 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4120 
4121 out:
4122 	mutex_unlock(&qp->mutex);
4123 	return err;
4124 }
4125 
4126 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
4127 					  struct ib_ucontext *context,
4128 					  struct ib_udata *udata)
4129 {
4130 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4131 	struct mlx5_ib_xrcd *xrcd;
4132 	int err;
4133 
4134 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
4135 		return ERR_PTR(-ENOSYS);
4136 
4137 	xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
4138 	if (!xrcd)
4139 		return ERR_PTR(-ENOMEM);
4140 
4141 	err = mlx5_core_xrcd_alloc(dev->mdev, &xrcd->xrcdn);
4142 	if (err) {
4143 		kfree(xrcd);
4144 		return ERR_PTR(-ENOMEM);
4145 	}
4146 
4147 	return &xrcd->ibxrcd;
4148 }
4149 
4150 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
4151 {
4152 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
4153 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
4154 	int err;
4155 
4156 	err = mlx5_core_xrcd_dealloc(dev->mdev, xrcdn);
4157 	if (err) {
4158 		mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
4159 		return err;
4160 	}
4161 
4162 	kfree(xrcd);
4163 
4164 	return 0;
4165 }
4166