xref: /linux/drivers/infiniband/hw/mlx5/qp.c (revision 827634added7f38b7d724cab1dccdb2b004c13c3)
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 "mlx5_ib.h"
36 #include "user.h"
37 
38 /* not supported currently */
39 static int wq_signature;
40 
41 enum {
42 	MLX5_IB_ACK_REQ_FREQ	= 8,
43 };
44 
45 enum {
46 	MLX5_IB_DEFAULT_SCHED_QUEUE	= 0x83,
47 	MLX5_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
48 	MLX5_IB_LINK_TYPE_IB		= 0,
49 	MLX5_IB_LINK_TYPE_ETH		= 1
50 };
51 
52 enum {
53 	MLX5_IB_SQ_STRIDE	= 6,
54 	MLX5_IB_CACHE_LINE_SIZE	= 64,
55 };
56 
57 static const u32 mlx5_ib_opcode[] = {
58 	[IB_WR_SEND]				= MLX5_OPCODE_SEND,
59 	[IB_WR_SEND_WITH_IMM]			= MLX5_OPCODE_SEND_IMM,
60 	[IB_WR_RDMA_WRITE]			= MLX5_OPCODE_RDMA_WRITE,
61 	[IB_WR_RDMA_WRITE_WITH_IMM]		= MLX5_OPCODE_RDMA_WRITE_IMM,
62 	[IB_WR_RDMA_READ]			= MLX5_OPCODE_RDMA_READ,
63 	[IB_WR_ATOMIC_CMP_AND_SWP]		= MLX5_OPCODE_ATOMIC_CS,
64 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= MLX5_OPCODE_ATOMIC_FA,
65 	[IB_WR_SEND_WITH_INV]			= MLX5_OPCODE_SEND_INVAL,
66 	[IB_WR_LOCAL_INV]			= MLX5_OPCODE_UMR,
67 	[IB_WR_FAST_REG_MR]			= MLX5_OPCODE_UMR,
68 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= MLX5_OPCODE_ATOMIC_MASKED_CS,
69 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= MLX5_OPCODE_ATOMIC_MASKED_FA,
70 	[MLX5_IB_WR_UMR]			= MLX5_OPCODE_UMR,
71 };
72 
73 
74 static int is_qp0(enum ib_qp_type qp_type)
75 {
76 	return qp_type == IB_QPT_SMI;
77 }
78 
79 static int is_qp1(enum ib_qp_type qp_type)
80 {
81 	return qp_type == IB_QPT_GSI;
82 }
83 
84 static int is_sqp(enum ib_qp_type qp_type)
85 {
86 	return is_qp0(qp_type) || is_qp1(qp_type);
87 }
88 
89 static void *get_wqe(struct mlx5_ib_qp *qp, int offset)
90 {
91 	return mlx5_buf_offset(&qp->buf, offset);
92 }
93 
94 static void *get_recv_wqe(struct mlx5_ib_qp *qp, int n)
95 {
96 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
97 }
98 
99 void *mlx5_get_send_wqe(struct mlx5_ib_qp *qp, int n)
100 {
101 	return get_wqe(qp, qp->sq.offset + (n << MLX5_IB_SQ_STRIDE));
102 }
103 
104 /**
105  * mlx5_ib_read_user_wqe() - Copy a user-space WQE to kernel space.
106  *
107  * @qp: QP to copy from.
108  * @send: copy from the send queue when non-zero, use the receive queue
109  *	  otherwise.
110  * @wqe_index:  index to start copying from. For send work queues, the
111  *		wqe_index is in units of MLX5_SEND_WQE_BB.
112  *		For receive work queue, it is the number of work queue
113  *		element in the queue.
114  * @buffer: destination buffer.
115  * @length: maximum number of bytes to copy.
116  *
117  * Copies at least a single WQE, but may copy more data.
118  *
119  * Return: the number of bytes copied, or an error code.
120  */
121 int mlx5_ib_read_user_wqe(struct mlx5_ib_qp *qp, int send, int wqe_index,
122 			  void *buffer, u32 length)
123 {
124 	struct ib_device *ibdev = qp->ibqp.device;
125 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
126 	struct mlx5_ib_wq *wq = send ? &qp->sq : &qp->rq;
127 	size_t offset;
128 	size_t wq_end;
129 	struct ib_umem *umem = qp->umem;
130 	u32 first_copy_length;
131 	int wqe_length;
132 	int ret;
133 
134 	if (wq->wqe_cnt == 0) {
135 		mlx5_ib_dbg(dev, "mlx5_ib_read_user_wqe for a QP with wqe_cnt == 0. qp_type: 0x%x\n",
136 			    qp->ibqp.qp_type);
137 		return -EINVAL;
138 	}
139 
140 	offset = wq->offset + ((wqe_index % wq->wqe_cnt) << wq->wqe_shift);
141 	wq_end = wq->offset + (wq->wqe_cnt << wq->wqe_shift);
142 
143 	if (send && length < sizeof(struct mlx5_wqe_ctrl_seg))
144 		return -EINVAL;
145 
146 	if (offset > umem->length ||
147 	    (send && offset + sizeof(struct mlx5_wqe_ctrl_seg) > umem->length))
148 		return -EINVAL;
149 
150 	first_copy_length = min_t(u32, offset + length, wq_end) - offset;
151 	ret = ib_umem_copy_from(buffer, umem, offset, first_copy_length);
152 	if (ret)
153 		return ret;
154 
155 	if (send) {
156 		struct mlx5_wqe_ctrl_seg *ctrl = buffer;
157 		int ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
158 
159 		wqe_length = ds * MLX5_WQE_DS_UNITS;
160 	} else {
161 		wqe_length = 1 << wq->wqe_shift;
162 	}
163 
164 	if (wqe_length <= first_copy_length)
165 		return first_copy_length;
166 
167 	ret = ib_umem_copy_from(buffer + first_copy_length, umem, wq->offset,
168 				wqe_length - first_copy_length);
169 	if (ret)
170 		return ret;
171 
172 	return wqe_length;
173 }
174 
175 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
176 {
177 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
178 	struct ib_event event;
179 
180 	if (type == MLX5_EVENT_TYPE_PATH_MIG)
181 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
182 
183 	if (ibqp->event_handler) {
184 		event.device     = ibqp->device;
185 		event.element.qp = ibqp;
186 		switch (type) {
187 		case MLX5_EVENT_TYPE_PATH_MIG:
188 			event.event = IB_EVENT_PATH_MIG;
189 			break;
190 		case MLX5_EVENT_TYPE_COMM_EST:
191 			event.event = IB_EVENT_COMM_EST;
192 			break;
193 		case MLX5_EVENT_TYPE_SQ_DRAINED:
194 			event.event = IB_EVENT_SQ_DRAINED;
195 			break;
196 		case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
197 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
198 			break;
199 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
200 			event.event = IB_EVENT_QP_FATAL;
201 			break;
202 		case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
203 			event.event = IB_EVENT_PATH_MIG_ERR;
204 			break;
205 		case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
206 			event.event = IB_EVENT_QP_REQ_ERR;
207 			break;
208 		case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
209 			event.event = IB_EVENT_QP_ACCESS_ERR;
210 			break;
211 		default:
212 			pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
213 			return;
214 		}
215 
216 		ibqp->event_handler(&event, ibqp->qp_context);
217 	}
218 }
219 
220 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
221 		       int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
222 {
223 	struct mlx5_general_caps *gen;
224 	int wqe_size;
225 	int wq_size;
226 
227 	gen = &dev->mdev->caps.gen;
228 	/* Sanity check RQ size before proceeding */
229 	if (cap->max_recv_wr  > gen->max_wqes)
230 		return -EINVAL;
231 
232 	if (!has_rq) {
233 		qp->rq.max_gs = 0;
234 		qp->rq.wqe_cnt = 0;
235 		qp->rq.wqe_shift = 0;
236 	} else {
237 		if (ucmd) {
238 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
239 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
240 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
241 			qp->rq.max_post = qp->rq.wqe_cnt;
242 		} else {
243 			wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
244 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
245 			wqe_size = roundup_pow_of_two(wqe_size);
246 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
247 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
248 			qp->rq.wqe_cnt = wq_size / wqe_size;
249 			if (wqe_size > gen->max_rq_desc_sz) {
250 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
251 					    wqe_size,
252 					    gen->max_rq_desc_sz);
253 				return -EINVAL;
254 			}
255 			qp->rq.wqe_shift = ilog2(wqe_size);
256 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
257 			qp->rq.max_post = qp->rq.wqe_cnt;
258 		}
259 	}
260 
261 	return 0;
262 }
263 
264 static int sq_overhead(enum ib_qp_type qp_type)
265 {
266 	int size = 0;
267 
268 	switch (qp_type) {
269 	case IB_QPT_XRC_INI:
270 		size += sizeof(struct mlx5_wqe_xrc_seg);
271 		/* fall through */
272 	case IB_QPT_RC:
273 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
274 			sizeof(struct mlx5_wqe_atomic_seg) +
275 			sizeof(struct mlx5_wqe_raddr_seg);
276 		break;
277 
278 	case IB_QPT_XRC_TGT:
279 		return 0;
280 
281 	case IB_QPT_UC:
282 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
283 			sizeof(struct mlx5_wqe_raddr_seg) +
284 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
285 			sizeof(struct mlx5_mkey_seg);
286 		break;
287 
288 	case IB_QPT_UD:
289 	case IB_QPT_SMI:
290 	case IB_QPT_GSI:
291 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
292 			sizeof(struct mlx5_wqe_datagram_seg);
293 		break;
294 
295 	case MLX5_IB_QPT_REG_UMR:
296 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
297 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
298 			sizeof(struct mlx5_mkey_seg);
299 		break;
300 
301 	default:
302 		return -EINVAL;
303 	}
304 
305 	return size;
306 }
307 
308 static int calc_send_wqe(struct ib_qp_init_attr *attr)
309 {
310 	int inl_size = 0;
311 	int size;
312 
313 	size = sq_overhead(attr->qp_type);
314 	if (size < 0)
315 		return size;
316 
317 	if (attr->cap.max_inline_data) {
318 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
319 			attr->cap.max_inline_data;
320 	}
321 
322 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
323 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN &&
324 	    ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
325 			return MLX5_SIG_WQE_SIZE;
326 	else
327 		return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
328 }
329 
330 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
331 			struct mlx5_ib_qp *qp)
332 {
333 	struct mlx5_general_caps *gen;
334 	int wqe_size;
335 	int wq_size;
336 
337 	gen = &dev->mdev->caps.gen;
338 	if (!attr->cap.max_send_wr)
339 		return 0;
340 
341 	wqe_size = calc_send_wqe(attr);
342 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
343 	if (wqe_size < 0)
344 		return wqe_size;
345 
346 	if (wqe_size > gen->max_sq_desc_sz) {
347 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
348 			    wqe_size, gen->max_sq_desc_sz);
349 		return -EINVAL;
350 	}
351 
352 	qp->max_inline_data = wqe_size - sq_overhead(attr->qp_type) -
353 		sizeof(struct mlx5_wqe_inline_seg);
354 	attr->cap.max_inline_data = qp->max_inline_data;
355 
356 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN)
357 		qp->signature_en = true;
358 
359 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
360 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
361 	if (qp->sq.wqe_cnt > gen->max_wqes) {
362 		mlx5_ib_dbg(dev, "wqe count(%d) exceeds limits(%d)\n",
363 			    qp->sq.wqe_cnt, gen->max_wqes);
364 		return -ENOMEM;
365 	}
366 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
367 	qp->sq.max_gs = attr->cap.max_send_sge;
368 	qp->sq.max_post = wq_size / wqe_size;
369 	attr->cap.max_send_wr = qp->sq.max_post;
370 
371 	return wq_size;
372 }
373 
374 static int set_user_buf_size(struct mlx5_ib_dev *dev,
375 			    struct mlx5_ib_qp *qp,
376 			    struct mlx5_ib_create_qp *ucmd)
377 {
378 	struct mlx5_general_caps *gen;
379 	int desc_sz = 1 << qp->sq.wqe_shift;
380 
381 	gen = &dev->mdev->caps.gen;
382 	if (desc_sz > gen->max_sq_desc_sz) {
383 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
384 			     desc_sz, gen->max_sq_desc_sz);
385 		return -EINVAL;
386 	}
387 
388 	if (ucmd->sq_wqe_count && ((1 << ilog2(ucmd->sq_wqe_count)) != ucmd->sq_wqe_count)) {
389 		mlx5_ib_warn(dev, "sq_wqe_count %d, sq_wqe_count %d\n",
390 			     ucmd->sq_wqe_count, ucmd->sq_wqe_count);
391 		return -EINVAL;
392 	}
393 
394 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
395 
396 	if (qp->sq.wqe_cnt > gen->max_wqes) {
397 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
398 			     qp->sq.wqe_cnt, gen->max_wqes);
399 		return -EINVAL;
400 	}
401 
402 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
403 		(qp->sq.wqe_cnt << 6);
404 
405 	return 0;
406 }
407 
408 static int qp_has_rq(struct ib_qp_init_attr *attr)
409 {
410 	if (attr->qp_type == IB_QPT_XRC_INI ||
411 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
412 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
413 	    !attr->cap.max_recv_wr)
414 		return 0;
415 
416 	return 1;
417 }
418 
419 static int first_med_uuar(void)
420 {
421 	return 1;
422 }
423 
424 static int next_uuar(int n)
425 {
426 	n++;
427 
428 	while (((n % 4) & 2))
429 		n++;
430 
431 	return n;
432 }
433 
434 static int num_med_uuar(struct mlx5_uuar_info *uuari)
435 {
436 	int n;
437 
438 	n = uuari->num_uars * MLX5_NON_FP_BF_REGS_PER_PAGE -
439 		uuari->num_low_latency_uuars - 1;
440 
441 	return n >= 0 ? n : 0;
442 }
443 
444 static int max_uuari(struct mlx5_uuar_info *uuari)
445 {
446 	return uuari->num_uars * 4;
447 }
448 
449 static int first_hi_uuar(struct mlx5_uuar_info *uuari)
450 {
451 	int med;
452 	int i;
453 	int t;
454 
455 	med = num_med_uuar(uuari);
456 	for (t = 0, i = first_med_uuar();; i = next_uuar(i)) {
457 		t++;
458 		if (t == med)
459 			return next_uuar(i);
460 	}
461 
462 	return 0;
463 }
464 
465 static int alloc_high_class_uuar(struct mlx5_uuar_info *uuari)
466 {
467 	int i;
468 
469 	for (i = first_hi_uuar(uuari); i < max_uuari(uuari); i = next_uuar(i)) {
470 		if (!test_bit(i, uuari->bitmap)) {
471 			set_bit(i, uuari->bitmap);
472 			uuari->count[i]++;
473 			return i;
474 		}
475 	}
476 
477 	return -ENOMEM;
478 }
479 
480 static int alloc_med_class_uuar(struct mlx5_uuar_info *uuari)
481 {
482 	int minidx = first_med_uuar();
483 	int i;
484 
485 	for (i = first_med_uuar(); i < first_hi_uuar(uuari); i = next_uuar(i)) {
486 		if (uuari->count[i] < uuari->count[minidx])
487 			minidx = i;
488 	}
489 
490 	uuari->count[minidx]++;
