xref: /openbmc/linux/drivers/infiniband/hw/mlx5/qp.c (revision 77d84ff8)
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 	int wqe_size;
162 	int wq_size;
163 
164 	/* Sanity check RQ size before proceeding */
165 	if (cap->max_recv_wr  > dev->mdev.caps.max_wqes)
166 		return -EINVAL;
167 
168 	if (!has_rq) {
169 		qp->rq.max_gs = 0;
170 		qp->rq.wqe_cnt = 0;
171 		qp->rq.wqe_shift = 0;
172 	} else {
173 		if (ucmd) {
174 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
175 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
176 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
177 			qp->rq.max_post = qp->rq.wqe_cnt;
178 		} else {
179 			wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
180 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
181 			wqe_size = roundup_pow_of_two(wqe_size);
182 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
183 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
184 			qp->rq.wqe_cnt = wq_size / wqe_size;
185 			if (wqe_size > dev->mdev.caps.max_rq_desc_sz) {
186 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
187 					    wqe_size,
188 					    dev->mdev.caps.max_rq_desc_sz);
189 				return -EINVAL;
190 			}
191 			qp->rq.wqe_shift = ilog2(wqe_size);
192 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
193 			qp->rq.max_post = qp->rq.wqe_cnt;
194 		}
195 	}
196 
197 	return 0;
198 }
199 
200 static int sq_overhead(enum ib_qp_type qp_type)
201 {
202 	int size = 0;
203 
204 	switch (qp_type) {
205 	case IB_QPT_XRC_INI:
206 		size += sizeof(struct mlx5_wqe_xrc_seg);
207 		/* fall through */
208 	case IB_QPT_RC:
209 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
210 			sizeof(struct mlx5_wqe_atomic_seg) +
211 			sizeof(struct mlx5_wqe_raddr_seg);
212 		break;
213 
214 	case IB_QPT_XRC_TGT:
215 		return 0;
216 
217 	case IB_QPT_UC:
218 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
219 			sizeof(struct mlx5_wqe_raddr_seg);
220 		break;
221 
222 	case IB_QPT_UD:
223 	case IB_QPT_SMI:
224 	case IB_QPT_GSI:
225 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
226 			sizeof(struct mlx5_wqe_datagram_seg);
227 		break;
228 
229 	case MLX5_IB_QPT_REG_UMR:
230 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
231 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
232 			sizeof(struct mlx5_mkey_seg);
233 		break;
234 
235 	default:
236 		return -EINVAL;
237 	}
238 
239 	return size;
240 }
241 
242 static int calc_send_wqe(struct ib_qp_init_attr *attr)
243 {
244 	int inl_size = 0;
245 	int size;
246 
247 	size = sq_overhead(attr->qp_type);
248 	if (size < 0)
249 		return size;
250 
251 	if (attr->cap.max_inline_data) {
252 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
253 			attr->cap.max_inline_data;
254 	}
255 
256 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
257 
258 	return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
259 }
260 
261 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
262 			struct mlx5_ib_qp *qp)
263 {
264 	int wqe_size;
265 	int wq_size;
266 
267 	if (!attr->cap.max_send_wr)
268 		return 0;
269 
270 	wqe_size = calc_send_wqe(attr);
271 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
272 	if (wqe_size < 0)
273 		return wqe_size;
274 
275 	if (wqe_size > dev->mdev.caps.max_sq_desc_sz) {
276 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
277 			    wqe_size, dev->mdev.caps.max_sq_desc_sz);
278 		return -EINVAL;
279 	}
280 
281 	qp->max_inline_data = wqe_size - sq_overhead(attr->qp_type) -
282 		sizeof(struct mlx5_wqe_inline_seg);
283 	attr->cap.max_inline_data = qp->max_inline_data;
284 
285 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
286 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
287 	if (qp->sq.wqe_cnt > dev->mdev.caps.max_wqes) {
288 		mlx5_ib_dbg(dev, "wqe count(%d) exceeds limits(%d)\n",
289 			    qp->sq.wqe_cnt, dev->mdev.caps.max_wqes);
290 		return -ENOMEM;
291 	}
292 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
293 	qp->sq.max_gs = attr->cap.max_send_sge;
294 	qp->sq.max_post = wq_size / wqe_size;
295 	attr->cap.max_send_wr = qp->sq.max_post;
296 
297 	return wq_size;
298 }
299 
300 static int set_user_buf_size(struct mlx5_ib_dev *dev,
301 			    struct mlx5_ib_qp *qp,
302 			    struct mlx5_ib_create_qp *ucmd)
303 {
304 	int desc_sz = 1 << qp->sq.wqe_shift;
305 
306 	if (desc_sz > dev->mdev.caps.max_sq_desc_sz) {
307 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
308 			     desc_sz, dev->mdev.caps.max_sq_desc_sz);
309 		return -EINVAL;
310 	}
311 
312 	if (ucmd->sq_wqe_count && ((1 << ilog2(ucmd->sq_wqe_count)) != ucmd->sq_wqe_count)) {
313 		mlx5_ib_warn(dev, "sq_wqe_count %d, sq_wqe_count %d\n",
314 			     ucmd->sq_wqe_count, ucmd->sq_wqe_count);
315 		return -EINVAL;
316 	}
317 
318 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
319 
320 	if (qp->sq.wqe_cnt > dev->mdev.caps.max_wqes) {
321 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
322 			     qp->sq.wqe_cnt, dev->mdev.caps.max_wqes);
323 		return -EINVAL;
324 	}
325 
326 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
327 		(qp->sq.wqe_cnt << 6);
328 
329 	return 0;
330 }
331 
332 static int qp_has_rq(struct ib_qp_init_attr *attr)
333 {
334 	if (attr->qp_type == IB_QPT_XRC_INI ||
335 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
336 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
337 	    !attr->cap.max_recv_wr)
338 		return 0;
339 
340 	return 1;
341 }
342 
343 static int alloc_high_class_uuar(struct mlx5_uuar_info *uuari)
344 {
345 	int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
346 	int start_uuar;
347 	int i;
348 
349 	start_uuar = nuuars - uuari->num_low_latency_uuars;
350 	for (i = start_uuar; i < nuuars; i++) {
351 		if (!test_bit(i, uuari->bitmap)) {
352 			set_bit(i, uuari->bitmap);
353 			uuari->count[i]++;
354 			return i;
355 		}
356 	}
357 
358 	return -ENOMEM;
359 }
360 
361 static int alloc_med_class_uuar(struct mlx5_uuar_info *uuari)
362 {
363 	int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
364 	int minidx = 1;
365 	int uuarn;
366 	int end;
367 	int i;
368 
369 	end = nuuars - uuari->num_low_latency_uuars;
370 
371 	for (i = 1; i < end; i++) {
372 		uuarn = i & 3;
373 		if (uuarn == 2 || uuarn == 3)
374 			continue;
375 
376 		if (uuari->count[i] < uuari->count[minidx])
377 			minidx = i;
378 	}
379 
380 	uuari->count[minidx]++;
381 	return minidx;
382 }
383 
384 static int alloc_uuar(struct mlx5_uuar_info *uuari,
385 		      enum mlx5_ib_latency_class lat)
386 {
387 	int uuarn = -EINVAL;
388 
389 	mutex_lock(&uuari->lock);
390 	switch (lat) {
391 	case MLX5_IB_LATENCY_CLASS_LOW:
392 		uuarn = 0;
393 		uuari->count[uuarn]++;
394 		break;
395 
396 	case MLX5_IB_LATENCY_CLASS_MEDIUM:
397 		uuarn = alloc_med_class_uuar(uuari);
398 		break;
399 
400 	case MLX5_IB_LATENCY_CLASS_HIGH:
401 		uuarn = alloc_high_class_uuar(uuari);
402 		break;
403 
404 	case MLX5_IB_LATENCY_CLASS_FAST_PATH:
405 		uuarn = 2;
406 		break;
407 	}
408 	mutex_unlock(&uuari->lock);
409 
410 	return uuarn;
411 }
412 
413 static void free_med_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
414 {
415 	clear_bit(uuarn, uuari->bitmap);
416 	--uuari->count[uuarn];
417 }
418 
419 static void free_high_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
420 {
421 	clear_bit(uuarn, uuari->bitmap);
422 	--uuari->count[uuarn];
423 }
424 
425 static void free_uuar(struct mlx5_uuar_info *uuari, int uuarn)
426 {
427 	int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
428 	int high_uuar = nuuars - uuari->num_low_latency_uuars;
429 
430 	mutex_lock(&uuari->lock);
431 	if (uuarn == 0) {
432 		--uuari->count[uuarn];
433 		goto out;
434 	}
435 
436 	if (uuarn < high_uuar) {
437 		free_med_class_uuar(uuari, uuarn);
438 		goto out;
439 	}
440 
441 	free_high_class_uuar(uuari, uuarn);
442 
443 out:
444 	mutex_unlock(&uuari->lock);
445 }
446 
447 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
448 {
449 	switch (state) {
450 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
451 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
452 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
453 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
454 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
455 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
456 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
457 	default:		return -1;
458 	}
459 }
460 
461 static int to_mlx5_st(enum ib_qp_type type)
462 {
463 	switch (type) {
464 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
465 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
466 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
467 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
468 	case IB_QPT_XRC_INI:
469 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
470 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
471 	case IB_QPT_GSI:		return MLX5_QP_ST_QP1;
472 	case IB_QPT_RAW_IPV6:		return MLX5_QP_ST_RAW_IPV6;
473 	case IB_QPT_RAW_ETHERTYPE:	return MLX5_QP_ST_RAW_ETHERTYPE;
474 	case IB_QPT_RAW_PACKET:
475 	case IB_QPT_MAX:
476 	default:		return -EINVAL;
477 	}
478 }
479 
480 static int uuarn_to_uar_index(struct mlx5_uuar_info *uuari, int uuarn)
481 {
482 	return uuari->uars[uuarn / MLX5_BF_REGS_PER_PAGE].index;
483 }
484 
485 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
486 			  struct mlx5_ib_qp *qp, struct ib_udata *udata,
487 			  struct mlx5_create_qp_mbox_in **in,
488 			  struct mlx5_ib_create_qp_resp *resp, int *inlen)
489 {
490 	struct mlx5_ib_ucontext *context;
491 	struct mlx5_ib_create_qp ucmd;
492 	int page_shift;
493 	int uar_index;
494 	int npages;
495 	u32 offset;
496 	int uuarn;
497 	int ncont;
498 	int err;
499 
500 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
501 	if (err) {
502 		mlx5_ib_dbg(dev, "copy failed\n");
503 		return err;
504 	}
505 
506 	context = to_mucontext(pd->uobject->context);
507 	/*
508 	 * TBD: should come from the verbs when we have the API
509 	 */
510 	uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_HIGH);
511 	if (uuarn < 0) {
512 		mlx5_ib_dbg(dev, "failed to allocate low latency UUAR\n");
513 		mlx5_ib_dbg(dev, "reverting to high latency\n");
514 		uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_LOW);
515 		if (uuarn < 0) {
516 			mlx5_ib_dbg(dev, "uuar allocation failed\n");
517 			return uuarn;
518 		}
519 	}
520 
521 	uar_index = uuarn_to_uar_index(&context->uuari, uuarn);
522 	mlx5_ib_dbg(dev, "uuarn 0x%x, uar_index 0x%x\n", uuarn, uar_index);
523 
524 	err = set_user_buf_size(dev, qp, &ucmd);
525 	if (err)
526 		goto err_uuar;
527 
528 	qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
529 			       qp->buf_size, 0, 0);
530 	if (IS_ERR(qp->umem)) {
531 		mlx5_ib_dbg(dev, "umem_get failed\n");
532 		err = PTR_ERR(qp->umem);
533 		goto err_uuar;
534 	}
535 
536 	mlx5_ib_cont_pages(qp->umem, ucmd.buf_addr, &npages, &page_shift,
537 			   &ncont, NULL);
538 	err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift, &offset);
539 	if (err) {
540 		mlx5_ib_warn(dev, "bad offset\n");
541 		goto err_umem;
542 	}
543 	mlx5_ib_dbg(dev, "addr 0x%llx, size %d, npages %d, page_shift %d, ncont %d, offset %d\n",
544 		    ucmd.buf_addr, qp->buf_size, npages, page_shift, ncont, offset);
545 
546 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * ncont;
547 	*in = mlx5_vzalloc(*inlen);
548 	if (!*in) {
549 		err = -ENOMEM;
550 		goto err_umem;
551 	}
552 	mlx5_ib_populate_pas(dev, qp->umem, page_shift, (*in)->pas, 0);
553 	(*in)->ctx.log_pg_sz_remote_qpn =
554 		cpu_to_be32((page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
555 	(*in)->ctx.params2 = cpu_to_be32(offset << 6);
556 
557 	(*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
558 	resp->uuar_index = uuarn;
559 	qp->uuarn = uuarn;
560 
561 	err = mlx5_ib_db_map_user(context, ucmd.db_addr, &qp->db);
562 	if (err) {
563 		mlx5_ib_dbg(dev, "map failed\n");
564 		goto err_free;
565 	}
566 
567 	err = ib_copy_to_udata(udata, resp, sizeof(*resp));
568 	if (err) {
569 		mlx5_ib_dbg(dev, "copy failed\n");
570 		goto err_unmap;
571 	}
572 	qp->create_type = MLX5_QP_USER;
573 
574 	return 0;
575 
576 err_unmap:
577 	mlx5_ib_db_unmap_user(context, &qp->db);
578 
579 err_free:
580 	mlx5_vfree(*in);
581 
582 err_umem:
583 	ib_umem_release(qp->umem);
584 
585 err_uuar:
586 	free_uuar(&context->uuari, uuarn);
587 	return err;
588 }
589 
590 static void destroy_qp_user(struct ib_pd *pd, struct mlx5_ib_qp *qp)
591 {
592 	struct mlx5_ib_ucontext *context;
593 
594 	context = to_mucontext(pd->uobject->context);
595 	mlx5_ib_db_unmap_user(context, &qp->db);
596 	ib_umem_release(qp->umem);
597 	free_uuar(&context->uuari, qp->uuarn);
598 }
599 
600 static int create_kernel_qp(struct mlx5_ib_dev *dev,
601 			    struct ib_qp_init_attr *init_attr,
602 			    struct mlx5_ib_qp *qp,
603 			    struct mlx5_create_qp_mbox_in **in, int *inlen)
604 {
605 	enum mlx5_ib_latency_class lc = MLX5_IB_LATENCY_CLASS_LOW;
606 	struct mlx5_uuar_info *uuari;
607 	int uar_index;
608 	int uuarn;
609 	int err;
610 
611 	uuari = &dev->mdev.priv.uuari;
612 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
613 		qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
614 
615 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
616 		lc = MLX5_IB_LATENCY_CLASS_FAST_PATH;
617 
618 	uuarn = alloc_uuar(uuari, lc);
619 	if (uuarn < 0) {
620 		mlx5_ib_dbg(dev, "\n");
621 		return -ENOMEM;
622 	}
623 
624 	qp->bf = &uuari->bfs[uuarn];
625 	uar_index = qp->bf->uar->index;
626 
627 	err = calc_sq_size(dev, init_attr, qp);
628 	if (err < 0) {
629 		mlx5_ib_dbg(dev, "err %d\n", err);
630 		goto err_uuar;
631 	}
632 
633 	qp->rq.offset = 0;
634 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
635 	qp->buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
636 
637 	err = mlx5_buf_alloc(&dev->mdev, qp->buf_size, PAGE_SIZE * 2, &qp->buf);
638 	if (err) {
639 		mlx5_ib_dbg(dev, "err %d\n", err);
640 		goto err_uuar;
641 	}
642 
643 	qp->sq.qend = mlx5_get_send_wqe(qp, qp->sq.wqe_cnt);
644 	*inlen = sizeof(**in) + sizeof(*(*in)->pas) * qp->buf.npages;
645 	*in = mlx5_vzalloc(*inlen);
646 	if (!*in) {
647 		err = -ENOMEM;
648 		goto err_buf;
649 	}
650 	(*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
651 	(*in)->ctx.log_pg_sz_remote_qpn =
652 		cpu_to_be32((qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
653 	/* Set "fast registration enabled" for all kernel QPs */
654 	(*in)->ctx.params1 |= cpu_to_be32(1 << 11);
655 	(*in)->ctx.sq_crq_size |= cpu_to_be16(1 << 4);
656 
657 	mlx5_fill_page_array(&qp->buf, (*in)->pas);
658 
659 	err = mlx5_db_alloc(&dev->mdev, &qp->db);
660 	if (err) {
661 		mlx5_ib_dbg(dev, "err %d\n", err);
662 		goto err_free;
663 	}
664 
665 	qp->db.db[0] = 0;
666 	qp->db.db[1] = 0;
667 
668 	qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
669 	qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
670 	qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
671 	qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
672 	qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
673 
674 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
675 	    !qp->sq.w_list || !qp->sq.wqe_head) {
676 		err = -ENOMEM;
677 		goto err_wrid;
678 	}
679 	qp->create_type = MLX5_QP_KERNEL;
680 
681 	return 0;
682 
683 err_wrid:
684 	mlx5_db_free(&dev->mdev, &qp->db);
685 	kfree(qp->sq.wqe_head);
686 	kfree(qp->sq.w_list);
687 	kfree(qp->sq.wrid);
688 	kfree(qp->sq.wr_data);
689 	kfree(qp->rq.wrid);
690 
691 err_free:
692 	mlx5_vfree(*in);
693 
694 err_buf:
695 	mlx5_buf_free(&dev->mdev, &qp->buf);
696 
697 err_uuar:
698 	free_uuar(&dev->mdev.priv.uuari, uuarn);
699 	return err;
700 }
701 
702 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
703 {
704 	mlx5_db_free(&dev->mdev, &qp->db);
705 	kfree(qp->sq.wqe_head);
706 	kfree(qp->sq.w_list);
707 	kfree(qp->sq.wrid);
708 	kfree(qp->sq.wr_data);
709 	kfree(qp->rq.wrid);
710 	mlx5_buf_free(&dev->mdev, &qp->buf);
711 	free_uuar(&dev->mdev.priv.uuari, qp->bf->uuarn);
712 }
713 
714 static __be32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
715 {
716 	if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
717 	    (attr->qp_type == IB_QPT_XRC_INI))
718 		return cpu_to_be32(MLX5_SRQ_RQ);
719 	else if (!qp->has_rq)
720 		return cpu_to_be32(MLX5_ZERO_LEN_RQ);
721 	else
722 		return cpu_to_be32(MLX5_NON_ZERO_RQ);
723 }
724 
725 static int is_connected(enum ib_qp_type qp_type)
726 {
727 	if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC)
728 		return 1;
729 
730 	return 0;
731 }
732 
733 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
734 			    struct ib_qp_init_attr *init_attr,
735 			    struct ib_udata *udata, struct mlx5_ib_qp *qp)
736 {
737 	struct mlx5_ib_resources *devr = &dev->devr;
738 	struct mlx5_ib_create_qp_resp resp;
739 	struct mlx5_create_qp_mbox_in *in;
740 	struct mlx5_ib_create_qp ucmd;
741 	int inlen = sizeof(*in);
742 	int err;
743 
744 	mutex_init(&qp->mutex);
745 	spin_lock_init(&qp->sq.lock);
746 	spin_lock_init(&qp->rq.lock);
747 
748 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
749 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
750 
751 	if (pd && pd->uobject) {
752 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
753 			mlx5_ib_dbg(dev, "copy failed\n");
754 			return -EFAULT;
755 		}
756 
757 		qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
758 		qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
759 	} else {
760 		qp->wq_sig = !!wq_signature;
761 	}
762 
