xref: /linux/drivers/infiniband/hw/mlx5/qp.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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_sqp(enum ib_qp_type qp_type)
80 {
81 	return is_qp0(qp_type) || is_qp1(qp_type);
82 }
83 
84 static void *get_wqe(struct mlx5_ib_qp *qp, int offset)
85 {
86 	return mlx5_buf_offset(&qp->buf, offset);
87 }
88 
89 static void *get_recv_wqe(struct mlx5_ib_qp *qp, int n)
90 {
91 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
92 }
93 
94 void *mlx5_get_send_wqe(struct mlx5_ib_qp *qp, int n)
95 {
96 	return get_wqe(qp, qp->sq.offset + (n << MLX5_IB_SQ_STRIDE));
97 }
98 
99 /**
100  * mlx5_ib_read_user_wqe() - Copy a user-space WQE to kernel space.
101  *
102  * @qp: QP to copy from.
103  * @send: copy from the send queue when non-zero, use the receive queue
104  *	  otherwise.
105  * @wqe_index:  index to start copying from. For send work queues, the
106  *		wqe_index is in units of MLX5_SEND_WQE_BB.
107  *		For receive work queue, it is the number of work queue
108  *		element in the queue.
109  * @buffer: destination buffer.
110  * @length: maximum number of bytes to copy.
111  *
112  * Copies at least a single WQE, but may copy more data.
113  *
114  * Return: the number of bytes copied, or an error code.
115  */
116 int mlx5_ib_read_user_wqe(struct mlx5_ib_qp *qp, int send, int wqe_index,
117 			  void *buffer, u32 length)
118 {
119 	struct ib_device *ibdev = qp->ibqp.device;
120 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
121 	struct mlx5_ib_wq *wq = send ? &qp->sq : &qp->rq;
122 	size_t offset;
123 	size_t wq_end;
124 	struct ib_umem *umem = qp->umem;
125 	u32 first_copy_length;
126 	int wqe_length;
127 	int ret;
128 
129 	if (wq->wqe_cnt == 0) {
130 		mlx5_ib_dbg(dev, "mlx5_ib_read_user_wqe for a QP with wqe_cnt == 0. qp_type: 0x%x\n",
131 			    qp->ibqp.qp_type);
132 		return -EINVAL;
133 	}
134 
135 	offset = wq->offset + ((wqe_index % wq->wqe_cnt) << wq->wqe_shift);
136 	wq_end = wq->offset + (wq->wqe_cnt << wq->wqe_shift);
137 
138 	if (send && length < sizeof(struct mlx5_wqe_ctrl_seg))
139 		return -EINVAL;
140 
141 	if (offset > umem->length ||
142 	    (send && offset + sizeof(struct mlx5_wqe_ctrl_seg) > umem->length))
143 		return -EINVAL;
144 
145 	first_copy_length = min_t(u32, offset + length, wq_end) - offset;
146 	ret = ib_umem_copy_from(buffer, umem, offset, first_copy_length);
147 	if (ret)
148 		return ret;
149 
150 	if (send) {
151 		struct mlx5_wqe_ctrl_seg *ctrl = buffer;
152 		int ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
153 
154 		wqe_length = ds * MLX5_WQE_DS_UNITS;
155 	} else {
156 		wqe_length = 1 << wq->wqe_shift;
157 	}
158 
159 	if (wqe_length <= first_copy_length)
160 		return first_copy_length;
161 
162 	ret = ib_umem_copy_from(buffer + first_copy_length, umem, wq->offset,
163 				wqe_length - first_copy_length);
164 	if (ret)
165 		return ret;
166 
167 	return wqe_length;
168 }
169 
170 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
171 {
172 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
173 	struct ib_event event;
174 
175 	if (type == MLX5_EVENT_TYPE_PATH_MIG)
176 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
177 
178 	if (ibqp->event_handler) {
179 		event.device     = ibqp->device;
180 		event.element.qp = ibqp;
181 		switch (type) {
182 		case MLX5_EVENT_TYPE_PATH_MIG:
183 			event.event = IB_EVENT_PATH_MIG;
184 			break;
185 		case MLX5_EVENT_TYPE_COMM_EST:
186 			event.event = IB_EVENT_COMM_EST;
187 			break;
188 		case MLX5_EVENT_TYPE_SQ_DRAINED:
189 			event.event = IB_EVENT_SQ_DRAINED;
190 			break;
191 		case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
192 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
193 			break;
194 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
195 			event.event = IB_EVENT_QP_FATAL;
196 			break;
197 		case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
198 			event.event = IB_EVENT_PATH_MIG_ERR;
199 			break;
200 		case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
201 			event.event = IB_EVENT_QP_REQ_ERR;
202 			break;
203 		case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
204 			event.event = IB_EVENT_QP_ACCESS_ERR;
205 			break;
206 		default:
207 			pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
208 			return;
209 		}
210 
211 		ibqp->event_handler(&event, ibqp->qp_context);
212 	}
213 }
214 
215 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
216 		       int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
217 {
218 	int wqe_size;
219 	int wq_size;
220 
221 	/* Sanity check RQ size before proceeding */
222 	if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
223 		return -EINVAL;
224 
225 	if (!has_rq) {
226 		qp->rq.max_gs = 0;
227 		qp->rq.wqe_cnt = 0;
228 		qp->rq.wqe_shift = 0;
229 	} else {
230 		if (ucmd) {
231 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
232 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
233 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
234 			qp->rq.max_post = qp->rq.wqe_cnt;
235 		} else {
236 			wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
237 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
238 			wqe_size = roundup_pow_of_two(wqe_size);
239 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
240 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
241 			qp->rq.wqe_cnt = wq_size / wqe_size;
242 			if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
243 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
244 					    wqe_size,
245 					    MLX5_CAP_GEN(dev->mdev,
246 							 max_wqe_sz_rq));
247 				return -EINVAL;
248 			}
249 			qp->rq.wqe_shift = ilog2(wqe_size);
250 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
251 			qp->rq.max_post = qp->rq.wqe_cnt;
252 		}
253 	}
254 
255 	return 0;
256 }
257 
258 static int sq_overhead(enum ib_qp_type qp_type)
259 {
260 	int size = 0;
261 
262 	switch (qp_type) {
263 	case IB_QPT_XRC_INI:
264 		size += sizeof(struct mlx5_wqe_xrc_seg);
265 		/* fall through */
266 	case IB_QPT_RC:
267 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
268 			sizeof(struct mlx5_wqe_atomic_seg) +
269 			sizeof(struct mlx5_wqe_raddr_seg);
270 		break;
271 
272 	case IB_QPT_XRC_TGT:
273 		return 0;
274 
275 	case IB_QPT_UC:
276 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
277 			sizeof(struct mlx5_wqe_raddr_seg) +
278 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
279 			sizeof(struct mlx5_mkey_seg);
280 		break;
281 
282 	case IB_QPT_UD:
283 	case IB_QPT_SMI:
284 	case IB_QPT_GSI:
285 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
286 			sizeof(struct mlx5_wqe_datagram_seg);
287 		break;
288 
289 	case MLX5_IB_QPT_REG_UMR:
290 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
291 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
292 			sizeof(struct mlx5_mkey_seg);
293 		break;
294 
295 	default:
296 		return -EINVAL;
297 	}
298 
299 	return size;
300 }
301 
302 static int calc_send_wqe(struct ib_qp_init_attr *attr)
303 {
304 	int inl_size = 0;
305 	int size;
306 
307 	size = sq_overhead(attr->qp_type);
308 	if (size < 0)
309 		return size;
310 
311 	if (attr->cap.max_inline_data) {
312 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
313 			attr->cap.max_inline_data;
314 	}
315 
316 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
317 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN &&
318 	    ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
319 			return MLX5_SIG_WQE_SIZE;
320 	else
321 		return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
322 }
323 
324 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
325 			struct mlx5_ib_qp *qp)
326 {
327 	int wqe_size;
328 	int wq_size;
329 
330 	if (!attr->cap.max_send_wr)
331 		return 0;
332 
333 	wqe_size = calc_send_wqe(attr);
334 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
335 	if (wqe_size < 0)
336 		return wqe_size;
337 
338 	if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
339 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
340 			    wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
341 		return -EINVAL;
342 	}
343 
344 	qp->max_inline_data = wqe_size - sq_overhead(attr->qp_type) -
345 		sizeof(struct mlx5_wqe_inline_seg);
346 	attr->cap.max_inline_data = qp->max_inline_data;
347 
348 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN)
349 		qp->signature_en = true;
350 
351 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
352 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
353 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
354 		mlx5_ib_dbg(dev, "wqe count(%d) exceeds limits(%d)\n",
355 			    qp->sq.wqe_cnt,
356 			    1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
357 		return -ENOMEM;
358 	}
359 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
360 	qp->sq.max_gs = attr->cap.max_send_sge;
361 	qp->sq.max_post = wq_size / wqe_size;
362 	attr->cap.max_send_wr = qp->sq.max_post;
363 
364 	return wq_size;
365 }
366 
367 static int set_user_buf_size(struct mlx5_ib_dev *dev,
368 			    struct mlx5_ib_qp *qp,
369 			    struct mlx5_ib_create_qp *ucmd)
370 {
371 	int desc_sz = 1 << qp->sq.wqe_shift;
372 
373 	if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
374 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
375 			     desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
376 		return -EINVAL;
377 	}
378 
379 	if (ucmd->sq_wqe_count && ((1 << ilog2(ucmd->sq_wqe_count)) != ucmd->sq_wqe_count)) {
380 		mlx5_ib_warn(dev, "sq_wqe_count %d, sq_wqe_count %d\n",
381 			     ucmd->sq_wqe_count, ucmd->sq_wqe_count);
382 		return -EINVAL;
383 	}
384 
385 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
386 
387 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
388 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
389 			     qp->sq.wqe_cnt,
390 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
391 		return -EINVAL;
392 	}
393 
394 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
395 		(qp->sq.wqe_cnt << 6);
396 
397 	return 0;
398 }
399 
400 static int qp_has_rq(struct ib_qp_init_attr *attr)
401 {
402 	if (attr->qp_type == IB_QPT_XRC_INI ||
403 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
404 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
405 	    !attr->cap.max_recv_wr)
406 		return 0;
407 
408 	return 1;
409 }
410 
411 static int first_med_uuar(void)
412 {
413 	return 1;
414 }
415 
416 static int next_uuar(int n)
417 {
418 	n++;
419 
420 	while (((n % 4) & 2))
421 		n++;
422 
423 	return n;
424 }
425 
426 static int num_med_uuar(struct mlx5_uuar_info *uuari)
427 {
428 	int n;
429 
430 	n = uuari->num_uars * MLX5_NON_FP_BF_REGS_PER_PAGE -
431 		uuari->num_low_latency_uuars - 1;
432 
433 	return n >= 0 ? n : 0;
434 }
435 
436 static int max_uuari(struct mlx5_uuar_info *uuari)
437 {
438 	return uuari->num_uars * 4;
439 }
440 
441 static int first_hi_uuar(struct mlx5_uuar_info *uuari)
442 {
443 	int med;
444 	int i;
445 	int t;
446 
447 	med = num_med_uuar(uuari);
448 	for (t = 0, i = first_med_uuar();; i = next_uuar(i)) {
449 		t++;
450 		if (t == med)
451 			return next_uuar(i);
452 	}
453 
454 	return 0;
455 }
456 
457 static int alloc_high_class_uuar(struct mlx5_uuar_info *uuari)
458 {
459 	int i;
460 
461 	for (i = first_hi_uuar(uuari); i < max_uuari(uuari); i = next_uuar(i)) {
462 		if (!test_bit(i, uuari->bitmap)) {
463 			set_bit(i, uuari->bitmap);
464 			uuari->count[i]++;
465 			return i;
466 		}
467 	}
468 
469 	return -ENOMEM;
470 }
471 
472 static int alloc_med_class_uuar(struct mlx5_uuar_info *uuari)
473 {
474 	int minidx = first_med_uuar();
475 	int i;
476 
477 	for (i = first_med_uuar(); i < first_hi_uuar(uuari); i = next_uuar(i)) {
478 		if (uuari->count[i] < uuari->count[minidx])
479 			minidx = i;
480 	}
481 
482 	uuari->count[minidx]++;
483 	return minidx;
484 }
485 
486 static int alloc_uuar(struct mlx5_uuar_info *uuari,
487 		      enum mlx5_ib_latency_class lat)
488 {
489 	int uuarn = -EINVAL;
490 
491 	mutex_lock(&uuari->lock);
492 	switch (lat) {
493 	case MLX5_IB_LATENCY_CLASS_LOW:
494 		uuarn = 0;
495 		uuari->count[uuarn]++;
496 		break;
497 
498 	case MLX5_IB_LATENCY_CLASS_MEDIUM:
499 		if (uuari->ver < 2)
500 			uuarn = -ENOMEM;
501 		else
502 			uuarn = alloc_med_class_uuar(uuari);
503 		break;
504 
505 	case MLX5_IB_LATENCY_CLASS_HIGH:
506 		if (uuari->ver < 2)
507 			uuarn = -ENOMEM;
508 		else
509 			uuarn = alloc_high_class_uuar(uuari);
510 		break;
511 
512 	case MLX5_IB_LATENCY_CLASS_FAST_PATH:
513 		uuarn = 2;
514 		break;
515 	}
516 	mutex_unlock(&uuari->lock);
517 
518 	return uuarn;
519 }
520 
521 static void free_med_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
522 {
523 	clear_bit(uuarn, uuari->bitmap);
524 	--uuari->count[uuarn];
525 }
526 
527 static void free_high_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
528 {
529 	clear_bit(uuarn, uuari->bitmap);
530 	--uuari->count[uuarn];
531 }
532 
533 static void free_uuar(struct mlx5_uuar_info *uuari, int uuarn)
534 {
535 	int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
536 	int high_uuar = nuuars - uuari->num_low_latency_uuars;
537 
538 	mutex_lock(&uuari->lock);
539 	if (uuarn == 0) {
540 		--uuari->count[uuarn];
541 		goto out;
542 	}
543 
544 	if (uuarn < high_uuar) {
545 		free_med_class_uuar(uuari, uuarn);
546 		goto out;
547 	}
548 
549 	free_high_class_uuar(uuari, uuarn);
550 
551 out:
552 	mutex_unlock(&uuari->lock);
553 }
554 
555 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
556 {
557 	switch (state) {
558 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
559 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
560 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
561 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
562 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
563 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
564 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
565 	default:		return -1;
566 	}
567 }
568 
569 static int to_mlx5_st(enum ib_qp_type type)
570 {
571 	switch (type) {
572 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
573 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
574 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
575 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