491 	return minidx;
492 }
493 
494 static int alloc_uuar(struct mlx5_uuar_info *uuari,
495 		      enum mlx5_ib_latency_class lat)
496 {
497 	int uuarn = -EINVAL;
498 
499 	mutex_lock(&uuari->lock);
500 	switch (lat) {
501 	case MLX5_IB_LATENCY_CLASS_LOW:
502 		uuarn = 0;
503 		uuari->count[uuarn]++;
504 		break;
505 
506 	case MLX5_IB_LATENCY_CLASS_MEDIUM:
507 		if (uuari->ver < 2)
508 			uuarn = -ENOMEM;
509 		else
510 			uuarn = alloc_med_class_uuar(uuari);
511 		break;
512 
513 	case MLX5_IB_LATENCY_CLASS_HIGH:
514 		if (uuari->ver < 2)
515 			uuarn = -ENOMEM;
516 		else
517 			uuarn = alloc_high_class_uuar(uuari);
518 		break;
519 
520 	case MLX5_IB_LATENCY_CLASS_FAST_PATH:
521 		uuarn = 2;
522 		break;
523 	}
524 	mutex_unlock(&uuari->lock);
525 
526 	return uuarn;
527 }
528 
529 static void free_med_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
530 {
531 	clear_bit(uuarn, uuari->bitmap);
532 	--uuari->count[uuarn];
533 }
534 
535 static void free_high_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
536 {
537 	clear_bit(uuarn, uuari->bitmap);
538 	--uuari->count[uuarn];
539 }
540 
541 static void free_uuar(struct mlx5_uuar_info *uuari, int uuarn)
542 {
543 	int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
544 	int high_uuar = nuuars - uuari->num_low_latency_uuars;
545 
546 	mutex_lock(&uuari->lock);
547 	if (uuarn == 0) {
548 		--uuari->count[uuarn];
549 		goto out;
550 	}
551 
552 	if (uuarn < high_uuar) {
553 		free_med_class_uuar(uuari, uuarn);
554 		goto out;
555 	}
556 
557 	free_high_class_uuar(uuari, uuarn);
558 
559 out:
560 	mutex_unlock(&uuari->lock);
561 }
562 
563 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
564 {
565 	switch (state) {
566 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
567 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
568 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
569 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
570 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
571 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
572 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
573 	default:		return -1;
574 	}
575 }
576 
577 static int to_mlx5_st(enum ib_qp_type type)
578 {
579 	switch (type) {
580 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
581 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
582 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
583 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
584 	case IB_QPT_XRC_INI:
585 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
586 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
587 	case IB_QPT_GSI:		return MLX5_QP_ST_QP1;
588 	case IB_QPT_RAW_IPV6:		return MLX5_QP_ST_RAW_IPV6;
589 	case IB_QPT_RAW_ETHERTYPE:	return MLX5_QP_ST_RAW_ETHERTYPE;
590 	case IB_QPT_RAW_PACKET:
591 	case IB_QPT_MAX:
592 	default:		return -EINVAL;
593 	}
594 }
595 
596 static int uuarn_to_uar_index(struct mlx5_uuar_info *uuari, int uuarn)
597 {
598 	return uuari->uars[uuarn / MLX5_BF_REGS_PER_PAGE].index;
599 }
600 
601 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
602 			  struct mlx5_ib_qp *qp, struct ib_udata *udata,
603 			  struct mlx5_create_qp_mbox_in **in,
604 			  struct mlx5_ib_create_qp_resp *resp, int *inlen)
605 {
606 	struct mlx5_ib_ucontext *context;
607 	struct mlx5_ib_create_qp ucmd;
608 	int page_shift = 0;
609 	int uar_index;
610 	int npages;
611 	u32 offset = 0;
612 	int uuarn;
613 	int ncont = 0;
614 	int err;
615 
616 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
617 	if (err) {
618 		mlx5_ib_dbg(dev, "copy failed\n");
619 		return err;
620 	}
621 
622 	context = to_mucontext(pd->uobject->context);
623 	/*
624 	 * TBD: should come from the verbs when we have the API
625 	 */
626 	uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_HIGH);
627 	if (uuarn < 0) {
628 		mlx5_ib_dbg(dev, "failed to allocate low latency UUAR\n");
629 		mlx5_ib_dbg(dev, "reverting to medium latency\n");
630 		uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_MEDIUM);
631 		if (uuarn < 0) {
632 			mlx5_ib_dbg(dev, "failed to allocate medium latency UUAR\n");
633 			mlx5_ib_dbg(dev, "reverting to high latency\n");
634 			uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_LOW);
635 			if (uuarn < 0) {
636 				mlx5_ib_warn(dev, "uuar allocation failed\n");
637 				return uuarn;
638 			}
639 		}
640 	}
641 
642 	uar_index = uuarn_to_uar_index(&context->uuari, uuarn);
643 	mlx5_ib_dbg(dev, "uuarn 0x%x, uar_index 0x%x\n", uuarn, uar_index);
644 
645 	qp->rq.offset = 0;
646 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
647 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
648 
649 	err = set_user_buf_size(dev, qp, &ucmd);
650 	if (err)
651 		goto err_uuar;
652 
653 	if (ucmd.buf_addr && qp->buf_size) {
654 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
655 				       qp->buf_size, 0, 0);
656 		if (IS_ERR(qp->umem)) {
657 			mlx5_ib_dbg(dev, "umem_get failed\n");
658 			err = PTR_ERR(qp->umem);
659 			goto err_uuar;
660 		}
661 	} else {
662 		qp->umem = NULL;
663 	}
664 
665 	if (qp->umem) {
666 		mlx5_ib_cont_pages(qp->umem, ucmd.buf_addr, &npages, &page_shift,
667 				   &ncont, NULL);
668 		err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift, &offset);
669 		if (err) {
670 			mlx5_ib_warn(dev, "bad offset\n");
671 			goto err_umem;
672 		}
673 		mlx5_ib_dbg(dev, "addr 0x%llx, size %d, npages %d, page_shift %d, ncont %d, offset %d\n",
674 			    ucmd.buf_addr, qp->buf_size, npages, page_shift, ncont, offset);
675 	}
676 
677 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * ncont;
678 	*in = mlx5_vzalloc(*inlen);
679 	if (!*in) {
680 		err = -ENOMEM;
681 		goto err_umem;
682 	}
683 	if (qp->umem)
684 		mlx5_ib_populate_pas(dev, qp->umem, page_shift, (*in)->pas, 0);
685 	(*in)->ctx.log_pg_sz_remote_qpn =
686 		cpu_to_be32((page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
687 	(*in)->ctx.params2 = cpu_to_be32(offset << 6);
688 
689 	(*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
690 	resp->uuar_index = uuarn;
691 	qp->uuarn = uuarn;
692 
693 	err = mlx5_ib_db_map_user(context, ucmd.db_addr, &qp->db);
694 	if (err) {
695 		mlx5_ib_dbg(dev, "map failed\n");
696 		goto err_free;
697 	}
698 
699 	err = ib_copy_to_udata(udata, resp, sizeof(*resp));
700 	if (err) {
701 		mlx5_ib_dbg(dev, "copy failed\n");
702 		goto err_unmap;
703 	}
704 	qp->create_type = MLX5_QP_USER;
705 
706 	return 0;
707 
708 err_unmap:
709 	mlx5_ib_db_unmap_user(context, &qp->db);
710 
711 err_free:
712 	kvfree(*in);
713 
714 err_umem:
715 	if (qp->umem)
716 		ib_umem_release(qp->umem);
717 
718 err_uuar:
719 	free_uuar(&context->uuari, uuarn);
720 	return err;
721 }
722 
723 static void destroy_qp_user(struct ib_pd *pd, struct mlx5_ib_qp *qp)
724 {
725 	struct mlx5_ib_ucontext *context;
726 
727 	context = to_mucontext(pd->uobject->context);
728 	mlx5_ib_db_unmap_user(context, &qp->db);
729 	if (qp->umem)
730 		ib_umem_release(qp->umem);
731 	free_uuar(&context->uuari, qp->uuarn);
732 }
733 
734 static int create_kernel_qp(struct mlx5_ib_dev *dev,
735 			    struct ib_qp_init_attr *init_attr,
736 			    struct mlx5_ib_qp *qp,
737 			    struct mlx5_create_qp_mbox_in **in, int *inlen)
738 {
739 	enum mlx5_ib_latency_class lc = MLX5_IB_LATENCY_CLASS_LOW;
740 	struct mlx5_uuar_info *uuari;
741 	int uar_index;
742 	int uuarn;
743 	int err;
744 
745 	uuari = &dev->mdev->priv.uuari;
746 	if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN | IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK))
747 		return -EINVAL;
748 
749 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
750 		lc = MLX5_IB_LATENCY_CLASS_FAST_PATH;
751 
752 	uuarn = alloc_uuar(uuari, lc);
753 	if (uuarn < 0) {
754 		mlx5_ib_dbg(dev, "\n");
755 		return -ENOMEM;
756 	}
757 
758 	qp->bf = &uuari->bfs[uuarn];
759 	uar_index = qp->bf->uar->index;
760 
761 	err = calc_sq_size(dev, init_attr, qp);
762 	if (err < 0) {
763 		mlx5_ib_dbg(dev, "err %d\n", err);
764 		goto err_uuar;
765 	}
766 
767 	qp->rq.offset = 0;
768 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
769 	qp->buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
770 
771 	err = mlx5_buf_alloc(dev->mdev, qp->buf_size, PAGE_SIZE * 2, &qp->buf);
772 	if (err) {
773 		mlx5_ib_dbg(dev, "err %d\n", err);
774 		goto err_uuar;
775 	}
776 
777 	qp->sq.qend = mlx5_get_send_wqe(qp, qp->sq.wqe_cnt);
778 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * qp->buf.npages;
779 	*in = mlx5_vzalloc(*inlen);
780 	if (!*in) {
781 		err = -ENOMEM;
782 		goto err_buf;
783 	}
784 	(*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
785 	(*in)->ctx.log_pg_sz_remote_qpn =
786 		cpu_to_be32((qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
787 	/* Set "fast registration enabled" for all kernel QPs */
788 	(*in)->ctx.params1 |= cpu_to_be32(1 << 11);
789 	(*in)->ctx.sq_crq_size |= cpu_to_be16(1 << 4);
790 
791 	mlx5_fill_page_array(&qp->buf, (*in)->pas);
792 
793 	err = mlx5_db_alloc(dev->mdev, &qp->db);
794 	if (err) {
795 		mlx5_ib_dbg(dev, "err %d\n", err);
796 		goto err_free;
797 	}
798 
799 	qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
800 	qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
801 	qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
802 	qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
803 	qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
804 
805 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
806 	    !qp->sq.w_list || !qp->sq.wqe_head) {
807 		err = -ENOMEM;
808 		goto err_wrid;
809 	}
810 	qp->create_type = MLX5_QP_KERNEL;
811 
812 	return 0;
813 
814 err_wrid:
815 	mlx5_db_free(dev->mdev, &qp->db);
816 	kfree(qp->sq.wqe_head);
817 	kfree(qp->sq.w_list);
818 	kfree(qp->sq.wrid);
819 	kfree(qp->sq.wr_data);
820 	kfree(qp->rq.wrid);
821 
822 err_free:
823 	kvfree(*in);
824 
825 err_buf:
826 	mlx5_buf_free(dev->mdev, &qp->buf);
827 
828 err_uuar:
829 	free_uuar(&dev->mdev->priv.uuari, uuarn);
830 	return err;
831 }
832 
833 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
834 {
835 	mlx5_db_free(dev->mdev, &qp->db);
836 	kfree(qp->sq.wqe_head);
837 	kfree(qp->sq.w_list);
838 	kfree(qp->sq.wrid);
839 	kfree(qp->sq.wr_data);
840 	kfree(qp->rq.wrid);
841 	mlx5_buf_free(dev->mdev, &qp->buf);
842 	free_uuar(&dev->mdev->priv.uuari, qp->bf->uuarn);
843 }
844 
845 static __be32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
846 {
847 	if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
848 	    (attr->qp_type == IB_QPT_XRC_INI))
849 		return cpu_to_be32(MLX5_SRQ_RQ);
850 	else if (!qp->has_rq)
851 		return cpu_to_be32(MLX5_ZERO_LEN_RQ);
852 	else
853 		return cpu_to_be32(MLX5_NON_ZERO_RQ);
854 }
855 
856 static int is_connected(enum ib_qp_type qp_type)
857 {
858 	if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC)
859 		return 1;
860 
861 	return 0;
862 }
863 
864 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
865 			    struct ib_qp_init_attr *init_attr,
866 			    struct ib_udata *udata, struct mlx5_ib_qp *qp)
867 {
868 	struct mlx5_ib_resources *devr = &dev->devr;
869 	struct mlx5_ib_create_qp_resp resp;
870 	struct mlx5_create_qp_mbox_in *in;
871 	struct mlx5_general_caps *gen;
872 	struct mlx5_ib_create_qp ucmd;
873 	int inlen = sizeof(*in);
874 	int err;
875 
876 	mlx5_ib_odp_create_qp(qp);
877 
878 	gen = &dev->mdev->caps.gen;
879 	mutex_init(&qp->mutex);
880 	spin_lock_init(&qp->sq.lock);
881 	spin_lock_init(&qp->rq.lock);
882 
883 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
884 		if (!(gen->flags & MLX5_DEV_CAP_FLAG_BLOCK_MCAST)) {
885 			mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
886 			return -EINVAL;
887 		} else {
888 			qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
889 		}
890 	}
891 
892 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
893 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
894 
895 	if (pd && pd->uobject) {
896 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
897 			mlx5_ib_dbg(dev, "copy failed\n");
898 			return -EFAULT;
899 		}
900 
901 		qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
902 		qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
903 	} else {
904 		qp->wq_sig = !!wq_signature;
905 	}
906 
907 	qp->has_rq = qp_has_rq(init_attr);
908 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
909 			  qp, (pd && pd->uobject) ? &ucmd : NULL);
910 	if (err) {
911 		mlx5_ib_dbg(dev, "err %d\n", err);
912 		return err;
913 	}
914 
915 	if (pd) {
916 		if (pd->uobject) {
917 			mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
918 			if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
919 			    ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
920 				mlx5_ib_dbg(dev, "invalid rq params\n");
921 				return -EINVAL;
922 			}
923 			if (ucmd.sq_wqe_count > gen->max_wqes) {
924 				mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
925 					    ucmd.sq_wqe_count, gen->max_wqes);
926 				return -EINVAL;
927 			}
928 			err = create_user_qp(dev, pd, qp, udata, &in, &resp, &inlen);
929 			if (err)
930 				mlx5_ib_dbg(dev, "err %d\n", err);
931 		} else {
932 			err = create_kernel_qp(dev, init_attr, qp, &in, &inlen);
933 			if (err)