763 	qp->has_rq = qp_has_rq(init_attr);
764 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
765 			  qp, (pd && pd->uobject) ? &ucmd : NULL);
766 	if (err) {
767 		mlx5_ib_dbg(dev, "err %d\n", err);
768 		return err;
769 	}
770 
771 	if (pd) {
772 		if (pd->uobject) {
773 			mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
774 			if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
775 			    ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
776 				mlx5_ib_dbg(dev, "invalid rq params\n");
777 				return -EINVAL;
778 			}
779 			if (ucmd.sq_wqe_count > dev->mdev.caps.max_wqes) {
780 				mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
781 					    ucmd.sq_wqe_count, dev->mdev.caps.max_wqes);
782 				return -EINVAL;
783 			}
784 			err = create_user_qp(dev, pd, qp, udata, &in, &resp, &inlen);
785 			if (err)
786 				mlx5_ib_dbg(dev, "err %d\n", err);
787 		} else {
788 			err = create_kernel_qp(dev, init_attr, qp, &in, &inlen);
789 			if (err)
790 				mlx5_ib_dbg(dev, "err %d\n", err);
791 			else
792 				qp->pa_lkey = to_mpd(pd)->pa_lkey;
793 		}
794 
795 		if (err)
796 			return err;
797 	} else {
798 		in = mlx5_vzalloc(sizeof(*in));
799 		if (!in)
800 			return -ENOMEM;
801 
802 		qp->create_type = MLX5_QP_EMPTY;
803 	}
804 
805 	if (is_sqp(init_attr->qp_type))
806 		qp->port = init_attr->port_num;
807 
808 	in->ctx.flags = cpu_to_be32(to_mlx5_st(init_attr->qp_type) << 16 |
809 				    MLX5_QP_PM_MIGRATED << 11);
810 
811 	if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
812 		in->ctx.flags_pd = cpu_to_be32(to_mpd(pd ? pd : devr->p0)->pdn);
813 	else
814 		in->ctx.flags_pd = cpu_to_be32(MLX5_QP_LAT_SENSITIVE);
815 
816 	if (qp->wq_sig)
817 		in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_ENABLE_SIG);
818 
819 	if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
820 		int rcqe_sz;
821 		int scqe_sz;
822 
823 		rcqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->recv_cq);
824 		scqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->send_cq);
825 
826 		if (rcqe_sz == 128)
827 			in->ctx.cs_res = MLX5_RES_SCAT_DATA64_CQE;
828 		else
829 			in->ctx.cs_res = MLX5_RES_SCAT_DATA32_CQE;
830 
831 		if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) {
832 			if (scqe_sz == 128)
833 				in->ctx.cs_req = MLX5_REQ_SCAT_DATA64_CQE;
834 			else
835 				in->ctx.cs_req = MLX5_REQ_SCAT_DATA32_CQE;
836 		}
837 	}
838 
839 	if (qp->rq.wqe_cnt) {
840 		in->ctx.rq_size_stride = (qp->rq.wqe_shift - 4);
841 		in->ctx.rq_size_stride |= ilog2(qp->rq.wqe_cnt) << 3;
842 	}
843 
844 	in->ctx.rq_type_srqn = get_rx_type(qp, init_attr);
845 
846 	if (qp->sq.wqe_cnt)
847 		in->ctx.sq_crq_size |= cpu_to_be16(ilog2(qp->sq.wqe_cnt) << 11);
848 	else
849 		in->ctx.sq_crq_size |= cpu_to_be16(0x8000);
850 
851 	/* Set default resources */
852 	switch (init_attr->qp_type) {
853 	case IB_QPT_XRC_TGT:
854 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
855 		in->ctx.cqn_send = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
856 		in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
857 		in->ctx.xrcd = cpu_to_be32(to_mxrcd(init_attr->xrcd)->xrcdn);
858 		break;
859 	case IB_QPT_XRC_INI:
860 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
861 		in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
862 		in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
863 		break;
864 	default:
865 		if (init_attr->srq) {
866 			in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x0)->xrcdn);
867 			in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(init_attr->srq)->msrq.srqn);
868 		} else {
869 			in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
870 			in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
871 		}
872 	}
873 
874 	if (init_attr->send_cq)
875 		in->ctx.cqn_send = cpu_to_be32(to_mcq(init_attr->send_cq)->mcq.cqn);
876 
877 	if (init_attr->recv_cq)
878 		in->ctx.cqn_recv = cpu_to_be32(to_mcq(init_attr->recv_cq)->mcq.cqn);
879 
880 	in->ctx.db_rec_addr = cpu_to_be64(qp->db.dma);
881 
882 	err = mlx5_core_create_qp(&dev->mdev, &qp->mqp, in, inlen);
883 	if (err) {
884 		mlx5_ib_dbg(dev, "create qp failed\n");
885 		goto err_create;
886 	}
887 
888 	mlx5_vfree(in);
889 	/* Hardware wants QPN written in big-endian order (after
890 	 * shifting) for send doorbell.  Precompute this value to save
891 	 * a little bit when posting sends.
892 	 */
893 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
894 
895 	qp->mqp.event = mlx5_ib_qp_event;
896 
897 	return 0;
898 
899 err_create:
900 	if (qp->create_type == MLX5_QP_USER)
901 		destroy_qp_user(pd, qp);
902 	else if (qp->create_type == MLX5_QP_KERNEL)
903 		destroy_qp_kernel(dev, qp);
904 
905 	mlx5_vfree(in);
906 	return err;
907 }
908 
909 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
910 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
911 {
912 	if (send_cq) {
913 		if (recv_cq) {
914 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
915 				spin_lock_irq(&send_cq->lock);
916 				spin_lock_nested(&recv_cq->lock,
917 						 SINGLE_DEPTH_NESTING);
918 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
919 				spin_lock_irq(&send_cq->lock);
920 				__acquire(&recv_cq->lock);
921 			} else {
922 				spin_lock_irq(&recv_cq->lock);
923 				spin_lock_nested(&send_cq->lock,
924 						 SINGLE_DEPTH_NESTING);
925 			}
926 		} else {
927 			spin_lock_irq(&send_cq->lock);
928 		}
929 	} else if (recv_cq) {
930 		spin_lock_irq(&recv_cq->lock);
931 	}
932 }
933 
934 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
935 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
936 {
937 	if (send_cq) {
938 		if (recv_cq) {
939 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
940 				spin_unlock(&recv_cq->lock);
941 				spin_unlock_irq(&send_cq->lock);
942 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
943 				__release(&recv_cq->lock);
944 				spin_unlock_irq(&send_cq->lock);
945 			} else {
946 				spin_unlock(&send_cq->lock);
947 				spin_unlock_irq(&recv_cq->lock);
948 			}
949 		} else {
950 			spin_unlock_irq(&send_cq->lock);
951 		}
952 	} else if (recv_cq) {
953 		spin_unlock_irq(&recv_cq->lock);
954 	}
955 }
956 
957 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
958 {
959 	return to_mpd(qp->ibqp.pd);
960 }
961 
962 static void get_cqs(struct mlx5_ib_qp *qp,
963 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
964 {
965 	switch (qp->ibqp.qp_type) {
966 	case IB_QPT_XRC_TGT:
967 		*send_cq = NULL;
968 		*recv_cq = NULL;
969 		break;
970 	case MLX5_IB_QPT_REG_UMR:
971 	case IB_QPT_XRC_INI:
972 		*send_cq = to_mcq(qp->ibqp.send_cq);
973 		*recv_cq = NULL;
974 		break;
975 
976 	case IB_QPT_SMI:
977 	case IB_QPT_GSI:
978 	case IB_QPT_RC:
979 	case IB_QPT_UC:
980 	case IB_QPT_UD:
981 	case IB_QPT_RAW_IPV6:
982 	case IB_QPT_RAW_ETHERTYPE:
983 		*send_cq = to_mcq(qp->ibqp.send_cq);
984 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
985 		break;
986 
987 	case IB_QPT_RAW_PACKET:
988 	case IB_QPT_MAX:
989 	default:
990 		*send_cq = NULL;
991 		*recv_cq = NULL;
992 		break;
993 	}
994 }
995 
996 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
997 {
998 	struct mlx5_ib_cq *send_cq, *recv_cq;
999 	struct mlx5_modify_qp_mbox_in *in;
1000 	int err;
1001 
1002 	in = kzalloc(sizeof(*in), GFP_KERNEL);
1003 	if (!in)
1004 		return;
1005 	if (qp->state != IB_QPS_RESET)
1006 		if (mlx5_core_qp_modify(&dev->mdev, to_mlx5_state(qp->state),
1007 					MLX5_QP_STATE_RST, in, sizeof(*in), &qp->mqp))
1008 			mlx5_ib_warn(dev, "mlx5_ib: modify QP %06x to RESET failed\n",
1009 				     qp->mqp.qpn);
1010 
1011 	get_cqs(qp, &send_cq, &recv_cq);
1012 
1013 	if (qp->create_type == MLX5_QP_KERNEL) {
1014 		mlx5_ib_lock_cqs(send_cq, recv_cq);
1015 		__mlx5_ib_cq_clean(recv_cq, qp->mqp.qpn,
1016 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1017 		if (send_cq != recv_cq)
1018 			__mlx5_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1019 		mlx5_ib_unlock_cqs(send_cq, recv_cq);
1020 	}
1021 
1022 	err = mlx5_core_destroy_qp(&dev->mdev, &qp->mqp);
1023 	if (err)
1024 		mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n", qp->mqp.qpn);
1025 	kfree(in);
1026 
1027 
1028 	if (qp->create_type == MLX5_QP_KERNEL)
1029 		destroy_qp_kernel(dev, qp);
1030 	else if (qp->create_type == MLX5_QP_USER)
1031 		destroy_qp_user(&get_pd(qp)->ibpd, qp);
1032 }
1033 
1034 static const char *ib_qp_type_str(enum ib_qp_type type)
1035 {
1036 	switch (type) {
1037 	case IB_QPT_SMI:
1038 		return "IB_QPT_SMI";
1039 	case IB_QPT_GSI:
1040 		return "IB_QPT_GSI";
1041 	case IB_QPT_RC:
1042 		return "IB_QPT_RC";
1043 	case IB_QPT_UC:
1044 		return "IB_QPT_UC";
1045 	case IB_QPT_UD:
1046 		return "IB_QPT_UD";
1047 	case IB_QPT_RAW_IPV6:
1048 		return "IB_QPT_RAW_IPV6";
1049 	case IB_QPT_RAW_ETHERTYPE:
1050 		return "IB_QPT_RAW_ETHERTYPE";
1051 	case IB_QPT_XRC_INI:
1052 		return "IB_QPT_XRC_INI";
1053 	case IB_QPT_XRC_TGT:
1054 		return "IB_QPT_XRC_TGT";
1055 	case IB_QPT_RAW_PACKET:
1056 		return "IB_QPT_RAW_PACKET";
1057 	case MLX5_IB_QPT_REG_UMR:
1058 		return "MLX5_IB_QPT_REG_UMR";
1059 	case IB_QPT_MAX:
1060 	default:
1061 		return "Invalid QP type";
1062 	}
1063 }
1064 
1065 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
1066 				struct ib_qp_init_attr *init_attr,
1067 				struct ib_udata *udata)
1068 {
1069 	struct mlx5_ib_dev *dev;
1070 	struct mlx5_ib_qp *qp;
1071 	u16 xrcdn = 0;
1072 	int err;
1073 
1074 	if (pd) {
1075 		dev = to_mdev(pd->device);
1076 	} else {
1077 		/* being cautious here */
1078 		if (init_attr->qp_type != IB_QPT_XRC_TGT &&
1079 		    init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
1080 			pr_warn("%s: no PD for transport %s\n", __func__,
1081 				ib_qp_type_str(init_attr->qp_type));
1082 			return ERR_PTR(-EINVAL);
1083 		}
1084 		dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
1085 	}
1086 
1087 	switch (init_attr->qp_type) {
1088 	case IB_QPT_XRC_TGT:
1089 	case IB_QPT_XRC_INI:
1090 		if (!(dev->mdev.caps.flags & MLX5_DEV_CAP_FLAG_XRC)) {
1091 			mlx5_ib_dbg(dev, "XRC not supported\n");
1092 			return ERR_PTR(-ENOSYS);
1093 		}
1094 		init_attr->recv_cq = NULL;
1095 		if (init_attr->qp_type == IB_QPT_XRC_TGT) {
1096 			xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1097 			init_attr->send_cq = NULL;
1098 		}
1099 
1100 		/* fall through */
1101 	case IB_QPT_RC:
1102 	case IB_QPT_UC:
1103 	case IB_QPT_UD:
1104 	case IB_QPT_SMI:
1105 	case IB_QPT_GSI:
1106 	case MLX5_IB_QPT_REG_UMR:
1107 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1108 		if (!qp)
1109 			return ERR_PTR(-ENOMEM);
1110 
1111 		err = create_qp_common(dev, pd, init_attr, udata, qp);
1112 		if (err) {
1113 			mlx5_ib_dbg(dev, "create_qp_common failed\n");
1114 			kfree(qp);
1115 			return ERR_PTR(err);
1116 		}
1117 
1118 		if (is_qp0(init_attr->qp_type))
1119 			qp->ibqp.qp_num = 0;
1120 		else if (is_qp1(init_attr->qp_type))
1121 			qp->ibqp.qp_num = 1;
1122 		else
1123 			qp->ibqp.qp_num = qp->mqp.qpn;
1124 
1125 		mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
1126 			    qp->ibqp.qp_num, qp->mqp.qpn, to_mcq(init_attr->recv_cq)->mcq.cqn,
1127 			    to_mcq(init_attr->send_cq)->mcq.cqn);
1128 
1129 		qp->xrcdn = xrcdn;
1130 
1131 		break;
1132 
1133 	case IB_QPT_RAW_IPV6:
1134 	case IB_QPT_RAW_ETHERTYPE:
1135 	case IB_QPT_RAW_PACKET:
1136 	case IB_QPT_MAX:
1137 	default:
1138 		mlx5_ib_dbg(dev, "unsupported qp type %d\n",
1139 			    init_attr->qp_type);
1140 		/* Don't support raw QPs */
1141 		return ERR_PTR(-EINVAL);
1142 	}
1143 
1144 	return &qp->ibqp;
1145 }
1146 
1147 int mlx5_ib_destroy_qp(struct ib_qp *qp)
1148 {
1149 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
1150 	struct mlx5_ib_qp *mqp = to_mqp(qp);
1151 
1152 	destroy_qp_common(dev, mqp);
1153 
1154 	kfree(mqp);
1155 
1156 	return 0;
1157 }
1158 
1159 static __be32 to_mlx5_access_flags(struct mlx5_ib_qp *qp, const struct ib_qp_attr *attr,
1160 				   int attr_mask)
1161 {
1162 	u32 hw_access_flags = 0;
1163 	u8 dest_rd_atomic;
1164 	u32 access_flags;
1165 
1166 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1167 		dest_rd_atomic = attr->max_dest_rd_atomic;
1168 	else
1169 		dest_rd_atomic = qp->resp_depth;
1170 
1171 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1172 		access_flags = attr->qp_access_flags;
1173 	else
1174 		access_flags = qp->atomic_rd_en;
1175 
1176 	if (!dest_rd_atomic)
1177 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1178 
1179 	if (access_flags & IB_ACCESS_REMOTE_READ)
1180 		hw_access_flags |= MLX5_QP_BIT_RRE;
1181 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1182 		hw_access_flags |= (MLX5_QP_BIT_RAE | MLX5_ATOMIC_MODE_CX);
1183 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1184 		hw_access_flags |= MLX5_QP_BIT_RWE;
1185 
1186 	return cpu_to_be32(hw_access_flags);
1187 }
1188 
1189 enum {
1190 	MLX5_PATH_FLAG_FL	= 1 << 0,
1191 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
1192 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
1193 };
1194 
1195 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
1196 {
1197 	if (rate == IB_RATE_PORT_CURRENT) {
1198 		return 0;
1199 	} else if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_300_GBPS) {
1200 		return -EINVAL;
1201 	} else {
1202 		while (rate != IB_RATE_2_5_GBPS &&
1203 		       !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
1204 			 dev->mdev.caps.stat_rate_support))
1205 			--rate;
1206 	}
1207 
1208 	return rate + MLX5_STAT_RATE_OFFSET;
1209 }
1210 
1211 static int mlx5_set_path(struct mlx5_ib_dev *dev, const struct ib_ah_attr *ah,
1212 			 struct mlx5_qp_path *path, u8 port, int attr_mask,
1213 			 u32 path_flags, const struct ib_qp_attr *attr)
1214 {
1215 	int err;
1216 
1217 	path->fl = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
1218 	path->free_ar = (path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x80 : 0;
1219 
1220 	if (attr_mask & IB_QP_PKEY_INDEX)
1221 		path->pkey_index = attr->pkey_index;
1222 
1223 	path->grh_mlid	= ah->src_path_bits & 0x7f;
1224 	path->rlid	= cpu_to_be16(ah->dlid);
1225 
1226 	if (ah->ah_flags & IB_AH_GRH) {
1227 		path->grh_mlid |= 1 << 7;
1228 		path->mgid_index = ah->grh.sgid_index;
1229 		path->hop_limit  = ah->grh.hop_limit;
1230 		path->tclass_flowlabel =
1231 			cpu_to_be32((ah->grh.traffic_class << 20) |
1232 				    (ah->grh.flow_label));
1233 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1234 	}
1235 
1236 	err = ib_rate_to_mlx5(dev, ah->static_rate);
1237 	if (err < 0)
1238 		return err;
1239 	path->static_rate = err;
1240 	path->port = port;
1241 
1242 	if (ah->ah_flags & IB_AH_GRH) {
1243 		if (ah->grh.sgid_index >= dev->mdev.caps.port[port - 1].gid_table_len) {
1244 			pr_err(KERN_ERR "sgid_index (%u) too large. max is %d\n",
1245 			       ah->grh.sgid_index, dev->mdev.caps.port[port - 1].gid_table_len);
1246 			return -EINVAL;
1247 		}
1248 
1249 		path->grh_mlid |= 1 << 7;
1250 		path->mgid_index = ah->grh.sgid_index;
1251 		path->hop_limit  = ah->grh.hop_limit;
1252 		path->tclass_flowlabel =
1253 			cpu_to_be32((ah->grh.traffic_class << 20) |
1254 				    (ah->grh.flow_label));
1255 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1256 	}
1257 
1258 	if (attr_mask & IB_QP_TIMEOUT)
1259 		path->ackto_lt = attr->timeout << 3;
1260 
1261 	path->sl = ah->sl & 0xf;
1262 
1263 	return 0;
1264 }
1265 
1266 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
1267 	[MLX5_QP_STATE_INIT] = {
1268 		[MLX5_QP_STATE_INIT] = {
1269 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
1270 					  MLX5_QP_OPTPAR_RAE		|
1271 					  MLX5_QP_OPTPAR_RWE		|
1272 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
1273 					  MLX5_QP_OPTPAR_PRI_PORT,
1274 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
1275 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
1276 					  MLX5_QP_OPTPAR_PRI_PORT,
1277 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
1278 					  MLX5_QP_OPTPAR_Q_KEY		|
1279 					  MLX5_QP_OPTPAR_PRI_PORT,
1280 		},
1281 		[MLX5_QP_STATE_RTR] = {
1282 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
1283 					  MLX5_QP_OPTPAR_RRE            |
1284 					  MLX5_QP_OPTPAR_RAE            |
1285 					  MLX5_QP_OPTPAR_RWE            |
1286 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1287 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
1288 					  MLX5_QP_OPTPAR_RWE            |
1289 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1290 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
1291 					  MLX5_QP_OPTPAR_Q_KEY,
1292 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
1293 					   MLX5_QP_OPTPAR_Q_KEY,
1294 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
1295 					  MLX5_QP_OPTPAR_RRE            |
1296 					  MLX5_QP_OPTPAR_RAE            |
1297 					  MLX5_QP_OPTPAR_RWE            |
1298 					  MLX5_QP_OPTPAR_PKEY_INDEX,
1299 		},
1300 	},
1301 	[MLX5_QP_STATE_RTR] = {
1302 		[MLX5_QP_STATE_RTS] = {
1303 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
1304 					  MLX5_QP_OPTPAR_RRE		|
1305 					  MLX5_QP_OPTPAR_RAE		|
1306 					  MLX5_QP_OPTPAR_RWE		|
1307 					  MLX5_QP_OPTPAR_PM_STATE	|
1308 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
1309 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
1310 					  MLX5_QP_OPTPAR_RWE		|
1311 					  MLX5_QP_OPTPAR_PM_STATE,
1312 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
1313 		},
1314 	},
1315 	[MLX5_QP_STATE_RTS] = {
1316 		[MLX5_QP_STATE_RTS] = {
1317 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
1318 					  MLX5_QP_OPTPAR_RAE		|