576 	case IB_QPT_XRC_INI:
577 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
578 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
579 	case IB_QPT_GSI:		return MLX5_QP_ST_QP1;
580 	case IB_QPT_RAW_IPV6:		return MLX5_QP_ST_RAW_IPV6;
581 	case IB_QPT_RAW_ETHERTYPE:	return MLX5_QP_ST_RAW_ETHERTYPE;
582 	case IB_QPT_RAW_PACKET:
583 	case IB_QPT_MAX:
584 	default:		return -EINVAL;
585 	}
586 }
587 
588 static int uuarn_to_uar_index(struct mlx5_uuar_info *uuari, int uuarn)
589 {
590 	return uuari->uars[uuarn / MLX5_BF_REGS_PER_PAGE].index;
591 }
592 
593 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
594 			  struct mlx5_ib_qp *qp, struct ib_udata *udata,
595 			  struct mlx5_create_qp_mbox_in **in,
596 			  struct mlx5_ib_create_qp_resp *resp, int *inlen)
597 {
598 	struct mlx5_ib_ucontext *context;
599 	struct mlx5_ib_create_qp ucmd;
600 	int page_shift = 0;
601 	int uar_index;
602 	int npages;
603 	u32 offset = 0;
604 	int uuarn;
605 	int ncont = 0;
606 	int err;
607 
608 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
609 	if (err) {
610 		mlx5_ib_dbg(dev, "copy failed\n");
611 		return err;
612 	}
613 
614 	context = to_mucontext(pd->uobject->context);
615 	/*
616 	 * TBD: should come from the verbs when we have the API
617 	 */
618 	uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_HIGH);
619 	if (uuarn < 0) {
620 		mlx5_ib_dbg(dev, "failed to allocate low latency UUAR\n");
621 		mlx5_ib_dbg(dev, "reverting to medium latency\n");
622 		uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_MEDIUM);
623 		if (uuarn < 0) {
624 			mlx5_ib_dbg(dev, "failed to allocate medium latency UUAR\n");
625 			mlx5_ib_dbg(dev, "reverting to high latency\n");
626 			uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_LOW);
627 			if (uuarn < 0) {
628 				mlx5_ib_warn(dev, "uuar allocation failed\n");
629 				return uuarn;
630 			}
631 		}
632 	}
633 
634 	uar_index = uuarn_to_uar_index(&context->uuari, uuarn);
635 	mlx5_ib_dbg(dev, "uuarn 0x%x, uar_index 0x%x\n", uuarn, uar_index);
636 
637 	qp->rq.offset = 0;
638 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
639 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
640 
641 	err = set_user_buf_size(dev, qp, &ucmd);
642 	if (err)
643 		goto err_uuar;
644 
645 	if (ucmd.buf_addr && qp->buf_size) {
646 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
647 				       qp->buf_size, 0, 0);
648 		if (IS_ERR(qp->umem)) {
649 			mlx5_ib_dbg(dev, "umem_get failed\n");
650 			err = PTR_ERR(qp->umem);
651 			goto err_uuar;
652 		}
653 	} else {
654 		qp->umem = NULL;
655 	}
656 
657 	if (qp->umem) {
658 		mlx5_ib_cont_pages(qp->umem, ucmd.buf_addr, &npages, &page_shift,
659 				   &ncont, NULL);
660 		err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift, &offset);
661 		if (err) {
662 			mlx5_ib_warn(dev, "bad offset\n");
663 			goto err_umem;
664 		}
665 		mlx5_ib_dbg(dev, "addr 0x%llx, size %d, npages %d, page_shift %d, ncont %d, offset %d\n",
666 			    ucmd.buf_addr, qp->buf_size, npages, page_shift, ncont, offset);
667 	}
668 
669 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * ncont;
670 	*in = mlx5_vzalloc(*inlen);
671 	if (!*in) {
672 		err = -ENOMEM;
673 		goto err_umem;
674 	}
675 	if (qp->umem)
676 		mlx5_ib_populate_pas(dev, qp->umem, page_shift, (*in)->pas, 0);
677 	(*in)->ctx.log_pg_sz_remote_qpn =
678 		cpu_to_be32((page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
679 	(*in)->ctx.params2 = cpu_to_be32(offset << 6);
680 
681 	(*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
682 	resp->uuar_index = uuarn;
683 	qp->uuarn = uuarn;
684 
685 	err = mlx5_ib_db_map_user(context, ucmd.db_addr, &qp->db);
686 	if (err) {
687 		mlx5_ib_dbg(dev, "map failed\n");
688 		goto err_free;
689 	}
690 
691 	err = ib_copy_to_udata(udata, resp, sizeof(*resp));
692 	if (err) {
693 		mlx5_ib_dbg(dev, "copy failed\n");
694 		goto err_unmap;
695 	}
696 	qp->create_type = MLX5_QP_USER;
697 
698 	return 0;
699 
700 err_unmap:
701 	mlx5_ib_db_unmap_user(context, &qp->db);
702 
703 err_free:
704 	kvfree(*in);
705 
706 err_umem:
707 	if (qp->umem)
708 		ib_umem_release(qp->umem);
709 
710 err_uuar:
711 	free_uuar(&context->uuari, uuarn);
712 	return err;
713 }
714 
715 static void destroy_qp_user(struct ib_pd *pd, struct mlx5_ib_qp *qp)
716 {
717 	struct mlx5_ib_ucontext *context;
718 
719 	context = to_mucontext(pd->uobject->context);
720 	mlx5_ib_db_unmap_user(context, &qp->db);
721 	if (qp->umem)
722 		ib_umem_release(qp->umem);
723 	free_uuar(&context->uuari, qp->uuarn);
724 }
725 
726 static int create_kernel_qp(struct mlx5_ib_dev *dev,
727 			    struct ib_qp_init_attr *init_attr,
728 			    struct mlx5_ib_qp *qp,
729 			    struct mlx5_create_qp_mbox_in **in, int *inlen)
730 {
731 	enum mlx5_ib_latency_class lc = MLX5_IB_LATENCY_CLASS_LOW;
732 	struct mlx5_uuar_info *uuari;
733 	int uar_index;
734 	int uuarn;
735 	int err;
736 
737 	uuari = &dev->mdev->priv.uuari;
738 	if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN | IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK))
739 		return -EINVAL;
740 
741 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
742 		lc = MLX5_IB_LATENCY_CLASS_FAST_PATH;
743 
744 	uuarn = alloc_uuar(uuari, lc);
745 	if (uuarn < 0) {
746 		mlx5_ib_dbg(dev, "\n");
747 		return -ENOMEM;
748 	}
749 
750 	qp->bf = &uuari->bfs[uuarn];
751 	uar_index = qp->bf->uar->index;
752 
753 	err = calc_sq_size(dev, init_attr, qp);
754 	if (err < 0) {
755 		mlx5_ib_dbg(dev, "err %d\n", err);
756 		goto err_uuar;
757 	}
758 
759 	qp->rq.offset = 0;
760 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
761 	qp->buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
762 
763 	err = mlx5_buf_alloc(dev->mdev, qp->buf_size, &qp->buf);
764 	if (err) {
765 		mlx5_ib_dbg(dev, "err %d\n", err);
766 		goto err_uuar;
767 	}
768 
769 	qp->sq.qend = mlx5_get_send_wqe(qp, qp->sq.wqe_cnt);
770 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * qp->buf.npages;
771 	*in = mlx5_vzalloc(*inlen);
772 	if (!*in) {
773 		err = -ENOMEM;
774 		goto err_buf;
775 	}
776 	(*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
777 	(*in)->ctx.log_pg_sz_remote_qpn =
778 		cpu_to_be32((qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
779 	/* Set "fast registration enabled" for all kernel QPs */
780 	(*in)->ctx.params1 |= cpu_to_be32(1 << 11);
781 	(*in)->ctx.sq_crq_size |= cpu_to_be16(1 << 4);
782 
783 	mlx5_fill_page_array(&qp->buf, (*in)->pas);
784 
785 	err = mlx5_db_alloc(dev->mdev, &qp->db);
786 	if (err) {
787 		mlx5_ib_dbg(dev, "err %d\n", err);
788 		goto err_free;
789 	}
790 
791 	qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
792 	qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
793 	qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
794 	qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
795 	qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
796 
797 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
798 	    !qp->sq.w_list || !qp->sq.wqe_head) {
799 		err = -ENOMEM;
800 		goto err_wrid;
801 	}
802 	qp->create_type = MLX5_QP_KERNEL;
803 
804 	return 0;
805 
806 err_wrid:
807 	mlx5_db_free(dev->mdev, &qp->db);
808 	kfree(qp->sq.wqe_head);
809 	kfree(qp->sq.w_list);
810 	kfree(qp->sq.wrid);
811 	kfree(qp->sq.wr_data);
812 	kfree(qp->rq.wrid);
813 
814 err_free:
815 	kvfree(*in);
816 
817 err_buf:
818 	mlx5_buf_free(dev->mdev, &qp->buf);
819 
820 err_uuar:
821 	free_uuar(&dev->mdev->priv.uuari, uuarn);
822 	return err;
823 }
824 
825 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
826 {
827 	mlx5_db_free(dev->mdev, &qp->db);
828 	kfree(qp->sq.wqe_head);
829 	kfree(qp->sq.w_list);
830 	kfree(qp->sq.wrid);
831 	kfree(qp->sq.wr_data);
832 	kfree(qp->rq.wrid);
833 	mlx5_buf_free(dev->mdev, &qp->buf);
834 	free_uuar(&dev->mdev->priv.uuari, qp->bf->uuarn);
835 }
836 
837 static __be32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
838 {
839 	if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
840 	    (attr->qp_type == IB_QPT_XRC_INI))
841 		return cpu_to_be32(MLX5_SRQ_RQ);
842 	else if (!qp->has_rq)
843 		return cpu_to_be32(MLX5_ZERO_LEN_RQ);
844 	else
845 		return cpu_to_be32(MLX5_NON_ZERO_RQ);
846 }
847 
848 static int is_connected(enum ib_qp_type qp_type)
849 {
850 	if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC)
851 		return 1;
852 
853 	return 0;
854 }
855 
856 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
857 			    struct ib_qp_init_attr *init_attr,
858 			    struct ib_udata *udata, struct mlx5_ib_qp *qp)
859 {
860 	struct mlx5_ib_resources *devr = &dev->devr;
861 	struct mlx5_core_dev *mdev = dev->mdev;
862 	struct mlx5_ib_create_qp_resp resp;
863 	struct mlx5_create_qp_mbox_in *in;
864 	struct mlx5_ib_create_qp ucmd;
865 	int inlen = sizeof(*in);
866 	int err;
867 
868 	mlx5_ib_odp_create_qp(qp);
869 
870 	mutex_init(&qp->mutex);
871 	spin_lock_init(&qp->sq.lock);
872 	spin_lock_init(&qp->rq.lock);
873 
874 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
875 		if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
876 			mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
877 			return -EINVAL;
878 		} else {
879 			qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
880 		}
881 	}
882 
883 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
884 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
885 
886 	if (pd && pd->uobject) {
887 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
888 			mlx5_ib_dbg(dev, "copy failed\n");
889 			return -EFAULT;
890 		}
891 
892 		qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
893 		qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
894 	} else {
895 		qp->wq_sig = !!wq_signature;
896 	}
897 
898 	qp->has_rq = qp_has_rq(init_attr);
899 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
900 			  qp, (pd && pd->uobject) ? &ucmd : NULL);
901 	if (err) {
902 		mlx5_ib_dbg(dev, "err %d\n", err);
903 		return err;
904 	}
905 
906 	if (pd) {
907 		if (pd->uobject) {
908 			__u32 max_wqes =
909 				1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
910 			mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
911 			if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
912 			    ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
913 				mlx5_ib_dbg(dev, "invalid rq params\n");
914 				return -EINVAL;
915 			}
916 			if (ucmd.sq_wqe_count > max_wqes) {
917 				mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
918 					    ucmd.sq_wqe_count, max_wqes);
919 				return -EINVAL;
920 			}
921 			err = create_user_qp(dev, pd, qp, udata, &in, &resp, &inlen);
922 			if (err)
923 				mlx5_ib_dbg(dev, "err %d\n", err);
924 		} else {
925 			err = create_kernel_qp(dev, init_attr, qp, &in, &inlen);
926 			if (err)
927 				mlx5_ib_dbg(dev, "err %d\n", err);
928 		}
929 
930 		if (err)
931 			return err;
932 	} else {
933 		in = mlx5_vzalloc(sizeof(*in));
934 		if (!in)
935 			return -ENOMEM;
936 
937 		qp->create_type = MLX5_QP_EMPTY;
938 	}
939 
940 	if (is_sqp(init_attr->qp_type))
941 		qp->port = init_attr->port_num;
942 
943 	in->ctx.flags = cpu_to_be32(to_mlx5_st(init_attr->qp_type) << 16 |
944 				    MLX5_QP_PM_MIGRATED << 11);
945 
946 	if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
947 		in->ctx.flags_pd = cpu_to_be32(to_mpd(pd ? pd : devr->p0)->pdn);
948 	else
949 		in->ctx.flags_pd = cpu_to_be32(MLX5_QP_LAT_SENSITIVE);
950 
951 	if (qp->wq_sig)
952 		in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_ENABLE_SIG);
953 
954 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
955 		in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_BLOCK_MCAST);
956 
957 	if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
958 		int rcqe_sz;
959 		int scqe_sz;
960 
961 		rcqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->recv_cq);
962 		scqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->send_cq);
963 
964 		if (rcqe_sz == 128)
965 			in->ctx.cs_res = MLX5_RES_SCAT_DATA64_CQE;
966 		else
967 			in->ctx.cs_res = MLX5_RES_SCAT_DATA32_CQE;
968 
969 		if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) {
970 			if (scqe_sz == 128)
971 				in->ctx.cs_req = MLX5_REQ_SCAT_DATA64_CQE;
972 			else
973 				in->ctx.cs_req = MLX5_REQ_SCAT_DATA32_CQE;
974 		}
975 	}
976 
977 	if (qp->rq.wqe_cnt) {
978 		in->ctx.rq_size_stride = (qp->rq.wqe_shift - 4);
979 		in->ctx.rq_size_stride |= ilog2(qp->rq.wqe_cnt) << 3;
980 	}
981 
982 	in->ctx.rq_type_srqn = get_rx_type(qp, init_attr);
983 
984 	if (qp->sq.wqe_cnt)
985 		in->ctx.sq_crq_size |= cpu_to_be16(ilog2(qp->sq.wqe_cnt) << 11);
986 	else
987 		in->ctx.sq_crq_size |= cpu_to_be16(0x8000);
988 
989 	/* Set default resources */
990 	switch (init_attr->qp_type) {
991 	case IB_QPT_XRC_TGT:
992 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
993 		in->ctx.cqn_send = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
994 		in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
995 		in->ctx.xrcd = cpu_to_be32(to_mxrcd(init_attr->xrcd)->xrcdn);
996 		break;
997 	case IB_QPT_XRC_INI:
998 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
999 		in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
1000 		in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
1001 		break;
1002 	default:
1003 		if (init_attr->srq) {
1004 			in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x0)->xrcdn);
1005 			in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(init_attr->srq)->msrq.srqn);
1006 		} else {
1007 			in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
1008 			in->ctx.rq_type_srqn |=
1009 				cpu_to_be32(to_msrq(devr->s1)->msrq.srqn);
1010 		}
1011 	}
1012 
1013 	if (init_attr->send_cq)
1014 		in->ctx.cqn_send = cpu_to_be32(to_mcq(init_attr->send_cq)->mcq.cqn);
1015 
1016 	if (init_attr->recv_cq)
1017 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(init_attr->recv_cq)->mcq.cqn);
1018 
1019 	in->ctx.db_rec_addr = cpu_to_be64(qp->db.dma);
1020 
1021 	err = mlx5_core_create_qp(dev->mdev, &qp->mqp, in, inlen);
1022 	if (err) {
1023 		mlx5_ib_dbg(dev, "create qp failed\n");
1024 		goto err_create;
1025 	}
1026 
1027 	kvfree(in);
1028 	/* Hardware wants QPN written in big-endian order (after
1029 	 * shifting) for send doorbell.  Precompute this value to save
1030 	 * a little bit when posting sends.