934 				mlx5_ib_dbg(dev, "err %d\n", err);
935 			else
936 				qp->pa_lkey = to_mpd(pd)->pa_lkey;
937 		}
938 
939 		if (err)
940 			return err;
941 	} else {
942 		in = mlx5_vzalloc(sizeof(*in));
943 		if (!in)
944 			return -ENOMEM;
945 
946 		qp->create_type = MLX5_QP_EMPTY;
947 	}
948 
949 	if (is_sqp(init_attr->qp_type))
950 		qp->port = init_attr->port_num;
951 
952 	in->ctx.flags = cpu_to_be32(to_mlx5_st(init_attr->qp_type) << 16 |
953 				    MLX5_QP_PM_MIGRATED << 11);
954 
955 	if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
956 		in->ctx.flags_pd = cpu_to_be32(to_mpd(pd ? pd : devr->p0)->pdn);
957 	else
958 		in->ctx.flags_pd = cpu_to_be32(MLX5_QP_LAT_SENSITIVE);
959 
960 	if (qp->wq_sig)
961 		in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_ENABLE_SIG);
962 
963 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
964 		in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_BLOCK_MCAST);
965 
966 	if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
967 		int rcqe_sz;
968 		int scqe_sz;
969 
970 		rcqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->recv_cq);
971 		scqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->send_cq);
972 
973 		if (rcqe_sz == 128)
974 			in->ctx.cs_res = MLX5_RES_SCAT_DATA64_CQE;
975 		else
976 			in->ctx.cs_res = MLX5_RES_SCAT_DATA32_CQE;
977 
978 		if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) {
979 			if (scqe_sz == 128)
980 				in->ctx.cs_req = MLX5_REQ_SCAT_DATA64_CQE;
981 			else
982 				in->ctx.cs_req = MLX5_REQ_SCAT_DATA32_CQE;
983 		}
984 	}
985 
986 	if (qp->rq.wqe_cnt) {
987 		in->ctx.rq_size_stride = (qp->rq.wqe_shift - 4);
988 		in->ctx.rq_size_stride |= ilog2(qp->rq.wqe_cnt) << 3;
989 	}
990 
991 	in->ctx.rq_type_srqn = get_rx_type(qp, init_attr);
992 
993 	if (qp->sq.wqe_cnt)
994 		in->ctx.sq_crq_size |= cpu_to_be16(ilog2(qp->sq.wqe_cnt) << 11);
995 	else
996 		in->ctx.sq_crq_size |= cpu_to_be16(0x8000);
997 
998 	/* Set default resources */
999 	switch (init_attr->qp_type) {
1000 	case IB_QPT_XRC_TGT:
1001 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
1002 		in->ctx.cqn_send = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
1003 		in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
1004 		in->ctx.xrcd = cpu_to_be32(to_mxrcd(init_attr->xrcd)->xrcdn);
1005 		break;
1006 	case IB_QPT_XRC_INI:
1007 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
1008 		in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
1009 		in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
1010 		break;
1011 	default:
1012 		if (init_attr->srq) {
1013 			in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x0)->xrcdn);
1014 			in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(init_attr->srq)->msrq.srqn);
1015 		} else {
1016 			in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
1017 			in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
1018 		}
1019 	}
1020 
1021 	if (init_attr->send_cq)
1022 		in->ctx.cqn_send = cpu_to_be32(to_mcq(init_attr->send_cq)->mcq.cqn);
1023 
1024 	if (init_attr->recv_cq)
1025 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(init_attr->recv_cq)->mcq.cqn);
1026 
1027 	in->ctx.db_rec_addr = cpu_to_be64(qp->db.dma);
1028 
1029 	err = mlx5_core_create_qp(dev->mdev, &qp->mqp, in, inlen);
1030 	if (err) {
1031 		mlx5_ib_dbg(dev, "create qp failed\n");
1032 		goto err_create;
1033 	}
1034 
1035 	kvfree(in);
1036 	/* Hardware wants QPN written in big-endian order (after
1037 	 * shifting) for send doorbell.  Precompute this value to save
1038 	 * a little bit when posting sends.
1039 	 */
1040 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1041 
1042 	qp->mqp.event = mlx5_ib_qp_event;
1043 
1044 	return 0;
1045 
1046 err_create:
1047 	if (qp->create_type == MLX5_QP_USER)
1048 		destroy_qp_user(pd, qp);
1049 	else if (qp->create_type == MLX5_QP_KERNEL)
1050 		destroy_qp_kernel(dev, qp);
1051 
1052 	kvfree(in);
1053 	return err;
1054 }
1055 
1056 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1057 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1058 {
1059 	if (send_cq) {
1060 		if (recv_cq) {
1061 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1062 				spin_lock_irq(&send_cq->lock);
1063 				spin_lock_nested(&recv_cq->lock,
1064 						 SINGLE_DEPTH_NESTING);
1065 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1066 				spin_lock_irq(&send_cq->lock);
1067 				__acquire(&recv_cq->lock);
1068 			} else {
1069 				spin_lock_irq(&recv_cq->lock);
1070 				spin_lock_nested(&send_cq->lock,
1071 						 SINGLE_DEPTH_NESTING);
1072 			}
1073 		} else {
1074 			spin_lock_irq(&send_cq->lock);
1075 			__acquire(&recv_cq->lock);
1076 		}
1077 	} else if (recv_cq) {
1078 		spin_lock_irq(&recv_cq->lock);
1079 		__acquire(&send_cq->lock);
1080 	} else {
1081 		__acquire(&send_cq->lock);
1082 		__acquire(&recv_cq->lock);
1083 	}
1084 }
1085 
1086 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1087 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1088 {
1089 	if (send_cq) {
1090 		if (recv_cq) {
1091 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1092 				spin_unlock(&recv_cq->lock);
1093 				spin_unlock_irq(&send_cq->lock);
1094 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1095 				__release(&recv_cq->lock);
1096 				spin_unlock_irq(&send_cq->lock);
1097 			} else {
1098 				spin_unlock(&send_cq->lock);
1099 				spin_unlock_irq(&recv_cq->lock);
1100 			}
1101 		} else {
1102 			__release(&recv_cq->lock);
1103 			spin_unlock_irq(&send_cq->lock);
1104 		}
1105 	} else if (recv_cq) {
1106 		__release(&send_cq->lock);
1107 		spin_unlock_irq(&recv_cq->lock);
1108 	} else {
1109 		__release(&recv_cq->lock);
1110 		__release(&send_cq->lock);
1111 	}
1112 }
1113 
1114 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
1115 {
1116 	return to_mpd(qp->ibqp.pd);
1117 }
1118 
1119 static void get_cqs(struct mlx5_ib_qp *qp,
1120 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
1121 {
1122 	switch (qp->ibqp.qp_type) {
1123 	case IB_QPT_XRC_TGT:
1124 		*send_cq = NULL;
1125 		*recv_cq = NULL;
1126 		break;
1127 	case MLX5_IB_QPT_REG_UMR:
1128 	case IB_QPT_XRC_INI:
1129 		*send_cq = to_mcq(qp->ibqp.send_cq);
1130 		*recv_cq = NULL;
1131 		break;
1132 
1133 	case IB_QPT_SMI:
1134 	case IB_QPT_GSI:
1135 	case IB_QPT_RC:
1136 	case IB_QPT_UC:
1137 	case IB_QPT_UD:
1138 	case IB_QPT_RAW_IPV6:
1139 	case IB_QPT_RAW_ETHERTYPE:
1140 		*send_cq = to_mcq(qp->ibqp.send_cq);
1141 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
1142 		break;
1143 
1144 	case IB_QPT_RAW_PACKET:
1145 	case IB_QPT_MAX:
1146 	default:
1147 		*send_cq = NULL;
1148 		*recv_cq = NULL;
1149 		break;
1150 	}
1151 }
1152 
1153 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1154 {
1155 	struct mlx5_ib_cq *send_cq, *recv_cq;
1156 	struct mlx5_modify_qp_mbox_in *in;
1157 	int err;
1158 
1159 	in = kzalloc(sizeof(*in), GFP_KERNEL);
1160 	if (!in)
1161 		return;
1162 
1163 	if (qp->state != IB_QPS_RESET) {
1164 		mlx5_ib_qp_disable_pagefaults(qp);
1165 		if (mlx5_core_qp_modify(dev->mdev, to_mlx5_state(qp->state),
1166 					MLX5_QP_STATE_RST, in, 0, &qp->mqp))
1167 			mlx5_ib_warn(dev, "mlx5_ib: modify QP %06x to RESET failed\n",
1168 				     qp->mqp.qpn);
1169 	}
1170 
1171 	get_cqs(qp, &send_cq, &recv_cq);
1172 
1173 	if (qp->create_type == MLX5_QP_KERNEL) {
1174 		mlx5_ib_lock_cqs(send_cq, recv_cq);
1175 		__mlx5_ib_cq_clean(recv_cq, qp->mqp.qpn,
1176 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1177 		if (send_cq != recv_cq)
1178 			__mlx5_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1179 		mlx5_ib_unlock_cqs(send_cq, recv_cq);
1180 	}
1181 
1182 	err = mlx5_core_destroy_qp(dev->mdev, &qp->mqp);
1183 	if (err)
1184 		mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n", qp->mqp.qpn);
1185 	kfree(in);
1186 
1187 
1188 	if (qp->create_type == MLX5_QP_KERNEL)
1189 		destroy_qp_kernel(dev, qp);
1190 	else if (qp->create_type == MLX5_QP_USER)
1191 		destroy_qp_user(&get_pd(qp)->ibpd, qp);
1192 }
1193 
1194 static const char *ib_qp_type_str(enum ib_qp_type type)
1195 {
1196 	switch (type) {
1197 	case IB_QPT_SMI:
1198 		return "IB_QPT_SMI";
1199 	case IB_QPT_GSI:
1200 		return "IB_QPT_GSI";
1201 	case IB_QPT_RC:
1202 		return "IB_QPT_RC";
1203 	case IB_QPT_UC:
1204 		return "IB_QPT_UC";
1205 	case IB_QPT_UD:
1206 		return "IB_QPT_UD";
1207 	case IB_QPT_RAW_IPV6:
1208 		return "IB_QPT_RAW_IPV6";
1209 	case IB_QPT_RAW_ETHERTYPE:
1210 		return "IB_QPT_RAW_ETHERTYPE";
1211 	case IB_QPT_XRC_INI:
1212 		return "IB_QPT_XRC_INI";
1213 	case IB_QPT_XRC_TGT:
1214 		return "IB_QPT_XRC_TGT";
1215 	case IB_QPT_RAW_PACKET:
1216 		return "IB_QPT_RAW_PACKET";
1217 	case MLX5_IB_QPT_REG_UMR:
1218 		return "MLX5_IB_QPT_REG_UMR";
1219 	case IB_QPT_MAX:
1220 	default:
1221 		return "Invalid QP type";
1222 	}
1223 }
1224 
1225 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
1226 				struct ib_qp_init_attr *init_attr,
1227 				struct ib_udata *udata)
1228 {
1229 	struct mlx5_general_caps *gen;
1230 	struct mlx5_ib_dev *dev;
1231 	struct mlx5_ib_qp *qp;
1232 	u16 xrcdn = 0;
1233 	int err;
1234 
1235 	if (pd) {
1236 		dev = to_mdev(pd->device);
1237 	} else {
1238 		/* being cautious here */
1239 		if (init_attr->qp_type != IB_QPT_XRC_TGT &&
1240 		    init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
1241 			pr_warn("%s: no PD for transport %s\n", __func__,
1242 				ib_qp_type_str(init_attr->qp_type));
1243 			return ERR_PTR(-EINVAL);
1244 		}
1245 		dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
1246 	}
1247 	gen = &dev->mdev->caps.gen;
1248 
1249 	switch (init_attr->qp_type) {
1250 	case IB_QPT_XRC_TGT:
1251 	case IB_QPT_XRC_INI:
1252 		if (!(gen->flags & MLX5_DEV_CAP_FLAG_XRC)) {
1253 			mlx5_ib_dbg(dev, "XRC not supported\n");
1254 			return ERR_PTR(-ENOSYS);
1255 		}
1256 		init_attr->recv_cq = NULL;
1257 		if (init_attr->qp_type == IB_QPT_XRC_TGT) {
1258 			xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1259 			init_attr->send_cq = NULL;
1260 		}
1261 
1262 		/* fall through */
1263 	case IB_QPT_RC:
1264 	case IB_QPT_UC:
1265 	case IB_QPT_UD:
1266 	case IB_QPT_SMI:
1267 	case IB_QPT_GSI:
1268 	case MLX5_IB_QPT_REG_UMR:
1269 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1270 		if (!qp)
1271 			return ERR_PTR(-ENOMEM);
1272 
1273 		err = create_qp_common(dev, pd, init_attr, udata, qp);
1274 		if (err) {
1275 			mlx5_ib_dbg(dev, "create_qp_common failed\n");
1276 			kfree(qp);
1277 			return ERR_PTR(err);
1278 		}
1279 
1280 		if (is_qp0(init_attr->qp_type))
1281 			qp->ibqp.qp_num = 0;
1282 		else if (is_qp1(init_attr->qp_type))
1283 			qp->ibqp.qp_num = 1;
1284 		else
1285 			qp->ibqp.qp_num = qp->mqp.qpn;
1286 
1287 		mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
1288 			    qp->ibqp.qp_num, qp->mqp.qpn, to_mcq(init_attr->recv_cq)->mcq.cqn,
1289 			    to_mcq(init_attr->send_cq)->mcq.cqn);
1290 
1291 		qp->xrcdn = xrcdn;
1292 
1293 		break;
1294 
1295 	case IB_QPT_RAW_IPV6:
1296 	case IB_QPT_RAW_ETHERTYPE:
1297 	case IB_QPT_RAW_PACKET:
1298 	case IB_QPT_MAX:
1299 	default:
1300 		mlx5_ib_dbg(dev, "unsupported qp type %d\n",
1301 			    init_attr->qp_type);
1302 		/* Don't support raw QPs */
1303 		return ERR_PTR(-EINVAL);
1304 	}
1305 
1306 	return &qp->ibqp;
1307 }
1308 
1309 int mlx5_ib_destroy_qp(struct ib_qp *qp)
1310 {
1311 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
1312 	struct mlx5_ib_qp *mqp = to_mqp(qp);
1313 
1314 	destroy_qp_common(dev, mqp);
1315 
1316 	kfree(mqp);
1317 
1318 	return 0;
1319 }
1320 
1321 static __be32 to_mlx5_access_flags(struct mlx5_ib_qp *qp, const struct ib_qp_attr *attr,
1322 				   int attr_mask)
1323 {
1324 	u32 hw_access_flags = 0;
1325 	u8 dest_rd_atomic;
1326 	u32 access_flags;
1327 
1328 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1329 		dest_rd_atomic = attr->max_dest_rd_atomic;
1330 	else
1331 		dest_rd_atomic = qp->resp_depth;
1332 
1333 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1334 		access_flags = attr->qp_access_flags;
1335 	else
1336 		access_flags = qp->atomic_rd_en;
1337 
1338 	if (!dest_rd_atomic)
1339 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1340 
1341 	if (access_flags & IB_ACCESS_REMOTE_READ)
1342 		hw_access_flags |= MLX5_QP_BIT_RRE;
1343 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1344 		hw_access_flags |= (MLX5_QP_BIT_RAE | MLX5_ATOMIC_MODE_CX);
1345 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1346 		hw_access_flags |= MLX5_QP_BIT_RWE;
1347 
1348 	return cpu_to_be32(hw_access_flags);
1349 }
1350 
1351 enum {
1352 	MLX5_PATH_FLAG_FL	= 1 << 0,
1353 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
1354 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
1355 };
1356 
1357 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
1358 {
1359 	struct mlx5_general_caps *gen;
1360 
1361 	gen = &dev->mdev->caps.gen;
1362 	if (rate == IB_RATE_PORT_CURRENT) {
1363 		return 0;
1364 	} else if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_300_GBPS) {
1365 		return -EINVAL;
1366 	} else {
1367 		while (rate != IB_RATE_2_5_GBPS &&
1368 		       !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
1369 			 gen->stat_rate_support))
1370 			--rate;
1371 	}
1372 
1373 	return rate + MLX5_STAT_RATE_OFFSET;
