xref: /openbmc/linux/drivers/infiniband/hw/mlx5/qp.c (revision 4a44a19b)
1 /*
2  * Copyright (c) 2013, Mellanox Technologies inc.  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 struct umr_wr {
74 	u64				virt_addr;
75 	struct ib_pd		       *pd;
76 	unsigned int			page_shift;
77 	unsigned int			npages;
78 	u32				length;
79 	int				access_flags;
80 	u32				mkey;
81 };
82 
83 static int is_qp0(enum ib_qp_type qp_type)
84 {
85 	return qp_type == IB_QPT_SMI;
86 }
87 
88 static int is_qp1(enum ib_qp_type qp_type)
89 {
90 	return qp_type == IB_QPT_GSI;
91 }
92 
93 static int is_sqp(enum ib_qp_type qp_type)
94 {
95 	return is_qp0(qp_type) || is_qp1(qp_type);
96 }
97 
98 static void *get_wqe(struct mlx5_ib_qp *qp, int offset)
99 {
100 	return mlx5_buf_offset(&qp->buf, offset);
101 }
102 
103 static void *get_recv_wqe(struct mlx5_ib_qp *qp, int n)
104 {
105 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
106 }
107 
108 void *mlx5_get_send_wqe(struct mlx5_ib_qp *qp, int n)
109 {
110 	return get_wqe(qp, qp->sq.offset + (n << MLX5_IB_SQ_STRIDE));
111 }
112 
113 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
114 {
115 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
116 	struct ib_event event;
117 
118 	if (type == MLX5_EVENT_TYPE_PATH_MIG)
119 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
120 
121 	if (ibqp->event_handler) {
122 		event.device     = ibqp->device;
123 		event.element.qp = ibqp;
124 		switch (type) {
125 		case MLX5_EVENT_TYPE_PATH_MIG:
126 			event.event = IB_EVENT_PATH_MIG;
127 			break;
128 		case MLX5_EVENT_TYPE_COMM_EST:
129 			event.event = IB_EVENT_COMM_EST;
130 			break;
131 		case MLX5_EVENT_TYPE_SQ_DRAINED:
132 			event.event = IB_EVENT_SQ_DRAINED;
133 			break;
134 		case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
135 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
136 			break;
137 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
138 			event.event = IB_EVENT_QP_FATAL;
139 			break;
140 		case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
141 			event.event = IB_EVENT_PATH_MIG_ERR;
142 			break;
143 		case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
144 			event.event = IB_EVENT_QP_REQ_ERR;
145 			break;
146 		case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
147 			event.event = IB_EVENT_QP_ACCESS_ERR;
148 			break;
149 		default:
150 			pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
151 			return;
152 		}
153 
154 		ibqp->event_handler(&event, ibqp->qp_context);
155 	}
156 }
157 
158 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
159 		       int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
160 {
161 	struct mlx5_general_caps *gen;
162 	int wqe_size;
163 	int wq_size;
164 
165 	gen = &dev->mdev->caps.gen;
166 	/* Sanity check RQ size before proceeding */
167 	if (cap->max_recv_wr  > gen->max_wqes)
168 		return -EINVAL;
169 
170 	if (!has_rq) {
171 		qp->rq.max_gs = 0;
172 		qp->rq.wqe_cnt = 0;
173 		qp->rq.wqe_shift = 0;
174 	} else {
175 		if (ucmd) {
176 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
177 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
178 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
179 			qp->rq.max_post = qp->rq.wqe_cnt;
180 		} else {
181 			wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
182 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
183 			wqe_size = roundup_pow_of_two(wqe_size);
184 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
185 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
186 			qp->rq.wqe_cnt = wq_size / wqe_size;
187 			if (wqe_size > gen->max_rq_desc_sz) {
188 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
189 					    wqe_size,
190 					    gen->max_rq_desc_sz);
191 				return -EINVAL;
192 			}
193 			qp->rq.wqe_shift = ilog2(wqe_size);
194 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
195 			qp->rq.max_post = qp->rq.wqe_cnt;
196 		}
197 	}
198 
199 	return 0;
200 }
201 
202 static int sq_overhead(enum ib_qp_type qp_type)
203 {
204 	int size = 0;
205 
206 	switch (qp_type) {
207 	case IB_QPT_XRC_INI:
208 		size += sizeof(struct mlx5_wqe_xrc_seg);
209 		/* fall through */
210 	case IB_QPT_RC:
211 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
212 			sizeof(struct mlx5_wqe_atomic_seg) +
213 			sizeof(struct mlx5_wqe_raddr_seg);
214 		break;
215 
216 	case IB_QPT_XRC_TGT:
217 		return 0;
218 
219 	case IB_QPT_UC:
220 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
221 			sizeof(struct mlx5_wqe_raddr_seg) +
222 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
223 			sizeof(struct mlx5_mkey_seg);
224 		break;
225 
226 	case IB_QPT_UD:
227 	case IB_QPT_SMI:
228 	case IB_QPT_GSI:
229 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
230 			sizeof(struct mlx5_wqe_datagram_seg);
231 		break;
232 
233 	case MLX5_IB_QPT_REG_UMR:
234 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
235 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
236 			sizeof(struct mlx5_mkey_seg);
237 		break;
238 
239 	default:
240 		return -EINVAL;
241 	}
242 
243 	return size;
244 }
245 
246 static int calc_send_wqe(struct ib_qp_init_attr *attr)
247 {
248 	int inl_size = 0;
249 	int size;
250 
251 	size = sq_overhead(attr->qp_type);
252 	if (size < 0)
253 		return size;
254 
255 	if (attr->cap.max_inline_data) {
256 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
257 			attr->cap.max_inline_data;
258 	}
259 
260 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
261 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN &&
262 	    ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
263 			return MLX5_SIG_WQE_SIZE;
264 	else
265 		return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
266 }
267 
268 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
269 			struct mlx5_ib_qp *qp)
270 {
271 	struct mlx5_general_caps *gen;
272 	int wqe_size;
273 	int wq_size;
274 
275 	gen = &dev->mdev->caps.gen;
276 	if (!attr->cap.max_send_wr)
277 		return 0;
278 
279 	wqe_size = calc_send_wqe(attr);
280 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
281 	if (wqe_size < 0)
282 		return wqe_size;
283 
284 	if (wqe_size > gen->max_sq_desc_sz) {
285 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
286 			    wqe_size, gen->max_sq_desc_sz);
287 		return -EINVAL;
288 	}
289 
290 	qp->max_inline_data = wqe_size - sq_overhead(attr->qp_type) -
291 		sizeof(struct mlx5_wqe_inline_seg);
292 	attr->cap.max_inline_data = qp->max_inline_data;
293 
294 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN)
295 		qp->signature_en = true;
296 
297 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
298 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
299 	if (qp->sq.wqe_cnt > gen->max_wqes) {
300 		mlx5_ib_dbg(dev, "wqe count(%d) exceeds limits(%d)\n",
301 			    qp->sq.wqe_cnt, gen->max_wqes);
302 		return -ENOMEM;
303 	}
304 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
305 	qp->sq.max_gs = attr->cap.max_send_sge;
306 	qp->sq.max_post = wq_size / wqe_size;
307 	attr->cap.max_send_wr = qp->sq.max_post;
308 
309 	return wq_size;
310 }
311 
312 static int set_user_buf_size(struct mlx5_ib_dev *dev,
313 			    struct mlx5_ib_qp *qp,
314 			    struct mlx5_ib_create_qp *ucmd)
315 {
316 	struct mlx5_general_caps *gen;
317 	int desc_sz = 1 << qp->sq.wqe_shift;
318 
319 	gen = &dev->mdev->caps.gen;
320 	if (desc_sz > gen->max_sq_desc_sz) {
321 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
322 			     desc_sz, gen->max_sq_desc_sz);
323 		return -EINVAL;
324 	}
325 
326 	if (ucmd->sq_wqe_count && ((1 << ilog2(ucmd->sq_wqe_count)) != ucmd->sq_wqe_count)) {
327 		mlx5_ib_warn(dev, "sq_wqe_count %d, sq_wqe_count %d\n",
328 			     ucmd->sq_wqe_count, ucmd->sq_wqe_count);
329 		return -EINVAL;
330 	}
331 
332 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
333 
334 	if (qp->sq.wqe_cnt > gen->max_wqes) {
335 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
336 			     qp->sq.wqe_cnt, gen->max_wqes);
337 		return -EINVAL;
338 	}
339 
340 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
341 		(qp->sq.wqe_cnt << 6);
342 
343 	return 0;
344 }
345 
346 static int qp_has_rq(struct ib_qp_init_attr *attr)
347 {
348 	if (attr->qp_type == IB_QPT_XRC_INI ||
349 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
350 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
351 	    !attr->cap.max_recv_wr)
352 		return 0;
353 
354 	return 1;
355 }
356 
357 static int first_med_uuar(void)
358 {
359 	return 1;
360 }
361 
362 static int next_uuar(int n)
363 {
364 	n++;
365 
366 	while (((n % 4) & 2))
367 		n++;
368 
369 	return n;
370 }
371 
372 static int num_med_uuar(struct mlx5_uuar_info *uuari)
373 {
374 	int n;
375 
376 	n = uuari->num_uars * MLX5_NON_FP_BF_REGS_PER_PAGE -
377 		uuari->num_low_latency_uuars - 1;
378 
379 	return n >= 0 ? n : 0;
380 }
381 
382 static int max_uuari(struct mlx5_uuar_info *uuari)
383 {
384 	return uuari->num_uars * 4;
385 }
386 
387 static int first_hi_uuar(struct mlx5_uuar_info *uuari)
388 {
389 	int med;
390 	int i;
391 	int t;
392 
393 	med = num_med_uuar(uuari);
394 	for (t = 0, i = first_med_uuar();; i = next_uuar(i)) {
395 		t++;
396 		if (t == med)
397 			return next_uuar(i);
398 	}
399 
400 	return 0;
401 }
402 
403 static int alloc_high_class_uuar(struct mlx5_uuar_info *uuari)
404 {
405 	int i;
406 
407 	for (i = first_hi_uuar(uuari); i < max_uuari(uuari); i = next_uuar(i)) {
408 		if (!test_bit(i, uuari->bitmap)) {
409 			set_bit(i, uuari->bitmap);
410 			uuari->count[i]++;
411 			return i;
412 		}
413 	}
414 
415 	return -ENOMEM;
416 }
417 
418 static int alloc_med_class_uuar(struct mlx5_uuar_info *uuari)
419 {
420 	int minidx = first_med_uuar();
421 	int i;
422 
423 	for (i = first_med_uuar(); i < first_hi_uuar(uuari); i = next_uuar(i)) {
424 		if (uuari->count[i] < uuari->count[minidx])
425 			minidx = i;
426 	}
427 
428 	uuari->count[minidx]++;
429 	return minidx;
430 }
431 
432 static int alloc_uuar(struct mlx5_uuar_info *uuari,
433 		      enum mlx5_ib_latency_class lat)
434 {
435 	int uuarn = -EINVAL;
436 
437 	mutex_lock(&uuari->lock);
438 	switch (lat) {
439 	case MLX5_IB_LATENCY_CLASS_LOW:
440 		uuarn = 0;
441 		uuari->count[uuarn]++;
442 		break;
443 
444 	case MLX5_IB_LATENCY_CLASS_MEDIUM:
445 		if (uuari->ver < 2)
446 			uuarn = -ENOMEM;
447 		else
448 			uuarn = alloc_med_class_uuar(uuari);
449 		break;
450 
451 	case MLX5_IB_LATENCY_CLASS_HIGH:
452 		if (uuari->ver < 2)
453 			uuarn = -ENOMEM;
454 		else
455 			uuarn = alloc_high_class_uuar(uuari);
456 		break;
457 
458 	case MLX5_IB_LATENCY_CLASS_FAST_PATH:
459 		uuarn = 2;
460 		break;
461 	}
462 	mutex_unlock(&uuari->lock);
463 
464 	return uuarn;
465 }
466 
467 static void free_med_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
468 {
469 	clear_bit(uuarn, uuari->bitmap);
470 	--uuari->count[uuarn];
471 }
472 
473 static void free_high_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
474 {
475 	clear_bit(uuarn, uuari->bitmap);
476 	--uuari->count[uuarn];
477 }
478 
479 static void free_uuar(struct mlx5_uuar_info *uuari, int uuarn)
480 {
481 	int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
482 	int high_uuar = nuuars - uuari->num_low_latency_uuars;
483 
484 	mutex_lock(&uuari->lock);
485 	if (uuarn == 0) {
486 		--uuari->count[uuarn];
487 		goto out;
488 	}
489 
490 	if (uuarn < high_uuar) {
491 		free_med_class_uuar(uuari, uuarn);
492 		goto out;
493 	}
494 
495 	free_high_class_uuar(uuari, uuarn);
496 
497 out:
498 	mutex_unlock(&uuari->lock);
499 }
500 
501 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
502 {
503 	switch (state) {
504 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
505 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
506 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
507 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
508 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
509 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
510 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
511 	default:		return -1;
512 	}
513 }
514 
515 static int to_mlx5_st(enum ib_qp_type type)
516 {
517 	switch (type) {
518 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
519 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
520 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
521 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
522 	case IB_QPT_XRC_INI:
523 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
524 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
525 	case IB_QPT_GSI:		return MLX5_QP_ST_QP1;
526 	case IB_QPT_RAW_IPV6:		return MLX5_QP_ST_RAW_IPV6;
527 	case IB_QPT_RAW_ETHERTYPE:	return MLX5_QP_ST_RAW_ETHERTYPE;
528 	case IB_QPT_RAW_PACKET:
529 	case IB_QPT_MAX:
530 	default:		return -EINVAL;
531 	}
532 }
533 
534 static int uuarn_to_uar_index(struct mlx5_uuar_info *uuari, int uuarn)
535 {
536 	return uuari->uars[uuarn / MLX5_BF_REGS_PER_PAGE].index;
537 }
538 
539 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
540 			  struct mlx5_ib_qp *qp, struct ib_udata *udata,
541 			  struct mlx5_create_qp_mbox_in **in,
542 			  struct mlx5_ib_create_qp_resp *resp, int *inlen)
543 {
544 	struct mlx5_ib_ucontext *context;
545 	struct mlx5_ib_create_qp ucmd;
546 	int page_shift = 0;
547 	int uar_index;
548 	int npages;
549 	u32 offset = 0;
550 	int uuarn;
