xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_qp.c (revision 77013d11e6483b970af25e13c9b892075742f7e5)
1 /*-
2  * Copyright (c) 2013-2021, Mellanox Technologies. All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  * $FreeBSD$
26  */
27 
28 #include <linux/module.h>
29 #include <rdma/ib_umem.h>
30 #include <rdma/ib_cache.h>
31 #include <rdma/ib_user_verbs.h>
32 #include <rdma/uverbs_ioctl.h>
33 #include "mlx5_ib.h"
34 
35 /* not supported currently */
36 static int wq_signature;
37 
38 enum {
39 	MLX5_IB_ACK_REQ_FREQ	= 8,
40 };
41 
42 enum {
43 	MLX5_IB_DEFAULT_SCHED_QUEUE	= 0x83,
44 	MLX5_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
45 	MLX5_IB_LINK_TYPE_IB		= 0,
46 	MLX5_IB_LINK_TYPE_ETH		= 1
47 };
48 
49 enum {
50 	MLX5_IB_SQ_STRIDE	= 6,
51 };
52 
53 static const u32 mlx5_ib_opcode[] = {
54 	[IB_WR_SEND]				= MLX5_OPCODE_SEND,
55 	[IB_WR_LSO]				= MLX5_OPCODE_LSO,
56 	[IB_WR_SEND_WITH_IMM]			= MLX5_OPCODE_SEND_IMM,
57 	[IB_WR_RDMA_WRITE]			= MLX5_OPCODE_RDMA_WRITE,
58 	[IB_WR_RDMA_WRITE_WITH_IMM]		= MLX5_OPCODE_RDMA_WRITE_IMM,
59 	[IB_WR_RDMA_READ]			= MLX5_OPCODE_RDMA_READ,
60 	[IB_WR_ATOMIC_CMP_AND_SWP]		= MLX5_OPCODE_ATOMIC_CS,
61 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= MLX5_OPCODE_ATOMIC_FA,
62 	[IB_WR_SEND_WITH_INV]			= MLX5_OPCODE_SEND_INVAL,
63 	[IB_WR_LOCAL_INV]			= MLX5_OPCODE_UMR,
64 	[IB_WR_REG_MR]				= MLX5_OPCODE_UMR,
65 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= MLX5_OPCODE_ATOMIC_MASKED_CS,
66 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= MLX5_OPCODE_ATOMIC_MASKED_FA,
67 	[MLX5_IB_WR_UMR]			= MLX5_OPCODE_UMR,
68 };
69 
70 struct mlx5_wqe_eth_pad {
71 	u8 rsvd0[16];
72 };
73 
74 enum raw_qp_set_mask_map {
75 	MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID		= 1UL << 0,
76 };
77 
78 struct mlx5_modify_raw_qp_param {
79 	u16 operation;
80 
81 	u32 set_mask; /* raw_qp_set_mask_map */
82 	u8 rq_q_ctr_id;
83 };
84 
85 static void get_cqs(enum ib_qp_type qp_type,
86 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
87 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq);
88 
89 static int is_qp0(enum ib_qp_type qp_type)
90 {
91 	return qp_type == IB_QPT_SMI;
92 }
93 
94 static int is_sqp(enum ib_qp_type qp_type)
95 {
96 	return is_qp0(qp_type) || is_qp1(qp_type);
97 }
98 
99 static void *get_wqe(struct mlx5_ib_qp *qp, int offset)
100 {
101 	return mlx5_buf_offset(&qp->buf, offset);
102 }
103 
104 static void *get_recv_wqe(struct mlx5_ib_qp *qp, int n)
105 {
106 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
107 }
108 
109 void *mlx5_get_send_wqe(struct mlx5_ib_qp *qp, int n)
110 {
111 	return get_wqe(qp, qp->sq.offset + (n << MLX5_IB_SQ_STRIDE));
112 }
113 
114 /**
115  * mlx5_ib_read_user_wqe() - Copy a user-space WQE to kernel space.
116  *
117  * @qp: QP to copy from.
118  * @send: copy from the send queue when non-zero, use the receive queue
119  *	  otherwise.
120  * @wqe_index:  index to start copying from. For send work queues, the
121  *		wqe_index is in units of MLX5_SEND_WQE_BB.
122  *		For receive work queue, it is the number of work queue
123  *		element in the queue.
124  * @buffer: destination buffer.