1031 	 */
1032 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1033 
1034 	qp->mqp.event = mlx5_ib_qp_event;
1035 
1036 	return 0;
1037 
1038 err_create:
1039 	if (qp->create_type == MLX5_QP_USER)
1040 		destroy_qp_user(pd, qp);
1041 	else if (qp->create_type == MLX5_QP_KERNEL)
1042 		destroy_qp_kernel(dev, qp);
1043 
1044 	kvfree(in);
1045 	return err;
1046 }
1047 
1048 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1049 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1050 {
1051 	if (send_cq) {
1052 		if (recv_cq) {
1053 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1054 				spin_lock_irq(&send_cq->lock);
1055 				spin_lock_nested(&recv_cq->lock,
1056 						 SINGLE_DEPTH_NESTING);
1057 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1058 				spin_lock_irq(&send_cq->lock);
1059 				__acquire(&recv_cq->lock);
1060 			} else {
1061 				spin_lock_irq(&recv_cq->lock);
1062 				spin_lock_nested(&send_cq->lock,
1063 						 SINGLE_DEPTH_NESTING);
1064 			}
1065 		} else {
1066 			spin_lock_irq(&send_cq->lock);
1067 			__acquire(&recv_cq->lock);
1068 		}
1069 	} else if (recv_cq) {
1070 		spin_lock_irq(&recv_cq->lock);
1071 		__acquire(&send_cq->lock);
1072 	} else {
1073 		__acquire(&send_cq->lock);
1074 		__acquire(&recv_cq->lock);
1075 	}
1076 }
1077 
1078 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1079 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1080 {
1081 	if (send_cq) {
1082 		if (recv_cq) {
1083 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1084 				spin_unlock(&recv_cq->lock);
1085 				spin_unlock_irq(&send_cq->lock);
1086 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1087 				__release(&recv_cq->lock);
1088 				spin_unlock_irq(&send_cq->lock);
1089 			} else {
1090 				spin_unlock(&send_cq->lock);
1091 				spin_unlock_irq(&recv_cq->lock);
1092 			}
1093 		} else {
1094 			__release(&recv_cq->lock);
1095 			spin_unlock_irq(&send_cq->lock);
1096 		}
1097 	} else if (recv_cq) {
1098 		__release(&send_cq->lock);
1099 		spin_unlock_irq(&recv_cq->lock);
1100 	} else {
1101 		__release(&recv_cq->lock);
1102 		__release(&send_cq->lock);
1103 	}
1104 }
1105 
1106 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
1107 {
1108 	return to_mpd(qp->ibqp.pd);
1109 }
1110 
1111 static void get_cqs(struct mlx5_ib_qp *qp,
1112 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
1113 {
1114 	switch (qp->ibqp.qp_type) {
1115 	case IB_QPT_XRC_TGT:
1116 		*send_cq = NULL;
1117 		*recv_cq = NULL;
1118 		break;
1119 	case MLX5_IB_QPT_REG_UMR:
1120 	case IB_QPT_XRC_INI:
1121 		*send_cq = to_mcq(qp->ibqp.send_cq);
1122 		*recv_cq = NULL;
1123 		break;
1124 
1125 	case IB_QPT_SMI:
1126 	case IB_QPT_GSI:
1127 	case IB_QPT_RC:
1128 	case IB_QPT_UC:
1129 	case IB_QPT_UD:
1130 	case IB_QPT_RAW_IPV6:
1131 	case IB_QPT_RAW_ETHERTYPE:
1132 		*send_cq = to_mcq(qp->ibqp.send_cq);
1133 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
1134 		break;
1135 
1136 	case IB_QPT_RAW_PACKET:
1137 	case IB_QPT_MAX:
1138 	default:
1139 		*send_cq = NULL;
1140 		*recv_cq = NULL;
1141 		break;
1142 	}
1143 }
1144 
1145 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1146 {
1147 	struct mlx5_ib_cq *send_cq, *recv_cq;
1148 	struct mlx5_modify_qp_mbox_in *in;
1149 	int err;
1150 
1151 	in = kzalloc(sizeof(*in), GFP_KERNEL);
1152 	if (!in)
1153 		return;
1154 
1155 	if (qp->state != IB_QPS_RESET) {
1156 		mlx5_ib_qp_disable_pagefaults(qp);
1157 		if (mlx5_core_qp_modify(dev->mdev, to_mlx5_state(qp->state),
1158 					MLX5_QP_STATE_RST, in, 0, &qp->mqp))
1159 			mlx5_ib_warn(dev, "mlx5_ib: modify QP %06x to RESET failed\n",
1160 				     qp->mqp.qpn);
1161 	}
1162 
1163 	get_cqs(qp, &send_cq, &recv_cq);
1164 
1165 	if (qp->create_type == MLX5_QP_KERNEL) {
1166 		mlx5_ib_lock_cqs(send_cq, recv_cq);
1167 		__mlx5_ib_cq_clean(recv_cq, qp->mqp.qpn,
1168 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1169 		if (send_cq != recv_cq)
1170 			__mlx5_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1171 		mlx5_ib_unlock_cqs(send_cq, recv_cq);
1172 	}
1173 
1174 	err = mlx5_core_destroy_qp(dev->mdev, &qp->mqp);
1175 	if (err)
1176 		mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n", qp->mqp.qpn);
1177 	kfree(in);
1178 
1179 
1180 	if (qp->create_type == MLX5_QP_KERNEL)
1181 		destroy_qp_kernel(dev, qp);
1182 	else if (qp->create_type == MLX5_QP_USER)
1183 		destroy_qp_user(&get_pd(qp)->ibpd, qp);
1184 }
1185 
1186 static const char *ib_qp_type_str(enum ib_qp_type type)
1187 {
1188 	switch (type) {
1189 	case IB_QPT_SMI:
1190 		return "IB_QPT_SMI";
1191 	case IB_QPT_GSI:
1192 		return "IB_QPT_GSI";
1193 	case IB_QPT_RC:
1194 		return "IB_QPT_RC";
1195 	case IB_QPT_UC:
1196 		return "IB_QPT_UC";
1197 	case IB_QPT_UD:
1198 		return "IB_QPT_UD";
1199 	case IB_QPT_RAW_IPV6:
1200 		return "IB_QPT_RAW_IPV6";
1201 	case IB_QPT_RAW_ETHERTYPE:
1202 		return "IB_QPT_RAW_ETHERTYPE";
1203 	case IB_QPT_XRC_INI:
1204 		return "IB_QPT_XRC_INI";
1205 	case IB_QPT_XRC_TGT:
1206 		return "IB_QPT_XRC_TGT";
1207 	case IB_QPT_RAW_PACKET:
1208 		return "IB_QPT_RAW_PACKET";
1209 	case MLX5_IB_QPT_REG_UMR:
1210 		return "MLX5_IB_QPT_REG_UMR";
1211 	case IB_QPT_MAX:
1212 	default:
1213 		return "Invalid QP type";
1214 	}
1215 }
1216 
1217 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
1218 				struct ib_qp_init_attr *init_attr,
1219 				struct ib_udata *udata)
1220 {
1221 	struct mlx5_ib_dev *dev;
1222 	struct mlx5_ib_qp *qp;
1223 	u16 xrcdn = 0;
1224 	int err;
1225 
1226 	if (pd) {
1227 		dev = to_mdev(pd->device);
1228 	} else {
1229 		/* being cautious here */
1230 		if (init_attr->qp_type != IB_QPT_XRC_TGT &&
1231 		    init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
1232 			pr_warn("%s: no PD for transport %s\n", __func__,
1233 				ib_qp_type_str(init_attr->qp_type));
1234 			return ERR_PTR(-EINVAL);
1235 		}
1236 		dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
1237 	}
1238 
1239 	switch (init_attr->qp_type) {
1240 	case IB_QPT_XRC_TGT:
1241 	case IB_QPT_XRC_INI:
1242 		if (!MLX5_CAP_GEN(dev->mdev, xrc)) {
1243 			mlx5_ib_dbg(dev, "XRC not supported\n");
1244 			return ERR_PTR(-ENOSYS);
1245 		}
1246 		init_attr->recv_cq = NULL;
1247 		if (init_attr->qp_type == IB_QPT_XRC_TGT) {
1248 			xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1249 			init_attr->send_cq = NULL;
1250 		}
1251 
1252 		/* fall through */
1253 	case IB_QPT_RC:
1254 	case IB_QPT_UC:
1255 	case IB_QPT_UD:
1256 	case IB_QPT_SMI:
1257 	case IB_QPT_GSI:
1258 	case MLX5_IB_QPT_REG_UMR:
1259 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1260 		if (!qp)
1261 			return ERR_PTR(-ENOMEM);
1262 
1263 		err = create_qp_common(dev, pd, init_attr, udata, qp);
1264 		if (err) {
1265 			mlx5_ib_dbg(dev, "create_qp_common failed\n");
1266 			kfree(qp);
1267 			return ERR_PTR(err);
1268 		}
1269 
1270 		if (is_qp0(init_attr->qp_type))
1271 			qp->ibqp.qp_num = 0;
1272 		else if (is_qp1(init_attr->qp_type))
1273 			qp->ibqp.qp_num = 1;
1274 		else
1275 			qp->ibqp.qp_num = qp->mqp.qpn;
1276 
1277 		mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
1278 			    qp->ibqp.qp_num, qp->mqp.qpn, to_mcq(init_attr->recv_cq)->mcq.cqn,
1279 			    to_mcq(init_attr->send_cq)->mcq.cqn);
1280 
1281 		qp->xrcdn = xrcdn;
1282 
1283 		break;
1284 
1285 	case IB_QPT_RAW_IPV6:
1286 	case IB_QPT_RAW_ETHERTYPE:
1287 	case IB_QPT_RAW_PACKET:
1288 	case IB_QPT_MAX:
1289 	default:
1290 		mlx5_ib_dbg(dev, "unsupported qp type %d\n",
1291 			    init_attr->qp_type);
1292 		/* Don't support raw QPs */
1293 		return ERR_PTR(-EINVAL);
1294 	}
1295 
1296 	return &qp->ibqp;
1297 }
1298 
1299 int mlx5_ib_destroy_qp(struct ib_qp *qp)
1300 {
1301 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
1302 	struct mlx5_ib_qp *mqp = to_mqp(qp);
1303 
1304 	destroy_qp_common(dev, mqp);
1305 
1306 	kfree(mqp);
1307 
1308 	return 0;
1309 }
1310 
1311 static __be32 to_mlx5_access_flags(struct mlx5_ib_qp *qp, const struct ib_qp_attr *attr,
1312 				   int attr_mask)
1313 {
1314 	u32 hw_access_flags = 0;
1315 	u8 dest_rd_atomic;
1316 	u32 access_flags;
1317 
1318 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1319 		dest_rd_atomic = attr->max_dest_rd_atomic;
1320 	else
1321 		dest_rd_atomic = qp->resp_depth;
1322 
1323 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1324 		access_flags = attr->qp_access_flags;
1325 	else
1326 		access_flags = qp->atomic_rd_en;
1327 
1328 	if (!dest_rd_atomic)
1329 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1330 
1331 	if (access_flags & IB_ACCESS_REMOTE_READ)
1332 		hw_access_flags |= MLX5_QP_BIT_RRE;
1333 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1334 		hw_access_flags |= (MLX5_QP_BIT_RAE | MLX5_ATOMIC_MODE_CX);
1335 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1336 		hw_access_flags |= MLX5_QP_BIT_RWE;
1337 
1338 	return cpu_to_be32(hw_access_flags);
1339 }
1340 
1341 enum {
1342 	MLX5_PATH_FLAG_FL	= 1 << 0,
1343 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
1344 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
1345 };
1346 
1347 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
1348 {
1349 	if (rate == IB_RATE_PORT_CURRENT) {
1350 		return 0;
1351 	} else if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_300_GBPS) {
1352 		return -EINVAL;
1353 	} else {
1354 		while (rate != IB_RATE_2_5_GBPS &&
1355 		       !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
1356 			 MLX5_CAP_GEN(dev->mdev, stat_rate_support)))
1357 			--rate;
1358 	}
1359 
1360 	return rate + MLX5_STAT_RATE_OFFSET;
1361 }
1362 
1363 static int mlx5_set_path(struct mlx5_ib_dev *dev, const struct ib_ah_attr *ah,
1364 			 struct mlx5_qp_path *path, u8 port, int attr_mask,
1365 			 u32 path_flags, const struct ib_qp_attr *attr)
1366 {
1367 	int err;
1368 
1369 	path->fl = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
1370 	path->free_ar = (path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x80 : 0;
1371 
1372 	if (attr_mask & IB_QP_PKEY_INDEX)
1373 		path->pkey_index = attr->pkey_index;
1374 
1375 	path->grh_mlid	= ah->src_path_bits & 0x7f;
1376 	path->rlid	= cpu_to_be16(ah->dlid);
1377 
1378 	if (ah->ah_flags & IB_AH_GRH) {
1379 		if (ah->grh.sgid_index >=
1380 		    dev->mdev->port_caps[port - 1].gid_table_len) {
1381 			pr_err("sgid_index (%u) too large. max is %d\n",
1382 			       ah->grh.sgid_index,
1383 			       dev->mdev->port_caps[port - 1].gid_table_len);
1384 			return -EINVAL;
1385 		}
1386 		path->grh_mlid |= 1 << 7;
1387 		path->mgid_index = ah->grh.sgid_index;
1388 		path->hop_limit  = ah->grh.hop_limit;
1389 		path->tclass_flowlabel =
1390 			cpu_to_be32((ah->grh.traffic_class << 20) |
1391 				    (ah->grh.flow_label));
1392 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1393 	}
1394 
1395 	err = ib_rate_to_mlx5(dev, ah->static_rate);
1396 	if (err < 0)
1397 		return err;
1398 	path->static_rate = err;
1399 	path->port = port;
1400 
1401 	if (attr_mask & IB_QP_TIMEOUT)
1402 		path->ackto_lt = attr->timeout << 3;
1403 
1404 	path->sl = ah->sl & 0xf;
1405 
1406 	return 0;
1407 }
1408 
1409 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
1410 	[MLX5_QP_STATE_INIT] = {
1411 		[MLX5_QP_STATE_INIT] = {
1412 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
1413 					  MLX5_QP_OPTPAR_RAE		|
1414 					  MLX5_QP_OPTPAR_RWE		|