1319 					  MLX5_QP_OPTPAR_RWE		|
1320 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
1321 					  MLX5_QP_OPTPAR_PM_STATE	|
1322 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
1323 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
1324 					  MLX5_QP_OPTPAR_PM_STATE	|
1325 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
1326 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
1327 					  MLX5_QP_OPTPAR_SRQN		|
1328 					  MLX5_QP_OPTPAR_CQN_RCV,
1329 		},
1330 	},
1331 	[MLX5_QP_STATE_SQER] = {
1332 		[MLX5_QP_STATE_RTS] = {
1333 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
1334 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
1335 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
1336 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
1337 					   MLX5_QP_OPTPAR_RWE		|
1338 					   MLX5_QP_OPTPAR_RAE		|
1339 					   MLX5_QP_OPTPAR_RRE,
1340 		},
1341 	},
1342 };
1343 
1344 static int ib_nr_to_mlx5_nr(int ib_mask)
1345 {
1346 	switch (ib_mask) {
1347 	case IB_QP_STATE:
1348 		return 0;
1349 	case IB_QP_CUR_STATE:
1350 		return 0;
1351 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
1352 		return 0;
1353 	case IB_QP_ACCESS_FLAGS:
1354 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
1355 			MLX5_QP_OPTPAR_RAE;
1356 	case IB_QP_PKEY_INDEX:
1357 		return MLX5_QP_OPTPAR_PKEY_INDEX;
1358 	case IB_QP_PORT:
1359 		return MLX5_QP_OPTPAR_PRI_PORT;
1360 	case IB_QP_QKEY:
1361 		return MLX5_QP_OPTPAR_Q_KEY;
1362 	case IB_QP_AV:
1363 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
1364 			MLX5_QP_OPTPAR_PRI_PORT;
1365 	case IB_QP_PATH_MTU:
1366 		return 0;
1367 	case IB_QP_TIMEOUT:
1368 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
1369 	case IB_QP_RETRY_CNT:
1370 		return MLX5_QP_OPTPAR_RETRY_COUNT;
1371 	case IB_QP_RNR_RETRY:
1372 		return MLX5_QP_OPTPAR_RNR_RETRY;
1373 	case IB_QP_RQ_PSN:
1374 		return 0;
1375 	case IB_QP_MAX_QP_RD_ATOMIC:
1376 		return MLX5_QP_OPTPAR_SRA_MAX;
1377 	case IB_QP_ALT_PATH:
1378 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
1379 	case IB_QP_MIN_RNR_TIMER:
1380 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
1381 	case IB_QP_SQ_PSN:
1382 		return 0;
1383 	case IB_QP_MAX_DEST_RD_ATOMIC:
1384 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
1385 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
1386 	case IB_QP_PATH_MIG_STATE:
1387 		return MLX5_QP_OPTPAR_PM_STATE;
1388 	case IB_QP_CAP:
1389 		return 0;
1390 	case IB_QP_DEST_QPN:
1391 		return 0;
1392 	}
1393 	return 0;
1394 }
1395 
1396 static int ib_mask_to_mlx5_opt(int ib_mask)
1397 {
1398 	int result = 0;
1399 	int i;
1400 
1401 	for (i = 0; i < 8 * sizeof(int); i++) {
1402 		if ((1 << i) & ib_mask)
1403 			result |= ib_nr_to_mlx5_nr(1 << i);
1404 	}
1405 
1406 	return result;
1407 }
1408 
1409 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
1410 			       const struct ib_qp_attr *attr, int attr_mask,
1411 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1412 {
1413 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1414 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
1415 	struct mlx5_ib_cq *send_cq, *recv_cq;
1416 	struct mlx5_qp_context *context;
1417 	struct mlx5_modify_qp_mbox_in *in;
1418 	struct mlx5_ib_pd *pd;
1419 	enum mlx5_qp_state mlx5_cur, mlx5_new;
1420 	enum mlx5_qp_optpar optpar;
1421 	int sqd_event;
1422 	int mlx5_st;
1423 	int err;
1424 
1425 	in = kzalloc(sizeof(*in), GFP_KERNEL);
1426 	if (!in)
1427 		return -ENOMEM;
1428 
1429 	context = &in->ctx;
1430 	err = to_mlx5_st(ibqp->qp_type);
1431 	if (err < 0)
1432 		goto out;
1433 
1434 	context->flags = cpu_to_be32(err << 16);
1435 
1436 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
1437 		context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
1438 	} else {
1439 		switch (attr->path_mig_state) {
1440 		case IB_MIG_MIGRATED:
1441 			context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
1442 			break;
1443 		case IB_MIG_REARM:
1444 			context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
1445 			break;
1446 		case IB_MIG_ARMED:
1447 			context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
1448 			break;
1449 		}
1450 	}
1451 
1452 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI) {
1453 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
1454 	} else if (ibqp->qp_type == IB_QPT_UD ||
1455 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
1456 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1457 	} else if (attr_mask & IB_QP_PATH_MTU) {
1458 		if (attr->path_mtu < IB_MTU_256 ||
1459 		    attr->path_mtu > IB_MTU_4096) {
1460 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
1461 			err = -EINVAL;
1462 			goto out;
1463 		}
1464 		context->mtu_msgmax = (attr->path_mtu << 5) | dev->mdev.caps.log_max_msg;
1465 	}
1466 
1467 	if (attr_mask & IB_QP_DEST_QPN)
1468 		context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
1469 
1470 	if (attr_mask & IB_QP_PKEY_INDEX)
1471 		context->pri_path.pkey_index = attr->pkey_index;
1472 
1473 	/* todo implement counter_index functionality */
1474 
1475 	if (is_sqp(ibqp->qp_type))
1476 		context->pri_path.port = qp->port;
1477 
1478 	if (attr_mask & IB_QP_PORT)
1479 		context->pri_path.port = attr->port_num;
1480 
1481 	if (attr_mask & IB_QP_AV) {
1482 		err = mlx5_set_path(dev, &attr->ah_attr, &context->pri_path,
1483 				    attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
1484 				    attr_mask, 0, attr);
1485 		if (err)
1486 			goto out;
1487 	}
1488 
1489 	if (attr_mask & IB_QP_TIMEOUT)
1490 		context->pri_path.ackto_lt |= attr->timeout << 3;
1491 
1492 	if (attr_mask & IB_QP_ALT_PATH) {
1493 		err = mlx5_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
1494 				    attr->alt_port_num, attr_mask, 0, attr);
1495 		if (err)
1496 			goto out;
1497 	}
1498 
1499 	pd = get_pd(qp);
1500 	get_cqs(qp, &send_cq, &recv_cq);
1501 
1502 	context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
1503 	context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
1504 	context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
1505 	context->params1  = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
1506 
1507 	if (attr_mask & IB_QP_RNR_RETRY)
1508 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1509 
1510 	if (attr_mask & IB_QP_RETRY_CNT)
1511 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1512 
1513 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1514 		if (attr->max_rd_atomic)
1515 			context->params1 |=
1516 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1517 	}
1518 
1519 	if (attr_mask & IB_QP_SQ_PSN)
1520 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1521 
1522 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1523 		if (attr->max_dest_rd_atomic)
1524 			context->params2 |=
1525 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1526 	}
1527 
1528 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
1529 		context->params2 |= to_mlx5_access_flags(qp, attr, attr_mask);
1530 
1531 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
1532 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1533 
1534 	if (attr_mask & IB_QP_RQ_PSN)
1535 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1536 
1537 	if (attr_mask & IB_QP_QKEY)
1538 		context->qkey = cpu_to_be32(attr->qkey);
1539 
1540 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1541 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1542 
1543 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1544 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1545 		sqd_event = 1;
1546 	else
1547 		sqd_event = 0;
1548 
1549 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1550 		context->sq_crq_size |= cpu_to_be16(1 << 4);
1551 
1552 
1553 	mlx5_cur = to_mlx5_state(cur_state);
1554 	mlx5_new = to_mlx5_state(new_state);
1555 	mlx5_st = to_mlx5_st(ibqp->qp_type);
1556 	if (mlx5_st < 0)
1557 		goto out;
1558 
1559 	optpar = ib_mask_to_mlx5_opt(attr_mask);
1560 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
1561 	in->optparam = cpu_to_be32(optpar);
1562 	err = mlx5_core_qp_modify(&dev->mdev, to_mlx5_state(cur_state),
1563 				  to_mlx5_state(new_state), in, sqd_event,
1564 				  &qp->mqp);
1565 	if (err)
1566 		goto out;
1567 
1568 	qp->state = new_state;
1569 
1570 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1571 		qp->atomic_rd_en = attr->qp_access_flags;
1572 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1573 		qp->resp_depth = attr->max_dest_rd_atomic;
1574 	if (attr_mask & IB_QP_PORT)
1575 		qp->port = attr->port_num;
1576 	if (attr_mask & IB_QP_ALT_PATH)
1577 		qp->alt_port = attr->alt_port_num;
1578 
1579 	/*
1580 	 * If we moved a kernel QP to RESET, clean up all old CQ
1581 	 * entries and reinitialize the QP.