125  * @length: maximum number of bytes to copy.
126  *
127  * Copies at least a single WQE, but may copy more data.
128  *
129  * Return: the number of bytes copied, or an error code.
130  */
131 int mlx5_ib_read_user_wqe(struct mlx5_ib_qp *qp, int send, int wqe_index,
132 			  void *buffer, u32 length,
133 			  struct mlx5_ib_qp_base *base)
134 {
135 	struct ib_device *ibdev = qp->ibqp.device;
136 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
137 	struct mlx5_ib_wq *wq = send ? &qp->sq : &qp->rq;
138 	size_t offset;
139 	size_t wq_end;
140 	struct ib_umem *umem = base->ubuffer.umem;
141 	u32 first_copy_length;
142 	int wqe_length;
143 	int ret;
144 
145 	if (wq->wqe_cnt == 0) {
146 		mlx5_ib_dbg(dev, "mlx5_ib_read_user_wqe for a QP with wqe_cnt == 0. qp_type: 0x%x\n",
147 			    qp->ibqp.qp_type);
148 		return -EINVAL;
149 	}
150 
151 	offset = wq->offset + ((wqe_index % wq->wqe_cnt) << wq->wqe_shift);
152 	wq_end = wq->offset + (wq->wqe_cnt << wq->wqe_shift);
153 
154 	if (send && length < sizeof(struct mlx5_wqe_ctrl_seg))
155 		return -EINVAL;
156 
157 	if (offset > umem->length ||
158 	    (send && offset + sizeof(struct mlx5_wqe_ctrl_seg) > umem->length))
159 		return -EINVAL;
160 
161 	first_copy_length = min_t(u32, offset + length, wq_end) - offset;
162 	ret = ib_umem_copy_from(buffer, umem, offset, first_copy_length);
163 	if (ret)
164 		return ret;
165 
166 	if (send) {
167 		struct mlx5_wqe_ctrl_seg *ctrl = buffer;
168 		int ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
169 
170 		wqe_length = ds * MLX5_WQE_DS_UNITS;
171 	} else {
172 		wqe_length = 1 << wq->wqe_shift;
173 	}
174 
175 	if (wqe_length <= first_copy_length)
176 		return first_copy_length;
177 
178 	ret = ib_umem_copy_from(buffer + first_copy_length, umem, wq->offset,
179 				wqe_length - first_copy_length);
180 	if (ret)
181 		return ret;
182 
183 	return wqe_length;
184 }
185 
186 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
187 {
188 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
189 	struct ib_event event;
190 
191 	if (type == MLX5_EVENT_TYPE_PATH_MIG) {
192 		/* This event is only valid for trans_qps */
193 		to_mibqp(qp)->port = to_mibqp(qp)->trans_qp.alt_port;
194 	}
195 
196 	if (ibqp->event_handler) {
197 		event.device     = ibqp->device;
198 		event.element.qp = ibqp;
199 		switch (type) {
200 		case MLX5_EVENT_TYPE_PATH_MIG:
201 			event.event = IB_EVENT_PATH_MIG;
202 			break;
203 		case MLX5_EVENT_TYPE_COMM_EST:
204 			event.event = IB_EVENT_COMM_EST;
205 			break;
206 		case MLX5_EVENT_TYPE_SQ_DRAINED:
207 			event.event = IB_EVENT_SQ_DRAINED;
208 			break;
209 		case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
210 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
211 			break;
212 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
213 			event.event = IB_EVENT_QP_FATAL;
214 			break;
215 		case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
216 			event.event = IB_EVENT_PATH_MIG_ERR;
217 			break;
218 		case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
219 			event.event = IB_EVENT_QP_REQ_ERR;
220 			break;
221 		case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
222 			event.event = IB_EVENT_QP_ACCESS_ERR;
223 			break;
224 		default:
225 			pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
226 			return;
227 		}
228 
229 		ibqp->event_handler(&event, ibqp->qp_context);
230 	}
231 }
232 
233 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
234 		       int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
235 {
236 	int wqe_size;
237 	int wq_size;
238 
239 	/* Sanity check RQ size before proceeding */
240 	if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
241 		return -EINVAL;
242 
243 	if (!has_rq) {
244 		qp->rq.max_gs = 0;
245 		qp->rq.wqe_cnt = 0;
246 		qp->rq.wqe_shift = 0;