1415 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
1416 					  MLX5_QP_OPTPAR_PRI_PORT,
1417 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
1418 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
1419 					  MLX5_QP_OPTPAR_PRI_PORT,
1420 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
1421 					  MLX5_QP_OPTPAR_Q_KEY		|
1422 					  MLX5_QP_OPTPAR_PRI_PORT,
1423 		},
1424 		[MLX5_QP_STATE_RTR] = {
1425 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
1426 					  MLX5_QP_OPTPAR_RRE            |
1427 					  MLX5_QP_OPTPAR_RAE            |
1428 					  MLX5_QP_OPTPAR_RWE            |
1429 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1430 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
1431 					  MLX5_QP_OPTPAR_RWE            |
1432 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1433 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
1434 					  MLX5_QP_OPTPAR_Q_KEY,
1435 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
1436 					   MLX5_QP_OPTPAR_Q_KEY,
1437 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
1438 					  MLX5_QP_OPTPAR_RRE            |
1439 					  MLX5_QP_OPTPAR_RAE            |
1440 					  MLX5_QP_OPTPAR_RWE            |
1441 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1442 		},
1443 	},
1444 	[MLX5_QP_STATE_RTR] = {
1445 		[MLX5_QP_STATE_RTS] = {
1446 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
1447 					  MLX5_QP_OPTPAR_RRE		|
1448 					  MLX5_QP_OPTPAR_RAE		|
1449 					  MLX5_QP_OPTPAR_RWE		|
1450 					  MLX5_QP_OPTPAR_PM_STATE	|
1451 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
1452 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
1453 					  MLX5_QP_OPTPAR_RWE		|
1454 					  MLX5_QP_OPTPAR_PM_STATE,
1455 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
1456 		},
1457 	},
1458 	[MLX5_QP_STATE_RTS] = {
1459 		[MLX5_QP_STATE_RTS] = {
1460 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
1461 					  MLX5_QP_OPTPAR_RAE		|
1462 					  MLX5_QP_OPTPAR_RWE		|
1463 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
1464 					  MLX5_QP_OPTPAR_PM_STATE	|
1465 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
1466 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
1467 					  MLX5_QP_OPTPAR_PM_STATE	|
1468 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
1469 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
1470 					  MLX5_QP_OPTPAR_SRQN		|
1471 					  MLX5_QP_OPTPAR_CQN_RCV,
1472 		},
1473 	},
1474 	[MLX5_QP_STATE_SQER] = {
1475 		[MLX5_QP_STATE_RTS] = {
1476 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
1477 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
1478 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
1479 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
1480 					   MLX5_QP_OPTPAR_RWE		|
1481 					   MLX5_QP_OPTPAR_RAE		|
1482 					   MLX5_QP_OPTPAR_RRE,
1483 		},
1484 	},
1485 };
1486 
1487 static int ib_nr_to_mlx5_nr(int ib_mask)
1488 {
1489 	switch (ib_mask) {
1490 	case IB_QP_STATE:
1491 		return 0;
1492 	case IB_QP_CUR_STATE:
1493 		return 0;
1494 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
1495 		return 0;
1496 	case IB_QP_ACCESS_FLAGS:
1497 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
1498 			MLX5_QP_OPTPAR_RAE;
1499 	case IB_QP_PKEY_INDEX:
1500 		return MLX5_QP_OPTPAR_PKEY_INDEX;
1501 	case IB_QP_PORT:
1502 		return MLX5_QP_OPTPAR_PRI_PORT;
1503 	case IB_QP_QKEY:
1504 		return MLX5_QP_OPTPAR_Q_KEY;
1505 	case IB_QP_AV:
1506 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
1507 			MLX5_QP_OPTPAR_PRI_PORT;
1508 	case IB_QP_PATH_MTU:
1509 		return 0;
1510 	case IB_QP_TIMEOUT:
1511 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
1512 	case IB_QP_RETRY_CNT:
1513 		return MLX5_QP_OPTPAR_RETRY_COUNT;
1514 	case IB_QP_RNR_RETRY:
1515 		return MLX5_QP_OPTPAR_RNR_RETRY;
1516 	case IB_QP_RQ_PSN:
1517 		return 0;
1518 	case IB_QP_MAX_QP_RD_ATOMIC:
1519 		return MLX5_QP_OPTPAR_SRA_MAX;
1520 	case IB_QP_ALT_PATH:
1521 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
1522 	case IB_QP_MIN_RNR_TIMER:
1523 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
1524 	case IB_QP_SQ_PSN:
1525 		return 0;
1526 	case IB_QP_MAX_DEST_RD_ATOMIC:
1527 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
1528 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
1529 	case IB_QP_PATH_MIG_STATE:
1530 		return MLX5_QP_OPTPAR_PM_STATE;
1531 	case IB_QP_CAP:
1532 		return 0;
1533 	case IB_QP_DEST_QPN:
1534 		return 0;
1535 	}
1536 	return 0;
1537 }
1538 
1539 static int ib_mask_to_mlx5_opt(int ib_mask)
1540 {
1541 	int result = 0;
1542 	int i;
1543 
1544 	for (i = 0; i < 8 * sizeof(int); i++) {
1545 		if ((1 << i) & ib_mask)
1546 			result |= ib_nr_to_mlx5_nr(1 << i);
1547 	}
1548 
1549 	return result;
1550 }
1551 
1552 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
1553 			       const struct ib_qp_attr *attr, int attr_mask,
1554 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1555 {
1556 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1557 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
1558 	struct mlx5_ib_cq *send_cq, *recv_cq;
1559 	struct mlx5_qp_context *context;
1560 	struct mlx5_modify_qp_mbox_in *in;
1561 	struct mlx5_ib_pd *pd;
1562 	enum mlx5_qp_state mlx5_cur, mlx5_new;
1563 	enum mlx5_qp_optpar optpar;
1564 	int sqd_event;
1565 	int mlx5_st;
1566 	int err;
1567 
1568 	in = kzalloc(sizeof(*in), GFP_KERNEL);
1569 	if (!in)
1570 		return -ENOMEM;
1571 
1572 	context = &in->ctx;
1573 	err = to_mlx5_st(ibqp->qp_type);
1574 	if (err < 0)
1575 		goto out;
1576 
1577 	context->flags = cpu_to_be32(err << 16);
1578 
1579 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
1580 		context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
1581 	} else {
1582 		switch (attr->path_mig_state) {
1583 		case IB_MIG_MIGRATED:
1584 			context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
1585 			break;
1586 		case IB_MIG_REARM:
1587 			context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
1588 			break;
1589 		case IB_MIG_ARMED:
1590 			context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
1591 			break;
1592 		}
1593 	}
1594 
1595 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI) {
1596 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
1597 	} else if (ibqp->qp_type == IB_QPT_UD ||
1598 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
1599 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1600 	} else if (attr_mask & IB_QP_PATH_MTU) {
1601 		if (attr->path_mtu < IB_MTU_256 ||
1602 		    attr->path_mtu > IB_MTU_4096) {
1603 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
1604 			err = -EINVAL;
1605 			goto out;
1606 		}
1607 		context->mtu_msgmax = (attr->path_mtu << 5) |
1608 				      (u8)MLX5_CAP_GEN(dev->mdev, log_max_msg);
1609 	}
1610 
1611 	if (attr_mask & IB_QP_DEST_QPN)
1612 		context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
1613 
1614 	if (attr_mask & IB_QP_PKEY_INDEX)
1615 		context->pri_path.pkey_index = attr->pkey_index;
1616 
1617 	/* todo implement counter_index functionality */
1618 
1619 	if (is_sqp(ibqp->qp_type))
1620 		context->pri_path.port = qp->port;
1621 
1622 	if (attr_mask & IB_QP_PORT)
1623 		context->pri_path.port = attr->port_num;
1624 
1625 	if (attr_mask & IB_QP_AV) {
1626 		err = mlx5_set_path(dev, &attr->ah_attr, &context->pri_path,
1627 				    attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
1628 				    attr_mask, 0, attr);
1629 		if (err)
1630 			goto out;
1631 	}
1632 
1633 	if (attr_mask & IB_QP_TIMEOUT)
1634 		context->pri_path.ackto_lt |= attr->timeout << 3;
1635 
1636 	if (attr_mask & IB_QP_ALT_PATH) {
1637 		err = mlx5_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
1638 				    attr->alt_port_num, attr_mask, 0, attr);
1639 		if (err)
1640 			goto out;
1641 	}
1642 
1643 	pd = get_pd(qp);
1644 	get_cqs(qp, &send_cq, &recv_cq);
1645 
1646 	context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
1647 	context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
1648 	context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
1649 	context->params1  = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
1650 
1651 	if (attr_mask & IB_QP_RNR_RETRY)
1652 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1653 
1654 	if (attr_mask & IB_QP_RETRY_CNT)
1655 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1656 
1657 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1658 		if (attr->max_rd_atomic)
1659 			context->params1 |=
1660 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1661 	}
1662 
1663 	if (attr_mask & IB_QP_SQ_PSN)
1664 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1665 
1666 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1667 		if (attr->max_dest_rd_atomic)
1668 			context->params2 |=
1669 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1670 	}
1671 
1672 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
1673 		context->params2 |= to_mlx5_access_flags(qp, attr, attr_mask);
1674 
1675 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
1676 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1677 
1678 	if (attr_mask & IB_QP_RQ_PSN)
1679 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1680 
1681 	if (attr_mask & IB_QP_QKEY)
1682 		context->qkey = cpu_to_be32(attr->qkey);
1683 
1684 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1685 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1686 
1687 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1688 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1689 		sqd_event = 1;
1690 	else
1691 		sqd_event = 0;
1692 
1693 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1694 		context->sq_crq_size |= cpu_to_be16(1 << 4);
1695 
1696 
1697 	mlx5_cur = to_mlx5_state(cur_state);
1698 	mlx5_new = to_mlx5_state(new_state);
1699 	mlx5_st = to_mlx5_st(ibqp->qp_type);
1700 	if (mlx5_st < 0)
1701 		goto out;
1702 
1703 	/* If moving to a reset or error state, we must disable page faults on
1704 	 * this QP and flush all current page faults. Otherwise a stale page
1705 	 * fault may attempt to work on this QP after it is reset and moved
1706 	 * again to RTS, and may cause the driver and the device to get out of
1707 	 * sync. */
1708 	if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1709 	    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1710 		mlx5_ib_qp_disable_pagefaults(qp);
1711 
1712 	optpar = ib_mask_to_mlx5_opt(attr_mask);
1713 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
1714 	in->optparam = cpu_to_be32(optpar);
1715 	err = mlx5_core_qp_modify(dev->mdev, to_mlx5_state(cur_state),
1716 				  to_mlx5_state(new_state), in, sqd_event,
1717 				  &qp->mqp);
1718 	if (err)
1719 		goto out;
1720 
1721 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1722 		mlx5_ib_qp_enable_pagefaults(qp);
1723 
1724 	qp->state = new_state;
1725 
1726 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1727 		qp->atomic_rd_en = attr->qp_access_flags;
1728 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1729 		qp->resp_depth = attr->max_dest_rd_atomic;
1730 	if (attr_mask & IB_QP_PORT)
1731 		qp->port = attr->port_num;
1732 	if (attr_mask & IB_QP_ALT_PATH)
1733 		qp->alt_port = attr->alt_port_num;
1734 
1735 	/*
1736 	 * If we moved a kernel QP to RESET, clean up all old CQ
1737 	 * entries and reinitialize the QP.