1582 	 */
1583 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1584 		mlx5_ib_cq_clean(recv_cq, qp->mqp.qpn,
1585 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
1586 		if (send_cq != recv_cq)
1587 			mlx5_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1588 
1589 		qp->rq.head = 0;
1590 		qp->rq.tail = 0;
1591 		qp->sq.head = 0;
1592 		qp->sq.tail = 0;
1593 		qp->sq.cur_post = 0;
1594 		qp->sq.last_poll = 0;
1595 		qp->db.db[MLX5_RCV_DBR] = 0;
1596 		qp->db.db[MLX5_SND_DBR] = 0;
1597 	}
1598 
1599 out:
1600 	kfree(in);
1601 	return err;
1602 }
1603 
1604 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1605 		      int attr_mask, struct ib_udata *udata)
1606 {
1607 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1608 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
1609 	enum ib_qp_state cur_state, new_state;
1610 	int err = -EINVAL;
1611 	int port;
1612 
1613 	mutex_lock(&qp->mutex);
1614 
1615 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1616 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1617 
1618 	if (ibqp->qp_type != MLX5_IB_QPT_REG_UMR &&
1619 	    !ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask))
1620 		goto out;
1621 
1622 	if ((attr_mask & IB_QP_PORT) &&
1623 	    (attr->port_num == 0 || attr->port_num > dev->mdev.caps.num_ports))
1624 		goto out;
1625 
1626 	if (attr_mask & IB_QP_PKEY_INDEX) {
1627 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1628 		if (attr->pkey_index >= dev->mdev.caps.port[port - 1].pkey_table_len)
1629 			goto out;
1630 	}
1631 
1632 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1633 	    attr->max_rd_atomic > dev->mdev.caps.max_ra_res_qp)
1634 		goto out;
1635 
1636 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1637 	    attr->max_dest_rd_atomic > dev->mdev.caps.max_ra_req_qp)
1638 		goto out;
1639 
1640 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1641 		err = 0;
1642 		goto out;
1643 	}
1644 
1645 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1646 
1647 out:
1648 	mutex_unlock(&qp->mutex);
1649 	return err;
1650 }
1651 
1652 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1653 {
1654 	struct mlx5_ib_cq *cq;
1655 	unsigned cur;
1656 
1657 	cur = wq->head - wq->tail;
1658 	if (likely(cur + nreq < wq->max_post))
1659 		return 0;
1660 
1661 	cq = to_mcq(ib_cq);
1662 	spin_lock(&cq->lock);
1663 	cur = wq->head - wq->tail;
1664 	spin_unlock(&cq->lock);
1665 
1666 	return cur + nreq >= wq->max_post;
1667 }
1668 
1669 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
1670 					  u64 remote_addr, u32 rkey)
1671 {
1672 	rseg->raddr    = cpu_to_be64(remote_addr);
1673 	rseg->rkey     = cpu_to_be32(rkey);
1674 	rseg->reserved = 0;
1675 }
1676 
1677 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
1678 			     struct ib_send_wr *wr)
1679 {
1680 	memcpy(&dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof(struct mlx5_av));
1681 	dseg->av.dqp_dct = cpu_to_be32(wr->wr.ud.remote_qpn | MLX5_EXTENDED_UD_AV);
1682 	dseg->av.key.qkey.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1683 }
1684 
1685 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
1686 {
1687 	dseg->byte_count = cpu_to_be32(sg->length);
1688 	dseg->lkey       = cpu_to_be32(sg->lkey);
1689 	dseg->addr       = cpu_to_be64(sg->addr);
1690 }
1691 
1692 static __be16 get_klm_octo(int npages)
1693 {
1694 	return cpu_to_be16(ALIGN(npages, 8) / 2);
1695 }
1696 
1697 static __be64 frwr_mkey_mask(void)
1698 {
1699 	u64 result;
1700 
1701 	result = MLX5_MKEY_MASK_LEN		|
1702 		MLX5_MKEY_MASK_PAGE_SIZE	|
1703 		MLX5_MKEY_MASK_START_ADDR	|
1704 		MLX5_MKEY_MASK_EN_RINVAL	|
1705 		MLX5_MKEY_MASK_KEY		|
1706 		MLX5_MKEY_MASK_LR		|
1707 		MLX5_MKEY_MASK_LW		|
1708 		MLX5_MKEY_MASK_RR		|
1709 		MLX5_MKEY_MASK_RW		|
1710 		MLX5_MKEY_MASK_A		|
1711 		MLX5_MKEY_MASK_SMALL_FENCE	|
1712 		MLX5_MKEY_MASK_FREE;
1713 
1714 	return cpu_to_be64(result);
1715 }
1716 
1717 static void set_frwr_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
1718 				 struct ib_send_wr *wr, int li)
1719 {
1720 	memset(umr, 0, sizeof(*umr));
1721 
1722 	if (li) {
1723 		umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
1724 		umr->flags = 1 << 7;
1725 		return;
1726 	}
1727 
1728 	umr->flags = (1 << 5); /* fail if not free */
1729 	umr->klm_octowords = get_klm_octo(wr->wr.fast_reg.page_list_len);
1730 	umr->mkey_mask = frwr_mkey_mask();
1731 }
1732 
1733 static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
1734 				struct ib_send_wr *wr)
1735 {
1736 	struct umr_wr *umrwr = (struct umr_wr *)&wr->wr.fast_reg;
1737 	u64 mask;
1738 
1739 	memset(umr, 0, sizeof(*umr));
1740 
1741 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UNREG)) {
1742 		umr->flags = 1 << 5; /* fail if not free */
1743 		umr->klm_octowords = get_klm_octo(umrwr->npages);
1744 		mask =  MLX5_MKEY_MASK_LEN		|
1745 			MLX5_MKEY_MASK_PAGE_SIZE	|
1746 			MLX5_MKEY_MASK_START_ADDR	|
1747 			MLX5_MKEY_MASK_PD		|
1748 			MLX5_MKEY_MASK_LR		|
1749 			MLX5_MKEY_MASK_LW		|
1750 			MLX5_MKEY_MASK_KEY		|
1751 			MLX5_MKEY_MASK_RR		|
1752 			MLX5_MKEY_MASK_RW		|
1753 			MLX5_MKEY_MASK_A		|
1754 			MLX5_MKEY_MASK_FREE;
1755 		umr->mkey_mask = cpu_to_be64(mask);
1756 	} else {
1757 		umr->flags = 2 << 5; /* fail if free */
1758 		mask = MLX5_MKEY_MASK_FREE;
1759 		umr->mkey_mask = cpu_to_be64(mask);
1760 	}
1761 
1762 	if (!wr->num_sge)
1763 		umr->flags |= (1 << 7); /* inline */
1764 }
1765 
1766 static u8 get_umr_flags(int acc)
1767 {
1768 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
1769 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
1770 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
1771 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
1772 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN | MLX5_ACCESS_MODE_MTT;
1773 }
1774 
1775 static void set_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr,
1776 			     int li, int *writ)
1777 {
1778 	memset(seg, 0, sizeof(*seg));
1779 	if (li) {
1780 		seg->status = 1 << 6;
1781 		return;
1782 	}
1783 
1784 	seg->flags = get_umr_flags(wr->wr.fast_reg.access_flags);
1785 	*writ = seg->flags & (MLX5_PERM_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE);
1786 	seg->qpn_mkey7_0 = cpu_to_be32((wr->wr.fast_reg.rkey & 0xff) | 0xffffff00);
1787 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
1788 	seg->start_addr = cpu_to_be64(wr->wr.fast_reg.iova_start);
1789 	seg->len = cpu_to_be64(wr->wr.fast_reg.length);
1790 	seg->xlt_oct_size = cpu_to_be32((wr->wr.fast_reg.page_list_len + 1) / 2);
1791 	seg->log2_page_size = wr->wr.fast_reg.page_shift;
1792 }
1793 
1794 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr)
1795 {
1796 	memset(seg, 0, sizeof(*seg));
1797 	if (wr->send_flags & MLX5_IB_SEND_UMR_UNREG) {
1798 		seg->status = 1 << 6;
1799 		return;
1800 	}
1801 
1802 	seg->flags = convert_access(wr->wr.fast_reg.access_flags);
1803 	seg->flags_pd = cpu_to_be32(to_mpd((struct ib_pd *)wr->wr.fast_reg.page_list)->pdn);
1804 	seg->start_addr = cpu_to_be64(wr->wr.fast_reg.iova_start);
1805 	seg->len = cpu_to_be64(wr->wr.fast_reg.length);
1806 	seg->log2_page_size = wr->wr.fast_reg.page_shift;
1807 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
1808 				       mlx5_mkey_variant(wr->wr.fast_reg.rkey));
1809 }
1810 
1811 static void set_frwr_pages(struct mlx5_wqe_data_seg *dseg,
1812 			   struct ib_send_wr *wr,
1813 			   struct mlx5_core_dev *mdev,
1814 			   struct mlx5_ib_pd *pd,
1815 			   int writ)
1816 {
1817 	struct mlx5_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
1818 	u64 *page_list = wr->wr.fast_reg.page_list->page_list;
1819 	u64 perm = MLX5_EN_RD | (writ ? MLX5_EN_WR : 0);
1820 	int i;
1821 
1822 	for (i = 0; i < wr->wr.fast_reg.page_list_len; i++)
1823 		mfrpl->mapped_page_list[i] = cpu_to_be64(page_list[i] | perm);
1824 	dseg->addr = cpu_to_be64(mfrpl->map);
1825 	dseg->byte_count = cpu_to_be32(ALIGN(sizeof(u64) * wr->wr.fast_reg.page_list_len, 64));
1826 	dseg->lkey = cpu_to_be32(pd->pa_lkey);
1827 }
1828 
1829 static __be32 send_ieth(struct ib_send_wr *wr)
1830 {
1831 	switch (wr->opcode) {
1832 	case IB_WR_SEND_WITH_IMM:
1833 	case IB_WR_RDMA_WRITE_WITH_IMM:
1834 		return wr->ex.imm_data;
1835 
1836 	case IB_WR_SEND_WITH_INV:
1837 		return cpu_to_be32(wr->ex.invalidate_rkey);
1838 
1839 	default:
1840 		return 0;
1841 	}
1842 }
1843 
1844 static u8 calc_sig(void *wqe, int size)
1845 {
1846 	u8 *p = wqe;
1847 	u8 res = 0;
1848 	int i;
1849 
1850 	for (i = 0; i < size; i++)
1851 		res ^= p[i];
1852 
1853 	return ~res;
1854 }
1855 
1856 static u8 wq_sig(void *wqe)
1857 {
1858 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
1859 }
1860 
1861 static int set_data_inl_seg(struct mlx5_ib_qp *qp, struct ib_send_wr *wr,
1862 			    void *wqe, int *sz)
1863 {
1864 	struct mlx5_wqe_inline_seg *seg;
1865 	void *qend = qp->sq.qend;
1866 	void *addr;
1867 	int inl = 0;
1868 	int copy;
1869 	int len;
1870 	int i;
1871 
1872 	seg = wqe;
1873 	wqe += sizeof(*seg);
1874 	for (i = 0; i < wr->num_sge; i++) {
1875 		addr = (void *)(unsigned long)(wr->sg_list[i].addr);
1876 		len  = wr->sg_list[i].length;
1877 		inl += len;
1878 
1879 		if (unlikely(inl > qp->max_inline_data))
1880 			return -ENOMEM;
1881 
1882 		if (unlikely(wqe + len > qend)) {
1883 			copy = qend - wqe;
1884 			memcpy(wqe, addr, copy);
1885 			addr += copy;
1886 			len -= copy;
1887 			wqe = mlx5_get_send_wqe(qp, 0);
1888 		}
1889 		memcpy(wqe, addr, len);
1890 		wqe += len;
1891 	}
1892 
1893 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
1894 
1895 	*sz = ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
1896 
1897 	return 0;
1898 }
1899 
1900 static int set_frwr_li_wr(void **seg, struct ib_send_wr *wr, int *size,
1901 			  struct mlx5_core_dev *mdev, struct mlx5_ib_pd *pd, struct mlx5_ib_qp *qp)
1902 {
1903 	int writ = 0;
1904 	int li;
1905 
1906 	li = wr->opcode == IB_WR_LOCAL_INV ? 1 : 0;
1907 	if (unlikely(wr->send_flags & IB_SEND_INLINE))
1908 		return -EINVAL;
1909 