247 		cap->max_recv_wr = 0;
248 		cap->max_recv_sge = 0;
249 	} else {
250 		if (ucmd) {
251 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
252 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
253 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
254 			qp->rq.max_post = qp->rq.wqe_cnt;
255 		} else {
256 			wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
257 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
258 			wqe_size = roundup_pow_of_two(wqe_size);
259 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
260 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
261 			qp->rq.wqe_cnt = wq_size / wqe_size;
262 			if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
263 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
264 					    wqe_size,
265 					    MLX5_CAP_GEN(dev->mdev,
266 							 max_wqe_sz_rq));
267 				return -EINVAL;
268 			}
269 			qp->rq.wqe_shift = ilog2(wqe_size);
270 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
271 			qp->rq.max_post = qp->rq.wqe_cnt;
272 		}
273 	}
274 
275 	return 0;
276 }
277 
278 static int sq_overhead(struct ib_qp_init_attr *attr)
279 {
280 	int size = 0;
281 
282 	switch (attr->qp_type) {
283 	case IB_QPT_XRC_INI:
284 		size += sizeof(struct mlx5_wqe_xrc_seg);
285 		/* fall through */
286 	case IB_QPT_RC:
287 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
288 			max(sizeof(struct mlx5_wqe_atomic_seg) +
289 			    sizeof(struct mlx5_wqe_raddr_seg),
290 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
291 			    sizeof(struct mlx5_mkey_seg));
292 		break;
293 
294 	case IB_QPT_XRC_TGT:
295 		return 0;
296 
297 	case IB_QPT_UC:
298 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
299 			max(sizeof(struct mlx5_wqe_raddr_seg),
300 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
301 			    sizeof(struct mlx5_mkey_seg));
302 		break;
303 
304 	case IB_QPT_UD:
305 		if (attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
306 			size += sizeof(struct mlx5_wqe_eth_pad) +
307 				sizeof(struct mlx5_wqe_eth_seg);
308 		/* fall through */
309 	case IB_QPT_SMI:
310 	case MLX5_IB_QPT_HW_GSI:
311 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
312 			sizeof(struct mlx5_wqe_datagram_seg);
313 		break;
314 
315 	case MLX5_IB_QPT_REG_UMR:
316 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
317 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
318 			sizeof(struct mlx5_mkey_seg);
319 		break;
320 
321 	default:
322 		return -EINVAL;
323 	}
324 
325 	return size;
326 }
327 
328 static int calc_send_wqe(struct ib_qp_init_attr *attr)
329 {
330 	int inl_size = 0;
331 	int size;
332 
333 	size = sq_overhead(attr);
334 	if (size < 0)
335 		return size;
336 
337 	if (attr->cap.max_inline_data) {
338 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
339 			attr->cap.max_inline_data;
340 	}
341 
342 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
343 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN &&
344 	    ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
345 			return MLX5_SIG_WQE_SIZE;
346 	else
347 		return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
348 }
349 
350 static int get_send_sge(struct ib_qp_init_attr *attr, int wqe_size)
351 {
352 	int max_sge;
353 
354 	if (attr->qp_type == IB_QPT_RC)
355 		max_sge = (min_t(int, wqe_size, 512) -
356 			   sizeof(struct mlx5_wqe_ctrl_seg) -
357 			   sizeof(struct mlx5_wqe_raddr_seg)) /
358 			sizeof(struct mlx5_wqe_data_seg);
359 	else if (attr->qp_type == IB_QPT_XRC_INI)
360 		max_sge = (min_t(int, wqe_size, 512) -
361 			   sizeof(struct mlx5_wqe_ctrl_seg) -
362 			   sizeof(struct mlx5_wqe_xrc_seg) -
363 			   sizeof(struct mlx5_wqe_raddr_seg)) /
364 			sizeof(struct mlx5_wqe_data_seg);
365 	else
366 		max_sge = (wqe_size - sq_overhead(attr)) /