1738 	 */
1739 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1740 		mlx5_ib_cq_clean(recv_cq, qp->mqp.qpn,
1741 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
1742 		if (send_cq != recv_cq)
1743 			mlx5_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1744 
1745 		qp->rq.head = 0;
1746 		qp->rq.tail = 0;
1747 		qp->sq.head = 0;
1748 		qp->sq.tail = 0;
1749 		qp->sq.cur_post = 0;
1750 		qp->sq.last_poll = 0;
1751 		qp->db.db[MLX5_RCV_DBR] = 0;
1752 		qp->db.db[MLX5_SND_DBR] = 0;
1753 	}
1754 
1755 out:
1756 	kfree(in);
1757 	return err;
1758 }
1759 
1760 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1761 		      int attr_mask, struct ib_udata *udata)
1762 {
1763 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1764 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
1765 	enum ib_qp_state cur_state, new_state;
1766 	int err = -EINVAL;
1767 	int port;
1768 
1769 	mutex_lock(&qp->mutex);
1770 
1771 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1772 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1773 
1774 	if (ibqp->qp_type != MLX5_IB_QPT_REG_UMR &&
1775 	    !ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask,
1776 				IB_LINK_LAYER_UNSPECIFIED))
1777 		goto out;
1778 
1779 	if ((attr_mask & IB_QP_PORT) &&
1780 	    (attr->port_num == 0 ||
1781 	     attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports)))
1782 		goto out;
1783 
1784 	if (attr_mask & IB_QP_PKEY_INDEX) {
1785 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1786 		if (attr->pkey_index >=
1787 		    dev->mdev->port_caps[port - 1].pkey_table_len)
1788 			goto out;
1789 	}
1790 
1791 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1792 	    attr->max_rd_atomic >
1793 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp)))
1794 		goto out;
1795 
1796 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1797 	    attr->max_dest_rd_atomic >
1798 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp)))
1799 		goto out;
1800 
1801 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1802 		err = 0;
1803 		goto out;
1804 	}
1805 
1806 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1807 
1808 out:
1809 	mutex_unlock(&qp->mutex);
1810 	return err;
1811 }
1812 
1813 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1814 {
1815 	struct mlx5_ib_cq *cq;
1816 	unsigned cur;
1817 
1818 	cur = wq->head - wq->tail;
1819 	if (likely(cur + nreq < wq->max_post))
1820 		return 0;
1821 
1822 	cq = to_mcq(ib_cq);
1823 	spin_lock(&cq->lock);
1824 	cur = wq->head - wq->tail;
1825 	spin_unlock(&cq->lock);
1826 
1827 	return cur + nreq >= wq->max_post;
1828 }
1829 
1830 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
1831 					  u64 remote_addr, u32 rkey)
1832 {
1833 	rseg->raddr    = cpu_to_be64(remote_addr);
1834 	rseg->rkey     = cpu_to_be32(rkey);
1835 	rseg->reserved = 0;
1836 }
1837 
1838 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
1839 			     struct ib_send_wr *wr)
1840 {
1841 	memcpy(&dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof(struct mlx5_av));
1842 	dseg->av.dqp_dct = cpu_to_be32(wr->wr.ud.remote_qpn | MLX5_EXTENDED_UD_AV);
1843 	dseg->av.key.qkey.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1844 }
1845 
1846 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
1847 {
1848 	dseg->byte_count = cpu_to_be32(sg->length);
1849 	dseg->lkey       = cpu_to_be32(sg->lkey);
1850 	dseg->addr       = cpu_to_be64(sg->addr);
1851 }
1852 
1853 static __be16 get_klm_octo(int npages)
1854 {
1855 	return cpu_to_be16(ALIGN(npages, 8) / 2);
1856 }
1857 
1858 static __be64 frwr_mkey_mask(void)
1859 {
1860 	u64 result;
1861 
1862 	result = MLX5_MKEY_MASK_LEN		|
1863 		MLX5_MKEY_MASK_PAGE_SIZE	|
1864 		MLX5_MKEY_MASK_START_ADDR	|
1865 		MLX5_MKEY_MASK_EN_RINVAL	|
1866 		MLX5_MKEY_MASK_KEY		|
1867 		MLX5_MKEY_MASK_LR		|
1868 		MLX5_MKEY_MASK_LW		|
1869 		MLX5_MKEY_MASK_RR		|
1870 		MLX5_MKEY_MASK_RW		|
1871 		MLX5_MKEY_MASK_A		|
1872 		MLX5_MKEY_MASK_SMALL_FENCE	|
1873 		MLX5_MKEY_MASK_FREE;
1874 
1875 	return cpu_to_be64(result);
1876 }
1877 
1878 static __be64 sig_mkey_mask(void)
1879 {
1880 	u64 result;
1881 
1882 	result = MLX5_MKEY_MASK_LEN		|
1883 		MLX5_MKEY_MASK_PAGE_SIZE	|
1884 		MLX5_MKEY_MASK_START_ADDR	|
1885 		MLX5_MKEY_MASK_EN_SIGERR	|
1886 		MLX5_MKEY_MASK_EN_RINVAL	|
1887 		MLX5_MKEY_MASK_KEY		|
1888 		MLX5_MKEY_MASK_LR		|
1889 		MLX5_MKEY_MASK_LW		|
1890 		MLX5_MKEY_MASK_RR		|
1891 		MLX5_MKEY_MASK_RW		|
1892 		MLX5_MKEY_MASK_SMALL_FENCE	|
1893 		MLX5_MKEY_MASK_FREE		|
1894 		MLX5_MKEY_MASK_BSF_EN;
1895 
1896 	return cpu_to_be64(result);
1897 }
1898 
1899 static void set_frwr_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
1900 				 struct ib_send_wr *wr, int li)
1901 {
1902 	memset(umr, 0, sizeof(*umr));
1903 
1904 	if (li) {
1905 		umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
1906 		umr->flags = 1 << 7;
1907 		return;
1908 	}
1909 
1910 	umr->flags = (1 << 5); /* fail if not free */
1911 	umr->klm_octowords = get_klm_octo(wr->wr.fast_reg.page_list_len);
1912 	umr->mkey_mask = frwr_mkey_mask();
1913 }
1914 
1915 static __be64 get_umr_reg_mr_mask(void)
1916 {
1917 	u64 result;
1918 
1919 	result = MLX5_MKEY_MASK_LEN		|
1920 		 MLX5_MKEY_MASK_PAGE_SIZE	|
1921 		 MLX5_MKEY_MASK_START_ADDR	|
1922 		 MLX5_MKEY_MASK_PD		|
1923 		 MLX5_MKEY_MASK_LR		|
1924 		 MLX5_MKEY_MASK_LW		|
1925 		 MLX5_MKEY_MASK_KEY		|
1926 		 MLX5_MKEY_MASK_RR		|
1927 		 MLX5_MKEY_MASK_RW		|
1928 		 MLX5_MKEY_MASK_A		|
1929 		 MLX5_MKEY_MASK_FREE;
1930 
1931 	return cpu_to_be64(result);
1932 }
1933 
1934 static __be64 get_umr_unreg_mr_mask(void)
1935 {
1936 	u64 result;
1937 
1938 	result = MLX5_MKEY_MASK_FREE;
1939 
1940 	return cpu_to_be64(result);
1941 }
1942 
1943 static __be64 get_umr_update_mtt_mask(void)
1944 {
1945 	u64 result;
1946 
1947 	result = MLX5_MKEY_MASK_FREE;
1948 
1949 	return cpu_to_be64(result);
1950 }
1951 
1952 static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
1953 				struct ib_send_wr *wr)
1954 {
1955 	struct mlx5_umr_wr *umrwr = (struct mlx5_umr_wr *)&wr->wr.fast_reg;
1956 
1957 	memset(umr, 0, sizeof(*umr));
1958 
1959 	if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
1960 		umr->flags = MLX5_UMR_CHECK_FREE; /* fail if free */
1961 	else
1962 		umr->flags = MLX5_UMR_CHECK_NOT_FREE; /* fail if not free */
1963 
1964 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UNREG)) {
1965 		umr->klm_octowords = get_klm_octo(umrwr->npages);
1966 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT) {
1967 			umr->mkey_mask = get_umr_update_mtt_mask();
1968 			umr->bsf_octowords = get_klm_octo(umrwr->target.offset);
1969 			umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
1970 		} else {
1971 			umr->mkey_mask = get_umr_reg_mr_mask();
1972 		}
1973 	} else {
1974 		umr->mkey_mask = get_umr_unreg_mr_mask();
1975 	}
1976 
1977 	if (!wr->num_sge)
1978 		umr->flags |= MLX5_UMR_INLINE;
1979 }
1980 
1981 static u8 get_umr_flags(int acc)
1982 {
1983 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
1984 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
1985 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
1986 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
1987 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
1988 }
1989 
1990 static void set_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr,
1991 			     int li, int *writ)
1992 {
1993 	memset(seg, 0, sizeof(*seg));
1994 	if (li) {
1995 		seg->status = MLX5_MKEY_STATUS_FREE;
1996 		return;
1997 	}
1998 
1999 	seg->flags = get_umr_flags(wr->wr.fast_reg.access_flags) |
2000 		     MLX5_ACCESS_MODE_MTT;
2001 	*writ = seg->flags & (MLX5_PERM_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE);
2002 	seg->qpn_mkey7_0 = cpu_to_be32((wr->wr.fast_reg.rkey & 0xff) | 0xffffff00);
2003 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
2004 	seg->start_addr = cpu_to_be64(wr->wr.fast_reg.iova_start);
2005 	seg->len = cpu_to_be64(wr->wr.fast_reg.length);
2006 	seg->xlt_oct_size = cpu_to_be32((wr->wr.fast_reg.page_list_len + 1) / 2);
2007 	seg->log2_page_size = wr->wr.fast_reg.page_shift;
2008 }
2009 
2010 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr)
2011 {
2012 	struct mlx5_umr_wr *umrwr = (struct mlx5_umr_wr *)&wr->wr.fast_reg;
2013 
2014 	memset(seg, 0, sizeof(*seg));
2015 	if (wr->send_flags & MLX5_IB_SEND_UMR_UNREG) {
2016 		seg->status = MLX5_MKEY_STATUS_FREE;
2017 		return;
2018 	}
2019 
2020 	seg->flags = convert_access(umrwr->access_flags);
2021 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT)) {
2022 		seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
2023 		seg->start_addr = cpu_to_be64(umrwr->target.virt_addr);
2024 	}
2025 	seg->len = cpu_to_be64(umrwr->length);
2026 	seg->log2_page_size = umrwr->page_shift;
2027 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
2028 				       mlx5_mkey_variant(umrwr->mkey));
2029 }
2030 
2031 static void set_frwr_pages(struct mlx5_wqe_data_seg *dseg,
2032 			   struct ib_send_wr *wr,
2033 			   struct mlx5_core_dev *mdev,
2034 			   struct mlx5_ib_pd *pd,
2035 			   int writ)
2036 {
2037 	struct mlx5_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
2038 	u64 *page_list = wr->wr.fast_reg.page_list->page_list;
2039 	u64 perm = MLX5_EN_RD | (writ ? MLX5_EN_WR : 0);
2040 	int i;
2041 
2042 	for (i = 0; i < wr->wr.fast_reg.page_list_len; i++)
2043 		mfrpl->mapped_page_list[i] = cpu_to_be64(page_list[i] | perm);
2044 	dseg->addr = cpu_to_be64(mfrpl->map);
2045 	dseg->byte_count = cpu_to_be32(ALIGN(sizeof(u64) * wr->wr.fast_reg.page_list_len, 64));
2046 	dseg->lkey = cpu_to_be32(pd->ibpd.local_dma_lkey);
2047 }
2048 
2049 static __be32 send_ieth(struct ib_send_wr *wr)
2050 {
2051 	switch (wr->opcode) {
2052 	case IB_WR_SEND_WITH_IMM:
2053 	case IB_WR_RDMA_WRITE_WITH_IMM:
2054 		return wr->ex.imm_data;
2055 
2056 	case IB_WR_SEND_WITH_INV:
2057 		return cpu_to_be32(wr->ex.invalidate_rkey);
2058 
2059 	default:
2060 		return 0;
2061 	}
2062 }
2063 
2064 static u8 calc_sig(void *wqe, int size)
2065 {
2066 	u8 *p = wqe;
2067 	u8 res = 0;
2068 	int i;
2069 
2070 	for (i = 0; i < size; i++)
2071 		res ^= p[i];
2072 
2073 	return ~res;
2074 }
2075 
2076 static u8 wq_sig(void *wqe)
2077 {
2078 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
2079 }
2080 
2081 static int set_data_inl_seg(struct mlx5_ib_qp *qp, struct ib_send_wr *wr,
2082 			    void *wqe, int *sz)
2083 {
2084 	struct mlx5_wqe_inline_seg *seg;
2085 	void *qend = qp->sq.qend;
2086 	void *addr;
2087 	int inl = 0;
2088 	int copy;
2089 	int len;
2090 	int i;
2091 
2092 	seg = wqe;
2093 	wqe += sizeof(*seg);
2094 	for (i = 0; i < wr->num_sge; i++) {
2095 		addr = (void *)(unsigned long)(wr->sg_list[i].addr);
2096 		len  = wr->sg_list[i].length;
2097 		inl += len;
2098 
2099 		if (unlikely(inl > qp->max_inline_data))
2100 			return -ENOMEM;
2101 
2102 		if (unlikely(wqe + len > qend)) {
2103 			copy = qend - wqe;
2104 			memcpy(wqe, addr, copy);
2105 			addr += copy;
2106 			len -= copy;
2107 			wqe = mlx5_get_send_wqe(qp, 0);
2108 		}
2109 		memcpy(wqe, addr, len);
2110 		wqe += len;
2111 	}
2112 
2113 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
2114 
2115 	*sz = ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
2116 
2117 	return 0;
2118 }
2119 
2120 static u16 prot_field_size(enum ib_signature_type type)