1374 }
1375 
1376 static int mlx5_set_path(struct mlx5_ib_dev *dev, const struct ib_ah_attr *ah,
1377 			 struct mlx5_qp_path *path, u8 port, int attr_mask,
1378 			 u32 path_flags, const struct ib_qp_attr *attr)
1379 {
1380 	struct mlx5_general_caps *gen;
1381 	int err;
1382 
1383 	gen = &dev->mdev->caps.gen;
1384 	path->fl = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
1385 	path->free_ar = (path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x80 : 0;
1386 
1387 	if (attr_mask & IB_QP_PKEY_INDEX)
1388 		path->pkey_index = attr->pkey_index;
1389 
1390 	path->grh_mlid	= ah->src_path_bits & 0x7f;
1391 	path->rlid	= cpu_to_be16(ah->dlid);
1392 
1393 	if (ah->ah_flags & IB_AH_GRH) {
1394 		if (ah->grh.sgid_index >= gen->port[port - 1].gid_table_len) {
1395 			pr_err(KERN_ERR "sgid_index (%u) too large. max is %d\n",
1396 			       ah->grh.sgid_index, gen->port[port - 1].gid_table_len);
1397 			return -EINVAL;
1398 		}
1399 		path->grh_mlid |= 1 << 7;
1400 		path->mgid_index = ah->grh.sgid_index;
1401 		path->hop_limit  = ah->grh.hop_limit;
1402 		path->tclass_flowlabel =
1403 			cpu_to_be32((ah->grh.traffic_class << 20) |
1404 				    (ah->grh.flow_label));
1405 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1406 	}
1407 
1408 	err = ib_rate_to_mlx5(dev, ah->static_rate);
1409 	if (err < 0)
1410 		return err;
1411 	path->static_rate = err;
1412 	path->port = port;
1413 
1414 	if (attr_mask & IB_QP_TIMEOUT)
1415 		path->ackto_lt = attr->timeout << 3;
1416 
1417 	path->sl = ah->sl & 0xf;
1418 
1419 	return 0;
1420 }
1421 
1422 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
1423 	[MLX5_QP_STATE_INIT] = {
1424 		[MLX5_QP_STATE_INIT] = {
1425 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
1426 					  MLX5_QP_OPTPAR_RAE		|
1427 					  MLX5_QP_OPTPAR_RWE		|
1428 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
1429 					  MLX5_QP_OPTPAR_PRI_PORT,
1430 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
1431 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
1432 					  MLX5_QP_OPTPAR_PRI_PORT,
1433 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
1434 					  MLX5_QP_OPTPAR_Q_KEY		|
1435 					  MLX5_QP_OPTPAR_PRI_PORT,
1436 		},
1437 		[MLX5_QP_STATE_RTR] = {
1438 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
1439 					  MLX5_QP_OPTPAR_RRE            |
1440 					  MLX5_QP_OPTPAR_RAE            |
1441 					  MLX5_QP_OPTPAR_RWE            |
1442 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1443 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
1444 					  MLX5_QP_OPTPAR_RWE            |
1445 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1446 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
1447 					  MLX5_QP_OPTPAR_Q_KEY,
1448 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
1449 					   MLX5_QP_OPTPAR_Q_KEY,
1450 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
1451 					  MLX5_QP_OPTPAR_RRE            |
1452 					  MLX5_QP_OPTPAR_RAE            |
1453 					  MLX5_QP_OPTPAR_RWE            |
1454 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1455 		},
1456 	},
1457 	[MLX5_QP_STATE_RTR] = {
1458 		[MLX5_QP_STATE_RTS] = {
1459 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
1460 					  MLX5_QP_OPTPAR_RRE		|
1461 					  MLX5_QP_OPTPAR_RAE		|
1462 					  MLX5_QP_OPTPAR_RWE		|
1463 					  MLX5_QP_OPTPAR_PM_STATE	|
1464 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
1465 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
1466 					  MLX5_QP_OPTPAR_RWE		|
1467 					  MLX5_QP_OPTPAR_PM_STATE,
1468 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
1469 		},
1470 	},
1471 	[MLX5_QP_STATE_RTS] = {
1472 		[MLX5_QP_STATE_RTS] = {
1473 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
1474 					  MLX5_QP_OPTPAR_RAE		|
1475 					  MLX5_QP_OPTPAR_RWE		|
1476 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
1477 					  MLX5_QP_OPTPAR_PM_STATE	|
1478 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
1479 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
1480 					  MLX5_QP_OPTPAR_PM_STATE	|
1481 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
1482 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
1483 					  MLX5_QP_OPTPAR_SRQN		|
1484 					  MLX5_QP_OPTPAR_CQN_RCV,
1485 		},
1486 	},
1487 	[MLX5_QP_STATE_SQER] = {
1488 		[MLX5_QP_STATE_RTS] = {
1489 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
1490 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
1491 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
1492 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
1493 					   MLX5_QP_OPTPAR_RWE		|
1494 					   MLX5_QP_OPTPAR_RAE		|
1495 					   MLX5_QP_OPTPAR_RRE,
1496 		},
1497 	},
1498 };
1499 
1500 static int ib_nr_to_mlx5_nr(int ib_mask)
1501 {
1502 	switch (ib_mask) {
1503 	case IB_QP_STATE:
1504 		return 0;
1505 	case IB_QP_CUR_STATE:
1506 		return 0;
1507 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
1508 		return 0;
1509 	case IB_QP_ACCESS_FLAGS:
1510 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
1511 			MLX5_QP_OPTPAR_RAE;
1512 	case IB_QP_PKEY_INDEX:
1513 		return MLX5_QP_OPTPAR_PKEY_INDEX;
1514 	case IB_QP_PORT:
1515 		return MLX5_QP_OPTPAR_PRI_PORT;
1516 	case IB_QP_QKEY:
1517 		return MLX5_QP_OPTPAR_Q_KEY;
1518 	case IB_QP_AV:
1519 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
1520 			MLX5_QP_OPTPAR_PRI_PORT;
1521 	case IB_QP_PATH_MTU:
1522 		return 0;
1523 	case IB_QP_TIMEOUT:
1524 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
1525 	case IB_QP_RETRY_CNT:
1526 		return MLX5_QP_OPTPAR_RETRY_COUNT;
1527 	case IB_QP_RNR_RETRY:
1528 		return MLX5_QP_OPTPAR_RNR_RETRY;
1529 	case IB_QP_RQ_PSN:
1530 		return 0;
1531 	case IB_QP_MAX_QP_RD_ATOMIC:
1532 		return MLX5_QP_OPTPAR_SRA_MAX;
1533 	case IB_QP_ALT_PATH:
1534 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
1535 	case IB_QP_MIN_RNR_TIMER:
1536 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
1537 	case IB_QP_SQ_PSN:
1538 		return 0;
1539 	case IB_QP_MAX_DEST_RD_ATOMIC:
1540 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
1541 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
1542 	case IB_QP_PATH_MIG_STATE:
1543 		return MLX5_QP_OPTPAR_PM_STATE;
1544 	case IB_QP_CAP:
1545 		return 0;
1546 	case IB_QP_DEST_QPN:
1547 		return 0;
1548 	}
1549 	return 0;
1550 }
1551 
1552 static int ib_mask_to_mlx5_opt(int ib_mask)
1553 {
1554 	int result = 0;
1555 	int i;
1556 
1557 	for (i = 0; i < 8 * sizeof(int); i++) {
1558 		if ((1 << i) & ib_mask)
1559 			result |= ib_nr_to_mlx5_nr(1 << i);
1560 	}
1561 
1562 	return result;
1563 }
1564 
1565 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
1566 			       const struct ib_qp_attr *attr, int attr_mask,
1567 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1568 {
1569 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1570 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
1571 	struct mlx5_ib_cq *send_cq, *recv_cq;
1572 	struct mlx5_qp_context *context;
1573 	struct mlx5_general_caps *gen;
1574 	struct mlx5_modify_qp_mbox_in *in;
1575 	struct mlx5_ib_pd *pd;
1576 	enum mlx5_qp_state mlx5_cur, mlx5_new;
1577 	enum mlx5_qp_optpar optpar;
1578 	int sqd_event;
1579 	int mlx5_st;
1580 	int err;
1581 
1582 	gen = &dev->mdev->caps.gen;
1583 	in = kzalloc(sizeof(*in), GFP_KERNEL);
1584 	if (!in)
1585 		return -ENOMEM;
1586 
1587 	context = &in->ctx;
1588 	err = to_mlx5_st(ibqp->qp_type);
1589 	if (err < 0)
1590 		goto out;
1591 
1592 	context->flags = cpu_to_be32(err << 16);
1593 
1594 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
1595 		context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
1596 	} else {
1597 		switch (attr->path_mig_state) {
1598 		case IB_MIG_MIGRATED:
1599 			context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
1600 			break;
1601 		case IB_MIG_REARM:
1602 			context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
1603 			break;
1604 		case IB_MIG_ARMED:
1605 			context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
1606 			break;
1607 		}
1608 	}
1609 
1610 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI) {
1611 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
1612 	} else if (ibqp->qp_type == IB_QPT_UD ||
1613 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
1614 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1615 	} else if (attr_mask & IB_QP_PATH_MTU) {
1616 		if (attr->path_mtu < IB_MTU_256 ||
1617 		    attr->path_mtu > IB_MTU_4096) {
1618 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
1619 			err = -EINVAL;
1620 			goto out;
1621 		}
1622 		context->mtu_msgmax = (attr->path_mtu << 5) | gen->log_max_msg;
1623 	}
1624 
1625 	if (attr_mask & IB_QP_DEST_QPN)
1626 		context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
1627 
1628 	if (attr_mask & IB_QP_PKEY_INDEX)
1629 		context->pri_path.pkey_index = attr->pkey_index;
1630 
1631 	/* todo implement counter_index functionality */
1632 
1633 	if (is_sqp(ibqp->qp_type))
1634 		context->pri_path.port = qp->port;
1635 
1636 	if (attr_mask & IB_QP_PORT)
1637 		context->pri_path.port = attr->port_num;
1638 
1639 	if (attr_mask & IB_QP_AV) {
1640 		err = mlx5_set_path(dev, &attr->ah_attr, &context->pri_path,
1641 				    attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
1642 				    attr_mask, 0, attr);
1643 		if (err)
1644 			goto out;
1645 	}
1646 
1647 	if (attr_mask & IB_QP_TIMEOUT)
1648 		context->pri_path.ackto_lt |= attr->timeout << 3;
1649 
1650 	if (attr_mask & IB_QP_ALT_PATH) {
1651 		err = mlx5_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
1652 				    attr->alt_port_num, attr_mask, 0, attr);
1653 		if (err)
1654 			goto out;
1655 	}
1656 
1657 	pd = get_pd(qp);
1658 	get_cqs(qp, &send_cq, &recv_cq);
1659 
1660 	context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
1661 	context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
1662 	context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
1663 	context->params1  = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
1664 
1665 	if (attr_mask & IB_QP_RNR_RETRY)
1666 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1667 
1668 	if (attr_mask & IB_QP_RETRY_CNT)
1669 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1670 
1671 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1672 		if (attr->max_rd_atomic)
1673 			context->params1 |=
1674 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1675 	}
1676 
1677 	if (attr_mask & IB_QP_SQ_PSN)
1678 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1679 
1680 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1681 		if (attr->max_dest_rd_atomic)
1682 			context->params2 |=
1683 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1684 	}
1685 
1686 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
1687 		context->params2 |= to_mlx5_access_flags(qp, attr, attr_mask);
1688 
1689 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
1690 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1691 
1692 	if (attr_mask & IB_QP_RQ_PSN)
1693 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1694 
1695 	if (attr_mask & IB_QP_QKEY)
1696 		context->qkey = cpu_to_be32(attr->qkey);
1697 
1698 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1699 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1700 
1701 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1702 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1703 		sqd_event = 1;
1704 	else
1705 		sqd_event = 0;
1706 
1707 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1708 		context->sq_crq_size |= cpu_to_be16(1 << 4);
1709 
1710 
1711 	mlx5_cur = to_mlx5_state(cur_state);
1712 	mlx5_new = to_mlx5_state(new_state);
1713 	mlx5_st = to_mlx5_st(ibqp->qp_type);
1714 	if (mlx5_st < 0)
1715 		goto out;
1716 
1717 	/* If moving to a reset or error state, we must disable page faults on
1718 	 * this QP and flush all current page faults. Otherwise a stale page
1719 	 * fault may attempt to work on this QP after it is reset and moved
1720 	 * again to RTS, and may cause the driver and the device to get out of
1721 	 * sync. */
1722 	if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1723 	    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1724 		mlx5_ib_qp_disable_pagefaults(qp);
1725 
1726 	optpar = ib_mask_to_mlx5_opt(attr_mask);
1727 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
1728 	in->optparam = cpu_to_be32(optpar);
1729 	err = mlx5_core_qp_modify(dev->mdev, to_mlx5_state(cur_state),
1730 				  to_mlx5_state(new_state), in, sqd_event,
1731 				  &qp->mqp);
1732 	if (err)
1733 		goto out;
1734 
1735 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1736 		mlx5_ib_qp_enable_pagefaults(qp);
1737 
1738 	qp->state = new_state;
1739 
1740 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1741 		qp->atomic_rd_en = attr->qp_access_flags;
1742 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1743 		qp->resp_depth = attr->max_dest_rd_atomic;
1744 	if (attr_mask & IB_QP_PORT)
1745 		qp->port = attr->port_num;
1746 	if (attr_mask & IB_QP_ALT_PATH)
1747 		qp->alt_port = attr->alt_port_num;
1748 
1749 	/*
1750 	 * If we moved a kernel QP to RESET, clean up all old CQ
1751 	 * entries and reinitialize the QP.