551 	int ncont = 0;
552 	int err;
553 
554 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
555 	if (err) {
556 		mlx5_ib_dbg(dev, "copy failed\n");
557 		return err;
558 	}
559 
560 	context = to_mucontext(pd->uobject->context);
561 	/*
562 	 * TBD: should come from the verbs when we have the API
563 	 */
564 	uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_HIGH);
565 	if (uuarn < 0) {
566 		mlx5_ib_dbg(dev, "failed to allocate low latency UUAR\n");
567 		mlx5_ib_dbg(dev, "reverting to medium latency\n");
568 		uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_MEDIUM);
569 		if (uuarn < 0) {
570 			mlx5_ib_dbg(dev, "failed to allocate medium latency UUAR\n");
571 			mlx5_ib_dbg(dev, "reverting to high latency\n");
572 			uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_LOW);
573 			if (uuarn < 0) {
574 				mlx5_ib_warn(dev, "uuar allocation failed\n");
575 				return uuarn;
576 			}
577 		}
578 	}
579 
580 	uar_index = uuarn_to_uar_index(&context->uuari, uuarn);
581 	mlx5_ib_dbg(dev, "uuarn 0x%x, uar_index 0x%x\n", uuarn, uar_index);
582 
583 	qp->rq.offset = 0;
584 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
585 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
586 
587 	err = set_user_buf_size(dev, qp, &ucmd);
588 	if (err)
589 		goto err_uuar;
590 
591 	if (ucmd.buf_addr && qp->buf_size) {
592 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
593 				       qp->buf_size, 0, 0);
594 		if (IS_ERR(qp->umem)) {
595 			mlx5_ib_dbg(dev, "umem_get failed\n");
596 			err = PTR_ERR(qp->umem);
597 			goto err_uuar;
598 		}
599 	} else {
600 		qp->umem = NULL;
601 	}
602 
603 	if (qp->umem) {
604 		mlx5_ib_cont_pages(qp->umem, ucmd.buf_addr, &npages, &page_shift,
605 				   &ncont, NULL);
606 		err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift, &offset);
607 		if (err) {
608 			mlx5_ib_warn(dev, "bad offset\n");
609 			goto err_umem;
610 		}
611 		mlx5_ib_dbg(dev, "addr 0x%llx, size %d, npages %d, page_shift %d, ncont %d, offset %d\n",
612 			    ucmd.buf_addr, qp->buf_size, npages, page_shift, ncont, offset);
613 	}
614 
615 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * ncont;
616 	*in = mlx5_vzalloc(*inlen);
617 	if (!*in) {
618 		err = -ENOMEM;
619 		goto err_umem;
620 	}
621 	if (qp->umem)
622 		mlx5_ib_populate_pas(dev, qp->umem, page_shift, (*in)->pas, 0);
623 	(*in)->ctx.log_pg_sz_remote_qpn =
624 		cpu_to_be32((page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
625 	(*in)->ctx.params2 = cpu_to_be32(offset << 6);
626 
627 	(*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
628 	resp->uuar_index = uuarn;
629 	qp->uuarn = uuarn;
630 
631 	err = mlx5_ib_db_map_user(context, ucmd.db_addr, &qp->db);
632 	if (err) {
633 		mlx5_ib_dbg(dev, "map failed\n");
634 		goto err_free;
635 	}
636 
637 	err = ib_copy_to_udata(udata, resp, sizeof(*resp));
638 	if (err) {
639 		mlx5_ib_dbg(dev, "copy failed\n");
640 		goto err_unmap;
641 	}
642 	qp->create_type = MLX5_QP_USER;
643 
644 	return 0;
645 
646 err_unmap:
647 	mlx5_ib_db_unmap_user(context, &qp->db);
648 
649 err_free:
650 	mlx5_vfree(*in);
651 
652 err_umem:
653 	if (qp->umem)
654 		ib_umem_release(qp->umem);
655 
656 err_uuar:
657 	free_uuar(&context->uuari, uuarn);
658 	return err;
659 }
660 
661 static void destroy_qp_user(struct ib_pd *pd, struct mlx5_ib_qp *qp)
662 {
663 	struct mlx5_ib_ucontext *context;
664 
665 	context = to_mucontext(pd->uobject->context);
666 	mlx5_ib_db_unmap_user(context, &qp->db);
667 	if (qp->umem)
668 		ib_umem_release(qp->umem);
669 	free_uuar(&context->uuari, qp->uuarn);
670 }
671 
672 static int create_kernel_qp(struct mlx5_ib_dev *dev,
673 			    struct ib_qp_init_attr *init_attr,
674 			    struct mlx5_ib_qp *qp,
675 			    struct mlx5_create_qp_mbox_in **in, int *inlen)
676 {
677 	enum mlx5_ib_latency_class lc = MLX5_IB_LATENCY_CLASS_LOW;
678 	struct mlx5_uuar_info *uuari;
679 	int uar_index;
680 	int uuarn;
681 	int err;
682 
683 	uuari = &dev->mdev->priv.uuari;
684 	if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN | IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK))
685 		return -EINVAL;
686 
687 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
688 		lc = MLX5_IB_LATENCY_CLASS_FAST_PATH;
689 
690 	uuarn = alloc_uuar(uuari, lc);
691 	if (uuarn < 0) {
692 		mlx5_ib_dbg(dev, "\n");
693 		return -ENOMEM;
694 	}
695 
696 	qp->bf = &uuari->bfs[uuarn];
697 	uar_index = qp->bf->uar->index;
698 
699 	err = calc_sq_size(dev, init_attr, qp);
700 	if (err < 0) {
701 		mlx5_ib_dbg(dev, "err %d\n", err);
702 		goto err_uuar;
703 	}
704 
705 	qp->rq.offset = 0;
706 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
707 	qp->buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
708 
709 	err = mlx5_buf_alloc(dev->mdev, qp->buf_size, PAGE_SIZE * 2, &qp->buf);
710 	if (err) {
711 		mlx5_ib_dbg(dev, "err %d\n", err);
712 		goto err_uuar;
713 	}
714 
715 	qp->sq.qend = mlx5_get_send_wqe(qp, qp->sq.wqe_cnt);
716 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * qp->buf.npages;
717 	*in = mlx5_vzalloc(*inlen);
718 	if (!*in) {
719 		err = -ENOMEM;
720 		goto err_buf;
721 	}
722 	(*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
723 	(*in)->ctx.log_pg_sz_remote_qpn =
724 		cpu_to_be32((qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
725 	/* Set "fast registration enabled" for all kernel QPs */
726 	(*in)->ctx.params1 |= cpu_to_be32(1 << 11);
727 	(*in)->ctx.sq_crq_size |= cpu_to_be16(1 << 4);
728 
729 	mlx5_fill_page_array(&qp->buf, (*in)->pas);
730 
731 	err = mlx5_db_alloc(dev->mdev, &qp->db);
732 	if (err) {
733 		mlx5_ib_dbg(dev, "err %d\n", err);
734 		goto err_free;
735 	}
736 
737 	qp->db.db[0] = 0;
738 	qp->db.db[1] = 0;
739 
740 	qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
741 	qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
742 	qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
743 	qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
744 	qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
745 
746 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
747 	    !qp->sq.w_list || !qp->sq.wqe_head) {
748 		err = -ENOMEM;
749 		goto err_wrid;
750 	}
751 	qp->create_type = MLX5_QP_KERNEL;
752 
753 	return 0;
754 
755 err_wrid:
756 	mlx5_db_free(dev->mdev, &qp->db);
757 	kfree(qp->sq.wqe_head);
758 	kfree(qp->sq.w_list);
759 	kfree(qp->sq.wrid);
760 	kfree(qp->sq.wr_data);
761 	kfree(qp->rq.wrid);
762 
763 err_free:
764 	mlx5_vfree(*in);
765 
766 err_buf:
767 	mlx5_buf_free(dev->mdev, &qp->buf);
768 
769 err_uuar:
770 	free_uuar(&dev->mdev->priv.uuari, uuarn);
771 	return err;
772 }
773 
774 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
775 {
776 	mlx5_db_free(dev->mdev, &qp->db);
777 	kfree(qp->sq.wqe_head);
778 	kfree(qp->sq.w_list);
779 	kfree(qp->sq.wrid);
780 	kfree(qp->sq.wr_data);
781 	kfree(qp->rq.wrid);
782 	mlx5_buf_free(dev->mdev, &qp->buf);
783 	free_uuar(&dev->mdev->priv.uuari, qp->bf->uuarn);
784 }
785 
786 static __be32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
787 {
788 	if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
789 	    (attr->qp_type == IB_QPT_XRC_INI))
790 		return cpu_to_be32(MLX5_SRQ_RQ);
791 	else if (!qp->has_rq)
792 		return cpu_to_be32(MLX5_ZERO_LEN_RQ);
793 	else
794 		return cpu_to_be32(MLX5_NON_ZERO_RQ);
795 }
796 
797 static int is_connected(enum ib_qp_type qp_type)
798 {
799 	if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC)
800 		return 1;
801 
802 	return 0;
803 }
804 
805 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
806 			    struct ib_qp_init_attr *init_attr,
807 			    struct ib_udata *udata, struct mlx5_ib_qp *qp)
808 {
809 	struct mlx5_ib_resources *devr = &dev->devr;
810 	struct mlx5_ib_create_qp_resp resp;
811 	struct mlx5_create_qp_mbox_in *in;
812 	struct mlx5_general_caps *gen;
813 	struct mlx5_ib_create_qp ucmd;
814 	int inlen = sizeof(*in);
815 	int err;
816 
817 	gen = &dev->mdev->caps.gen;
818 	mutex_init(&qp->mutex);
819 	spin_lock_init(&qp->sq.lock);
820 	spin_lock_init(&qp->rq.lock);
821 
822 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
823 		if (!(gen->flags & MLX5_DEV_CAP_FLAG_BLOCK_MCAST)) {
824 			mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
825 			return -EINVAL;
826 		} else {
827 			qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
828 		}
829 	}
830 
831 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
832 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
833 
834 	if (pd && pd->uobject) {
835 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
836 			mlx5_ib_dbg(dev, "copy failed\n");
837 			return -EFAULT;
838 		}
839 
840 		qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
841 		qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
842 	} else {
843 		qp->wq_sig = !!wq_signature;
844 	}
845 
846 	qp->has_rq = qp_has_rq(init_attr);
847 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
848 			  qp, (pd && pd->uobject) ? &ucmd : NULL);
849 	if (err) {
850 		mlx5_ib_dbg(dev, "err %d\n", err);
851 		return err;
852 	}
853 
854 	if (pd) {
855 		if (pd->uobject) {
856 			mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
857 			if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
858 			    ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
859 				mlx5_ib_dbg(dev, "invalid rq params\n");
860 				return -EINVAL;
861 			}
862 			if (ucmd.sq_wqe_count > gen->max_wqes) {
863 				mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
864 					    ucmd.sq_wqe_count, gen->max_wqes);
865 				return -EINVAL;
866 			}
867 			err = create_user_qp(dev, pd, qp, udata, &in, &resp, &inlen);
868 			if (err)
869 				mlx5_ib_dbg(dev, "err %d\n", err);
870 		} else {
871 			err = create_kernel_qp(dev, init_attr, qp, &in, &inlen);
872 			if (err)
873 				mlx5_ib_dbg(dev, "err %d\n", err);
874 			else
875 				qp->pa_lkey = to_mpd(pd)->pa_lkey;
876 		}
877 
878 		if (err)
879 			return err;
880 	} else {
881 		in = mlx5_vzalloc(sizeof(*in));
882 		if (!in)
883 			return -ENOMEM;
884 
885 		qp->create_type = MLX5_QP_EMPTY;
886 	}
887 
888 	if (is_sqp(init_attr->qp_type))
889 		qp->port = init_attr->port_num;
890 
891 	in->ctx.flags = cpu_to_be32(to_mlx5_st(init_attr->qp_type) << 16 |
892 				    MLX5_QP_PM_MIGRATED << 11);
893 
894 	if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
895 		in->ctx.flags_pd = cpu_to_be32(to_mpd(pd ? pd : devr->p0)->pdn);
896 	else
897 		in->ctx.flags_pd = cpu_to_be32(MLX5_QP_LAT_SENSITIVE);
898 
899 	if (qp->wq_sig)
900 		in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_ENABLE_SIG);
901 
902 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
903 		in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_BLOCK_MCAST);
904 
905 	if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
906 		int rcqe_sz;
907 		int scqe_sz;
908 
909 		rcqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->recv_cq);
910 		scqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->send_cq);
911 
912 		if (rcqe_sz == 128)
913 			in->ctx.cs_res = MLX5_RES_SCAT_DATA64_CQE;
914 		else
915 			in->ctx.cs_res = MLX5_RES_SCAT_DATA32_CQE;
916 
917 		if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) {
918 			if (scqe_sz == 128)
919 				in->ctx.cs_req = MLX5_REQ_SCAT_DATA64_CQE;
920 			else
921 				in->ctx.cs_req = MLX5_REQ_SCAT_DATA32_CQE;
922 		}
923 	}
924 
925 	if (qp->rq.wqe_cnt) {
926 		in->ctx.rq_size_stride = (qp->rq.wqe_shift - 4);
927 		in->ctx.rq_size_stride |= ilog2(qp->rq.wqe_cnt) << 3;
928 	}
929 
930 	in->ctx.rq_type_srqn = get_rx_type(qp, init_attr);
931 
932 	if (qp->sq.wqe_cnt)
933 		in->ctx.sq_crq_size |= cpu_to_be16(ilog2(qp->sq.wqe_cnt) << 11);
934 	else
935 		in->ctx.sq_crq_size |= cpu_to_be16(0x8000);
936 
937 	/* Set default resources */
938 	switch (init_attr->qp_type) {
939 	case IB_QPT_XRC_TGT:
940 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
941 		in->ctx.cqn_send = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
942 		in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
943 		in->ctx.xrcd = cpu_to_be32(to_mxrcd(init_attr->xrcd)->xrcdn);
944 		break;
945 	case IB_QPT_XRC_INI:
946 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
947 		in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
948 		in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
949 		break;
950 	default:
951 		if (init_attr->srq) {
952 			in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x0)->xrcdn);
953 			in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(init_attr->srq)->msrq.srqn);
954 		} else {
955 			in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
956 			in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
957 		}
958 	}
959 
960 	if (init_attr->send_cq)
961 		in->ctx.cqn_send = cpu_to_be32(to_mcq(init_attr->send_cq)->mcq.cqn);
962 
963 	if (init_attr->recv_cq)
964 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(init_attr->recv_cq)->mcq.cqn);
965 
966 	in->ctx.db_rec_addr = cpu_to_be64(qp->db.dma);
967 
968 	err = mlx5_core_create_qp(dev->mdev, &qp->mqp, in, inlen);
969 	if (err) {
970 		mlx5_ib_dbg(dev, "create qp failed\n");
971 		goto err_create;
972 	}
973 
974 	mlx5_vfree(in);
975 	/* Hardware wants QPN written in big-endian order (after
976 	 * shifting) for send doorbell.  Precompute this value to save
977 	 * a little bit when posting sends.