2121 {
2122 	switch (type) {
2123 	case IB_SIG_TYPE_T10_DIF:
2124 		return MLX5_DIF_SIZE;
2125 	default:
2126 		return 0;
2127 	}
2128 }
2129 
2130 static u8 bs_selector(int block_size)
2131 {
2132 	switch (block_size) {
2133 	case 512:	    return 0x1;
2134 	case 520:	    return 0x2;
2135 	case 4096:	    return 0x3;
2136 	case 4160:	    return 0x4;
2137 	case 1073741824:    return 0x5;
2138 	default:	    return 0;
2139 	}
2140 }
2141 
2142 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
2143 			      struct mlx5_bsf_inl *inl)
2144 {
2145 	/* Valid inline section and allow BSF refresh */
2146 	inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
2147 				       MLX5_BSF_REFRESH_DIF);
2148 	inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
2149 	inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
2150 	/* repeating block */
2151 	inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
2152 	inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
2153 			MLX5_DIF_CRC : MLX5_DIF_IPCS;
2154 
2155 	if (domain->sig.dif.ref_remap)
2156 		inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
2157 
2158 	if (domain->sig.dif.app_escape) {
2159 		if (domain->sig.dif.ref_escape)
2160 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
2161 		else
2162 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
2163 	}
2164 
2165 	inl->dif_app_bitmask_check =
2166 		cpu_to_be16(domain->sig.dif.apptag_check_mask);
2167 }
2168 
2169 static int mlx5_set_bsf(struct ib_mr *sig_mr,
2170 			struct ib_sig_attrs *sig_attrs,
2171 			struct mlx5_bsf *bsf, u32 data_size)
2172 {
2173 	struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
2174 	struct mlx5_bsf_basic *basic = &bsf->basic;
2175 	struct ib_sig_domain *mem = &sig_attrs->mem;
2176 	struct ib_sig_domain *wire = &sig_attrs->wire;
2177 
2178 	memset(bsf, 0, sizeof(*bsf));
2179 
2180 	/* Basic + Extended + Inline */
2181 	basic->bsf_size_sbs = 1 << 7;
2182 	/* Input domain check byte mask */
2183 	basic->check_byte_mask = sig_attrs->check_mask;
2184 	basic->raw_data_size = cpu_to_be32(data_size);
2185 
2186 	/* Memory domain */
2187 	switch (sig_attrs->mem.sig_type) {
2188 	case IB_SIG_TYPE_NONE:
2189 		break;
2190 	case IB_SIG_TYPE_T10_DIF:
2191 		basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
2192 		basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
2193 		mlx5_fill_inl_bsf(mem, &bsf->m_inl);
2194 		break;
2195 	default:
2196 		return -EINVAL;
2197 	}
2198 
2199 	/* Wire domain */
2200 	switch (sig_attrs->wire.sig_type) {
2201 	case IB_SIG_TYPE_NONE:
2202 		break;
2203 	case IB_SIG_TYPE_T10_DIF:
2204 		if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
2205 		    mem->sig_type == wire->sig_type) {
2206 			/* Same block structure */
2207 			basic->bsf_size_sbs |= 1 << 4;
2208 			if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
2209 				basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
2210 			if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
2211 				basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
2212 			if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
2213 				basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
2214 		} else
2215 			basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
2216 
2217 		basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
2218 		mlx5_fill_inl_bsf(wire, &bsf->w_inl);
2219 		break;
2220 	default:
2221 		return -EINVAL;
2222 	}
2223 
2224 	return 0;
2225 }
2226 
2227 static int set_sig_data_segment(struct ib_send_wr *wr, struct mlx5_ib_qp *qp,
2228 				void **seg, int *size)
2229 {
2230 	struct ib_sig_attrs *sig_attrs = wr->wr.sig_handover.sig_attrs;
2231 	struct ib_mr *sig_mr = wr->wr.sig_handover.sig_mr;
2232 	struct mlx5_bsf *bsf;
2233 	u32 data_len = wr->sg_list->length;
2234 	u32 data_key = wr->sg_list->lkey;
2235 	u64 data_va = wr->sg_list->addr;
2236 	int ret;
2237 	int wqe_size;
2238 
2239 	if (!wr->wr.sig_handover.prot ||
2240 	    (data_key == wr->wr.sig_handover.prot->lkey &&
2241 	     data_va == wr->wr.sig_handover.prot->addr &&
2242 	     data_len == wr->wr.sig_handover.prot->length)) {
2243 		/**
2244 		 * Source domain doesn't contain signature information
2245 		 * or data and protection are interleaved in memory.
2246 		 * So need construct:
2247 		 *                  ------------------
2248 		 *                 |     data_klm     |
2249 		 *                  ------------------
2250 		 *                 |       BSF        |
2251 		 *                  ------------------
2252 		 **/
2253 		struct mlx5_klm *data_klm = *seg;
2254 
2255 		data_klm->bcount = cpu_to_be32(data_len);
2256 		data_klm->key = cpu_to_be32(data_key);
2257 		data_klm->va = cpu_to_be64(data_va);
2258 		wqe_size = ALIGN(sizeof(*data_klm), 64);
2259 	} else {
2260 		/**
2261 		 * Source domain contains signature information
2262 		 * So need construct a strided block format:
2263 		 *               ---------------------------
2264 		 *              |     stride_block_ctrl     |
2265 		 *               ---------------------------
2266 		 *              |          data_klm         |
2267 		 *               ---------------------------
2268 		 *              |          prot_klm         |
2269 		 *               ---------------------------
2270 		 *              |             BSF           |
2271 		 *               ---------------------------
2272 		 **/
2273 		struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
2274 		struct mlx5_stride_block_entry *data_sentry;
2275 		struct mlx5_stride_block_entry *prot_sentry;
2276 		u32 prot_key = wr->wr.sig_handover.prot->lkey;
2277 		u64 prot_va = wr->wr.sig_handover.prot->addr;
2278 		u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
2279 		int prot_size;
2280 
2281 		sblock_ctrl = *seg;
2282 		data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
2283 		prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
2284 
2285 		prot_size = prot_field_size(sig_attrs->mem.sig_type);
2286 		if (!prot_size) {
2287 			pr_err("Bad block size given: %u\n", block_size);
2288 			return -EINVAL;
2289 		}
2290 		sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
2291 							    prot_size);
2292 		sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
2293 		sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
2294 		sblock_ctrl->num_entries = cpu_to_be16(2);
2295 
2296 		data_sentry->bcount = cpu_to_be16(block_size);
2297 		data_sentry->key = cpu_to_be32(data_key);
2298 		data_sentry->va = cpu_to_be64(data_va);
2299 		data_sentry->stride = cpu_to_be16(block_size);
2300 
2301 		prot_sentry->bcount = cpu_to_be16(prot_size);
2302 		prot_sentry->key = cpu_to_be32(prot_key);
2303 		prot_sentry->va = cpu_to_be64(prot_va);
2304 		prot_sentry->stride = cpu_to_be16(prot_size);
2305 
2306 		wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
2307 				 sizeof(*prot_sentry), 64);
2308 	}
2309 
2310 	*seg += wqe_size;
2311 	*size += wqe_size / 16;
2312 	if (unlikely((*seg == qp->sq.qend)))
2313 		*seg = mlx5_get_send_wqe(qp, 0);
2314 
2315 	bsf = *seg;
2316 	ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
2317 	if (ret)
2318 		return -EINVAL;
2319 
2320 	*seg += sizeof(*bsf);
2321 	*size += sizeof(*bsf) / 16;
2322 	if (unlikely((*seg == qp->sq.qend)))
2323 		*seg = mlx5_get_send_wqe(qp, 0);
2324 
2325 	return 0;
2326 }
2327 
2328 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
2329 				 struct ib_send_wr *wr, u32 nelements,
2330 				 u32 length, u32 pdn)
2331 {
2332 	struct ib_mr *sig_mr = wr->wr.sig_handover.sig_mr;
2333 	u32 sig_key = sig_mr->rkey;
2334 	u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
2335 
2336 	memset(seg, 0, sizeof(*seg));
2337 
2338 	seg->flags = get_umr_flags(wr->wr.sig_handover.access_flags) |
2339 				   MLX5_ACCESS_MODE_KLM;
2340 	seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
2341 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
2342 				    MLX5_MKEY_BSF_EN | pdn);
2343 	seg->len = cpu_to_be64(length);
2344 	seg->xlt_oct_size = cpu_to_be32(be16_to_cpu(get_klm_octo(nelements)));
2345 	seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
2346 }
2347 
2348 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
2349 				struct ib_send_wr *wr, u32 nelements)
2350 {
2351 	memset(umr, 0, sizeof(*umr));
2352 
2353 	umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
2354 	umr->klm_octowords = get_klm_octo(nelements);
2355 	umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
2356 	umr->mkey_mask = sig_mkey_mask();
2357 }
2358 
2359 
2360 static int set_sig_umr_wr(struct ib_send_wr *wr, struct mlx5_ib_qp *qp,
2361 			  void **seg, int *size)
2362 {
2363 	struct mlx5_ib_mr *sig_mr = to_mmr(wr->wr.sig_handover.sig_mr);
2364 	u32 pdn = get_pd(qp)->pdn;
2365 	u32 klm_oct_size;
2366 	int region_len, ret;
2367 
2368 	if (unlikely(wr->num_sge != 1) ||
2369 	    unlikely(wr->wr.sig_handover.access_flags &
2370 		     IB_ACCESS_REMOTE_ATOMIC) ||
2371 	    unlikely(!sig_mr->sig) || unlikely(!qp->signature_en) ||
2372 	    unlikely(!sig_mr->sig->sig_status_checked))
2373 		return -EINVAL;
2374 
2375 	/* length of the protected region, data + protection */
2376 	region_len = wr->sg_list->length;
2377 	if (wr->wr.sig_handover.prot &&
2378 	    (wr->wr.sig_handover.prot->lkey != wr->sg_list->lkey  ||
2379 	     wr->wr.sig_handover.prot->addr != wr->sg_list->addr  ||
2380 	     wr->wr.sig_handover.prot->length != wr->sg_list->length))
2381 		region_len += wr->wr.sig_handover.prot->length;
2382 
2383 	/**
2384 	 * KLM octoword size - if protection was provided
2385 	 * then we use strided block format (3 octowords),
2386 	 * else we use single KLM (1 octoword)
2387 	 **/
2388 	klm_oct_size = wr->wr.sig_handover.prot ? 3 : 1;
2389 
2390 	set_sig_umr_segment(*seg, wr, klm_oct_size);
2391 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2392 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2393 	if (unlikely((*seg == qp->sq.qend)))
2394 		*seg = mlx5_get_send_wqe(qp, 0);
2395 
2396 	set_sig_mkey_segment(*seg, wr, klm_oct_size, region_len, pdn);
2397 	*seg += sizeof(struct mlx5_mkey_seg);
2398 	*size += sizeof(struct mlx5_mkey_seg) / 16;
2399 	if (unlikely((*seg == qp->sq.qend)))
2400 		*seg = mlx5_get_send_wqe(qp, 0);
2401 
2402 	ret = set_sig_data_segment(wr, qp, seg, size);
2403 	if (ret)
2404 		return ret;
2405 
2406 	sig_mr->sig->sig_status_checked = false;
2407 	return 0;
2408 }
2409 
2410 static int set_psv_wr(struct ib_sig_domain *domain,
2411 		      u32 psv_idx, void **seg, int *size)
2412 {
2413 	struct mlx5_seg_set_psv *psv_seg = *seg;
2414 
2415 	memset(psv_seg, 0, sizeof(*psv_seg));
2416 	psv_seg->psv_num = cpu_to_be32(psv_idx);
2417 	switch (domain->sig_type) {
2418 	case IB_SIG_TYPE_NONE:
2419 		break;
2420 	case IB_SIG_TYPE_T10_DIF:
2421 		psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
2422 						     domain->sig.dif.app_tag);
2423 		psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
2424 		break;
2425 	default:
2426 		pr_err("Bad signature type given.\n");
2427 		return 1;
2428 	}
2429 
2430 	*seg += sizeof(*psv_seg);
2431 	*size += sizeof(*psv_seg) / 16;
2432 
2433 	return 0;
2434 }
2435 
2436 static int set_frwr_li_wr(void **seg, struct ib_send_wr *wr, int *size,
2437 			  struct mlx5_core_dev *mdev, struct mlx5_ib_pd *pd, struct mlx5_ib_qp *qp)