367 			sizeof(struct mlx5_wqe_data_seg);
368 
369 	return min_t(int, max_sge, wqe_size - sq_overhead(attr) /
370 		     sizeof(struct mlx5_wqe_data_seg));
371 }
372 
373 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
374 			struct mlx5_ib_qp *qp)
375 {
376 	int wqe_size;
377 	int wq_size;
378 
379 	if (!attr->cap.max_send_wr)
380 		return 0;
381 
382 	wqe_size = calc_send_wqe(attr);
383 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
384 	if (wqe_size < 0)
385 		return wqe_size;
386 
387 	if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
388 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
389 			    wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
390 		return -EINVAL;
391 	}
392 
393 	qp->max_inline_data = wqe_size - sq_overhead(attr) -
394 			      sizeof(struct mlx5_wqe_inline_seg);
395 	attr->cap.max_inline_data = qp->max_inline_data;
396 
397 	if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN)
398 		qp->signature_en = true;
399 
400 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
401 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
402 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
403 		mlx5_ib_dbg(dev, "wqe count(%d) exceeds limits(%d)\n",
404 			    qp->sq.wqe_cnt,
405 			    1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
406 		return -ENOMEM;
407 	}
408 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
409 	qp->sq.max_gs = get_send_sge(attr, wqe_size);
410 	if (qp->sq.max_gs < attr->cap.max_send_sge)
411 		return -ENOMEM;
412 
413 	attr->cap.max_send_sge = qp->sq.max_gs;
414 	qp->sq.max_post = wq_size / wqe_size;
415 	attr->cap.max_send_wr = qp->sq.max_post;
416 
417 	return wq_size;
418 }
419 
420 static int set_user_buf_size(struct mlx5_ib_dev *dev,
421 			    struct mlx5_ib_qp *qp,
422 			    struct mlx5_ib_create_qp *ucmd,
423 			    struct mlx5_ib_qp_base *base,
424 			    struct ib_qp_init_attr *attr)
425 {
426 	int desc_sz = 1 << qp->sq.wqe_shift;
427 
428 	if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
429 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
430 			     desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
431 		return -EINVAL;
432 	}
433 
434 	if (ucmd->sq_wqe_count && ((1 << ilog2(ucmd->sq_wqe_count)) != ucmd->sq_wqe_count)) {
435 		mlx5_ib_warn(dev, "sq_wqe_count %d, sq_wqe_count %d\n",
436 			     ucmd->sq_wqe_count, ucmd->sq_wqe_count);
437 		return -EINVAL;
438 	}
439 
440 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
441 
442 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
443 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
444 			     qp->sq.wqe_cnt,
445 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
446 		return -EINVAL;
447 	}
448 
449 	if (attr->qp_type == IB_QPT_RAW_PACKET) {
450 		base->ubuffer.buf_size = qp->rq.wqe_cnt << qp->rq.wqe_shift;
451 		qp->raw_packet_qp.sq.ubuffer.buf_size = qp->sq.wqe_cnt << 6;
452 	} else {
453 		base->ubuffer.buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
454 					 (qp->sq.wqe_cnt << 6);
455 	}
456 
457 	return 0;
458 }
459 
460 static int qp_has_rq(struct ib_qp_init_attr *attr)
461 {
462 	if (attr->qp_type == IB_QPT_XRC_INI ||
463 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
464 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
465 	    !attr->cap.max_recv_wr)
466 		return 0;
467 
468 	return 1;
469 }
470 
471 enum {
472 	/* this is the first blue flame register in the array of bfregs assigned
473 	 * to a processes. Since we do not use it for blue flame but rather
474 	 * regular 64 bit doorbells, we do not need a lock for maintaiing
475 	 * "odd/even" order
476 	 */
477 	NUM_NON_BLUE_FLAME_BFREGS = 1,
478 };
479 
480 static int max_bfregs(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi)
481 {
482 	return get_num_static_uars(dev, bfregi) * MLX5_NON_FP_BFREGS_PER_UAR;