1752 	 */
1753 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1754 		mlx5_ib_cq_clean(recv_cq, qp->mqp.qpn,
1755 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
1756 		if (send_cq != recv_cq)
1757 			mlx5_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1758 
1759 		qp->rq.head = 0;
1760 		qp->rq.tail = 0;
1761 		qp->sq.head = 0;
1762 		qp->sq.tail = 0;
1763 		qp->sq.cur_post = 0;
1764 		qp->sq.last_poll = 0;
1765 		qp->db.db[MLX5_RCV_DBR] = 0;
1766 		qp->db.db[MLX5_SND_DBR] = 0;
1767 	}
1768 
1769 out:
1770 	kfree(in);
1771 	return err;
1772 }
1773 
1774 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1775 		      int attr_mask, struct ib_udata *udata)
1776 {
1777 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1778 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
1779 	enum ib_qp_state cur_state, new_state;
1780 	struct mlx5_general_caps *gen;
1781 	int err = -EINVAL;
1782 	int port;
1783 
1784 	gen = &dev->mdev->caps.gen;
1785 	mutex_lock(&qp->mutex);
1786 
1787 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1788 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1789 
1790 	if (ibqp->qp_type != MLX5_IB_QPT_REG_UMR &&
1791 	    !ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask,
1792 				IB_LINK_LAYER_UNSPECIFIED))
1793 		goto out;
1794 
1795 	if ((attr_mask & IB_QP_PORT) &&
1796 	    (attr->port_num == 0 || attr->port_num > gen->num_ports))
1797 		goto out;
1798 
1799 	if (attr_mask & IB_QP_PKEY_INDEX) {
1800 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1801 		if (attr->pkey_index >= gen->port[port - 1].pkey_table_len)
1802 			goto out;
1803 	}
1804 
1805 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1806 	    attr->max_rd_atomic > (1 << gen->log_max_ra_res_qp))
1807 		goto out;
1808 
1809 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1810 	    attr->max_dest_rd_atomic > (1 << gen->log_max_ra_req_qp))
1811 		goto out;
1812 
1813 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1814 		err = 0;
1815 		goto out;
1816 	}
1817 
1818 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1819 
1820 out:
1821 	mutex_unlock(&qp->mutex);
1822 	return err;
1823 }
1824 
1825 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1826 {
1827 	struct mlx5_ib_cq *cq;
1828 	unsigned cur;
1829 
1830 	cur = wq->head - wq->tail;
1831 	if (likely(cur + nreq < wq->max_post))
1832 		return 0;
1833 
1834 	cq = to_mcq(ib_cq);
1835 	spin_lock(&cq->lock);
1836 	cur = wq->head - wq->tail;
1837 	spin_unlock(&cq->lock);
1838 
1839 	return cur + nreq >= wq->max_post;
1840 }
1841 
1842 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
1843 					  u64 remote_addr, u32 rkey)
1844 {
1845 	rseg->raddr    = cpu_to_be64(remote_addr);
1846 	rseg->rkey     = cpu_to_be32(rkey);
1847 	rseg->reserved = 0;
1848 }
1849 
1850 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
1851 			     struct ib_send_wr *wr)
1852 {
1853 	memcpy(&dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof(struct mlx5_av));
1854 	dseg->av.dqp_dct = cpu_to_be32(wr->wr.ud.remote_qpn | MLX5_EXTENDED_UD_AV);
1855 	dseg->av.key.qkey.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1856 }
1857 
1858 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
1859 {
1860 	dseg->byte_count = cpu_to_be32(sg->length);
1861 	dseg->lkey       = cpu_to_be32(sg->lkey);
1862 	dseg->addr       = cpu_to_be64(sg->addr);
1863 }
1864 
1865 static __be16 get_klm_octo(int npages)
1866 {
1867 	return cpu_to_be16(ALIGN(npages, 8) / 2);
1868 }
1869 
1870 static __be64 frwr_mkey_mask(void)
1871 {
1872 	u64 result;
1873 
1874 	result = MLX5_MKEY_MASK_LEN		|
1875 		MLX5_MKEY_MASK_PAGE_SIZE	|
1876 		MLX5_MKEY_MASK_START_ADDR	|
1877 		MLX5_MKEY_MASK_EN_RINVAL	|
1878 		MLX5_MKEY_MASK_KEY		|
1879 		MLX5_MKEY_MASK_LR		|
1880 		MLX5_MKEY_MASK_LW		|
1881 		MLX5_MKEY_MASK_RR		|
1882 		MLX5_MKEY_MASK_RW		|
1883 		MLX5_MKEY_MASK_A		|
1884 		MLX5_MKEY_MASK_SMALL_FENCE	|
1885 		MLX5_MKEY_MASK_FREE;
1886 
1887 	return cpu_to_be64(result);
1888 }
1889 
1890 static __be64 sig_mkey_mask(void)
1891 {
1892 	u64 result;
1893 
1894 	result = MLX5_MKEY_MASK_LEN		|
1895 		MLX5_MKEY_MASK_PAGE_SIZE	|
1896 		MLX5_MKEY_MASK_START_ADDR	|
1897 		MLX5_MKEY_MASK_EN_SIGERR	|
1898 		MLX5_MKEY_MASK_EN_RINVAL	|
1899 		MLX5_MKEY_MASK_KEY		|
1900 		MLX5_MKEY_MASK_LR		|
1901 		MLX5_MKEY_MASK_LW		|
1902 		MLX5_MKEY_MASK_RR		|
1903 		MLX5_MKEY_MASK_RW		|
1904 		MLX5_MKEY_MASK_SMALL_FENCE	|
1905 		MLX5_MKEY_MASK_FREE		|
1906 		MLX5_MKEY_MASK_BSF_EN;
1907 
1908 	return cpu_to_be64(result);
1909 }
1910 
1911 static void set_frwr_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
1912 				 struct ib_send_wr *wr, int li)
1913 {
1914 	memset(umr, 0, sizeof(*umr));
1915 
1916 	if (li) {
1917 		umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
1918 		umr->flags = 1 << 7;
1919 		return;
1920 	}
1921 
1922 	umr->flags = (1 << 5); /* fail if not free */
1923 	umr->klm_octowords = get_klm_octo(wr->wr.fast_reg.page_list_len);
1924 	umr->mkey_mask = frwr_mkey_mask();
1925 }
1926 
1927 static __be64 get_umr_reg_mr_mask(void)
1928 {
1929 	u64 result;
1930 
1931 	result = MLX5_MKEY_MASK_LEN		|
1932 		 MLX5_MKEY_MASK_PAGE_SIZE	|
1933 		 MLX5_MKEY_MASK_START_ADDR	|
1934 		 MLX5_MKEY_MASK_PD		|
1935 		 MLX5_MKEY_MASK_LR		|
1936 		 MLX5_MKEY_MASK_LW		|
1937 		 MLX5_MKEY_MASK_KEY		|
1938 		 MLX5_MKEY_MASK_RR		|
1939 		 MLX5_MKEY_MASK_RW		|
1940 		 MLX5_MKEY_MASK_A		|
1941 		 MLX5_MKEY_MASK_FREE;
1942 
1943 	return cpu_to_be64(result);
1944 }
1945 
1946 static __be64 get_umr_unreg_mr_mask(void)
1947 {
1948 	u64 result;
1949 
1950 	result = MLX5_MKEY_MASK_FREE;
1951 
1952 	return cpu_to_be64(result);
1953 }
1954 
1955 static __be64 get_umr_update_mtt_mask(void)
1956 {
1957 	u64 result;
1958 
1959 	result = MLX5_MKEY_MASK_FREE;
1960 
1961 	return cpu_to_be64(result);
1962 }
1963 
1964 static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
1965 				struct ib_send_wr *wr)
1966 {
1967 	struct mlx5_umr_wr *umrwr = (struct mlx5_umr_wr *)&wr->wr.fast_reg;
1968 
1969 	memset(umr, 0, sizeof(*umr));
1970 
1971 	if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
1972 		umr->flags = MLX5_UMR_CHECK_FREE; /* fail if free */
1973 	else
1974 		umr->flags = MLX5_UMR_CHECK_NOT_FREE; /* fail if not free */
1975 
1976 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UNREG)) {
1977 		umr->klm_octowords = get_klm_octo(umrwr->npages);
1978 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT) {
1979 			umr->mkey_mask = get_umr_update_mtt_mask();
1980 			umr->bsf_octowords = get_klm_octo(umrwr->target.offset);
1981 			umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
1982 		} else {
1983 			umr->mkey_mask = get_umr_reg_mr_mask();
1984 		}
1985 	} else {
1986 		umr->mkey_mask = get_umr_unreg_mr_mask();
1987 	}
1988 
1989 	if (!wr->num_sge)
1990 		umr->flags |= MLX5_UMR_INLINE;
1991 }
1992 
1993 static u8 get_umr_flags(int acc)
1994 {
1995 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
1996 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
1997 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
1998 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
1999 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
2000 }
2001 
2002 static void set_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr,
2003 			     int li, int *writ)
2004 {
2005 	memset(seg, 0, sizeof(*seg));
2006 	if (li) {
2007 		seg->status = MLX5_MKEY_STATUS_FREE;
2008 		return;
2009 	}
2010 
2011 	seg->flags = get_umr_flags(wr->wr.fast_reg.access_flags) |
2012 		     MLX5_ACCESS_MODE_MTT;
2013 	*writ = seg->flags & (MLX5_PERM_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE);
2014 	seg->qpn_mkey7_0 = cpu_to_be32((wr->wr.fast_reg.rkey & 0xff) | 0xffffff00);
2015 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
2016 	seg->start_addr = cpu_to_be64(wr->wr.fast_reg.iova_start);
2017 	seg->len = cpu_to_be64(wr->wr.fast_reg.length);
2018 	seg->xlt_oct_size = cpu_to_be32((wr->wr.fast_reg.page_list_len + 1) / 2);
2019 	seg->log2_page_size = wr->wr.fast_reg.page_shift;
2020 }
2021 
2022 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr)
2023 {
2024 	struct mlx5_umr_wr *umrwr = (struct mlx5_umr_wr *)&wr->wr.fast_reg;
2025 
2026 	memset(seg, 0, sizeof(*seg));
2027 	if (wr->send_flags & MLX5_IB_SEND_UMR_UNREG) {
2028 		seg->status = MLX5_MKEY_STATUS_FREE;
2029 		return;
2030 	}
2031 
2032 	seg->flags = convert_access(umrwr->access_flags);
2033 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT)) {
2034 		seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
2035 		seg->start_addr = cpu_to_be64(umrwr->target.virt_addr);
2036 	}
2037 	seg->len = cpu_to_be64(umrwr->length);
2038 	seg->log2_page_size = umrwr->page_shift;
2039 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
2040 				       mlx5_mkey_variant(umrwr->mkey));
2041 }
2042 
2043 static void set_frwr_pages(struct mlx5_wqe_data_seg *dseg,
2044 			   struct ib_send_wr *wr,
2045 			   struct mlx5_core_dev *mdev,
2046 			   struct mlx5_ib_pd *pd,
2047 			   int writ)
2048 {
2049 	struct mlx5_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
2050 	u64 *page_list = wr->wr.fast_reg.page_list->page_list;
2051 	u64 perm = MLX5_EN_RD | (writ ? MLX5_EN_WR : 0);
2052 	int i;
2053 
2054 	for (i = 0; i < wr->wr.fast_reg.page_list_len; i++)
2055 		mfrpl->mapped_page_list[i] = cpu_to_be64(page_list[i] | perm);
2056 	dseg->addr = cpu_to_be64(mfrpl->map);
2057 	dseg->byte_count = cpu_to_be32(ALIGN(sizeof(u64) * wr->wr.fast_reg.page_list_len, 64));
2058 	dseg->lkey = cpu_to_be32(pd->pa_lkey);
2059 }
2060 
2061 static __be32 send_ieth(struct ib_send_wr *wr)
2062 {
2063 	switch (wr->opcode) {
2064 	case IB_WR_SEND_WITH_IMM:
2065 	case IB_WR_RDMA_WRITE_WITH_IMM:
2066 		return wr->ex.imm_data;
2067 
2068 	case IB_WR_SEND_WITH_INV:
2069 		return cpu_to_be32(wr->ex.invalidate_rkey);
2070 
2071 	default:
2072 		return 0;
2073 	}
2074 }
2075 
2076 static u8 calc_sig(void *wqe, int size)
2077 {
2078 	u8 *p = wqe;
2079 	u8 res = 0;
2080 	int i;
2081 
2082 	for (i = 0; i < size; i++)
2083 		res ^= p[i];
2084 
2085 	return ~res;
2086 }
2087 
2088 static u8 wq_sig(void *wqe)
2089 {
2090 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
2091 }
2092 
2093 static int set_data_inl_seg(struct mlx5_ib_qp *qp, struct ib_send_wr *wr,
2094 			    void *wqe, int *sz)
2095 {
2096 	struct mlx5_wqe_inline_seg *seg;
2097 	void *qend = qp->sq.qend;
2098 	void *addr;
2099 	int inl = 0;
2100 	int copy;
2101 	int len;
2102 	int i;
2103 
2104 	seg = wqe;
2105 	wqe += sizeof(*seg);
2106 	for (i = 0; i < wr->num_sge; i++) {
2107 		addr = (void *)(unsigned long)(wr->sg_list[i].addr);
2108 		len  = wr->sg_list[i].length;
2109 		inl += len;
2110 
2111 		if (unlikely(inl > qp->max_inline_data))
2112 			return -ENOMEM;
2113 
2114 		if (unlikely(wqe + len > qend)) {
2115 			copy = qend - wqe;
2116 			memcpy(wqe, addr, copy);
2117 			addr += copy;
2118 			len -= copy;
2119 			wqe = mlx5_get_send_wqe(qp, 0);
2120 		}
2121 		memcpy(wqe, addr, len);
2122 		wqe += len;
2123 	}
2124 
2125 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
2126 
2127 	*sz = ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
2128 
2129 	return 0;
2130 }
2131 
2132 static u16 prot_field_size(enum ib_signature_type type)
2133 {
2134 	switch (type) {
2135 	case IB_SIG_TYPE_T10_DIF:
2136 		return MLX5_DIF_SIZE;
2137 	default:
2138 		return 0;
2139 	}
2140 }
2141 
2142 static u8 bs_selector(int block_size)
2143 {
2144 	switch (block_size) {
2145 	case 512:	    return 0x1;
2146 	case 520:	    return 0x2;
2147 	case 4096:	    return 0x3;
2148 	case 4160:	    return 0x4;
2149 	case 1073741824:    return 0x5;
2150 	default:	    return 0;
2151 	}
2152 }
2153 
2154 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
2155 			      struct mlx5_bsf_inl *inl)
2156 {
2157 	/* Valid inline section and allow BSF refresh */
2158 	inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
2159 				       MLX5_BSF_REFRESH_DIF);
2160 	inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
2161 	inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
2162 	/* repeating block */
2163 	inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
2164 	inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
2165 			MLX5_DIF_CRC : MLX5_DIF_IPCS;
2166 
2167 	if (domain->sig.dif.ref_remap)
2168 		inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
2169 
2170 	if (domain->sig.dif.app_escape) {
2171 		if (domain->sig.dif.ref_escape)
2172 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
2173 		else
2174 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
2175 	}
2176 
2177 	inl->dif_app_bitmask_check =
2178 		cpu_to_be16(domain->sig.dif.apptag_check_mask);
2179 }
2180 
2181 static int mlx5_set_bsf(struct ib_mr *sig_mr,
2182 			struct ib_sig_attrs *sig_attrs,
2183 			struct mlx5_bsf *bsf, u32 data_size)
2184 {
2185 	struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
2186 	struct mlx5_bsf_basic *basic = &bsf->basic;
2187 	struct ib_sig_domain *mem = &sig_attrs->mem;
2188 	struct ib_sig_domain *wire = &sig_attrs->wire;
2189 
2190 	memset(bsf, 0, sizeof(*bsf));
2191 
2192 	/* Basic + Extended + Inline */
2193 	basic->bsf_size_sbs = 1 << 7;
2194 	/* Input domain check byte mask */
2195 	basic->check_byte_mask = sig_attrs->check_mask;
2196 	basic->raw_data_size = cpu_to_be32(data_size);
2197 
2198 	/* Memory domain */
2199 	switch (sig_attrs->mem.sig_type) {
2200 	case IB_SIG_TYPE_NONE:
2201 		break;
2202 	case IB_SIG_TYPE_T10_DIF:
2203 		basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
2204 		basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
2205 		mlx5_fill_inl_bsf(mem, &bsf->m_inl);
2206 		break;
2207 	default:
2208 		return -EINVAL;
2209 	}
2210 
2211 	/* Wire domain */
2212 	switch (sig_attrs->wire.sig_type) {
2213 	case IB_SIG_TYPE_NONE:
2214 		break;
2215 	case IB_SIG_TYPE_T10_DIF:
2216 		if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
2217 		    mem->sig_type == wire->sig_type) {
2218 			/* Same block structure */
2219 			basic->bsf_size_sbs |= 1 << 4;
2220 			if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
2221 				basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
2222 			if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
2223 				basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
2224 			if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
2225 				basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
2226 		} else
2227 			basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
2228 
2229 		basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
2230 		mlx5_fill_inl_bsf(wire, &bsf->w_inl);
2231 		break;
2232 	default:
2233 		return -EINVAL;
2234 	}
2235 
2236 	return 0;
2237 }
2238 
2239 static int set_sig_data_segment(struct ib_send_wr *wr, struct mlx5_ib_qp *qp,
2240 				void **seg, int *size)
2241 {
2242 	struct ib_sig_attrs *sig_attrs = wr->wr.sig_handover.sig_attrs;
2243 	struct ib_mr *sig_mr = wr->wr.sig_handover.sig_mr;
2244 	struct mlx5_bsf *bsf;
2245 	u32 data_len = wr->sg_list->length;
2246 	u32 data_key = wr->sg_list->lkey;
2247 	u64 data_va = wr->sg_list->addr;
2248 	int ret;
2249 	int wqe_size;
2250 
2251 	if (!wr->wr.sig_handover.prot ||
2252 	    (data_key == wr->wr.sig_handover.prot->lkey &&
2253 	     data_va == wr->wr.sig_handover.prot->addr &&
2254 	     data_len == wr->wr.sig_handover.prot->length)) {
2255 		/**
2256 		 * Source domain doesn't contain signature information
2257 		 * or data and protection are interleaved in memory.
2258 		 * So need construct:
2259 		 *                  ------------------
2260 		 *                 |     data_klm     |
2261 		 *                  ------------------
2262 		 *                 |       BSF        |
2263 		 *                  ------------------
2264 		 **/
2265 		struct mlx5_klm *data_klm = *seg;
2266 
2267 		data_klm->bcount = cpu_to_be32(data_len);
2268 		data_klm->key = cpu_to_be32(data_key);
2269 		data_klm->va = cpu_to_be64(data_va);
2270 		wqe_size = ALIGN(sizeof(*data_klm), 64);
2271 	} else {
2272 		/**
2273 		 * Source domain contains signature information
2274 		 * So need construct a strided block format:
2275 		 *               ---------------------------
2276 		 *              |     stride_block_ctrl     |
2277 		 *               ---------------------------
2278 		 *              |          data_klm         |
2279 		 *               ---------------------------
2280 		 *              |          prot_klm         |
2281 		 *               ---------------------------
2282 		 *              |             BSF           |
2283 		 *               ---------------------------
2284 		 **/
2285 		struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
2286 		struct mlx5_stride_block_entry *data_sentry;
2287 		struct mlx5_stride_block_entry *prot_sentry;
2288 		u32 prot_key = wr->wr.sig_handover.prot->lkey;
2289 		u64 prot_va = wr->wr.sig_handover.prot->addr;
2290 		u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
2291 		int prot_size;
2292 
2293 		sblock_ctrl = *seg;
2294 		data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
2295 		prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
2296 
2297 		prot_size = prot_field_size(sig_attrs->mem.sig_type);
2298 		if (!prot_size) {
2299 			pr_err("Bad block size given: %u\n", block_size);
2300 			return -EINVAL;
2301 		}
2302 		sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
2303 							    prot_size);
2304 		sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
2305 		sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
2306 		sblock_ctrl->num_entries = cpu_to_be16(2);
2307 
2308 		data_sentry->bcount = cpu_to_be16(block_size);
2309 		data_sentry->key = cpu_to_be32(data_key);
2310 		data_sentry->va = cpu_to_be64(data_va);
2311 		data_sentry->stride = cpu_to_be16(block_size);
2312 
2313 		prot_sentry->bcount = cpu_to_be16(prot_size);
2314 		prot_sentry->key = cpu_to_be32(prot_key);
2315 		prot_sentry->va = cpu_to_be64(prot_va);
2316 		prot_sentry->stride = cpu_to_be16(prot_size);
2317 
2318 		wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
2319 				 sizeof(*prot_sentry), 64);
2320 	}
2321 
2322 	*seg += wqe_size;
2323 	*size += wqe_size / 16;
2324 	if (unlikely((*seg == qp->sq.qend)))
2325 		*seg = mlx5_get_send_wqe(qp, 0);
2326 
2327 	bsf = *seg;
2328 	ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
2329 	if (ret)
2330 		return -EINVAL;
2331 
2332 	*seg += sizeof(*bsf);
2333 	*size += sizeof(*bsf) / 16;
2334 	if (unlikely((*seg == qp->sq.qend)))
2335 		*seg = mlx5_get_send_wqe(qp, 0);
2336 
2337 	return 0;
2338 }
2339 
2340 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
2341 				 struct ib_send_wr *wr, u32 nelements,
2342 				 u32 length, u32 pdn)
2343 {
2344 	struct ib_mr *sig_mr = wr->wr.sig_handover.sig_mr;
2345 	u32 sig_key = sig_mr->rkey;
2346 	u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
2347 
2348 	memset(seg, 0, sizeof(*seg));
2349 
2350 	seg->flags = get_umr_flags(wr->wr.sig_handover.access_flags) |
2351 				   MLX5_ACCESS_MODE_KLM;
2352 	seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
2353 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
2354 				    MLX5_MKEY_BSF_EN | pdn);
2355 	seg->len = cpu_to_be64(length);
2356 	seg->xlt_oct_size = cpu_to_be32(be16_to_cpu(get_klm_octo(nelements)));
2357 	seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
2358 }
2359 
2360 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
2361 				struct ib_send_wr *wr, u32 nelements)
2362 {
2363 	memset(umr, 0, sizeof(*umr));
2364 
2365 	umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
2366 	umr->klm_octowords = get_klm_octo(nelements);
2367 	umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
2368 	umr->mkey_mask = sig_mkey_mask();
2369 }
2370 
2371 
2372 static int set_sig_umr_wr(struct ib_send_wr *wr, struct mlx5_ib_qp *qp,
2373 			  void **seg, int *size)
2374 {
2375 	struct mlx5_ib_mr *sig_mr = to_mmr(wr->wr.sig_handover.sig_mr);
2376 	u32 pdn = get_pd(qp)->pdn;
2377 	u32 klm_oct_size;
2378 	int region_len, ret;
2379 
2380 	if (unlikely(wr->num_sge != 1) ||
2381 	    unlikely(wr->wr.sig_handover.access_flags &
2382 		     IB_ACCESS_REMOTE_ATOMIC) ||
2383 	    unlikely(!sig_mr->sig) || unlikely(!qp->signature_en) ||
2384 	    unlikely(!sig_mr->sig->sig_status_checked))
2385 		return -EINVAL;
2386 
2387 	/* length of the protected region, data + protection */
2388 	region_len = wr->sg_list->length;
2389 	if (wr->wr.sig_handover.prot &&
2390 	    (wr->wr.sig_handover.prot->lkey != wr->sg_list->lkey  ||
2391 	     wr->wr.sig_handover.prot->addr != wr->sg_list->addr  ||
2392 	     wr->wr.sig_handover.prot->length != wr->sg_list->length))
2393 		region_len += wr->wr.sig_handover.prot->length;
2394 
2395 	/**
2396 	 * KLM octoword size - if protection was provided
2397 	 * then we use strided block format (3 octowords),
2398 	 * else we use single KLM (1 octoword)
2399 	 **/
2400 	klm_oct_size = wr->wr.sig_handover.prot ? 3 : 1;
2401 
2402 	set_sig_umr_segment(*seg, wr, klm_oct_size);
2403 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2404 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2405 	if (unlikely((*seg == qp->sq.qend)))
2406 		*seg = mlx5_get_send_wqe(qp, 0);
2407 
2408 	set_sig_mkey_segment(*seg, wr, klm_oct_size, region_len, pdn);
2409 	*seg += sizeof(struct mlx5_mkey_seg);
2410 	*size += sizeof(struct mlx5_mkey_seg) / 16;
2411 	if (unlikely((*seg == qp->sq.qend)))
2412 		*seg = mlx5_get_send_wqe(qp, 0);
2413 
2414 	ret = set_sig_data_segment(wr, qp, seg, size);
2415 	if (ret)
2416 		return ret;
2417 
2418 	sig_mr->sig->sig_status_checked = false;
2419 	return 0;
2420 }
2421 
2422 static int set_psv_wr(struct ib_sig_domain *domain,
2423 		      u32 psv_idx, void **seg, int *size)
2424 {
2425 	struct mlx5_seg_set_psv *psv_seg = *seg;
2426 
2427 	memset(psv_seg, 0, sizeof(*psv_seg));
2428 	psv_seg->psv_num = cpu_to_be32(psv_idx);
2429 	switch (domain->sig_type) {
2430 	case IB_SIG_TYPE_NONE:
2431 		break;
2432 	case IB_SIG_TYPE_T10_DIF:
2433 		psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
2434 						     domain->sig.dif.app_tag);
2435 		psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
2436 		break;
2437 	default:
2438 		pr_err("Bad signature type given.\n");
2439 		return 1;
2440 	}
2441 
2442 	*seg += sizeof(*psv_seg);
2443 	*size += sizeof(*psv_seg) / 16;
2444 
2445 	return 0;
2446 }
2447 
2448 static int set_frwr_li_wr(void **seg, struct ib_send_wr *wr, int *size,
2449 			  struct mlx5_core_dev *mdev, struct mlx5_ib_pd *pd, struct mlx5_ib_qp *qp)
2450 {
2451 	int writ = 0;