2438 {
2439 	int writ = 0;
2440 	int li;
2441 
2442 	li = wr->opcode == IB_WR_LOCAL_INV ? 1 : 0;
2443 	if (unlikely(wr->send_flags & IB_SEND_INLINE))
2444 		return -EINVAL;
2445 
2446 	set_frwr_umr_segment(*seg, wr, li);
2447 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2448 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2449 	if (unlikely((*seg == qp->sq.qend)))
2450 		*seg = mlx5_get_send_wqe(qp, 0);
2451 	set_mkey_segment(*seg, wr, li, &writ);
2452 	*seg += sizeof(struct mlx5_mkey_seg);
2453 	*size += sizeof(struct mlx5_mkey_seg) / 16;
2454 	if (unlikely((*seg == qp->sq.qend)))
2455 		*seg = mlx5_get_send_wqe(qp, 0);
2456 	if (!li) {
2457 		if (unlikely(wr->wr.fast_reg.page_list_len >
2458 			     wr->wr.fast_reg.page_list->max_page_list_len))
2459 			return	-ENOMEM;
2460 
2461 		set_frwr_pages(*seg, wr, mdev, pd, writ);
2462 		*seg += sizeof(struct mlx5_wqe_data_seg);
2463 		*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
2464 	}
2465 	return 0;
2466 }
2467 
2468 static void dump_wqe(struct mlx5_ib_qp *qp, int idx, int size_16)
2469 {
2470 	__be32 *p = NULL;
2471 	int tidx = idx;
2472 	int i, j;
2473 
2474 	pr_debug("dump wqe at %p\n", mlx5_get_send_wqe(qp, tidx));
2475 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
2476 		if ((i & 0xf) == 0) {
2477 			void *buf = mlx5_get_send_wqe(qp, tidx);
2478 			tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
2479 			p = buf;
2480 			j = 0;
2481 		}
2482 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
2483 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
2484 			 be32_to_cpu(p[j + 3]));
2485 	}
2486 }
2487 
2488 static void mlx5_bf_copy(u64 __iomem *dst, u64 *src,
2489 			 unsigned bytecnt, struct mlx5_ib_qp *qp)
2490 {
2491 	while (bytecnt > 0) {
2492 		__iowrite64_copy(dst++, src++, 8);
2493 		__iowrite64_copy(dst++, src++, 8);
2494 		__iowrite64_copy(dst++, src++, 8);
2495 		__iowrite64_copy(dst++, src++, 8);
2496 		__iowrite64_copy(dst++, src++, 8);
2497 		__iowrite64_copy(dst++, src++, 8);
2498 		__iowrite64_copy(dst++, src++, 8);
2499 		__iowrite64_copy(dst++, src++, 8);
2500 		bytecnt -= 64;
2501 		if (unlikely(src == qp->sq.qend))
2502 			src = mlx5_get_send_wqe(qp, 0);
2503 	}
2504 }
2505 
2506 static u8 get_fence(u8 fence, struct ib_send_wr *wr)
2507 {
2508 	if (unlikely(wr->opcode == IB_WR_LOCAL_INV &&
2509 		     wr->send_flags & IB_SEND_FENCE))
2510 		return MLX5_FENCE_MODE_STRONG_ORDERING;
2511 
2512 	if (unlikely(fence)) {
2513 		if (wr->send_flags & IB_SEND_FENCE)
2514 			return MLX5_FENCE_MODE_SMALL_AND_FENCE;
2515 		else
2516 			return fence;
2517 
2518 	} else {
2519 		return 0;
2520 	}
2521 }
2522 
2523 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
2524 		     struct mlx5_wqe_ctrl_seg **ctrl,
2525 		     struct ib_send_wr *wr, unsigned *idx,
2526 		     int *size, int nreq)
2527 {
2528 	int err = 0;
2529 
2530 	if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq))) {
2531 		err = -ENOMEM;
2532 		return err;
2533 	}
2534 
2535 	*idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
2536 	*seg = mlx5_get_send_wqe(qp, *idx);
2537 	*ctrl = *seg;
2538 	*(uint32_t *)(*seg + 8) = 0;
2539 	(*ctrl)->imm = send_ieth(wr);
2540 	(*ctrl)->fm_ce_se = qp->sq_signal_bits |
2541 		(wr->send_flags & IB_SEND_SIGNALED ?
2542 		 MLX5_WQE_CTRL_CQ_UPDATE : 0) |
2543 		(wr->send_flags & IB_SEND_SOLICITED ?
2544 		 MLX5_WQE_CTRL_SOLICITED : 0);
2545 
2546 	*seg += sizeof(**ctrl);
2547 	*size = sizeof(**ctrl) / 16;
2548 
2549 	return err;
2550 }
2551 
2552 static void finish_wqe(struct mlx5_ib_qp *qp,
2553 		       struct mlx5_wqe_ctrl_seg *ctrl,
2554 		       u8 size, unsigned idx, u64 wr_id,
2555 		       int nreq, u8 fence, u8 next_fence,
2556 		       u32 mlx5_opcode)
2557 {
2558 	u8 opmod = 0;
2559 
2560 	ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
2561 					     mlx5_opcode | ((u32)opmod << 24));
2562 	ctrl->qpn_ds = cpu_to_be32(size | (qp->mqp.qpn << 8));
2563 	ctrl->fm_ce_se |= fence;
2564 	qp->fm_cache = next_fence;
2565 	if (unlikely(qp->wq_sig))
2566 		ctrl->signature = wq_sig(ctrl);
2567 
2568 	qp->sq.wrid[idx] = wr_id;
2569 	qp->sq.w_list[idx].opcode = mlx5_opcode;
2570 	qp->sq.wqe_head[idx] = qp->sq.head + nreq;
2571 	qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
2572 	qp->sq.w_list[idx].next = qp->sq.cur_post;
2573 }
2574 
2575 
2576 int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2577 		      struct ib_send_wr **bad_wr)
2578 {
2579 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
2580 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2581 	struct mlx5_core_dev *mdev = dev->mdev;
2582 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2583 	struct mlx5_ib_mr *mr;
2584 	struct mlx5_wqe_data_seg *dpseg;
2585 	struct mlx5_wqe_xrc_seg *xrc;
2586 	struct mlx5_bf *bf = qp->bf;
2587 	int uninitialized_var(size);
2588 	void *qend = qp->sq.qend;
2589 	unsigned long flags;
2590 	unsigned idx;
2591 	int err = 0;
2592 	int inl = 0;
2593 	int num_sge;
2594 	void *seg;
2595 	int nreq;
2596 	int i;
2597 	u8 next_fence = 0;
2598 	u8 fence;
2599 
2600 	spin_lock_irqsave(&qp->sq.lock, flags);
2601 
2602 	for (nreq = 0; wr; nreq++, wr = wr->next) {
2603 		if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
2604 			mlx5_ib_warn(dev, "\n");
2605 			err = -EINVAL;
2606 			*bad_wr = wr;
2607 			goto out;
2608 		}
2609 
2610 		fence = qp->fm_cache;
2611 		num_sge = wr->num_sge;
2612 		if (unlikely(num_sge > qp->sq.max_gs)) {
2613 			mlx5_ib_warn(dev, "\n");
2614 			err = -ENOMEM;
2615 			*bad_wr = wr;
2616 			goto out;
2617 		}
2618 
2619 		err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, nreq);
2620 		if (err) {
2621 			mlx5_ib_warn(dev, "\n");
2622 			err = -ENOMEM;
2623 			*bad_wr = wr;
2624 			goto out;
2625 		}
2626 
2627 		switch (ibqp->qp_type) {
2628 		case IB_QPT_XRC_INI:
2629 			xrc = seg;
2630 			xrc->xrc_srqn = htonl(wr->xrc_remote_srq_num);
2631 			seg += sizeof(*xrc);
2632 			size += sizeof(*xrc) / 16;
2633 			/* fall through */
2634 		case IB_QPT_RC:
2635 			switch (wr->opcode) {
2636 			case IB_WR_RDMA_READ:
2637 			case IB_WR_RDMA_WRITE:
2638 			case IB_WR_RDMA_WRITE_WITH_IMM:
2639 				set_raddr_seg(seg, wr->wr.rdma.remote_addr,
2640 					      wr->wr.rdma.rkey);
2641 				seg += sizeof(struct mlx5_wqe_raddr_seg);
2642 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
2643 				break;
2644 
2645 			case IB_WR_ATOMIC_CMP_AND_SWP:
2646 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2647 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2648 				mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
2649 				err = -ENOSYS;
2650 				*bad_wr = wr;
2651 				goto out;
2652 
2653 			case IB_WR_LOCAL_INV:
2654 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2655 				qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
2656 				ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
2657 				err = set_frwr_li_wr(&seg, wr, &size, mdev, to_mpd(ibqp->pd), qp);
2658 				if (err) {
2659 					mlx5_ib_warn(dev, "\n");
2660 					*bad_wr = wr;
2661 					goto out;
2662 				}
2663 				num_sge = 0;
2664 				break;
2665 
2666 			case IB_WR_FAST_REG_MR:
2667 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2668 				qp->sq.wr_data[idx] = IB_WR_FAST_REG_MR;
2669 				ctrl->imm = cpu_to_be32(wr->wr.fast_reg.rkey);
2670 				err = set_frwr_li_wr(&seg, wr, &size, mdev, to_mpd(ibqp->pd), qp);
2671 				if (err) {
2672 					mlx5_ib_warn(dev, "\n");
2673 					*bad_wr = wr;
2674 					goto out;
2675 				}
2676 				num_sge = 0;
2677 				break;
2678 
2679 			case IB_WR_REG_SIG_MR:
2680 				qp->sq.wr_data[idx] = IB_WR_REG_SIG_MR;
2681 				mr = to_mmr(wr->wr.sig_handover.sig_mr);
2682 
2683 				ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
2684 				err = set_sig_umr_wr(wr, qp, &seg, &size);
2685 				if (err) {
2686 					mlx5_ib_warn(dev, "\n");
2687 					*bad_wr = wr;
2688 					goto out;
2689 				}
2690 
2691 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2692 					   nreq, get_fence(fence, wr),
2693 					   next_fence, MLX5_OPCODE_UMR);
2694 				/*
2695 				 * SET_PSV WQEs are not signaled and solicited
2696 				 * on error
2697 				 */
2698 				wr->send_flags &= ~IB_SEND_SIGNALED;
2699 				wr->send_flags |= IB_SEND_SOLICITED;
2700 				err = begin_wqe(qp, &seg, &ctrl, wr,
2701 						&idx, &size, nreq);
2702 				if (err) {
2703 					mlx5_ib_warn(dev, "\n");
2704 					err = -ENOMEM;
2705 					*bad_wr = wr;
2706 					goto out;
2707 				}
2708 
2709 				err = set_psv_wr(&wr->wr.sig_handover.sig_attrs->mem,
2710 						 mr->sig->psv_memory.psv_idx, &seg,
2711 						 &size);
2712 				if (err) {
2713 					mlx5_ib_warn(dev, "\n");
2714 					*bad_wr = wr;
2715 					goto out;
2716 				}
2717 
2718 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2719 					   nreq, get_fence(fence, wr),
2720 					   next_fence, MLX5_OPCODE_SET_PSV);
2721 				err = begin_wqe(qp, &seg, &ctrl, wr,
2722 						&idx, &size, nreq);
2723 				if (err) {
2724 					mlx5_ib_warn(dev, "\n");
2725 					err = -ENOMEM;
2726 					*bad_wr = wr;
2727 					goto out;
2728 				}
2729 
2730 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2731 				err = set_psv_wr(&wr->wr.sig_handover.sig_attrs->wire,
2732 						 mr->sig->psv_wire.psv_idx, &seg,
2733 						 &size);
2734 				if (err) {
2735 					mlx5_ib_warn(dev, "\n");
2736 					*bad_wr = wr;
2737 					goto out;
2738 				}
2739 
2740 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2741 					   nreq, get_fence(fence, wr),
2742 					   next_fence, MLX5_OPCODE_SET_PSV);
2743 				num_sge = 0;
2744 				goto skip_psv;
2745 
2746 			default:
2747 				break;
2748 			}
2749 			break;
2750 
2751 		case IB_QPT_UC:
2752 			switch (wr->opcode) {
2753 			case IB_WR_RDMA_WRITE:
2754 			case IB_WR_RDMA_WRITE_WITH_IMM:
2755 				set_raddr_seg(seg, wr->wr.rdma.remote_addr,
2756 					      wr->wr.rdma.rkey);
2757 				seg  += sizeof(struct mlx5_wqe_raddr_seg);
2758 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
2759 				break;
2760 
2761 			default:
2762 				break;
2763 			}
2764 			break;
2765 
2766 		case IB_QPT_UD:
2767 		case IB_QPT_SMI:
2768 		case IB_QPT_GSI:
2769 			set_datagram_seg(seg, wr);
2770 			seg += sizeof(struct mlx5_wqe_datagram_seg);
2771 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
2772 			if (unlikely((seg == qend)))
2773 				seg = mlx5_get_send_wqe(qp, 0);
2774 			break;
2775 
2776 		case MLX5_IB_QPT_REG_UMR:
2777 			if (wr->opcode != MLX5_IB_WR_UMR) {
2778 				err = -EINVAL;
2779 				mlx5_ib_warn(dev, "bad opcode\n");
2780 				goto out;
2781 			}
2782 			qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
2783 			ctrl->imm = cpu_to_be32(wr->wr.fast_reg.rkey);
2784 			set_reg_umr_segment(seg, wr);
2785 			seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2786 			size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2787 			if (unlikely((seg == qend)))
2788 				seg = mlx5_get_send_wqe(qp, 0);
2789 			set_reg_mkey_segment(seg, wr);
2790 			seg += sizeof(struct mlx5_mkey_seg);
2791 			size += sizeof(struct mlx5_mkey_seg) / 16;
2792 			if (unlikely((seg == qend)))
2793 				seg = mlx5_get_send_wqe(qp, 0);
2794 			break;
2795 
2796 		default:
2797 			break;
2798 		}
2799 
2800 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
2801 			int uninitialized_var(sz);
2802 