483 }
484 
485 static int num_med_bfreg(struct mlx5_ib_dev *dev,
486 			 struct mlx5_bfreg_info *bfregi)
487 {
488 	int n;
489 
490 	n = max_bfregs(dev, bfregi) - bfregi->num_low_latency_bfregs -
491 	    NUM_NON_BLUE_FLAME_BFREGS;
492 
493 	return n >= 0 ? n : 0;
494 }
495 
496 static int first_med_bfreg(struct mlx5_ib_dev *dev,
497 			   struct mlx5_bfreg_info *bfregi)
498 {
499 	return num_med_bfreg(dev, bfregi) ? 1 : -ENOMEM;
500 }
501 
502 static int first_hi_bfreg(struct mlx5_ib_dev *dev,
503 			  struct mlx5_bfreg_info *bfregi)
504 {
505 	int med;
506 
507 	med = num_med_bfreg(dev, bfregi);
508 	return ++med;
509 }
510 
511 static int alloc_high_class_bfreg(struct mlx5_ib_dev *dev,
512 				  struct mlx5_bfreg_info *bfregi)
513 {
514 	int i;
515 
516 	for (i = first_hi_bfreg(dev, bfregi); i < max_bfregs(dev, bfregi); i++) {
517 		if (!bfregi->count[i]) {
518 			bfregi->count[i]++;
519 			return i;
520 		}
521 	}
522 
523 	return -ENOMEM;
524 }
525 
526 static int alloc_med_class_bfreg(struct mlx5_ib_dev *dev,
527 				 struct mlx5_bfreg_info *bfregi)
528 {
529 	int minidx = first_med_bfreg(dev, bfregi);
530 	int i;
531 
532 	if (minidx < 0)
533 		return minidx;
534 
535 	for (i = minidx; i < first_hi_bfreg(dev, bfregi); i++) {
536 		if (bfregi->count[i] < bfregi->count[minidx])
537 			minidx = i;
538 		if (!bfregi->count[minidx])
539 			break;
540 	}
541 
542 	bfregi->count[minidx]++;
543 	return minidx;
544 }
545 
546 static int alloc_bfreg(struct mlx5_ib_dev *dev,
547 		       struct mlx5_bfreg_info *bfregi)
548 {
549 	int bfregn = -ENOMEM;
550 
551 	if (bfregi->lib_uar_dyn)
552 		return -EINVAL;
553 
554 	mutex_lock(&bfregi->lock);
555 	if (bfregi->ver >= 2) {
556 		bfregn = alloc_high_class_bfreg(dev, bfregi);
557 		if (bfregn < 0)
558 			bfregn = alloc_med_class_bfreg(dev, bfregi);
559 	}
560 
561 	if (bfregn < 0) {
562 		BUILD_BUG_ON(NUM_NON_BLUE_FLAME_BFREGS != 1);
563 		bfregn = 0;
564 		bfregi->count[bfregn]++;
565 	}
566 	mutex_unlock(&bfregi->lock);
567 
568 	return bfregn;
569 }
570 
571 void mlx5_ib_free_bfreg(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi, int bfregn)
572 {
573 	mutex_lock(&bfregi->lock);
574 	bfregi->count[bfregn]--;
575 	mutex_unlock(&bfregi->lock);
576 }
577 
578 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
579 {
580 	switch (state) {
581 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
582 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
583 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
584 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
585 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
586 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
587 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
588 	default:		return -1;
589 	}
590 }
591 
592 static int to_mlx5_st(enum ib_qp_type type)
593 {
594 	switch (type) {
595 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
596 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
597 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
598 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
599 	case IB_QPT_XRC_INI:
600 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
601 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
602 	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
603 	case IB_QPT_RAW_IPV6:		return MLX5_QP_ST_RAW_IPV6;
604 	case IB_QPT_RAW_PACKET:
605 	case IB_QPT_RAW_ETHERTYPE:	return MLX5_QP_ST_RAW_ETHERTYPE;
606 	case IB_QPT_MAX:
607 	default:		return -EINVAL;
608 	}
609 }
610 
611 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
612 			     struct mlx5_ib_cq *recv_cq);
613 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
614 			       struct mlx5_ib_cq *recv_cq);
615 
616 int bfregn_to_uar_index(struct mlx5_ib_dev *dev,
617 			struct mlx5_bfreg_info *bfregi, u32 bfregn,