2452 	int li;
2453 
2454 	li = wr->opcode == IB_WR_LOCAL_INV ? 1 : 0;
2455 	if (unlikely(wr->send_flags & IB_SEND_INLINE))
2456 		return -EINVAL;
2457 
2458 	set_frwr_umr_segment(*seg, wr, li);
2459 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2460 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2461 	if (unlikely((*seg == qp->sq.qend)))
2462 		*seg = mlx5_get_send_wqe(qp, 0);
2463 	set_mkey_segment(*seg, wr, li, &writ);
2464 	*seg += sizeof(struct mlx5_mkey_seg);
2465 	*size += sizeof(struct mlx5_mkey_seg) / 16;
2466 	if (unlikely((*seg == qp->sq.qend)))
2467 		*seg = mlx5_get_send_wqe(qp, 0);
2468 	if (!li) {
2469 		if (unlikely(wr->wr.fast_reg.page_list_len >
2470 			     wr->wr.fast_reg.page_list->max_page_list_len))
2471 			return	-ENOMEM;
2472 
2473 		set_frwr_pages(*seg, wr, mdev, pd, writ);
2474 		*seg += sizeof(struct mlx5_wqe_data_seg);
2475 		*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
2476 	}
2477 	return 0;
2478 }
2479 
2480 static void dump_wqe(struct mlx5_ib_qp *qp, int idx, int size_16)
2481 {
2482 	__be32 *p = NULL;
2483 	int tidx = idx;
2484 	int i, j;
2485 
2486 	pr_debug("dump wqe at %p\n", mlx5_get_send_wqe(qp, tidx));
2487 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
2488 		if ((i & 0xf) == 0) {
2489 			void *buf = mlx5_get_send_wqe(qp, tidx);
2490 			tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
2491 			p = buf;
2492 			j = 0;
2493 		}
2494 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
2495 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
2496 			 be32_to_cpu(p[j + 3]));
2497 	}
2498 }
2499 
2500 static void mlx5_bf_copy(u64 __iomem *dst, u64 *src,
2501 			 unsigned bytecnt, struct mlx5_ib_qp *qp)
2502 {
2503 	while (bytecnt > 0) {
2504 		__iowrite64_copy(dst++, src++, 8);
2505 		__iowrite64_copy(dst++, src++, 8);
2506 		__iowrite64_copy(dst++, src++, 8);
2507 		__iowrite64_copy(dst++, src++, 8);
2508 		__iowrite64_copy(dst++, src++, 8);
2509 		__iowrite64_copy(dst++, src++, 8);
2510 		__iowrite64_copy(dst++, src++, 8);
2511 		__iowrite64_copy(dst++, src++, 8);
2512 		bytecnt -= 64;
2513 		if (unlikely(src == qp->sq.qend))
2514 			src = mlx5_get_send_wqe(qp, 0);
2515 	}
2516 }
2517 
2518 static u8 get_fence(u8 fence, struct ib_send_wr *wr)
2519 {
2520 	if (unlikely(wr->opcode == IB_WR_LOCAL_INV &&
2521 		     wr->send_flags & IB_SEND_FENCE))
2522 		return MLX5_FENCE_MODE_STRONG_ORDERING;
2523 
2524 	if (unlikely(fence)) {
2525 		if (wr->send_flags & IB_SEND_FENCE)
2526 			return MLX5_FENCE_MODE_SMALL_AND_FENCE;
2527 		else
2528 			return fence;
2529 
2530 	} else {
2531 		return 0;
2532 	}
2533 }
2534 
2535 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
2536 		     struct mlx5_wqe_ctrl_seg **ctrl,
2537 		     struct ib_send_wr *wr, unsigned *idx,
2538 		     int *size, int nreq)
2539 {
2540 	int err = 0;
2541 
2542 	if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq))) {
2543 		err = -ENOMEM;
2544 		return err;
2545 	}
2546 
2547 	*idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
2548 	*seg = mlx5_get_send_wqe(qp, *idx);
2549 	*ctrl = *seg;
2550 	*(uint32_t *)(*seg + 8) = 0;
2551 	(*ctrl)->imm = send_ieth(wr);
2552 	(*ctrl)->fm_ce_se = qp->sq_signal_bits |
2553 		(wr->send_flags & IB_SEND_SIGNALED ?
2554 		 MLX5_WQE_CTRL_CQ_UPDATE : 0) |
2555 		(wr->send_flags & IB_SEND_SOLICITED ?
2556 		 MLX5_WQE_CTRL_SOLICITED : 0);
2557 
2558 	*seg += sizeof(**ctrl);
2559 	*size = sizeof(**ctrl) / 16;
2560 
2561 	return err;
2562 }
2563 
2564 static void finish_wqe(struct mlx5_ib_qp *qp,
2565 		       struct mlx5_wqe_ctrl_seg *ctrl,
2566 		       u8 size, unsigned idx, u64 wr_id,
2567 		       int nreq, u8 fence, u8 next_fence,
2568 		       u32 mlx5_opcode)
2569 {
2570 	u8 opmod = 0;
2571 
2572 	ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
2573 					     mlx5_opcode | ((u32)opmod << 24));
2574 	ctrl->qpn_ds = cpu_to_be32(size | (qp->mqp.qpn << 8));
2575 	ctrl->fm_ce_se |= fence;
2576 	qp->fm_cache = next_fence;
2577 	if (unlikely(qp->wq_sig))
2578 		ctrl->signature = wq_sig(ctrl);
2579 
2580 	qp->sq.wrid[idx] = wr_id;
2581 	qp->sq.w_list[idx].opcode = mlx5_opcode;
2582 	qp->sq.wqe_head[idx] = qp->sq.head + nreq;
2583 	qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
2584 	qp->sq.w_list[idx].next = qp->sq.cur_post;
2585 }
2586 
2587 
2588 int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2589 		      struct ib_send_wr **bad_wr)
2590 {
2591 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
2592 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2593 	struct mlx5_core_dev *mdev = dev->mdev;
2594 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2595 	struct mlx5_ib_mr *mr;
2596 	struct mlx5_wqe_data_seg *dpseg;
2597 	struct mlx5_wqe_xrc_seg *xrc;
2598 	struct mlx5_bf *bf = qp->bf;
2599 	int uninitialized_var(size);
2600 	void *qend = qp->sq.qend;
2601 	unsigned long flags;
2602 	unsigned idx;
2603 	int err = 0;
2604 	int inl = 0;
2605 	int num_sge;
2606 	void *seg;
2607 	int nreq;
2608 	int i;
2609 	u8 next_fence = 0;
2610 	u8 fence;
2611 
2612 	spin_lock_irqsave(&qp->sq.lock, flags);
2613 
2614 	for (nreq = 0; wr; nreq++, wr = wr->next) {
2615 		if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
2616 			mlx5_ib_warn(dev, "\n");
2617 			err = -EINVAL;
2618 			*bad_wr = wr;
2619 			goto out;
2620 		}
2621 
2622 		fence = qp->fm_cache;
2623 		num_sge = wr->num_sge;
2624 		if (unlikely(num_sge > qp->sq.max_gs)) {
2625 			mlx5_ib_warn(dev, "\n");
2626 			err = -ENOMEM;
2627 			*bad_wr = wr;
2628 			goto out;
2629 		}
2630 
2631 		err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, nreq);
2632 		if (err) {
2633 			mlx5_ib_warn(dev, "\n");
2634 			err = -ENOMEM;
2635 			*bad_wr = wr;
2636 			goto out;
2637 		}
2638 
2639 		switch (ibqp->qp_type) {
2640 		case IB_QPT_XRC_INI:
2641 			xrc = seg;
2642 			xrc->xrc_srqn = htonl(wr->xrc_remote_srq_num);
2643 			seg += sizeof(*xrc);
2644 			size += sizeof(*xrc) / 16;
2645 			/* fall through */
2646 		case IB_QPT_RC:
2647 			switch (wr->opcode) {
2648 			case IB_WR_RDMA_READ:
2649 			case IB_WR_RDMA_WRITE:
2650 			case IB_WR_RDMA_WRITE_WITH_IMM:
2651 				set_raddr_seg(seg, wr->wr.rdma.remote_addr,
2652 					      wr->wr.rdma.rkey);
2653 				seg += sizeof(struct mlx5_wqe_raddr_seg);
2654 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
2655 				break;
2656 
2657 			case IB_WR_ATOMIC_CMP_AND_SWP:
2658 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2659 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2660 				mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
2661 				err = -ENOSYS;
2662 				*bad_wr = wr;
2663 				goto out;
2664 
2665 			case IB_WR_LOCAL_INV:
2666 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2667 				qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
2668 				ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
2669 				err = set_frwr_li_wr(&seg, wr, &size, mdev, to_mpd(ibqp->pd), qp);
2670 				if (err) {
2671 					mlx5_ib_warn(dev, "\n");
2672 					*bad_wr = wr;
2673 					goto out;
2674 				}
2675 				num_sge = 0;
2676 				break;
2677 
2678 			case IB_WR_FAST_REG_MR:
2679 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2680 				qp->sq.wr_data[idx] = IB_WR_FAST_REG_MR;
2681 				ctrl->imm = cpu_to_be32(wr->wr.fast_reg.rkey);
2682 				err = set_frwr_li_wr(&seg, wr, &size, mdev, to_mpd(ibqp->pd), qp);
2683 				if (err) {
2684 					mlx5_ib_warn(dev, "\n");
2685 					*bad_wr = wr;
2686 					goto out;
2687 				}
2688 				num_sge = 0;
2689 				break;
2690 
2691 			case IB_WR_REG_SIG_MR:
2692 				qp->sq.wr_data[idx] = IB_WR_REG_SIG_MR;
2693 				mr = to_mmr(wr->wr.sig_handover.sig_mr);
2694 
2695 				ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
2696 				err = set_sig_umr_wr(wr, qp, &seg, &size);
2697 				if (err) {
2698 					mlx5_ib_warn(dev, "\n");
2699 					*bad_wr = wr;
2700 					goto out;
2701 				}
2702 
2703 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2704 					   nreq, get_fence(fence, wr),
2705 					   next_fence, MLX5_OPCODE_UMR);
2706 				/*
2707 				 * SET_PSV WQEs are not signaled and solicited
2708 				 * on error
2709 				 */
2710 				wr->send_flags &= ~IB_SEND_SIGNALED;
2711 				wr->send_flags |= IB_SEND_SOLICITED;
2712 				err = begin_wqe(qp, &seg, &ctrl, wr,
2713 						&idx, &size, nreq);
2714 				if (err) {
2715 					mlx5_ib_warn(dev, "\n");
2716 					err = -ENOMEM;
2717 					*bad_wr = wr;
2718 					goto out;
2719 				}
2720 
2721 				err = set_psv_wr(&wr->wr.sig_handover.sig_attrs->mem,
2722 						 mr->sig->psv_memory.psv_idx, &seg,
2723 						 &size);
2724 				if (err) {
2725 					mlx5_ib_warn(dev, "\n");
2726 					*bad_wr = wr;
2727 					goto out;
2728 				}
2729 
2730 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2731 					   nreq, get_fence(fence, wr),
2732 					   next_fence, MLX5_OPCODE_SET_PSV);
2733 				err = begin_wqe(qp, &seg, &ctrl, wr,
2734 						&idx, &size, nreq);
2735 				if (err) {
2736 					mlx5_ib_warn(dev, "\n");
2737 					err = -ENOMEM;
2738 					*bad_wr = wr;
2739 					goto out;
2740 				}
2741 
2742 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2743 				err = set_psv_wr(&wr->wr.sig_handover.sig_attrs->wire,
2744 						 mr->sig->psv_wire.psv_idx, &seg,
2745 						 &size);
2746 				if (err) {
2747 					mlx5_ib_warn(dev, "\n");
2748 					*bad_wr = wr;
2749 					goto out;
2750 				}
2751 
2752 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2753 					   nreq, get_fence(fence, wr),
2754 					   next_fence, MLX5_OPCODE_SET_PSV);
2755 				num_sge = 0;
2756 				goto skip_psv;
2757 
2758 			default:
2759 				break;
2760 			}
2761 			break;
2762 
2763 		case IB_QPT_UC:
2764 			switch (wr->opcode) {
2765 			case IB_WR_RDMA_WRITE:
2766 			case IB_WR_RDMA_WRITE_WITH_IMM:
2767 				set_raddr_seg(seg, wr->wr.rdma.remote_addr,
2768 					      wr->wr.rdma.rkey);
2769 				seg  += sizeof(struct mlx5_wqe_raddr_seg);
2770 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
2771 				break;
2772 
2773 			default:
2774 				break;
2775 			}
2776 			break;
2777 
2778 		case IB_QPT_UD:
2779 		case IB_QPT_SMI:
2780 		case IB_QPT_GSI:
2781 			set_datagram_seg(seg, wr);
2782 			seg += sizeof(struct mlx5_wqe_datagram_seg);
2783 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
2784 			if (unlikely((seg == qend)))
2785 				seg = mlx5_get_send_wqe(qp, 0);
2786 			break;
2787 
2788 		case MLX5_IB_QPT_REG_UMR:
2789 			if (wr->opcode != MLX5_IB_WR_UMR) {
2790 				err = -EINVAL;
2791 				mlx5_ib_warn(dev, "bad opcode\n");
2792 				goto out;
2793 			}
2794 			qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
2795 			ctrl->imm = cpu_to_be32(wr->wr.fast_reg.rkey);
2796 			set_reg_umr_segment(seg, wr);
2797 			seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2798 			size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2799 			if (unlikely((seg == qend)))
2800 				seg = mlx5_get_send_wqe(qp, 0);
2801 			set_reg_mkey_segment(seg, wr);
2802 			seg += sizeof(struct mlx5_mkey_seg);
2803 			size += sizeof(struct mlx5_mkey_seg) / 16;
2804 			if (unlikely((seg == qend)))
2805 				seg = mlx5_get_send_wqe(qp, 0);
2806 			break;
2807 
2808 		default:
2809 			break;
2810 		}
2811 
2812 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
2813 			int uninitialized_var(sz);
2814 
2815 			err = set_data_inl_seg(qp, wr, seg, &sz);