1910 	set_frwr_umr_segment(*seg, wr, li);
1911 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
1912 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
1913 	if (unlikely((*seg == qp->sq.qend)))
1914 		*seg = mlx5_get_send_wqe(qp, 0);
1915 	set_mkey_segment(*seg, wr, li, &writ);
1916 	*seg += sizeof(struct mlx5_mkey_seg);
1917 	*size += sizeof(struct mlx5_mkey_seg) / 16;
1918 	if (unlikely((*seg == qp->sq.qend)))
1919 		*seg = mlx5_get_send_wqe(qp, 0);
1920 	if (!li) {
1921 		if (unlikely(wr->wr.fast_reg.page_list_len >
1922 			     wr->wr.fast_reg.page_list->max_page_list_len))
1923 			return	-ENOMEM;
1924 
1925 		set_frwr_pages(*seg, wr, mdev, pd, writ);
1926 		*seg += sizeof(struct mlx5_wqe_data_seg);
1927 		*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
1928 	}
1929 	return 0;
1930 }
1931 
1932 static void dump_wqe(struct mlx5_ib_qp *qp, int idx, int size_16)
1933 {
1934 	__be32 *p = NULL;
1935 	int tidx = idx;
1936 	int i, j;
1937 
1938 	pr_debug("dump wqe at %p\n", mlx5_get_send_wqe(qp, tidx));
1939 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
1940 		if ((i & 0xf) == 0) {
1941 			void *buf = mlx5_get_send_wqe(qp, tidx);
1942 			tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
1943 			p = buf;
1944 			j = 0;
1945 		}
1946 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
1947 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
1948 			 be32_to_cpu(p[j + 3]));
1949 	}
1950 }
1951 
1952 static void mlx5_bf_copy(u64 __iomem *dst, u64 *src,
1953 			 unsigned bytecnt, struct mlx5_ib_qp *qp)
1954 {
1955 	while (bytecnt > 0) {
1956 		__iowrite64_copy(dst++, src++, 8);
1957 		__iowrite64_copy(dst++, src++, 8);
1958 		__iowrite64_copy(dst++, src++, 8);
1959 		__iowrite64_copy(dst++, src++, 8);
1960 		__iowrite64_copy(dst++, src++, 8);
1961 		__iowrite64_copy(dst++, src++, 8);
1962 		__iowrite64_copy(dst++, src++, 8);
1963 		__iowrite64_copy(dst++, src++, 8);
1964 		bytecnt -= 64;
1965 		if (unlikely(src == qp->sq.qend))
1966 			src = mlx5_get_send_wqe(qp, 0);
1967 	}
1968 }
1969 
1970 static u8 get_fence(u8 fence, struct ib_send_wr *wr)
1971 {
1972 	if (unlikely(wr->opcode == IB_WR_LOCAL_INV &&
1973 		     wr->send_flags & IB_SEND_FENCE))
1974 		return MLX5_FENCE_MODE_STRONG_ORDERING;
1975 
1976 	if (unlikely(fence)) {
1977 		if (wr->send_flags & IB_SEND_FENCE)
1978 			return MLX5_FENCE_MODE_SMALL_AND_FENCE;
1979 		else
1980 			return fence;
1981 
1982 	} else {
1983 		return 0;
1984 	}
1985 }
1986 
1987 int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1988 		      struct ib_send_wr **bad_wr)
1989 {
1990 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
1991 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1992 	struct mlx5_core_dev *mdev = &dev->mdev;
1993 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
1994 	struct mlx5_wqe_data_seg *dpseg;
1995 	struct mlx5_wqe_xrc_seg *xrc;
1996 	struct mlx5_bf *bf = qp->bf;
1997 	int uninitialized_var(size);
1998 	void *qend = qp->sq.qend;
1999 	unsigned long flags;
2000 	u32 mlx5_opcode;
2001 	unsigned idx;
2002 	int err = 0;
2003 	int inl = 0;
2004 	int num_sge;
2005 	void *seg;
2006 	int nreq;
2007 	int i;
2008 	u8 next_fence = 0;
2009 	u8 opmod = 0;
2010 	u8 fence;
2011 
2012 	spin_lock_irqsave(&qp->sq.lock, flags);
2013 
2014 	for (nreq = 0; wr; nreq++, wr = wr->next) {
2015 		if (unlikely(wr->opcode >= sizeof(mlx5_ib_opcode) / sizeof(mlx5_ib_opcode[0]))) {
2016 			mlx5_ib_warn(dev, "\n");
2017 			err = -EINVAL;
2018 			*bad_wr = wr;
2019 			goto out;
2020 		}
2021 
2022 		if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq))) {
2023 			mlx5_ib_warn(dev, "\n");
2024 			err = -ENOMEM;
2025 			*bad_wr = wr;
2026 			goto out;
2027 		}
2028 
2029 		fence = qp->fm_cache;
2030 		num_sge = wr->num_sge;
2031 		if (unlikely(num_sge > qp->sq.max_gs)) {
2032 			mlx5_ib_warn(dev, "\n");
2033 			err = -ENOMEM;
2034 			*bad_wr = wr;
2035 			goto out;
2036 		}
2037 
2038 		idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
2039 		seg = mlx5_get_send_wqe(qp, idx);
2040 		ctrl = seg;
2041 		*(uint32_t *)(seg + 8) = 0;
2042 		ctrl->imm = send_ieth(wr);
2043 		ctrl->fm_ce_se = qp->sq_signal_bits |
2044 			(wr->send_flags & IB_SEND_SIGNALED ?
2045 			 MLX5_WQE_CTRL_CQ_UPDATE : 0) |
2046 			(wr->send_flags & IB_SEND_SOLICITED ?
2047 			 MLX5_WQE_CTRL_SOLICITED : 0);
2048 
2049 		seg += sizeof(*ctrl);
2050 		size = sizeof(*ctrl) / 16;
2051 
2052 		switch (ibqp->qp_type) {
2053 		case IB_QPT_XRC_INI:
2054 			xrc = seg;
2055 			xrc->xrc_srqn = htonl(wr->xrc_remote_srq_num);
2056 			seg += sizeof(*xrc);
2057 			size += sizeof(*xrc) / 16;
2058 			/* fall through */
2059 		case IB_QPT_RC:
2060 			switch (wr->opcode) {
2061 			case IB_WR_RDMA_READ:
2062 			case IB_WR_RDMA_WRITE:
2063 			case IB_WR_RDMA_WRITE_WITH_IMM:
2064 				set_raddr_seg(seg, wr->wr.rdma.remote_addr,
2065 					      wr->wr.rdma.rkey);
2066 				seg  += sizeof(struct mlx5_wqe_raddr_seg);
2067 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
2068 				break;
2069 
2070 			case IB_WR_ATOMIC_CMP_AND_SWP:
2071 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2072 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2073 				mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
2074 				err = -ENOSYS;
2075 				*bad_wr = wr;
2076 				goto out;
2077 
2078 			case IB_WR_LOCAL_INV:
2079 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2080 				qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
2081 				ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
2082 				err = set_frwr_li_wr(&seg, wr, &size, mdev, to_mpd(ibqp->pd), qp);
2083 				if (err) {
2084 					mlx5_ib_warn(dev, "\n");
2085 					*bad_wr = wr;
2086 					goto out;
2087 				}
2088 				num_sge = 0;
2089 				break;
2090 
2091 			case IB_WR_FAST_REG_MR:
2092 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2093 				qp->sq.wr_data[idx] = IB_WR_FAST_REG_MR;
2094 				ctrl->imm = cpu_to_be32(wr->wr.fast_reg.rkey);
2095 				err = set_frwr_li_wr(&seg, wr, &size, mdev, to_mpd(ibqp->pd), qp);
2096 				if (err) {
2097 					mlx5_ib_warn(dev, "\n");
2098 					*bad_wr = wr;
2099 					goto out;
2100 				}
2101 				num_sge = 0;
2102 				break;
2103 
2104 			default:
2105 				break;
2106 			}
2107 			break;
2108 
2109 		case IB_QPT_UC:
2110 			switch (wr->opcode) {
2111 			case IB_WR_RDMA_WRITE:
2112 			case IB_WR_RDMA_WRITE_WITH_IMM:
2113 				set_raddr_seg(seg, wr->wr.rdma.remote_addr,
2114 					      wr->wr.rdma.rkey);
2115 				seg  += sizeof(struct mlx5_wqe_raddr_seg);
2116 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
2117 				break;
2118 
2119 			default:
2120 				break;
2121 			}
2122 			break;
2123 
2124 		case IB_QPT_UD:
2125 		case IB_QPT_SMI:
2126 		case IB_QPT_GSI:
2127 			set_datagram_seg(seg, wr);
2128 			seg  += sizeof(struct mlx5_wqe_datagram_seg);
2129 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
2130 			if (unlikely((seg == qend)))
2131 				seg = mlx5_get_send_wqe(qp, 0);
2132 			break;
2133 
2134 		case MLX5_IB_QPT_REG_UMR:
2135 			if (wr->opcode != MLX5_IB_WR_UMR) {
2136 				err = -EINVAL;
2137 				mlx5_ib_warn(dev, "bad opcode\n");
2138 				goto out;
2139 			}
2140 			qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
2141 			ctrl->imm = cpu_to_be32(wr->wr.fast_reg.rkey);
2142 			set_reg_umr_segment(seg, wr);
2143 			seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2144 			size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2145 			if (unlikely((seg == qend)))
2146 				seg = mlx5_get_send_wqe(qp, 0);
2147 			set_reg_mkey_segment(seg, wr);
2148 			seg += sizeof(struct mlx5_mkey_seg);
2149 			size += sizeof(struct mlx5_mkey_seg) / 16;
2150 			if (unlikely((seg == qend)))
2151 				seg = mlx5_get_send_wqe(qp, 0);
2152 			break;
2153 
2154 		default:
2155 			break;
2156 		}
2157 
2158 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
2159 			int uninitialized_var(sz);
2160 
2161 			err = set_data_inl_seg(qp, wr, seg, &sz);
2162 			if (unlikely(err)) {
2163 				mlx5_ib_warn(dev, "\n");
2164 				*bad_wr = wr;
2165 				goto out;
2166 			}
2167 			inl = 1;
2168 			size += sz;
2169 		} else {
2170 			dpseg = seg;
2171 			for (i = 0; i < num_sge; i++) {
2172 				if (unlikely(dpseg == qend)) {
2173 					seg = mlx5_get_send_wqe(qp, 0);
2174 					dpseg = seg;
2175 				}
2176 				if (likely(wr->sg_list[i].length)) {
2177 					set_data_ptr_seg(dpseg, wr->sg_list + i);
2178 					size += sizeof(struct mlx5_wqe_data_seg) / 16;
2179 					dpseg++;
2180 				}
2181 			}
2182 		}
2183 
2184 		mlx5_opcode = mlx5_ib_opcode[wr->opcode];
2185 		ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8)	|
2186 						     mlx5_opcode			|
2187 						     ((u32)opmod << 24));
2188 		ctrl->qpn_ds = cpu_to_be32(size | (qp->mqp.qpn << 8));
2189 		ctrl->fm_ce_se |= get_fence(fence, wr);
2190 		qp->fm_cache = next_fence;
2191 		if (unlikely(qp->wq_sig))
2192 			ctrl->signature = wq_sig(ctrl);
2193 
2194 		qp->sq.wrid[idx] = wr->wr_id;
2195 		qp->sq.w_list[idx].opcode = mlx5_opcode;
2196 		qp->sq.wqe_head[idx] = qp->sq.head + nreq;
2197 		qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
2198 		qp->sq.w_list[idx].next = qp->sq.cur_post;
2199 
2200 		if (0)
2201 			dump_wqe(qp, idx, size);
2202 	}
2203 
2204 out:
2205 	if (likely(nreq)) {
2206 		qp->sq.head += nreq;
2207 
2208 		/* Make sure that descriptors are written before
2209 		 * updating doorbell record and ringing the doorbell
2210 		 */
2211 		wmb();
2212 
2213 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
2214 
2215 		if (bf->need_lock)
2216 			spin_lock(&bf->lock);
2217 
2218 		/* TBD enable WC */
2219 		if (0 && nreq == 1 && bf->uuarn && inl && size > 1 && size <= bf->buf_size / 16) {
2220 			mlx5_bf_copy(bf->reg + bf->offset, (u64 *)ctrl, ALIGN(size * 16, 64), qp);
2221 			/* wc_wmb(); */
2222 		} else {
2223 			mlx5_write64((__be32 *)ctrl, bf->regreg + bf->offset,
2224 				     MLX5_GET_DOORBELL_LOCK(&bf->lock32));
2225 			/* Make sure doorbells don't leak out of SQ spinlock
2226 			 * and reach the HCA out of order.