978 	 */
979 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
980 
981 	qp->mqp.event = mlx5_ib_qp_event;
982 
983 	return 0;
984 
985 err_create:
986 	if (qp->create_type == MLX5_QP_USER)
987 		destroy_qp_user(pd, qp);
988 	else if (qp->create_type == MLX5_QP_KERNEL)
989 		destroy_qp_kernel(dev, qp);
990 
991 	mlx5_vfree(in);
992 	return err;
993 }
994 
995 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
996 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
997 {
998 	if (send_cq) {
999 		if (recv_cq) {
1000 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1001 				spin_lock_irq(&send_cq->lock);
1002 				spin_lock_nested(&recv_cq->lock,
1003 						 SINGLE_DEPTH_NESTING);
1004 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1005 				spin_lock_irq(&send_cq->lock);
1006 				__acquire(&recv_cq->lock);
1007 			} else {
1008 				spin_lock_irq(&recv_cq->lock);
1009 				spin_lock_nested(&send_cq->lock,
1010 						 SINGLE_DEPTH_NESTING);
1011 			}
1012 		} else {
1013 			spin_lock_irq(&send_cq->lock);
1014 		}
1015 	} else if (recv_cq) {
1016 		spin_lock_irq(&recv_cq->lock);
1017 	}
1018 }
1019 
1020 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1021 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1022 {
1023 	if (send_cq) {
1024 		if (recv_cq) {
1025 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1026 				spin_unlock(&recv_cq->lock);
1027 				spin_unlock_irq(&send_cq->lock);
1028 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1029 				__release(&recv_cq->lock);
1030 				spin_unlock_irq(&send_cq->lock);
1031 			} else {
1032 				spin_unlock(&send_cq->lock);
1033 				spin_unlock_irq(&recv_cq->lock);
1034 			}
1035 		} else {
1036 			spin_unlock_irq(&send_cq->lock);
1037 		}
1038 	} else if (recv_cq) {
1039 		spin_unlock_irq(&recv_cq->lock);
1040 	}
1041 }
1042 
1043 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
1044 {
1045 	return to_mpd(qp->ibqp.pd);
1046 }
1047 
1048 static void get_cqs(struct mlx5_ib_qp *qp,
1049 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
1050 {
1051 	switch (qp->ibqp.qp_type) {
1052 	case IB_QPT_XRC_TGT:
1053 		*send_cq = NULL;
1054 		*recv_cq = NULL;
1055 		break;
1056 	case MLX5_IB_QPT_REG_UMR:
1057 	case IB_QPT_XRC_INI:
1058 		*send_cq = to_mcq(qp->ibqp.send_cq);
1059 		*recv_cq = NULL;
1060 		break;
1061 
1062 	case IB_QPT_SMI:
1063 	case IB_QPT_GSI:
1064 	case IB_QPT_RC:
1065 	case IB_QPT_UC:
1066 	case IB_QPT_UD:
1067 	case IB_QPT_RAW_IPV6:
1068 	case IB_QPT_RAW_ETHERTYPE:
1069 		*send_cq = to_mcq(qp->ibqp.send_cq);
1070 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
1071 		break;
1072 
1073 	case IB_QPT_RAW_PACKET:
1074 	case IB_QPT_MAX:
1075 	default:
1076 		*send_cq = NULL;
1077 		*recv_cq = NULL;
1078 		break;
1079 	}
1080 }
1081 
1082 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1083 {
1084 	struct mlx5_ib_cq *send_cq, *recv_cq;
1085 	struct mlx5_modify_qp_mbox_in *in;
1086 	int err;
1087 
1088 	in = kzalloc(sizeof(*in), GFP_KERNEL);
1089 	if (!in)
1090 		return;
1091 	if (qp->state != IB_QPS_RESET)
1092 		if (mlx5_core_qp_modify(dev->mdev, to_mlx5_state(qp->state),
1093 					MLX5_QP_STATE_RST, in, sizeof(*in), &qp->mqp))
1094 			mlx5_ib_warn(dev, "mlx5_ib: modify QP %06x to RESET failed\n",
1095 				     qp->mqp.qpn);
1096 
1097 	get_cqs(qp, &send_cq, &recv_cq);
1098 
1099 	if (qp->create_type == MLX5_QP_KERNEL) {
1100 		mlx5_ib_lock_cqs(send_cq, recv_cq);
1101 		__mlx5_ib_cq_clean(recv_cq, qp->mqp.qpn,
1102 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1103 		if (send_cq != recv_cq)
1104 			__mlx5_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1105 		mlx5_ib_unlock_cqs(send_cq, recv_cq);
1106 	}
1107 
1108 	err = mlx5_core_destroy_qp(dev->mdev, &qp->mqp);
1109 	if (err)
1110 		mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n", qp->mqp.qpn);
1111 	kfree(in);
1112 
1113 
1114 	if (qp->create_type == MLX5_QP_KERNEL)
1115 		destroy_qp_kernel(dev, qp);
1116 	else if (qp->create_type == MLX5_QP_USER)
1117 		destroy_qp_user(&get_pd(qp)->ibpd, qp);
1118 }
1119 
1120 static const char *ib_qp_type_str(enum ib_qp_type type)
1121 {
1122 	switch (type) {
1123 	case IB_QPT_SMI:
1124 		return "IB_QPT_SMI";
1125 	case IB_QPT_GSI:
1126 		return "IB_QPT_GSI";
1127 	case IB_QPT_RC:
1128 		return "IB_QPT_RC";
1129 	case IB_QPT_UC:
1130 		return "IB_QPT_UC";
1131 	case IB_QPT_UD:
1132 		return "IB_QPT_UD";
1133 	case IB_QPT_RAW_IPV6:
1134 		return "IB_QPT_RAW_IPV6";
1135 	case IB_QPT_RAW_ETHERTYPE:
1136 		return "IB_QPT_RAW_ETHERTYPE";
1137 	case IB_QPT_XRC_INI:
1138 		return "IB_QPT_XRC_INI";
1139 	case IB_QPT_XRC_TGT:
1140 		return "IB_QPT_XRC_TGT";
1141 	case IB_QPT_RAW_PACKET:
1142 		return "IB_QPT_RAW_PACKET";
1143 	case MLX5_IB_QPT_REG_UMR:
1144 		return "MLX5_IB_QPT_REG_UMR";
1145 	case IB_QPT_MAX:
1146 	default:
1147 		return "Invalid QP type";
1148 	}
1149 }
1150 
1151 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
1152 				struct ib_qp_init_attr *init_attr,
1153 				struct ib_udata *udata)
1154 {
1155 	struct mlx5_general_caps *gen;
1156 	struct mlx5_ib_dev *dev;
1157 	struct mlx5_ib_qp *qp;
1158 	u16 xrcdn = 0;
1159 	int err;
1160 
1161 	if (pd) {
1162 		dev = to_mdev(pd->device);
1163 	} else {
1164 		/* being cautious here */
1165 		if (init_attr->qp_type != IB_QPT_XRC_TGT &&
1166 		    init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
1167 			pr_warn("%s: no PD for transport %s\n", __func__,
1168 				ib_qp_type_str(init_attr->qp_type));
1169 			return ERR_PTR(-EINVAL);
1170 		}
1171 		dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
1172 	}
1173 	gen = &dev->mdev->caps.gen;
1174 
1175 	switch (init_attr->qp_type) {
1176 	case IB_QPT_XRC_TGT:
1177 	case IB_QPT_XRC_INI:
1178 		if (!(gen->flags & MLX5_DEV_CAP_FLAG_XRC)) {
1179 			mlx5_ib_dbg(dev, "XRC not supported\n");
1180 			return ERR_PTR(-ENOSYS);
1181 		}
1182 		init_attr->recv_cq = NULL;
1183 		if (init_attr->qp_type == IB_QPT_XRC_TGT) {
1184 			xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1185 			init_attr->send_cq = NULL;
1186 		}
1187 
1188 		/* fall through */
1189 	case IB_QPT_RC:
1190 	case IB_QPT_UC:
1191 	case IB_QPT_UD:
1192 	case IB_QPT_SMI:
1193 	case IB_QPT_GSI:
1194 	case MLX5_IB_QPT_REG_UMR:
1195 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1196 		if (!qp)
1197 			return ERR_PTR(-ENOMEM);
1198 
1199 		err = create_qp_common(dev, pd, init_attr, udata, qp);
1200 		if (err) {
1201 			mlx5_ib_dbg(dev, "create_qp_common failed\n");
1202 			kfree(qp);
1203 			return ERR_PTR(err);
1204 		}
1205 
1206 		if (is_qp0(init_attr->qp_type))
1207 			qp->ibqp.qp_num = 0;
1208 		else if (is_qp1(init_attr->qp_type))
1209 			qp->ibqp.qp_num = 1;
1210 		else
1211 			qp->ibqp.qp_num = qp->mqp.qpn;
1212 
1213 		mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
1214 			    qp->ibqp.qp_num, qp->mqp.qpn, to_mcq(init_attr->recv_cq)->mcq.cqn,
1215 			    to_mcq(init_attr->send_cq)->mcq.cqn);
1216 
1217 		qp->xrcdn = xrcdn;
1218 
1219 		break;
1220 
1221 	case IB_QPT_RAW_IPV6:
1222 	case IB_QPT_RAW_ETHERTYPE:
1223 	case IB_QPT_RAW_PACKET:
1224 	case IB_QPT_MAX:
1225 	default:
1226 		mlx5_ib_dbg(dev, "unsupported qp type %d\n",
1227 			    init_attr->qp_type);
1228 		/* Don't support raw QPs */
1229 		return ERR_PTR(-EINVAL);
1230 	}
1231 
1232 	return &qp->ibqp;
1233 }
1234 
1235 int mlx5_ib_destroy_qp(struct ib_qp *qp)
1236 {
1237 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
1238 	struct mlx5_ib_qp *mqp = to_mqp(qp);
1239 
1240 	destroy_qp_common(dev, mqp);
1241 
1242 	kfree(mqp);
1243 
1244 	return 0;
1245 }
1246 
1247 static __be32 to_mlx5_access_flags(struct mlx5_ib_qp *qp, const struct ib_qp_attr *attr,
1248 				   int attr_mask)
1249 {
1250 	u32 hw_access_flags = 0;
1251 	u8 dest_rd_atomic;
1252 	u32 access_flags;
1253 
1254 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1255 		dest_rd_atomic = attr->max_dest_rd_atomic;
1256 	else
1257 		dest_rd_atomic = qp->resp_depth;
1258 
1259 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1260 		access_flags = attr->qp_access_flags;
1261 	else
1262 		access_flags = qp->atomic_rd_en;
1263 
1264 	if (!dest_rd_atomic)
1265 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1266 
1267 	if (access_flags & IB_ACCESS_REMOTE_READ)
1268 		hw_access_flags |= MLX5_QP_BIT_RRE;
1269 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1270 		hw_access_flags |= (MLX5_QP_BIT_RAE | MLX5_ATOMIC_MODE_CX);
1271 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1272 		hw_access_flags |= MLX5_QP_BIT_RWE;
1273 
1274 	return cpu_to_be32(hw_access_flags);
1275 }
1276 
1277 enum {
1278 	MLX5_PATH_FLAG_FL	= 1 << 0,
1279 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
1280 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
1281 };
1282 
1283 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
1284 {
1285 	struct mlx5_general_caps *gen;
1286 
1287 	gen = &dev->mdev->caps.gen;
1288 	if (rate == IB_RATE_PORT_CURRENT) {
1289 		return 0;
1290 	} else if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_300_GBPS) {
1291 		return -EINVAL;
1292 	} else {
1293 		while (rate != IB_RATE_2_5_GBPS &&
1294 		       !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
1295 			 gen->stat_rate_support))
1296 			--rate;
1297 	}
1298 
1299 	return rate + MLX5_STAT_RATE_OFFSET;
1300 }
1301 
1302 static int mlx5_set_path(struct mlx5_ib_dev *dev, const struct ib_ah_attr *ah,
1303 			 struct mlx5_qp_path *path, u8 port, int attr_mask,
1304 			 u32 path_flags, const struct ib_qp_attr *attr)
1305 {
1306 	struct mlx5_general_caps *gen;
1307 	int err;
1308 
1309 	gen = &dev->mdev->caps.gen;
1310 	path->fl = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
1311 	path->free_ar = (path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x80 : 0;
1312 
1313 	if (attr_mask & IB_QP_PKEY_INDEX)
1314 		path->pkey_index = attr->pkey_index;
1315 
1316 	path->grh_mlid	= ah->src_path_bits & 0x7f;
1317 	path->rlid	= cpu_to_be16(ah->dlid);
1318 
1319 	if (ah->ah_flags & IB_AH_GRH) {
1320 		if (ah->grh.sgid_index >= gen->port[port - 1].gid_table_len) {
1321 			pr_err(KERN_ERR "sgid_index (%u) too large. max is %d\n",
1322 			       ah->grh.sgid_index, gen->port[port - 1].gid_table_len);
1323 			return -EINVAL;
1324 		}
1325 		path->grh_mlid |= 1 << 7;
1326 		path->mgid_index = ah->grh.sgid_index;
1327 		path->hop_limit  = ah->grh.hop_limit;
1328 		path->tclass_flowlabel =
1329 			cpu_to_be32((ah->grh.traffic_class << 20) |
1330 				    (ah->grh.flow_label));
1331 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1332 	}
1333 
1334 	err = ib_rate_to_mlx5(dev, ah->static_rate);
1335 	if (err < 0)
1336 		return err;
1337 	path->static_rate = err;
1338 	path->port = port;
1339 
1340 	if (attr_mask & IB_QP_TIMEOUT)
1341 		path->ackto_lt = attr->timeout << 3;
1342 
1343 	path->sl = ah->sl & 0xf;
1344 
1345 	return 0;
1346 }
1347 
1348 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
1349 	[MLX5_QP_STATE_INIT] = {