2816 			if (unlikely(err)) {
2817 				mlx5_ib_warn(dev, "\n");
2818 				*bad_wr = wr;
2819 				goto out;
2820 			}
2821 			inl = 1;
2822 			size += sz;
2823 		} else {
2824 			dpseg = seg;
2825 			for (i = 0; i < num_sge; i++) {
2826 				if (unlikely(dpseg == qend)) {
2827 					seg = mlx5_get_send_wqe(qp, 0);
2828 					dpseg = seg;
2829 				}
2830 				if (likely(wr->sg_list[i].length)) {
2831 					set_data_ptr_seg(dpseg, wr->sg_list + i);
2832 					size += sizeof(struct mlx5_wqe_data_seg) / 16;
2833 					dpseg++;
2834 				}
2835 			}
2836 		}
2837 
2838 		finish_wqe(qp, ctrl, size, idx, wr->wr_id, nreq,
2839 			   get_fence(fence, wr), next_fence,
2840 			   mlx5_ib_opcode[wr->opcode]);
2841 skip_psv:
2842 		if (0)
2843 			dump_wqe(qp, idx, size);
2844 	}
2845 
2846 out:
2847 	if (likely(nreq)) {
2848 		qp->sq.head += nreq;
2849 
2850 		/* Make sure that descriptors are written before
2851 		 * updating doorbell record and ringing the doorbell
2852 		 */
2853 		wmb();
2854 
2855 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
2856 
2857 		/* Make sure doorbell record is visible to the HCA before
2858 		 * we hit doorbell */
2859 		wmb();
2860 
2861 		if (bf->need_lock)
2862 			spin_lock(&bf->lock);
2863 		else
2864 			__acquire(&bf->lock);
2865 
2866 		/* TBD enable WC */
2867 		if (0 && nreq == 1 && bf->uuarn && inl && size > 1 && size <= bf->buf_size / 16) {
2868 			mlx5_bf_copy(bf->reg + bf->offset, (u64 *)ctrl, ALIGN(size * 16, 64), qp);
2869 			/* wc_wmb(); */
2870 		} else {
2871 			mlx5_write64((__be32 *)ctrl, bf->regreg + bf->offset,
2872 				     MLX5_GET_DOORBELL_LOCK(&bf->lock32));
2873 			/* Make sure doorbells don't leak out of SQ spinlock
2874 			 * and reach the HCA out of order.
2875 			 */
2876 			mmiowb();
2877 		}
2878 		bf->offset ^= bf->buf_size;
2879 		if (bf->need_lock)
2880 			spin_unlock(&bf->lock);
2881 		else
2882 			__release(&bf->lock);
2883 	}
2884 
2885 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2886 
2887 	return err;
2888 }
2889 
2890 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
2891 {
2892 	sig->signature = calc_sig(sig, size);
2893 }
2894 
2895 int mlx5_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2896 		      struct ib_recv_wr **bad_wr)
2897 {
2898 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2899 	struct mlx5_wqe_data_seg *scat;
2900 	struct mlx5_rwqe_sig *sig;
2901 	unsigned long flags;
2902 	int err = 0;
2903 	int nreq;
2904 	int ind;
2905 	int i;
2906 
2907 	spin_lock_irqsave(&qp->rq.lock, flags);
2908 
2909 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
2910 
2911 	for (nreq = 0; wr; nreq++, wr = wr->next) {
2912 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2913 			err = -ENOMEM;
2914 			*bad_wr = wr;
2915 			goto out;
2916 		}
2917 
2918 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2919 			err = -EINVAL;
2920 			*bad_wr = wr;
2921 			goto out;
2922 		}
2923 
2924 		scat = get_recv_wqe(qp, ind);
2925 		if (qp->wq_sig)
2926 			scat++;
2927 
2928 		for (i = 0; i < wr->num_sge; i++)
2929 			set_data_ptr_seg(scat + i, wr->sg_list + i);
2930 
2931 		if (i < qp->rq.max_gs) {
2932 			scat[i].byte_count = 0;
2933 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
2934 			scat[i].addr       = 0;
2935 		}
2936 
2937 		if (qp->wq_sig) {
2938 			sig = (struct mlx5_rwqe_sig *)scat;
2939 			set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
2940 		}
2941 
2942 		qp->rq.wrid[ind] = wr->wr_id;
2943 
2944 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
2945 	}
2946 
2947 out:
2948 	if (likely(nreq)) {
2949 		qp->rq.head += nreq;
2950 
2951 		/* Make sure that descriptors are written before
2952 		 * doorbell record.
2953 		 */
2954 		wmb();
2955 
2956 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
2957 	}
2958 
2959 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2960 
2961 	return err;
2962 }
2963 
2964 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
2965 {
2966 	switch (mlx5_state) {
2967 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
2968 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
2969 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
2970 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
2971 	case MLX5_QP_STATE_SQ_DRAINING:
2972 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
2973 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
2974 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
2975 	default:		     return -1;
2976 	}
2977 }
2978 
2979 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
2980 {
2981 	switch (mlx5_mig_state) {
2982 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
2983 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
2984 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
2985 	default: return -1;
2986 	}
2987 }
2988 
2989 static int to_ib_qp_access_flags(int mlx5_flags)
2990 {
2991 	int ib_flags = 0;
2992 
2993 	if (mlx5_flags & MLX5_QP_BIT_RRE)
2994 		ib_flags |= IB_ACCESS_REMOTE_READ;
2995 	if (mlx5_flags & MLX5_QP_BIT_RWE)
2996 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
2997 	if (mlx5_flags & MLX5_QP_BIT_RAE)
2998 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
2999 
3000 	return ib_flags;
3001 }
3002 
3003 static void to_ib_ah_attr(struct mlx5_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
3004 				struct mlx5_qp_path *path)
3005 {
3006 	struct mlx5_core_dev *dev = ibdev->mdev;
3007 
3008 	memset(ib_ah_attr, 0, sizeof(*ib_ah_attr));
3009 	ib_ah_attr->port_num	  = path->port;
3010 
3011 	if (ib_ah_attr->port_num == 0 ||
3012 	    ib_ah_attr->port_num > dev->caps.gen.num_ports)
3013 		return;
3014 
3015 	ib_ah_attr->sl = path->sl & 0xf;
3016 
3017 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
3018 	ib_ah_attr->src_path_bits = path->grh_mlid & 0x7f;
3019 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
3020 	ib_ah_attr->ah_flags      = (path->grh_mlid & (1 << 7)) ? IB_AH_GRH : 0;
3021 	if (ib_ah_attr->ah_flags) {
3022 		ib_ah_attr->grh.sgid_index = path->mgid_index;
3023 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
3024 		ib_ah_attr->grh.traffic_class =
3025 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3026 		ib_ah_attr->grh.flow_label =
3027 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3028 		memcpy(ib_ah_attr->grh.dgid.raw,
3029 		       path->rgid, sizeof(ib_ah_attr->grh.dgid.raw));
3030 	}
3031 }
3032 
3033 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3034 		     struct ib_qp_init_attr *qp_init_attr)
3035 {
3036 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3037 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3038 	struct mlx5_query_qp_mbox_out *outb;
3039 	struct mlx5_qp_context *context;
3040 	int mlx5_state;
3041 	int err = 0;
3042 
3043 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
3044 	/*
3045 	 * Wait for any outstanding page faults, in case the user frees memory
3046 	 * based upon this query's result.
3047 	 */
3048 	flush_workqueue(mlx5_ib_page_fault_wq);
3049 #endif
3050 
3051 	mutex_lock(&qp->mutex);
3052 	outb = kzalloc(sizeof(*outb), GFP_KERNEL);
3053 	if (!outb) {
3054 		err = -ENOMEM;
3055 		goto out;
3056 	}
3057 	context = &outb->ctx;
3058 	err = mlx5_core_qp_query(dev->mdev, &qp->mqp, outb, sizeof(*outb));
3059 	if (err)
3060 		goto out_free;
3061 
3062 	mlx5_state = be32_to_cpu(context->flags) >> 28;
3063 
3064 	qp->state		     = to_ib_qp_state(mlx5_state);
3065 	qp_attr->qp_state	     = qp->state;
3066 	qp_attr->path_mtu	     = context->mtu_msgmax >> 5;
3067 	qp_attr->path_mig_state	     =
3068 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
3069 	qp_attr->qkey		     = be32_to_cpu(context->qkey);
3070 	qp_attr->rq_psn		     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
3071 	qp_attr->sq_psn		     = be32_to_cpu(context->next_send_psn) & 0xffffff;
3072 	qp_attr->dest_qp_num	     = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
3073 	qp_attr->qp_access_flags     =
3074 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
3075 
3076 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3077 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
3078 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
3079 		qp_attr->alt_pkey_index = context->alt_path.pkey_index & 0x7f;
3080 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
3081 	}
3082 
3083 	qp_attr->pkey_index = context->pri_path.pkey_index & 0x7f;
3084 	qp_attr->port_num = context->pri_path.port;
3085 
3086 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3087 	qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
3088 
3089 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
3090 
3091 	qp_attr->max_dest_rd_atomic =
3092 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
3093 	qp_attr->min_rnr_timer	    =
3094 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
3095 	qp_attr->timeout	    = context->pri_path.ackto_lt >> 3;
3096 	qp_attr->retry_cnt	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
3097 	qp_attr->rnr_retry	    = (be32_to_cpu(context->params1) >> 13) & 0x7;
3098 	qp_attr->alt_timeout	    = context->alt_path.ackto_lt >> 3;
3099 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
3100 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
3101 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
3102 
3103 	if (!ibqp->uobject) {
3104 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
3105 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
3106 	} else {
3107 		qp_attr->cap.max_send_wr  = 0;
3108 		qp_attr->cap.max_send_sge = 0;
3109 	}
3110 
3111 	/* We don't support inline sends for kernel QPs (yet), and we
3112 	 * don't know what userspace's value should be.
3113 	 */
3114 	qp_attr->cap.max_inline_data = 0;
3115 
3116 	qp_init_attr->cap	     = qp_attr->cap;
3117 
3118 	qp_init_attr->create_flags = 0;
3119 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3120 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3121 
3122 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
3123 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3124 
3125 out_free:
3126 	kfree(outb);
3127 
3128 out:
3129 	mutex_unlock(&qp->mutex);
3130 	return err;
3131 }
3132 
3133 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
3134 					  struct ib_ucontext *context,
3135 					  struct ib_udata *udata)
3136 {
3137 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3138 	struct mlx5_general_caps *gen;
3139 	struct mlx5_ib_xrcd *xrcd;
3140 	int err;
3141 
3142 	gen = &dev->mdev->caps.gen;
3143 	if (!(gen->flags & MLX5_DEV_CAP_FLAG_XRC))
3144 		return ERR_PTR(-ENOSYS);
3145 
3146 	xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
3147 	if (!xrcd)
3148 		return ERR_PTR(-ENOMEM);
3149 
3150 	err = mlx5_core_xrcd_alloc(dev->mdev, &xrcd->xrcdn);
3151 	if (err) {
3152 		kfree(xrcd);
3153 		return ERR_PTR(-ENOMEM);
3154 	}
3155 
3156 	return &xrcd->ibxrcd;
3157 }
3158 
3159 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
3160 {
3161 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
3162 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
3163 	int err;
3164 
3165 	err = mlx5_core_xrcd_dealloc(dev->mdev, xrcdn);
3166 	if (err) {
3167 		mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
3168 		return err;
3169 	}
3170 
3171 	kfree(xrcd);
3172 
3173 	return 0;
3174 }
3175