2227 			 */
2228 			mmiowb();
2229 		}
2230 		bf->offset ^= bf->buf_size;
2231 		if (bf->need_lock)
2232 			spin_unlock(&bf->lock);
2233 	}
2234 
2235 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2236 
2237 	return err;
2238 }
2239 
2240 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
2241 {
2242 	sig->signature = calc_sig(sig, size);
2243 }
2244 
2245 int mlx5_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2246 		      struct ib_recv_wr **bad_wr)
2247 {
2248 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2249 	struct mlx5_wqe_data_seg *scat;
2250 	struct mlx5_rwqe_sig *sig;
2251 	unsigned long flags;
2252 	int err = 0;
2253 	int nreq;
2254 	int ind;
2255 	int i;
2256 
2257 	spin_lock_irqsave(&qp->rq.lock, flags);
2258 
2259 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
2260 
2261 	for (nreq = 0; wr; nreq++, wr = wr->next) {
2262 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2263 			err = -ENOMEM;
2264 			*bad_wr = wr;
2265 			goto out;
2266 		}
2267 
2268 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2269 			err = -EINVAL;
2270 			*bad_wr = wr;
2271 			goto out;
2272 		}
2273 
2274 		scat = get_recv_wqe(qp, ind);
2275 		if (qp->wq_sig)
2276 			scat++;
2277 
2278 		for (i = 0; i < wr->num_sge; i++)
2279 			set_data_ptr_seg(scat + i, wr->sg_list + i);
2280 
2281 		if (i < qp->rq.max_gs) {
2282 			scat[i].byte_count = 0;
2283 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
2284 			scat[i].addr       = 0;
2285 		}
2286 
2287 		if (qp->wq_sig) {
2288 			sig = (struct mlx5_rwqe_sig *)scat;
2289 			set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
2290 		}
2291 
2292 		qp->rq.wrid[ind] = wr->wr_id;
2293 
2294 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
2295 	}
2296 
2297 out:
2298 	if (likely(nreq)) {
2299 		qp->rq.head += nreq;
2300 
2301 		/* Make sure that descriptors are written before
2302 		 * doorbell record.
2303 		 */
2304 		wmb();
2305 
2306 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
2307 	}
2308 
2309 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2310 
2311 	return err;
2312 }
2313 
2314 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
2315 {
2316 	switch (mlx5_state) {
2317 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
2318 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
2319 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
2320 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
2321 	case MLX5_QP_STATE_SQ_DRAINING:
2322 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
2323 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
2324 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
2325 	default:		     return -1;
2326 	}
2327 }
2328 
2329 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
2330 {
2331 	switch (mlx5_mig_state) {
2332 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
2333 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
2334 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
2335 	default: return -1;
2336 	}
2337 }
2338 
2339 static int to_ib_qp_access_flags(int mlx5_flags)
2340 {
2341 	int ib_flags = 0;
2342 
2343 	if (mlx5_flags & MLX5_QP_BIT_RRE)
2344 		ib_flags |= IB_ACCESS_REMOTE_READ;
2345 	if (mlx5_flags & MLX5_QP_BIT_RWE)
2346 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
2347 	if (mlx5_flags & MLX5_QP_BIT_RAE)
2348 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
2349 
2350 	return ib_flags;
2351 }
2352 
2353 static void to_ib_ah_attr(struct mlx5_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
2354 				struct mlx5_qp_path *path)
2355 {
2356 	struct mlx5_core_dev *dev = &ibdev->mdev;
2357 
2358 	memset(ib_ah_attr, 0, sizeof(*ib_ah_attr));
2359 	ib_ah_attr->port_num	  = path->port;
2360 
2361 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
2362 		return;
2363 
2364 	ib_ah_attr->sl = path->sl & 0xf;
2365 
2366 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
2367 	ib_ah_attr->src_path_bits = path->grh_mlid & 0x7f;
2368 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
2369 	ib_ah_attr->ah_flags      = (path->grh_mlid & (1 << 7)) ? IB_AH_GRH : 0;
2370 	if (ib_ah_attr->ah_flags) {
2371 		ib_ah_attr->grh.sgid_index = path->mgid_index;
2372 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
2373 		ib_ah_attr->grh.traffic_class =
2374 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
2375 		ib_ah_attr->grh.flow_label =
2376 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
2377 		memcpy(ib_ah_attr->grh.dgid.raw,
2378 		       path->rgid, sizeof(ib_ah_attr->grh.dgid.raw));
2379 	}
2380 }
2381 
2382 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
2383 		     struct ib_qp_init_attr *qp_init_attr)
2384 {
2385 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2386 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2387 	struct mlx5_query_qp_mbox_out *outb;
2388 	struct mlx5_qp_context *context;
2389 	int mlx5_state;
2390 	int err = 0;
2391 
2392 	mutex_lock(&qp->mutex);
2393 	outb = kzalloc(sizeof(*outb), GFP_KERNEL);
2394 	if (!outb) {
2395 		err = -ENOMEM;
2396 		goto out;
2397 	}
2398 	context = &outb->ctx;
2399 	err = mlx5_core_qp_query(&dev->mdev, &qp->mqp, outb, sizeof(*outb));
2400 	if (err)
2401 		goto out_free;
2402 
2403 	mlx5_state = be32_to_cpu(context->flags) >> 28;
2404 
2405 	qp->state		     = to_ib_qp_state(mlx5_state);
2406 	qp_attr->qp_state	     = qp->state;
2407 	qp_attr->path_mtu	     = context->mtu_msgmax >> 5;
2408 	qp_attr->path_mig_state	     =
2409 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
2410 	qp_attr->qkey		     = be32_to_cpu(context->qkey);
2411 	qp_attr->rq_psn		     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
2412 	qp_attr->sq_psn		     = be32_to_cpu(context->next_send_psn) & 0xffffff;
2413 	qp_attr->dest_qp_num	     = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
2414 	qp_attr->qp_access_flags     =
2415 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
2416 
2417 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
2418 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
2419 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
2420 		qp_attr->alt_pkey_index = context->alt_path.pkey_index & 0x7f;
2421 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
2422 	}
2423 
2424 	qp_attr->pkey_index = context->pri_path.pkey_index & 0x7f;
2425 	qp_attr->port_num = context->pri_path.port;
2426 
2427 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
2428 	qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
2429 
2430 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
2431 
2432 	qp_attr->max_dest_rd_atomic =
2433 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
2434 	qp_attr->min_rnr_timer	    =
2435 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
2436 	qp_attr->timeout	    = context->pri_path.ackto_lt >> 3;
2437 	qp_attr->retry_cnt	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
2438 	qp_attr->rnr_retry	    = (be32_to_cpu(context->params1) >> 13) & 0x7;
2439 	qp_attr->alt_timeout	    = context->alt_path.ackto_lt >> 3;
2440 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
2441 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
2442 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
2443 
2444 	if (!ibqp->uobject) {
2445 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
2446 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
2447 	} else {
2448 		qp_attr->cap.max_send_wr  = 0;
2449 		qp_attr->cap.max_send_sge = 0;
2450 	}
2451 
2452 	/* We don't support inline sends for kernel QPs (yet), and we
2453 	 * don't know what userspace's value should be.
2454 	 */
2455 	qp_attr->cap.max_inline_data = 0;
2456 
2457 	qp_init_attr->cap	     = qp_attr->cap;
2458 
2459 	qp_init_attr->create_flags = 0;
2460 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
2461 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
2462 
2463 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
2464 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
2465 
2466 out_free:
2467 	kfree(outb);
2468 
2469 out:
2470 	mutex_unlock(&qp->mutex);
2471 	return err;
2472 }
2473 
2474 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
2475 					  struct ib_ucontext *context,
2476 					  struct ib_udata *udata)
2477 {
2478 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
2479 	struct mlx5_ib_xrcd *xrcd;
2480 	int err;
2481 
2482 	if (!(dev->mdev.caps.flags & MLX5_DEV_CAP_FLAG_XRC))
2483 		return ERR_PTR(-ENOSYS);
2484 
2485 	xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
2486 	if (!xrcd)
2487 		return ERR_PTR(-ENOMEM);
2488 
2489 	err = mlx5_core_xrcd_alloc(&dev->mdev, &xrcd->xrcdn);
2490 	if (err) {
2491 		kfree(xrcd);
2492 		return ERR_PTR(-ENOMEM);
2493 	}
2494 
2495 	return &xrcd->ibxrcd;
2496 }
2497 
2498 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
2499 {
2500 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
2501 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
2502 	int err;
2503 
2504 	err = mlx5_core_xrcd_dealloc(&dev->mdev, xrcdn);
2505 	if (err) {
2506 		mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
2507 		return err;
2508 	}
2509 
2510 	kfree(xrcd);
2511 
2512 	return 0;
2513 }
2514