1350 		[MLX5_QP_STATE_INIT] = {
1351 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
1352 					  MLX5_QP_OPTPAR_RAE		|
1353 					  MLX5_QP_OPTPAR_RWE		|
1354 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
1355 					  MLX5_QP_OPTPAR_PRI_PORT,
1356 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
1357 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
1358 					  MLX5_QP_OPTPAR_PRI_PORT,
1359 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
1360 					  MLX5_QP_OPTPAR_Q_KEY		|
1361 					  MLX5_QP_OPTPAR_PRI_PORT,
1362 		},
1363 		[MLX5_QP_STATE_RTR] = {
1364 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
1365 					  MLX5_QP_OPTPAR_RRE            |
1366 					  MLX5_QP_OPTPAR_RAE            |
1367 					  MLX5_QP_OPTPAR_RWE            |
1368 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1369 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
1370 					  MLX5_QP_OPTPAR_RWE            |
1371 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1372 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
1373 					  MLX5_QP_OPTPAR_Q_KEY,
1374 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
1375 					   MLX5_QP_OPTPAR_Q_KEY,
1376 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
1377 					  MLX5_QP_OPTPAR_RRE            |
1378 					  MLX5_QP_OPTPAR_RAE            |
1379 					  MLX5_QP_OPTPAR_RWE            |
1380 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1381 		},
1382 	},
1383 	[MLX5_QP_STATE_RTR] = {
1384 		[MLX5_QP_STATE_RTS] = {
1385 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
1386 					  MLX5_QP_OPTPAR_RRE		|
1387 					  MLX5_QP_OPTPAR_RAE		|
1388 					  MLX5_QP_OPTPAR_RWE		|
1389 					  MLX5_QP_OPTPAR_PM_STATE	|
1390 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
1391 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
1392 					  MLX5_QP_OPTPAR_RWE		|
1393 					  MLX5_QP_OPTPAR_PM_STATE,
1394 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
1395 		},
1396 	},
1397 	[MLX5_QP_STATE_RTS] = {
1398 		[MLX5_QP_STATE_RTS] = {
1399 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
1400 					  MLX5_QP_OPTPAR_RAE		|
1401 					  MLX5_QP_OPTPAR_RWE		|
1402 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
1403 					  MLX5_QP_OPTPAR_PM_STATE	|
1404 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
1405 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
1406 					  MLX5_QP_OPTPAR_PM_STATE	|
1407 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
1408 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
1409 					  MLX5_QP_OPTPAR_SRQN		|
1410 					  MLX5_QP_OPTPAR_CQN_RCV,
1411 		},
1412 	},
1413 	[MLX5_QP_STATE_SQER] = {
1414 		[MLX5_QP_STATE_RTS] = {
1415 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
1416 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
1417 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
1418 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
1419 					   MLX5_QP_OPTPAR_RWE		|
1420 					   MLX5_QP_OPTPAR_RAE		|
1421 					   MLX5_QP_OPTPAR_RRE,
1422 		},
1423 	},
1424 };
1425 
1426 static int ib_nr_to_mlx5_nr(int ib_mask)
1427 {
1428 	switch (ib_mask) {
1429 	case IB_QP_STATE:
1430 		return 0;
1431 	case IB_QP_CUR_STATE:
1432 		return 0;
1433 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
1434 		return 0;
1435 	case IB_QP_ACCESS_FLAGS:
1436 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
1437 			MLX5_QP_OPTPAR_RAE;
1438 	case IB_QP_PKEY_INDEX:
1439 		return MLX5_QP_OPTPAR_PKEY_INDEX;
1440 	case IB_QP_PORT:
1441 		return MLX5_QP_OPTPAR_PRI_PORT;
1442 	case IB_QP_QKEY:
1443 		return MLX5_QP_OPTPAR_Q_KEY;
1444 	case IB_QP_AV:
1445 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
1446 			MLX5_QP_OPTPAR_PRI_PORT;
1447 	case IB_QP_PATH_MTU:
1448 		return 0;
1449 	case IB_QP_TIMEOUT:
1450 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
1451 	case IB_QP_RETRY_CNT:
1452 		return MLX5_QP_OPTPAR_RETRY_COUNT;
1453 	case IB_QP_RNR_RETRY:
1454 		return MLX5_QP_OPTPAR_RNR_RETRY;
1455 	case IB_QP_RQ_PSN:
1456 		return 0;
1457 	case IB_QP_MAX_QP_RD_ATOMIC:
1458 		return MLX5_QP_OPTPAR_SRA_MAX;
1459 	case IB_QP_ALT_PATH:
1460 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
1461 	case IB_QP_MIN_RNR_TIMER:
1462 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
1463 	case IB_QP_SQ_PSN:
1464 		return 0;
1465 	case IB_QP_MAX_DEST_RD_ATOMIC:
1466 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
1467 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
1468 	case IB_QP_PATH_MIG_STATE:
1469 		return MLX5_QP_OPTPAR_PM_STATE;
1470 	case IB_QP_CAP:
1471 		return 0;
1472 	case IB_QP_DEST_QPN:
1473 		return 0;
1474 	}
1475 	return 0;
1476 }
1477 
1478 static int ib_mask_to_mlx5_opt(int ib_mask)
1479 {
1480 	int result = 0;
1481 	int i;
1482 
1483 	for (i = 0; i < 8 * sizeof(int); i++) {
1484 		if ((1 << i) & ib_mask)
1485 			result |= ib_nr_to_mlx5_nr(1 << i);
1486 	}
1487 
1488 	return result;
1489 }
1490 
1491 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
1492 			       const struct ib_qp_attr *attr, int attr_mask,
1493 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1494 {
1495 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1496 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
1497 	struct mlx5_ib_cq *send_cq, *recv_cq;
1498 	struct mlx5_qp_context *context;
1499 	struct mlx5_general_caps *gen;
1500 	struct mlx5_modify_qp_mbox_in *in;
1501 	struct mlx5_ib_pd *pd;
1502 	enum mlx5_qp_state mlx5_cur, mlx5_new;
1503 	enum mlx5_qp_optpar optpar;
1504 	int sqd_event;
1505 	int mlx5_st;
1506 	int err;
1507 
1508 	gen = &dev->mdev->caps.gen;
1509 	in = kzalloc(sizeof(*in), GFP_KERNEL);
1510 	if (!in)
1511 		return -ENOMEM;
1512 
1513 	context = &in->ctx;
1514 	err = to_mlx5_st(ibqp->qp_type);
1515 	if (err < 0)
1516 		goto out;
1517 
1518 	context->flags = cpu_to_be32(err << 16);
1519 
1520 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
1521 		context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
1522 	} else {
1523 		switch (attr->path_mig_state) {
1524 		case IB_MIG_MIGRATED:
1525 			context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
1526 			break;
1527 		case IB_MIG_REARM:
1528 			context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
1529 			break;
1530 		case IB_MIG_ARMED:
1531 			context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
1532 			break;
1533 		}
1534 	}
1535 
1536 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI) {
1537 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
1538 	} else if (ibqp->qp_type == IB_QPT_UD ||
1539 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
1540 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1541 	} else if (attr_mask & IB_QP_PATH_MTU) {
1542 		if (attr->path_mtu < IB_MTU_256 ||
1543 		    attr->path_mtu > IB_MTU_4096) {
1544 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
1545 			err = -EINVAL;
1546 			goto out;
1547 		}
1548 		context->mtu_msgmax = (attr->path_mtu << 5) | gen->log_max_msg;
1549 	}
1550 
1551 	if (attr_mask & IB_QP_DEST_QPN)
1552 		context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
1553 
1554 	if (attr_mask & IB_QP_PKEY_INDEX)
1555 		context->pri_path.pkey_index = attr->pkey_index;
1556 
1557 	/* todo implement counter_index functionality */
1558 
1559 	if (is_sqp(ibqp->qp_type))
1560 		context->pri_path.port = qp->port;
1561 
1562 	if (attr_mask & IB_QP_PORT)
1563 		context->pri_path.port = attr->port_num;
1564 
1565 	if (attr_mask & IB_QP_AV) {
1566 		err = mlx5_set_path(dev, &attr->ah_attr, &context->pri_path,
1567 				    attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
1568 				    attr_mask, 0, attr);
1569 		if (err)
1570 			goto out;
1571 	}
1572 
1573 	if (attr_mask & IB_QP_TIMEOUT)
1574 		context->pri_path.ackto_lt |= attr->timeout << 3;
1575 
1576 	if (attr_mask & IB_QP_ALT_PATH) {
1577 		err = mlx5_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
1578 				    attr->alt_port_num, attr_mask, 0, attr);
1579 		if (err)
1580 			goto out;
1581 	}
1582 
1583 	pd = get_pd(qp);
1584 	get_cqs(qp, &send_cq, &recv_cq);
1585 
1586 	context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
1587 	context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
1588 	context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
1589 	context->params1  = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
1590 
1591 	if (attr_mask & IB_QP_RNR_RETRY)
1592 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1593 
1594 	if (attr_mask & IB_QP_RETRY_CNT)
1595 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1596 
1597 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1598 		if (attr->max_rd_atomic)
1599 			context->params1 |=
1600 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1601 	}
1602 
1603 	if (attr_mask & IB_QP_SQ_PSN)
1604 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1605 
1606 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1607 		if (attr->max_dest_rd_atomic)
1608 			context->params2 |=
1609 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1610 	}
1611 
1612 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
1613 		context->params2 |= to_mlx5_access_flags(qp, attr, attr_mask);
1614 
1615 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
1616 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1617 
1618 	if (attr_mask & IB_QP_RQ_PSN)
1619 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1620 
1621 	if (attr_mask & IB_QP_QKEY)
1622 		context->qkey = cpu_to_be32(attr->qkey);
1623 
1624 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1625 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1626 
1627 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1628 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1629 		sqd_event = 1;
1630 	else
1631 		sqd_event = 0;
1632 
1633 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1634 		context->sq_crq_size |= cpu_to_be16(1 << 4);
1635 
1636 
1637 	mlx5_cur = to_mlx5_state(cur_state);
1638 	mlx5_new = to_mlx5_state(new_state);
1639 	mlx5_st = to_mlx5_st(ibqp->qp_type);
1640 	if (mlx5_st < 0)
1641 		goto out;
1642 
1643 	optpar = ib_mask_to_mlx5_opt(attr_mask);
1644 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
1645 	in->optparam = cpu_to_be32(optpar);
1646 	err = mlx5_core_qp_modify(dev->mdev, to_mlx5_state(cur_state),
1647 				  to_mlx5_state(new_state), in, sqd_event,
1648 				  &qp->mqp);
1649 	if (err)
1650 		goto out;
1651 
1652 	qp->state = new_state;
1653 
1654 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1655 		qp->atomic_rd_en = attr->qp_access_flags;
1656 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1657 		qp->resp_depth = attr->max_dest_rd_atomic;
1658 	if (attr_mask & IB_QP_PORT)
1659 		qp->port = attr->port_num;
1660 	if (attr_mask & IB_QP_ALT_PATH)
1661 		qp->alt_port = attr->alt_port_num;
1662 
1663 	/*
1664 	 * If we moved a kernel QP to RESET, clean up all old CQ
1665 	 * entries and reinitialize the QP.