2803 			err = set_data_inl_seg(qp, wr, seg, &sz);
2804 			if (unlikely(err)) {
2805 				mlx5_ib_warn(dev, "\n");
2806 				*bad_wr = wr;
2807 				goto out;
2808 			}
2809 			inl = 1;
2810 			size += sz;
2811 		} else {
2812 			dpseg = seg;
2813 			for (i = 0; i < num_sge; i++) {
2814 				if (unlikely(dpseg == qend)) {
2815 					seg = mlx5_get_send_wqe(qp, 0);
2816 					dpseg = seg;
2817 				}
2818 				if (likely(wr->sg_list[i].length)) {
2819 					set_data_ptr_seg(dpseg, wr->sg_list + i);
2820 					size += sizeof(struct mlx5_wqe_data_seg) / 16;
2821 					dpseg++;
2822 				}
2823 			}
2824 		}
2825 
2826 		finish_wqe(qp, ctrl, size, idx, wr->wr_id, nreq,
2827 			   get_fence(fence, wr), next_fence,
2828 			   mlx5_ib_opcode[wr->opcode]);
2829 skip_psv:
2830 		if (0)
2831 			dump_wqe(qp, idx, size);
2832 	}
2833 
2834 out:
2835 	if (likely(nreq)) {
2836 		qp->sq.head += nreq;
2837 
2838 		/* Make sure that descriptors are written before
2839 		 * updating doorbell record and ringing the doorbell
2840 		 */
2841 		wmb();
2842 
2843 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
2844 
2845 		/* Make sure doorbell record is visible to the HCA before
2846 		 * we hit doorbell */
2847 		wmb();
2848 
2849 		if (bf->need_lock)
2850 			spin_lock(&bf->lock);
2851 		else
2852 			__acquire(&bf->lock);
2853 
2854 		/* TBD enable WC */
2855 		if (0 && nreq == 1 && bf->uuarn && inl && size > 1 && size <= bf->buf_size / 16) {
2856 			mlx5_bf_copy(bf->reg + bf->offset, (u64 *)ctrl, ALIGN(size * 16, 64), qp);
2857 			/* wc_wmb(); */
2858 		} else {
2859 			mlx5_write64((__be32 *)ctrl, bf->regreg + bf->offset,
2860 				     MLX5_GET_DOORBELL_LOCK(&bf->lock32));
2861 			/* Make sure doorbells don't leak out of SQ spinlock
2862 			 * and reach the HCA out of order.
2863 			 */
2864 			mmiowb();
2865 		}
2866 		bf->offset ^= bf->buf_size;
2867 		if (bf->need_lock)
2868 			spin_unlock(&bf->lock);
2869 		else
2870 			__release(&bf->lock);
2871 	}
2872 
2873 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2874 
2875 	return err;
2876 }
2877 
2878 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
2879 {
2880 	sig->signature = calc_sig(sig, size);
2881 }
2882 
2883 int mlx5_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2884 		      struct ib_recv_wr **bad_wr)
2885 {
2886 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2887 	struct mlx5_wqe_data_seg *scat;
2888 	struct mlx5_rwqe_sig *sig;
2889 	unsigned long flags;
2890 	int err = 0;
2891 	int nreq;
2892 	int ind;
2893 	int i;
2894 
2895 	spin_lock_irqsave(&qp->rq.lock, flags);
2896 
2897 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
2898 
2899 	for (nreq = 0; wr; nreq++, wr = wr->next) {
2900 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2901 			err = -ENOMEM;
2902 			*bad_wr = wr;
2903 			goto out;
2904 		}
2905 
2906 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2907 			err = -EINVAL;
2908 			*bad_wr = wr;
2909 			goto out;
2910 		}
2911 
2912 		scat = get_recv_wqe(qp, ind);
2913 		if (qp->wq_sig)
2914 			scat++;
2915 
2916 		for (i = 0; i < wr->num_sge; i++)
2917 			set_data_ptr_seg(scat + i, wr->sg_list + i);
2918 
2919 		if (i < qp->rq.max_gs) {
2920 			scat[i].byte_count = 0;
2921 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
2922 			scat[i].addr       = 0;
2923 		}
2924 
2925 		if (qp->wq_sig) {
2926 			sig = (struct mlx5_rwqe_sig *)scat;
2927 			set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
2928 		}
2929 
2930 		qp->rq.wrid[ind] = wr->wr_id;
2931 
2932 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
2933 	}
2934 
2935 out:
2936 	if (likely(nreq)) {
2937 		qp->rq.head += nreq;
2938 
2939 		/* Make sure that descriptors are written before
2940 		 * doorbell record.
2941 		 */
2942 		wmb();
2943 
2944 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
2945 	}
2946 
2947 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2948 
2949 	return err;
2950 }
2951 
2952 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
2953 {
2954 	switch (mlx5_state) {
2955 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
2956 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
2957 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
2958 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
2959 	case MLX5_QP_STATE_SQ_DRAINING:
2960 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
2961 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
2962 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
2963 	default:		     return -1;
2964 	}
2965 }
2966 
2967 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
2968 {
2969 	switch (mlx5_mig_state) {
2970 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
2971 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
2972 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
2973 	default: return -1;
2974 	}
2975 }
2976 
2977 static int to_ib_qp_access_flags(int mlx5_flags)
2978 {
2979 	int ib_flags = 0;
2980 
2981 	if (mlx5_flags & MLX5_QP_BIT_RRE)
2982 		ib_flags |= IB_ACCESS_REMOTE_READ;
2983 	if (mlx5_flags & MLX5_QP_BIT_RWE)
2984 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
2985 	if (mlx5_flags & MLX5_QP_BIT_RAE)
2986 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
2987 
2988 	return ib_flags;
2989 }
2990 
2991 static void to_ib_ah_attr(struct mlx5_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
2992 				struct mlx5_qp_path *path)
2993 {
2994 	struct mlx5_core_dev *dev = ibdev->mdev;
2995 
2996 	memset(ib_ah_attr, 0, sizeof(*ib_ah_attr));
2997 	ib_ah_attr->port_num	  = path->port;
2998 
2999 	if (ib_ah_attr->port_num == 0 ||
3000 	    ib_ah_attr->port_num > MLX5_CAP_GEN(dev, num_ports))
3001 		return;
3002 
3003 	ib_ah_attr->sl = path->sl & 0xf;
3004 
3005 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
3006 	ib_ah_attr->src_path_bits = path->grh_mlid & 0x7f;
3007 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
3008 	ib_ah_attr->ah_flags      = (path->grh_mlid & (1 << 7)) ? IB_AH_GRH : 0;
3009 	if (ib_ah_attr->ah_flags) {
3010 		ib_ah_attr->grh.sgid_index = path->mgid_index;
3011 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
3012 		ib_ah_attr->grh.traffic_class =
3013 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3014 		ib_ah_attr->grh.flow_label =
3015 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3016 		memcpy(ib_ah_attr->grh.dgid.raw,
3017 		       path->rgid, sizeof(ib_ah_attr->grh.dgid.raw));
3018 	}
3019 }
3020 
3021 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3022 		     struct ib_qp_init_attr *qp_init_attr)
3023 {
3024 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3025 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3026 	struct mlx5_query_qp_mbox_out *outb;
3027 	struct mlx5_qp_context *context;
3028 	int mlx5_state;
3029 	int err = 0;
3030 
3031 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
3032 	/*
3033 	 * Wait for any outstanding page faults, in case the user frees memory
3034 	 * based upon this query's result.
3035 	 */
3036 	flush_workqueue(mlx5_ib_page_fault_wq);
3037 #endif
3038 
3039 	mutex_lock(&qp->mutex);
3040 	outb = kzalloc(sizeof(*outb), GFP_KERNEL);
3041 	if (!outb) {
3042 		err = -ENOMEM;
3043 		goto out;
3044 	}
3045 	context = &outb->ctx;
3046 	err = mlx5_core_qp_query(dev->mdev, &qp->mqp, outb, sizeof(*outb));
3047 	if (err)
3048 		goto out_free;
3049 
3050 	mlx5_state = be32_to_cpu(context->flags) >> 28;
3051 
3052 	qp->state		     = to_ib_qp_state(mlx5_state);
3053 	qp_attr->qp_state	     = qp->state;
3054 	qp_attr->path_mtu	     = context->mtu_msgmax >> 5;
3055 	qp_attr->path_mig_state	     =
3056 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
3057 	qp_attr->qkey		     = be32_to_cpu(context->qkey);
3058 	qp_attr->rq_psn		     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
3059 	qp_attr->sq_psn		     = be32_to_cpu(context->next_send_psn) & 0xffffff;
3060 	qp_attr->dest_qp_num	     = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
3061 	qp_attr->qp_access_flags     =
3062 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
3063 
3064 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3065 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
3066 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
3067 		qp_attr->alt_pkey_index = context->alt_path.pkey_index & 0x7f;
3068 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
3069 	}
3070 
3071 	qp_attr->pkey_index = context->pri_path.pkey_index & 0x7f;
3072 	qp_attr->port_num = context->pri_path.port;
3073 
3074 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3075 	qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
3076 
3077 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
3078 
3079 	qp_attr->max_dest_rd_atomic =
3080 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
3081 	qp_attr->min_rnr_timer	    =
3082 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
3083 	qp_attr->timeout	    = context->pri_path.ackto_lt >> 3;
3084 	qp_attr->retry_cnt	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
3085 	qp_attr->rnr_retry	    = (be32_to_cpu(context->params1) >> 13) & 0x7;
3086 	qp_attr->alt_timeout	    = context->alt_path.ackto_lt >> 3;
3087 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
3088 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
3089 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
3090 
3091 	if (!ibqp->uobject) {
3092 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
3093 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
3094 	} else {
3095 		qp_attr->cap.max_send_wr  = 0;
3096 		qp_attr->cap.max_send_sge = 0;
3097 	}
3098 
3099 	/* We don't support inline sends for kernel QPs (yet), and we
3100 	 * don't know what userspace's value should be.
3101 	 */
3102 	qp_attr->cap.max_inline_data = 0;
3103 
3104 	qp_init_attr->cap	     = qp_attr->cap;
3105 
3106 	qp_init_attr->create_flags = 0;
3107 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3108 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3109 
3110 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
3111 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3112 
3113 out_free:
3114 	kfree(outb);
3115 
3116 out:
3117 	mutex_unlock(&qp->mutex);
3118 	return err;
3119 }
3120 
3121 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
3122 					  struct ib_ucontext *context,
3123 					  struct ib_udata *udata)
3124 {
3125 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3126 	struct mlx5_ib_xrcd *xrcd;
3127 	int err;
3128 
3129 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
3130 		return ERR_PTR(-ENOSYS);
3131 
3132 	xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
3133 	if (!xrcd)
3134 		return ERR_PTR(-ENOMEM);
3135 
3136 	err = mlx5_core_xrcd_alloc(dev->mdev, &xrcd->xrcdn);
3137 	if (err) {
3138 		kfree(xrcd);
3139 		return ERR_PTR(-ENOMEM);
3140 	}
3141 
3142 	return &xrcd->ibxrcd;
3143 }
3144 
3145 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
3146 {
3147 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
3148 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
3149 	int err;
3150 
3151 	err = mlx5_core_xrcd_dealloc(dev->mdev, xrcdn);
3152 	if (err) {
3153 		mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
3154 		return err;
3155 	}
3156 
3157 	kfree(xrcd);
3158 
3159 	return 0;
3160 }
3161