1666 	 */
1667 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1668 		mlx5_ib_cq_clean(recv_cq, qp->mqp.qpn,
1669 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
1670 		if (send_cq != recv_cq)
1671 			mlx5_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1672 
1673 		qp->rq.head = 0;
1674 		qp->rq.tail = 0;
1675 		qp->sq.head = 0;
1676 		qp->sq.tail = 0;
1677 		qp->sq.cur_post = 0;
1678 		qp->sq.last_poll = 0;
1679 		qp->db.db[MLX5_RCV_DBR] = 0;
1680 		qp->db.db[MLX5_SND_DBR] = 0;
1681 	}
1682 
1683 out:
1684 	kfree(in);
1685 	return err;
1686 }
1687 
1688 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1689 		      int attr_mask, struct ib_udata *udata)
1690 {
1691 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1692 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
1693 	enum ib_qp_state cur_state, new_state;
1694 	struct mlx5_general_caps *gen;
1695 	int err = -EINVAL;
1696 	int port;
1697 
1698 	gen = &dev->mdev->caps.gen;
1699 	mutex_lock(&qp->mutex);
1700 
1701 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1702 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1703 
1704 	if (ibqp->qp_type != MLX5_IB_QPT_REG_UMR &&
1705 	    !ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask,
1706 				IB_LINK_LAYER_UNSPECIFIED))
1707 		goto out;
1708 
1709 	if ((attr_mask & IB_QP_PORT) &&
1710 	    (attr->port_num == 0 || attr->port_num > gen->num_ports))
1711 		goto out;
1712 
1713 	if (attr_mask & IB_QP_PKEY_INDEX) {
1714 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1715 		if (attr->pkey_index >= gen->port[port - 1].pkey_table_len)
1716 			goto out;
1717 	}
1718 
1719 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1720 	    attr->max_rd_atomic > (1 << gen->log_max_ra_res_qp))
1721 		goto out;
1722 
1723 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1724 	    attr->max_dest_rd_atomic > (1 << gen->log_max_ra_req_qp))
1725 		goto out;
1726 
1727 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1728 		err = 0;
1729 		goto out;
1730 	}
1731 
1732 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1733 
1734 out:
1735 	mutex_unlock(&qp->mutex);
1736 	return err;
1737 }
1738 
1739 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1740 {
1741 	struct mlx5_ib_cq *cq;
1742 	unsigned cur;
1743 
1744 	cur = wq->head - wq->tail;
1745 	if (likely(cur + nreq < wq->max_post))
1746 		return 0;
1747 
1748 	cq = to_mcq(ib_cq);
1749 	spin_lock(&cq->lock);
1750 	cur = wq->head - wq->tail;
1751 	spin_unlock(&cq->lock);
1752 
1753 	return cur + nreq >= wq->max_post;
1754 }
1755 
1756 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
1757 					  u64 remote_addr, u32 rkey)
1758 {
1759 	rseg->raddr    = cpu_to_be64(remote_addr);
1760 	rseg->rkey     = cpu_to_be32(rkey);
1761 	rseg->reserved = 0;
1762 }
1763 
1764 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
1765 			     struct ib_send_wr *wr)
1766 {
1767 	memcpy(&dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof(struct mlx5_av));
1768 	dseg->av.dqp_dct = cpu_to_be32(wr->wr.ud.remote_qpn | MLX5_EXTENDED_UD_AV);
1769 	dseg->av.key.qkey.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1770 }
1771 
1772 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
1773 {
1774 	dseg->byte_count = cpu_to_be32(sg->length);
1775 	dseg->lkey       = cpu_to_be32(sg->lkey);
1776 	dseg->addr       = cpu_to_be64(sg->addr);
1777 }
1778 
1779 static __be16 get_klm_octo(int npages)
1780 {
1781 	return cpu_to_be16(ALIGN(npages, 8) / 2);
1782 }
1783 
1784 static __be64 frwr_mkey_mask(void)
1785 {
1786 	u64 result;
1787 
1788 	result = MLX5_MKEY_MASK_LEN		|
1789 		MLX5_MKEY_MASK_PAGE_SIZE	|
1790 		MLX5_MKEY_MASK_START_ADDR	|
1791 		MLX5_MKEY_MASK_EN_RINVAL	|
1792 		MLX5_MKEY_MASK_KEY		|
1793 		MLX5_MKEY_MASK_LR		|
1794 		MLX5_MKEY_MASK_LW		|
1795 		MLX5_MKEY_MASK_RR		|
1796 		MLX5_MKEY_MASK_RW		|
1797 		MLX5_MKEY_MASK_A		|
1798 		MLX5_MKEY_MASK_SMALL_FENCE	|
1799 		MLX5_MKEY_MASK_FREE;
1800 
1801 	return cpu_to_be64(result);
1802 }
1803 
1804 static __be64 sig_mkey_mask(void)
1805 {
1806 	u64 result;
1807 
1808 	result = MLX5_MKEY_MASK_LEN		|
1809 		MLX5_MKEY_MASK_PAGE_SIZE	|
1810 		MLX5_MKEY_MASK_START_ADDR	|
1811 		MLX5_MKEY_MASK_EN_SIGERR	|
1812 		MLX5_MKEY_MASK_EN_RINVAL	|
1813 		MLX5_MKEY_MASK_KEY		|
1814 		MLX5_MKEY_MASK_LR		|
1815 		MLX5_MKEY_MASK_LW		|
1816 		MLX5_MKEY_MASK_RR		|
1817 		MLX5_MKEY_MASK_RW		|
1818 		MLX5_MKEY_MASK_SMALL_FENCE	|
1819 		MLX5_MKEY_MASK_FREE		|
1820 		MLX5_MKEY_MASK_BSF_EN;
1821 
1822 	return cpu_to_be64(result);
1823 }
1824 
1825 static void set_frwr_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
1826 				 struct ib_send_wr *wr, int li)
1827 {
1828 	memset(umr, 0, sizeof(*umr));
1829 
1830 	if (li) {
1831 		umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
1832 		umr->flags = 1 << 7;
1833 		return;
1834 	}
1835 
1836 	umr->flags = (1 << 5); /* fail if not free */
1837 	umr->klm_octowords = get_klm_octo(wr->wr.fast_reg.page_list_len);
1838 	umr->mkey_mask = frwr_mkey_mask();
1839 }
1840 
1841 static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
1842 				struct ib_send_wr *wr)
1843 {
1844 	struct umr_wr *umrwr = (struct umr_wr *)&wr->wr.fast_reg;
1845 	u64 mask;
1846 
1847 	memset(umr, 0, sizeof(*umr));
1848 
1849 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UNREG)) {
1850 		umr->flags = 1 << 5; /* fail if not free */
1851 		umr->klm_octowords = get_klm_octo(umrwr->npages);
1852 		mask =  MLX5_MKEY_MASK_LEN		|
1853 			MLX5_MKEY_MASK_PAGE_SIZE	|
1854 			MLX5_MKEY_MASK_START_ADDR	|
1855 			MLX5_MKEY_MASK_PD		|
1856 			MLX5_MKEY_MASK_LR		|
1857 			MLX5_MKEY_MASK_LW		|
1858 			MLX5_MKEY_MASK_KEY		|
1859 			MLX5_MKEY_MASK_RR		|
1860 			MLX5_MKEY_MASK_RW		|
1861 			MLX5_MKEY_MASK_A		|
1862 			MLX5_MKEY_MASK_FREE;
1863 		umr->mkey_mask = cpu_to_be64(mask);
1864 	} else {
1865 		umr->flags = 2 << 5; /* fail if free */
1866 		mask = MLX5_MKEY_MASK_FREE;
1867 		umr->mkey_mask = cpu_to_be64(mask);
1868 	}
1869 
1870 	if (!wr->num_sge)
1871 		umr->flags |= (1 << 7); /* inline */
1872 }
1873 
1874 static u8 get_umr_flags(int acc)
1875 {
1876 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
1877 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
1878 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
1879 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
1880 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
1881 }
1882 
1883 static void set_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr,
1884 			     int li, int *writ)
1885 {
1886 	memset(seg, 0, sizeof(*seg));
1887 	if (li) {
1888 		seg->status = 1 << 6;
1889 		return;
1890 	}
1891 
1892 	seg->flags = get_umr_flags(wr->wr.fast_reg.access_flags) |
1893 		     MLX5_ACCESS_MODE_MTT;
1894 	*writ = seg->flags & (MLX5_PERM_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE);
1895 	seg->qpn_mkey7_0 = cpu_to_be32((wr->wr.fast_reg.rkey & 0xff) | 0xffffff00);
1896 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
1897 	seg->start_addr = cpu_to_be64(wr->wr.fast_reg.iova_start);
1898 	seg->len = cpu_to_be64(wr->wr.fast_reg.length);
1899 	seg->xlt_oct_size = cpu_to_be32((wr->wr.fast_reg.page_list_len + 1) / 2);
1900 	seg->log2_page_size = wr->wr.fast_reg.page_shift;
1901 }
1902 
1903 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr)
1904 {
1905 	memset(seg, 0, sizeof(*seg));
1906 	if (wr->send_flags & MLX5_IB_SEND_UMR_UNREG) {
1907 		seg->status = 1 << 6;
1908 		return;
1909 	}
1910 
1911 	seg->flags = convert_access(wr->wr.fast_reg.access_flags);
1912 	seg->flags_pd = cpu_to_be32(to_mpd((struct ib_pd *)wr->wr.fast_reg.page_list)->pdn);
1913 	seg->start_addr = cpu_to_be64(wr->wr.fast_reg.iova_start);
1914 	seg->len = cpu_to_be64(wr->wr.fast_reg.length);
1915 	seg->log2_page_size = wr->wr.fast_reg.page_shift;
1916 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
1917 				       mlx5_mkey_variant(wr->wr.fast_reg.rkey));
1918 }
1919 
1920 static void set_frwr_pages(struct mlx5_wqe_data_seg *dseg,
1921 			   struct ib_send_wr *wr,
1922 			   struct mlx5_core_dev *mdev,
1923 			   struct mlx5_ib_pd *pd,
1924 			   int writ)
1925 {
1926 	struct mlx5_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
1927 	u64 *page_list = wr->wr.fast_reg.page_list->page_list;
1928 	u64 perm = MLX5_EN_RD | (writ ? MLX5_EN_WR : 0);
1929 	int i;
1930 
1931 	for (i = 0; i < wr->wr.fast_reg.page_list_len; i++)
1932 		mfrpl->mapped_page_list[i] = cpu_to_be64(page_list[i] | perm);
1933 	dseg->addr = cpu_to_be64(mfrpl->map);
1934 	dseg->byte_count = cpu_to_be32(ALIGN(sizeof(u64) * wr->wr.fast_reg.page_list_len, 64));
1935 	dseg->lkey = cpu_to_be32(pd->pa_lkey);
1936 }
1937 
1938 static __be32 send_ieth(struct ib_send_wr *wr)
1939 {
1940 	switch (wr->opcode) {
1941 	case IB_WR_SEND_WITH_IMM:
1942 	case IB_WR_RDMA_WRITE_WITH_IMM:
1943 		return wr->ex.imm_data;
1944 
1945 	case IB_WR_SEND_WITH_INV:
1946 		return cpu_to_be32(wr->ex.invalidate_rkey);
1947 
1948 	default:
1949 		return 0;
1950 	}
1951 }
1952 
1953 static u8 calc_sig(void *wqe, int size)
1954 {
1955 	u8 *p = wqe;
1956 	u8 res = 0;
1957 	int i;
1958 
1959 	for (i = 0; i < size; i++)
1960 		res ^= p[i];
1961 
1962 	return ~res;
1963 }
1964 
1965 static u8 wq_sig(void *wqe)
1966 {
1967 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
1968 }
1969 
1970 static int set_data_inl_seg(struct mlx5_ib_qp *qp, struct ib_send_wr *wr,
1971 			    void *wqe, int *sz)
1972 {
1973 	struct mlx5_wqe_inline_seg *seg;
1974 	void *qend = qp->sq.qend;
1975 	void *addr;
1976 	int inl = 0;
1977 	int copy;
1978 	int len;
1979 	int i;
1980 
1981 	seg = wqe;
1982 	wqe += sizeof(*seg);
1983 	for (i = 0; i < wr->num_sge; i++) {
1984 		addr = (void *)(unsigned long)(wr->sg_list[i].addr);
1985 		len  = wr->sg_list[i].length;
1986 		inl += len;
1987 
1988 		if (unlikely(inl > qp->max_inline_data))
1989 			return -ENOMEM;
1990 
1991 		if (unlikely(wqe + len > qend)) {
1992 			copy = qend - wqe;
1993 			memcpy(wqe, addr, copy);
1994 			addr += copy;
1995 			len -= copy;
1996 			wqe = mlx5_get_send_wqe(qp, 0);
1997 		}
1998 		memcpy(wqe, addr, len);
1999 		wqe += len;
2000 	}
2001 
2002 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
2003 
2004 	*sz = ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
2005 
2006 	return 0;
2007 }
2008 
2009 static u16 prot_field_size(enum ib_signature_type type)
2010 {
2011 	switch (type) {
2012 	case IB_SIG_TYPE_T10_DIF:
2013 		return MLX5_DIF_SIZE;
2014 	default:
2015 		return 0;
2016 	}
2017 }
2018 
2019 static u8 bs_selector(int block_size)
2020 {
2021 	switch (block_size) {
2022 	case 512:	    return 0x1;
2023 	case 520:	    return 0x2;
2024 	case 4096:	    return 0x3;
2025 	case 4160:	    return 0x4;
2026 	case 1073741824:    return 0x5;
2027 	default:	    return 0;
2028 	}
2029 }
2030 
2031 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
2032 			      struct mlx5_bsf_inl *inl)
2033 {
2034 	/* Valid inline section and allow BSF refresh */
2035 	inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
2036 				       MLX5_BSF_REFRESH_DIF);
2037 	inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
2038 	inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
2039 	/* repeating block */
2040 	inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
2041 	inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
2042 			MLX5_DIF_CRC : MLX5_DIF_IPCS;
2043 
2044 	if (domain->sig.dif.ref_remap)
2045 		inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
2046 
2047 	if (domain->sig.dif.app_escape) {
2048 		if (domain->sig.dif.ref_escape)
2049 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
2050 		else
2051 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
2052 	}
2053 
2054 	inl->dif_app_bitmask_check =
2055 		cpu_to_be16(domain->sig.dif.apptag_check_mask);
2056 }
2057 
2058 static int mlx5_set_bsf(struct ib_mr *sig_mr,
2059 			struct ib_sig_attrs *sig_attrs,
2060 			struct mlx5_bsf *bsf, u32 data_size)
2061 {
2062 	struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
2063 	struct mlx5_bsf_basic *basic = &bsf->basic;
2064 	struct ib_sig_domain *mem = &sig_attrs->mem;
2065 	struct ib_sig_domain *wire = &sig_attrs->wire;
2066 
2067 	memset(bsf, 0, sizeof(*bsf));
2068 
2069 	/* Basic + Extended + Inline */
2070 	basic->bsf_size_sbs = 1 << 7;
2071 	/* Input domain check byte mask */
2072 	basic->check_byte_mask = sig_attrs->check_mask;
2073 	basic->raw_data_size = cpu_to_be32(data_size);
2074 
2075 	/* Memory domain */
2076 	switch (sig_attrs->mem.sig_type) {
2077 	case IB_SIG_TYPE_NONE:
2078 		break;
2079 	case IB_SIG_TYPE_T10_DIF:
2080 		basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
2081 		basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
2082 		mlx5_fill_inl_bsf(mem, &bsf->m_inl);
2083 		break;
2084 	default:
2085 		return -EINVAL;
2086 	}
2087 
2088 	/* Wire domain */
2089 	switch (sig_attrs->wire.sig_type) {
2090 	case IB_SIG_TYPE_NONE:
2091 		break;
2092 	case IB_SIG_TYPE_T10_DIF:
2093 		if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
2094 		    mem->sig_type == wire->sig_type) {
2095 			/* Same block structure */
2096 			basic->bsf_size_sbs |= 1 << 4;
2097 			if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
2098 				basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
2099 			if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
2100 				basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
2101 			if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
2102 				basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
2103 		} else
2104 			basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
2105 
2106 		basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
2107 		mlx5_fill_inl_bsf(wire, &bsf->w_inl);
2108 		break;
2109 	default:
2110 		return -EINVAL;
2111 	}
2112 
2113 	return 0;
2114 }
2115 
2116 static int set_sig_data_segment(struct ib_send_wr *wr, struct mlx5_ib_qp *qp,
2117 				void **seg, int *size)
2118 {
2119 	struct ib_sig_attrs *sig_attrs = wr->wr.sig_handover.sig_attrs;
2120 	struct ib_mr *sig_mr = wr->wr.sig_handover.sig_mr;
2121 	struct mlx5_bsf *bsf;
2122 	u32 data_len = wr->sg_list->length;
2123 	u32 data_key = wr->sg_list->lkey;
2124 	u64 data_va = wr->sg_list->addr;
2125 	int ret;
2126 	int wqe_size;
2127 
2128 	if (!wr->wr.sig_handover.prot ||
2129 	    (data_key == wr->wr.sig_handover.prot->lkey &&
2130 	     data_va == wr->wr.sig_handover.prot->addr &&
2131 	     data_len == wr->wr.sig_handover.prot->length)) {
2132 		/**
2133 		 * Source domain doesn't contain signature information
2134 		 * or data and protection are interleaved in memory.
2135 		 * So need construct:
2136 		 *                  ------------------
2137 		 *                 |     data_klm     |
2138 		 *                  ------------------
2139 		 *                 |       BSF        |
2140 		 *                  ------------------
2141 		 **/
2142 		struct mlx5_klm *data_klm = *seg;
2143 
2144 		data_klm->bcount = cpu_to_be32(data_len);
2145 		data_klm->key = cpu_to_be32(data_key);
2146 		data_klm->va = cpu_to_be64(data_va);
2147 		wqe_size = ALIGN(sizeof(*data_klm), 64);
2148 	} else {
2149 		/**
2150 		 * Source domain contains signature information
2151 		 * So need construct a strided block format:
2152 		 *               ---------------------------
2153 		 *              |     stride_block_ctrl     |
2154 		 *               ---------------------------
2155 		 *              |          data_klm         |
2156 		 *               ---------------------------
2157 		 *              |          prot_klm         |
2158 		 *               ---------------------------
2159 		 *              |             BSF           |
2160 		 *               ---------------------------
2161 		 **/
2162 		struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
2163 		struct mlx5_stride_block_entry *data_sentry;
2164 		struct mlx5_stride_block_entry *prot_sentry;
2165 		u32 prot_key = wr->wr.sig_handover.prot->lkey;
2166 		u64 prot_va = wr->wr.sig_handover.prot->addr;
2167 		u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
2168 		int prot_size;
2169 
2170 		sblock_ctrl = *seg;
2171 		data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
2172 		prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
2173 
2174 		prot_size = prot_field_size(sig_attrs->mem.sig_type);
2175 		if (!prot_size) {
2176 			pr_err("Bad block size given: %u\n", block_size);
2177 			return -EINVAL;
2178 		}
2179 		sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
2180 							    prot_size);
2181 		sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
2182 		sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
2183 		sblock_ctrl->num_entries = cpu_to_be16(2);
2184 
2185 		data_sentry->bcount = cpu_to_be16(block_size);
2186 		data_sentry->key = cpu_to_be32(data_key);
2187 		data_sentry->va = cpu_to_be64(data_va);
2188 		data_sentry->stride = cpu_to_be16(block_size);
2189 
2190 		prot_sentry->bcount = cpu_to_be16(prot_size);
2191 		prot_sentry->key = cpu_to_be32(prot_key);
2192 		prot_sentry->va = cpu_to_be64(prot_va);
2193 		prot_sentry->stride = cpu_to_be16(prot_size);
2194 
2195 		wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
2196 				 sizeof(*prot_sentry), 64);
2197 	}
2198 
2199 	*seg += wqe_size;
2200 	*size += wqe_size / 16;
2201 	if (unlikely((*seg == qp->sq.qend)))
2202 		*seg = mlx5_get_send_wqe(qp, 0);
2203 
2204 	bsf = *seg;
2205 	ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
2206 	if (ret)
2207 		return -EINVAL;
2208 
2209 	*seg += sizeof(*bsf);
2210 	*size += sizeof(*bsf) / 16;
2211 	if (unlikely((*seg == qp->sq.qend)))
2212 		*seg = mlx5_get_send_wqe(qp, 0);
2213 
2214 	return 0;
2215 }
2216 
2217 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
2218 				 struct ib_send_wr *wr, u32 nelements,
2219 				 u32 length, u32 pdn)
2220 {
2221 	struct ib_mr *sig_mr = wr->wr.sig_handover.sig_mr;
2222 	u32 sig_key = sig_mr->rkey;
2223 	u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
2224 
2225 	memset(seg, 0, sizeof(*seg));
2226 
2227 	seg->flags = get_umr_flags(wr->wr.sig_handover.access_flags) |
2228 				   MLX5_ACCESS_MODE_KLM;
2229 	seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
2230 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
2231 				    MLX5_MKEY_BSF_EN | pdn);
2232 	seg->len = cpu_to_be64(length);
2233 	seg->xlt_oct_size = cpu_to_be32(be16_to_cpu(get_klm_octo(nelements)));
2234 	seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
2235 }
2236 
2237 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
2238 				struct ib_send_wr *wr, u32 nelements)
2239 {
2240 	memset(umr, 0, sizeof(*umr));
2241 
2242 	umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
2243 	umr->klm_octowords = get_klm_octo(nelements);
2244 	umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
2245 	umr->mkey_mask = sig_mkey_mask();
2246 }
2247 
2248 
2249 static int set_sig_umr_wr(struct ib_send_wr *wr, struct mlx5_ib_qp *qp,
2250 			  void **seg, int *size)
2251 {
2252 	struct mlx5_ib_mr *sig_mr = to_mmr(wr->wr.sig_handover.sig_mr);
2253 	u32 pdn = get_pd(qp)->pdn;
2254 	u32 klm_oct_size;
2255 	int region_len, ret;
2256 
2257 	if (unlikely(wr->num_sge != 1) ||
2258 	    unlikely(wr->wr.sig_handover.access_flags &
2259 		     IB_ACCESS_REMOTE_ATOMIC) ||
2260 	    unlikely(!sig_mr->sig) || unlikely(!qp->signature_en) ||
2261 	    unlikely(!sig_mr->sig->sig_status_checked))
2262 		return -EINVAL;
2263 
2264 	/* length of the protected region, data + protection */
2265 	region_len = wr->sg_list->length;
2266 	if (wr->wr.sig_handover.prot &&
2267 	    (wr->wr.sig_handover.prot->lkey != wr->sg_list->lkey  ||
2268 	     wr->wr.sig_handover.prot->addr != wr->sg_list->addr  ||
2269 	     wr->wr.sig_handover.prot->length != wr->sg_list->length))
2270 		region_len += wr->wr.sig_handover.prot->length;
2271 
2272 	/**
2273 	 * KLM octoword size - if protection was provided
2274 	 * then we use strided block format (3 octowords),
2275 	 * else we use single KLM (1 octoword)
2276 	 **/
2277 	klm_oct_size = wr->wr.sig_handover.prot ? 3 : 1;
2278 
2279 	set_sig_umr_segment(*seg, wr, klm_oct_size);
2280 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2281 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2282 	if (unlikely((*seg == qp->sq.qend)))
2283 		*seg = mlx5_get_send_wqe(qp, 0);
2284 
2285 	set_sig_mkey_segment(*seg, wr, klm_oct_size, region_len, pdn);
2286 	*seg += sizeof(struct mlx5_mkey_seg);
2287 	*size += sizeof(struct mlx5_mkey_seg) / 16;
2288 	if (unlikely((*seg == qp->sq.qend)))
2289 		*seg = mlx5_get_send_wqe(qp, 0);
2290 
2291 	ret = set_sig_data_segment(wr, qp, seg, size);
2292 	if (ret)
2293 		return ret;
2294 
2295 	sig_mr->sig->sig_status_checked = false;
2296 	return 0;
2297 }
2298 
2299 static int set_psv_wr(struct ib_sig_domain *domain,
2300 		      u32 psv_idx, void **seg, int *size)
2301 {
2302 	struct mlx5_seg_set_psv *psv_seg = *seg;
2303 
2304 	memset(psv_seg, 0, sizeof(*psv_seg));
2305 	psv_seg->psv_num = cpu_to_be32(psv_idx);
2306 	switch (domain->sig_type) {
2307 	case IB_SIG_TYPE_NONE:
2308 		break;
2309 	case IB_SIG_TYPE_T10_DIF:
2310 		psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
2311 						     domain->sig.dif.app_tag);
2312 		psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
2313 		break;
2314 	default:
2315 		pr_err("Bad signature type given.\n");
2316 		return 1;
2317 	}
2318 
2319 	*seg += sizeof(*psv_seg);
2320 	*size += sizeof(*psv_seg) / 16;
2321 
2322 	return 0;
2323 }
2324 
2325 static int set_frwr_li_wr(void **seg, struct ib_send_wr *wr, int *size,
2326 			  struct mlx5_core_dev *mdev, struct mlx5_ib_pd *pd, struct mlx5_ib_qp *qp)
2327 {
2328 	int writ = 0;
2329 	int li;
2330 
2331 	li = wr->opcode == IB_WR_LOCAL_INV ? 1 : 0;
2332 	if (unlikely(wr->send_flags & IB_SEND_INLINE))
2333 		return -EINVAL;
2334 
2335 	set_frwr_umr_segment(*seg, wr, li);
2336 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2337 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2338 	if (unlikely((*seg == qp->sq.qend)))
2339 		*seg = mlx5_get_send_wqe(qp, 0);
2340 	set_mkey_segment(*seg, wr, li, &writ);
2341 	*seg += sizeof(struct mlx5_mkey_seg);
2342 	*size += sizeof(struct mlx5_mkey_seg) / 16;
2343 	if (unlikely((*seg == qp->sq.qend)))
2344 		*seg = mlx5_get_send_wqe(qp, 0);
2345 	if (!li) {
2346 		if (unlikely(wr->wr.fast_reg.page_list_len >
2347 			     wr->wr.fast_reg.page_list->max_page_list_len))
2348 			return	-ENOMEM;
2349 
2350 		set_frwr_pages(*seg, wr, mdev, pd, writ);
2351 		*seg += sizeof(struct mlx5_wqe_data_seg);
2352 		*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
2353 	}
2354 	return 0;
2355 }
2356 
2357 static void dump_wqe(struct mlx5_ib_qp *qp, int idx, int size_16)
2358 {
2359 	__be32 *p = NULL;
2360 	int tidx = idx;
2361 	int i, j;
2362 
2363 	pr_debug("dump wqe at %p\n", mlx5_get_send_wqe(qp, tidx));
2364 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
2365 		if ((i & 0xf) == 0) {
2366 			void *buf = mlx5_get_send_wqe(qp, tidx);
2367 			tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
2368 			p = buf;
2369 			j = 0;
2370 		}
2371 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
2372 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
2373 			 be32_to_cpu(p[j + 3]));
2374 	}
2375 }
2376 
2377 static void mlx5_bf_copy(u64 __iomem *dst, u64 *src,
2378 			 unsigned bytecnt, struct mlx5_ib_qp *qp)
2379 {
2380 	while (bytecnt > 0) {
2381 		__iowrite64_copy(dst++, src++, 8);
2382 		__iowrite64_copy(dst++, src++, 8);
2383 		__iowrite64_copy(dst++, src++, 8);
2384 		__iowrite64_copy(dst++, src++, 8);
2385 		__iowrite64_copy(dst++, src++, 8);
2386 		__iowrite64_copy(dst++, src++, 8);
2387 		__iowrite64_copy(dst++, src++, 8);
2388 		__iowrite64_copy(dst++, src++, 8);
2389 		bytecnt -= 64;
2390 		if (unlikely(src == qp->sq.qend))
2391 			src = mlx5_get_send_wqe(qp, 0);
2392 	}
2393 }
2394 
2395 static u8 get_fence(u8 fence, struct ib_send_wr *wr)
2396 {
2397 	if (unlikely(wr->opcode == IB_WR_LOCAL_INV &&
2398 		     wr->send_flags & IB_SEND_FENCE))
2399 		return MLX5_FENCE_MODE_STRONG_ORDERING;
2400 
2401 	if (unlikely(fence)) {
2402 		if (wr->send_flags & IB_SEND_FENCE)
2403 			return MLX5_FENCE_MODE_SMALL_AND_FENCE;
2404 		else
2405 			return fence;
2406 
2407 	} else {
2408 		return 0;
2409 	}
2410 }
2411 
2412 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
2413 		     struct mlx5_wqe_ctrl_seg **ctrl,
2414 		     struct ib_send_wr *wr, int *idx,
2415 		     int *size, int nreq)
2416 {
2417 	int err = 0;
2418 
2419 	if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq))) {
2420 		err = -ENOMEM;
2421 		return err;
2422 	}
2423 
2424 	*idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
2425 	*seg = mlx5_get_send_wqe(qp, *idx);
2426 	*ctrl = *seg;
2427 	*(uint32_t *)(*seg + 8) = 0;
2428 	(*ctrl)->imm = send_ieth(wr);
2429 	(*ctrl)->fm_ce_se = qp->sq_signal_bits |
2430 		(wr->send_flags & IB_SEND_SIGNALED ?
2431 		 MLX5_WQE_CTRL_CQ_UPDATE : 0) |
2432 		(wr->send_flags & IB_SEND_SOLICITED ?
2433 		 MLX5_WQE_CTRL_SOLICITED : 0);
2434 
2435 	*seg += sizeof(**ctrl);
2436 	*size = sizeof(**ctrl) / 16;
2437 
2438 	return err;
2439 }
2440 
2441 static void finish_wqe(struct mlx5_ib_qp *qp,
2442 		       struct mlx5_wqe_ctrl_seg *ctrl,
2443 		       u8 size, unsigned idx, u64 wr_id,
2444 		       int nreq, u8 fence, u8 next_fence,
2445 		       u32 mlx5_opcode)
2446 {
2447 	u8 opmod = 0;
2448 
2449 	ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
2450 					     mlx5_opcode | ((u32)opmod << 24));
2451 	ctrl->qpn_ds = cpu_to_be32(size | (qp->mqp.qpn << 8));
2452 	ctrl->fm_ce_se |= fence;
2453 	qp->fm_cache = next_fence;
2454 	if (unlikely(qp->wq_sig))
2455 		ctrl->signature = wq_sig(ctrl);
2456 
2457 	qp->sq.wrid[idx] = wr_id;
2458 	qp->sq.w_list[idx].opcode = mlx5_opcode;
2459 	qp->sq.wqe_head[idx] = qp->sq.head + nreq;
2460 	qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
2461 	qp->sq.w_list[idx].next = qp->sq.cur_post;
2462 }
2463 
2464 
2465 int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2466 		      struct ib_send_wr **bad_wr)
2467 {
2468 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
2469 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2470 	struct mlx5_core_dev *mdev = dev->mdev;
2471 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2472 	struct mlx5_ib_mr *mr;
2473 	struct mlx5_wqe_data_seg *dpseg;
2474 	struct mlx5_wqe_xrc_seg *xrc;
2475 	struct mlx5_bf *bf = qp->bf;
2476 	int uninitialized_var(size);
2477 	void *qend = qp->sq.qend;
2478 	unsigned long flags;
2479 	unsigned idx;
2480 	int err = 0;
2481 	int inl = 0;
2482 	int num_sge;
2483 	void *seg;
2484 	int nreq;
2485 	int i;
2486 	u8 next_fence = 0;
2487 	u8 fence;
2488 
2489 	spin_lock_irqsave(&qp->sq.lock, flags);
2490 
2491 	for (nreq = 0; wr; nreq++, wr = wr->next) {
2492 		if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
2493 			mlx5_ib_warn(dev, "\n");
2494 			err = -EINVAL;
2495 			*bad_wr = wr;
2496 			goto out;
2497 		}
2498 
2499 		fence = qp->fm_cache;
2500 		num_sge = wr->num_sge;
2501 		if (unlikely(num_sge > qp->sq.max_gs)) {
2502 			mlx5_ib_warn(dev, "\n");
2503 			err = -ENOMEM;
2504 			*bad_wr = wr;
2505 			goto out;
2506 		}
2507 
2508 		err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, nreq);
2509 		if (err) {
2510 			mlx5_ib_warn(dev, "\n");
2511 			err = -ENOMEM;
2512 			*bad_wr = wr;
2513 			goto out;
2514 		}
2515 
2516 		switch (ibqp->qp_type) {
2517 		case IB_QPT_XRC_INI:
2518 			xrc = seg;
2519 			xrc->xrc_srqn = htonl(wr->xrc_remote_srq_num);
2520 			seg += sizeof(*xrc);
2521 			size += sizeof(*xrc) / 16;
2522 			/* fall through */
2523 		case IB_QPT_RC:
2524 			switch (wr->opcode) {
2525 			case IB_WR_RDMA_READ:
2526 			case IB_WR_RDMA_WRITE:
2527 			case IB_WR_RDMA_WRITE_WITH_IMM:
2528 				set_raddr_seg(seg, wr->wr.rdma.remote_addr,
2529 					      wr->wr.rdma.rkey);
2530 				seg += sizeof(struct mlx5_wqe_raddr_seg);
2531 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
2532 				break;
2533 
2534 			case IB_WR_ATOMIC_CMP_AND_SWP:
2535 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2536 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2537 				mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
2538 				err = -ENOSYS;
2539 				*bad_wr = wr;
2540 				goto out;
2541 
2542 			case IB_WR_LOCAL_INV:
2543 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2544 				qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
2545 				ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
2546 				err = set_frwr_li_wr(&seg, wr, &size, mdev, to_mpd(ibqp->pd), qp);
2547 				if (err) {
2548 					mlx5_ib_warn(dev, "\n");
2549 					*bad_wr = wr;
2550 					goto out;
2551 				}
2552 				num_sge = 0;
2553 				break;
2554 
2555 			case IB_WR_FAST_REG_MR:
2556 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2557 				qp->sq.wr_data[idx] = IB_WR_FAST_REG_MR;
2558 				ctrl->imm = cpu_to_be32(wr->wr.fast_reg.rkey);
2559 				err = set_frwr_li_wr(&seg, wr, &size, mdev, to_mpd(ibqp->pd), qp);
2560 				if (err) {
2561 					mlx5_ib_warn(dev, "\n");
2562 					*bad_wr = wr;
2563 					goto out;
2564 				}
2565 				num_sge = 0;
2566 				break;
2567 
2568 			case IB_WR_REG_SIG_MR:
2569 				qp->sq.wr_data[idx] = IB_WR_REG_SIG_MR;
2570 				mr = to_mmr(wr->wr.sig_handover.sig_mr);
2571 
2572 				ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
2573 				err = set_sig_umr_wr(wr, qp, &seg, &size);
2574 				if (err) {
2575 					mlx5_ib_warn(dev, "\n");
2576 					*bad_wr = wr;
2577 					goto out;
2578 				}
2579 
2580 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2581 					   nreq, get_fence(fence, wr),
2582 					   next_fence, MLX5_OPCODE_UMR);
2583 				/*
2584 				 * SET_PSV WQEs are not signaled and solicited
2585 				 * on error
2586 				 */
2587 				wr->send_flags &= ~IB_SEND_SIGNALED;
2588 				wr->send_flags |= IB_SEND_SOLICITED;
2589 				err = begin_wqe(qp, &seg, &ctrl, wr,
2590 						&idx, &size, nreq);
2591 				if (err) {
2592 					mlx5_ib_warn(dev, "\n");
2593 					err = -ENOMEM;
2594 					*bad_wr = wr;
2595 					goto out;
2596 				}
2597 
2598 				err = set_psv_wr(&wr->wr.sig_handover.sig_attrs->mem,
2599 						 mr->sig->psv_memory.psv_idx, &seg,
2600 						 &size);
2601 				if (err) {
2602 					mlx5_ib_warn(dev, "\n");
2603 					*bad_wr = wr;
2604 					goto out;
2605 				}
2606 
2607 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2608 					   nreq, get_fence(fence, wr),
2609 					   next_fence, MLX5_OPCODE_SET_PSV);
2610 				err = begin_wqe(qp, &seg, &ctrl, wr,
2611 						&idx, &size, nreq);
2612 				if (err) {
2613 					mlx5_ib_warn(dev, "\n");
2614 					err = -ENOMEM;
2615 					*bad_wr = wr;
2616 					goto out;
2617 				}
2618 
2619 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2620 				err = set_psv_wr(&wr->wr.sig_handover.sig_attrs->wire,
2621 						 mr->sig->psv_wire.psv_idx, &seg,
2622 						 &size);
2623 				if (err) {
2624 					mlx5_ib_warn(dev, "\n");
2625 					*bad_wr = wr;
2626 					goto out;
2627 				}
2628 
2629 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2630 					   nreq, get_fence(fence, wr),
2631 					   next_fence, MLX5_OPCODE_SET_PSV);
2632 				num_sge = 0;
2633 				goto skip_psv;
2634 
2635 			default:
2636 				break;
2637 			}
2638 			break;
2639 
2640 		case IB_QPT_UC:
2641 			switch (wr->opcode) {
2642 			case IB_WR_RDMA_WRITE:
2643 			case IB_WR_RDMA_WRITE_WITH_IMM:
2644 				set_raddr_seg(seg, wr->wr.rdma.remote_addr,
2645 					      wr->wr.rdma.rkey);
2646 				seg  += sizeof(struct mlx5_wqe_raddr_seg);
2647 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
2648 				break;
2649 
2650 			default:
2651 				break;
2652 			}
2653 			break;
2654 
2655 		case IB_QPT_UD:
2656 		case IB_QPT_SMI:
2657 		case IB_QPT_GSI:
2658 			set_datagram_seg(seg, wr);
2659 			seg += sizeof(struct mlx5_wqe_datagram_seg);
2660 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
2661 			if (unlikely((seg == qend)))
2662 				seg = mlx5_get_send_wqe(qp, 0);
2663 			break;
2664 
2665 		case MLX5_IB_QPT_REG_UMR:
2666 			if (wr->opcode != MLX5_IB_WR_UMR) {
2667 				err = -EINVAL;
2668 				mlx5_ib_warn(dev, "bad opcode\n");
2669 				goto out;
2670 			}
2671 			qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
2672 			ctrl->imm = cpu_to_be32(wr->wr.fast_reg.rkey);
2673 			set_reg_umr_segment(seg, wr);
2674 			seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2675 			size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2676 			if (unlikely((seg == qend)))
2677 				seg = mlx5_get_send_wqe(qp, 0);
2678 			set_reg_mkey_segment(seg, wr);
2679 			seg += sizeof(struct mlx5_mkey_seg);
2680 			size += sizeof(struct mlx5_mkey_seg) / 16;
2681 			if (unlikely((seg == qend)))
2682 				seg = mlx5_get_send_wqe(qp, 0);
2683 			break;
2684 
2685 		default:
2686 			break;
2687 		}
2688 
2689 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
2690 			int uninitialized_var(sz);
2691 
2692 			err = set_data_inl_seg(qp, wr, seg, &sz);
2693 			if (unlikely(err)) {
2694 				mlx5_ib_warn(dev, "\n");
2695 				*bad_wr = wr;
2696 				goto out;
2697 			}
2698 			inl = 1;
2699 			size += sz;
2700 		} else {
2701 			dpseg = seg;
2702 			for (i = 0; i < num_sge; i++) {
2703 				if (unlikely(dpseg == qend)) {
2704 					seg = mlx5_get_send_wqe(qp, 0);
2705 					dpseg = seg;
2706 				}
2707 				if (likely(wr->sg_list[i].length)) {
2708 					set_data_ptr_seg(dpseg, wr->sg_list + i);
2709 					size += sizeof(struct mlx5_wqe_data_seg) / 16;
2710 					dpseg++;
2711 				}
2712 			}
2713 		}
2714 
2715 		finish_wqe(qp, ctrl, size, idx, wr->wr_id, nreq,
2716 			   get_fence(fence, wr), next_fence,
2717 			   mlx5_ib_opcode[wr->opcode]);
2718 skip_psv:
2719 		if (0)
2720 			dump_wqe(qp, idx, size);
2721 	}
2722 
2723 out:
2724 	if (likely(nreq)) {
2725 		qp->sq.head += nreq;
2726 
2727 		/* Make sure that descriptors are written before
2728 		 * updating doorbell record and ringing the doorbell
2729 		 */
2730 		wmb();
2731 
2732 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
2733 
2734 		/* Make sure doorbell record is visible to the HCA before
2735 		 * we hit doorbell */
2736 		wmb();
2737 
2738 		if (bf->need_lock)
2739 			spin_lock(&bf->lock);
2740 
2741 		/* TBD enable WC */
2742 		if (0 && nreq == 1 && bf->uuarn && inl && size > 1 && size <= bf->buf_size / 16) {
2743 			mlx5_bf_copy(bf->reg + bf->offset, (u64 *)ctrl, ALIGN(size * 16, 64), qp);
2744 			/* wc_wmb(); */
2745 		} else {
2746 			mlx5_write64((__be32 *)ctrl, bf->regreg + bf->offset,
2747 				     MLX5_GET_DOORBELL_LOCK(&bf->lock32));
2748 			/* Make sure doorbells don't leak out of SQ spinlock
2749 			 * and reach the HCA out of order.
2750 			 */
2751 			mmiowb();
2752 		}
2753 		bf->offset ^= bf->buf_size;
2754 		if (bf->need_lock)
2755 			spin_unlock(&bf->lock);
2756 	}
2757 
2758 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2759 
2760 	return err;
2761 }
2762 
2763 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
2764 {
2765 	sig->signature = calc_sig(sig, size);
2766 }
2767 
2768 int mlx5_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2769 		      struct ib_recv_wr **bad_wr)
2770 {
2771 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2772 	struct mlx5_wqe_data_seg *scat;
2773 	struct mlx5_rwqe_sig *sig;
2774 	unsigned long flags;
2775 	int err = 0;
2776 	int nreq;
2777 	int ind;
2778 	int i;
2779 
2780 	spin_lock_irqsave(&qp->rq.lock, flags);
2781 
2782 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
2783 
2784 	for (nreq = 0; wr; nreq++, wr = wr->next) {
2785 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2786 			err = -ENOMEM;
2787 			*bad_wr = wr;
2788 			goto out;
2789 		}
2790 
2791 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2792 			err = -EINVAL;
2793 			*bad_wr = wr;
2794 			goto out;
2795 		}
2796 
2797 		scat = get_recv_wqe(qp, ind);
2798 		if (qp->wq_sig)
2799 			scat++;
2800 
2801 		for (i = 0; i < wr->num_sge; i++)
2802 			set_data_ptr_seg(scat + i, wr->sg_list + i);
2803 
2804 		if (i < qp->rq.max_gs) {
2805 			scat[i].byte_count = 0;
2806 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
2807 			scat[i].addr       = 0;
2808 		}
2809 
2810 		if (qp->wq_sig) {
2811 			sig = (struct mlx5_rwqe_sig *)scat;
2812 			set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
2813 		}
2814 
2815 		qp->rq.wrid[ind] = wr->wr_id;
2816 
2817 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
2818 	}
2819 
2820 out:
2821 	if (likely(nreq)) {
2822 		qp->rq.head += nreq;
2823 
2824 		/* Make sure that descriptors are written before
2825 		 * doorbell record.
2826 		 */
2827 		wmb();
2828 
2829 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
2830 	}
2831 
2832 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2833 
2834 	return err;
2835 }
2836 
2837 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
2838 {
2839 	switch (mlx5_state) {
2840 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
2841 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
2842 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
2843 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
2844 	case MLX5_QP_STATE_SQ_DRAINING:
2845 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
2846 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
2847 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
2848 	default:		     return -1;
2849 	}
2850 }
2851 
2852 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
2853 {
2854 	switch (mlx5_mig_state) {
2855 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
2856 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
2857 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
2858 	default: return -1;
2859 	}
2860 }
2861 
2862 static int to_ib_qp_access_flags(int mlx5_flags)
2863 {
2864 	int ib_flags = 0;
2865 
2866 	if (mlx5_flags & MLX5_QP_BIT_RRE)
2867 		ib_flags |= IB_ACCESS_REMOTE_READ;
2868 	if (mlx5_flags & MLX5_QP_BIT_RWE)
2869 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
2870 	if (mlx5_flags & MLX5_QP_BIT_RAE)
2871 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
2872 
2873 	return ib_flags;
2874 }
2875 
2876 static void to_ib_ah_attr(struct mlx5_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
2877 				struct mlx5_qp_path *path)
2878 {
2879 	struct mlx5_core_dev *dev = ibdev->mdev;
2880 
2881 	memset(ib_ah_attr, 0, sizeof(*ib_ah_attr));
2882 	ib_ah_attr->port_num	  = path->port;
2883 
2884 	if (ib_ah_attr->port_num == 0 ||
2885 	    ib_ah_attr->port_num > dev->caps.gen.num_ports)
2886 		return;
2887 
2888 	ib_ah_attr->sl = path->sl & 0xf;
2889 
2890 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
2891 	ib_ah_attr->src_path_bits = path->grh_mlid & 0x7f;
2892 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
2893 	ib_ah_attr->ah_flags      = (path->grh_mlid & (1 << 7)) ? IB_AH_GRH : 0;
2894 	if (ib_ah_attr->ah_flags) {
2895 		ib_ah_attr->grh.sgid_index = path->mgid_index;
2896 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
2897 		ib_ah_attr->grh.traffic_class =
2898 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
2899 		ib_ah_attr->grh.flow_label =
2900 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
2901 		memcpy(ib_ah_attr->grh.dgid.raw,
2902 		       path->rgid, sizeof(ib_ah_attr->grh.dgid.raw));
2903 	}
2904 }
2905 
2906 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
2907 		     struct ib_qp_init_attr *qp_init_attr)
2908 {
2909 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2910 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2911 	struct mlx5_query_qp_mbox_out *outb;
2912 	struct mlx5_qp_context *context;
2913 	int mlx5_state;
2914 	int err = 0;
2915 
2916 	mutex_lock(&qp->mutex);
2917 	outb = kzalloc(sizeof(*outb), GFP_KERNEL);
2918 	if (!outb) {
2919 		err = -ENOMEM;
2920 		goto out;
2921 	}
2922 	context = &outb->ctx;
2923 	err = mlx5_core_qp_query(dev->mdev, &qp->mqp, outb, sizeof(*outb));
2924 	if (err)
2925 		goto out_free;
2926 
2927 	mlx5_state = be32_to_cpu(context->flags) >> 28;
2928 
2929 	qp->state		     = to_ib_qp_state(mlx5_state);
2930 	qp_attr->qp_state	     = qp->state;
2931 	qp_attr->path_mtu	     = context->mtu_msgmax >> 5;
2932 	qp_attr->path_mig_state	     =
2933 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
2934 	qp_attr->qkey		     = be32_to_cpu(context->qkey);
2935 	qp_attr->rq_psn		     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
2936 	qp_attr->sq_psn		     = be32_to_cpu(context->next_send_psn) & 0xffffff;
2937 	qp_attr->dest_qp_num	     = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
2938 	qp_attr->qp_access_flags     =
2939 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
2940 
2941 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
2942 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
2943 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
2944 		qp_attr->alt_pkey_index = context->alt_path.pkey_index & 0x7f;
2945 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
2946 	}
2947 
2948 	qp_attr->pkey_index = context->pri_path.pkey_index & 0x7f;
2949 	qp_attr->port_num = context->pri_path.port;
2950 
2951 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
2952 	qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
2953 
2954 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
2955 
2956 	qp_attr->max_dest_rd_atomic =
2957 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
2958 	qp_attr->min_rnr_timer	    =
2959 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
2960 	qp_attr->timeout	    = context->pri_path.ackto_lt >> 3;
2961 	qp_attr->retry_cnt	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
2962 	qp_attr->rnr_retry	    = (be32_to_cpu(context->params1) >> 13) & 0x7;
2963 	qp_attr->alt_timeout	    = context->alt_path.ackto_lt >> 3;
2964 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
2965 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
2966 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
2967 
2968 	if (!ibqp->uobject) {
2969 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
2970 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
2971 	} else {
2972 		qp_attr->cap.max_send_wr  = 0;
2973 		qp_attr->cap.max_send_sge = 0;
2974 	}
2975 
2976 	/* We don't support inline sends for kernel QPs (yet), and we
2977 	 * don't know what userspace's value should be.
2978 	 */
2979 	qp_attr->cap.max_inline_data = 0;
2980 
2981 	qp_init_attr->cap	     = qp_attr->cap;
2982 
2983 	qp_init_attr->create_flags = 0;
2984 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
2985 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
2986 
2987 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
2988 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
2989 
2990 out_free:
2991 	kfree(outb);
2992 
2993 out:
2994 	mutex_unlock(&qp->mutex);
2995 	return err;
2996 }
2997 
2998 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
2999 					  struct ib_ucontext *context,
3000 					  struct ib_udata *udata)
3001 {
3002 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
3003 	struct mlx5_general_caps *gen;
3004 	struct mlx5_ib_xrcd *xrcd;
3005 	int err;
3006 
3007 	gen = &dev->mdev->caps.gen;
3008 	if (!(gen->flags & MLX5_DEV_CAP_FLAG_XRC))
3009 		return ERR_PTR(-ENOSYS);
3010 
3011 	xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
3012 	if (!xrcd)
3013 		return ERR_PTR(-ENOMEM);
3014 
3015 	err = mlx5_core_xrcd_alloc(dev->mdev, &xrcd->xrcdn);
3016 	if (err) {
3017 		kfree(xrcd);
3018 		return ERR_PTR(-ENOMEM);
3019 	}
3020 
3021 	return &xrcd->ibxrcd;
3022 }
3023 
3024 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
3025 {
3026 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
3027 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
3028 	int err;
3029 
3030 	err = mlx5_core_xrcd_dealloc(dev->mdev, xrcdn);
3031 	if (err) {
3032 		mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
3033 		return err;
3034 	}
3035 
3036 	kfree(xrcd);
3037 
3038 	return 0;
3039 }
3040