618 			bool dyn_bfreg)
619 {
620 	unsigned int bfregs_per_sys_page;
621 	u32 index_of_sys_page;
622 	u32 offset;
623 
624 	if (bfregi->lib_uar_dyn)
625 		return -EINVAL;
626 
627 	bfregs_per_sys_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k) *
628 				MLX5_NON_FP_BFREGS_PER_UAR;
629 	index_of_sys_page = bfregn / bfregs_per_sys_page;
630 
631 	if (dyn_bfreg) {
632 		index_of_sys_page += bfregi->num_static_sys_pages;
633 
634 		if (index_of_sys_page >= bfregi->num_sys_pages)
635 			return -EINVAL;
636 
637 		if (bfregn > bfregi->num_dyn_bfregs ||
638 		    bfregi->sys_pages[index_of_sys_page] == MLX5_IB_INVALID_UAR_INDEX) {
639 			mlx5_ib_dbg(dev, "Invalid dynamic uar index\n");
640 			return -EINVAL;
641 		}
642 	}
643 
644 	offset = bfregn % bfregs_per_sys_page / MLX5_NON_FP_BFREGS_PER_UAR;
645 	return bfregi->sys_pages[index_of_sys_page] + offset;
646 }
647 
648 static int mlx5_ib_umem_get(struct mlx5_ib_dev *dev,
649 			    struct ib_pd *pd,
650 			    unsigned long addr, size_t size,
651 			    struct ib_umem **umem,
652 			    int *npages, int *page_shift, int *ncont,
653 			    u32 *offset)
654 {
655 	int err;
656 
657 	*umem = ib_umem_get(pd->uobject->context, addr, size, 0, 0);
658 	if (IS_ERR(*umem)) {
659 		mlx5_ib_dbg(dev, "umem_get failed\n");
660 		return PTR_ERR(*umem);
661 	}
662 
663 	mlx5_ib_cont_pages(*umem, addr, 0, npages, page_shift, ncont, NULL);
664 
665 	err = mlx5_ib_get_buf_offset(addr, *page_shift, offset);
666 	if (err) {
667 		mlx5_ib_warn(dev, "bad offset\n");
668 		goto err_umem;
669 	}
670 
671 	mlx5_ib_dbg(dev, "addr 0x%lx, size %zu, npages %d, page_shift %d, ncont %d, offset %d\n",
672 		    addr, size, *npages, *page_shift, *ncont, *offset);
673 
674 	return 0;
675 
676 err_umem:
677 	ib_umem_release(*umem);
678 	*umem = NULL;
679 
680 	return err;
681 }
682 
683 static void destroy_user_rq(struct ib_pd *pd, struct mlx5_ib_rwq *rwq,
684 			    struct ib_udata *udata)
685 {
686 	struct mlx5_ib_ucontext *context =
687 		rdma_udata_to_drv_context(
688 			udata,
689 			struct mlx5_ib_ucontext,
690 			ibucontext);
691 
692 	mlx5_ib_db_unmap_user(context, &rwq->db);
693 	if (rwq->umem)
694 		ib_umem_release(rwq->umem);
695 }
696 
697 static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
698 			  struct mlx5_ib_rwq *rwq,
699 			  struct mlx5_ib_create_wq *ucmd)
700 {
701 	struct mlx5_ib_ucontext *context;
702 	int page_shift = 0;
703 	int npages;
704 	u32 offset = 0;
705 	int ncont = 0;
706 	int err;
707 
708 	if (!ucmd->buf_addr)
709 		return -EINVAL;
710 
711 	context = to_mucontext(pd->uobject->context);
712 	rwq->umem = ib_umem_get(pd->uobject->context, ucmd->buf_addr,
713 			       rwq->buf_size, 0, 0);
714 	if (IS_ERR(rwq->umem)) {
715 		mlx5_ib_dbg(dev, "umem_get failed\n");
716 		err = PTR_ERR(rwq->umem);
717 		return err;
718 	}
719 
720 	mlx5_ib_cont_pages(rwq->umem, ucmd->buf_addr, 0, &npages, &page_shift,
721 			   &ncont, NULL);
722 	err = mlx5_ib_get_buf_offset(ucmd->buf_addr, page_shift,
723 				     &rwq->rq_page_offset);
724 	if (err) {
725 		mlx5_ib_warn(dev, "bad offset\n");
726 		goto err_umem;
727 	}
728 
729 	rwq->rq_num_pas = ncont;
730 	rwq->page_shift = page_shift;
731 	rwq->log_page_size =  page_shift - MLX5_ADAPTER_PAGE_SHIFT;
732 	rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
733 
734 	mlx5_ib_dbg(dev, "addr 0x%llx, size %zd, npages %d, page_shift %d, ncont %d, offset %d\n",
735 		    (unsigned long long)ucmd->buf_addr, rwq->buf_size,
736 		    npages, page_shift, ncont, offset);
737 
738 	err = mlx5_ib_db_map_user(context, ucmd->db_addr, &rwq->db);
739 	if (err) {
740 		mlx5_ib_dbg(dev, "map failed\n");
741 		goto err_umem;
742 	}
743 
744 	rwq->create_type = MLX5_WQ_USER;
745 	return 0;
746 
747 err_umem:
748 	ib_umem_release(rwq->umem);
749 	return err;
750 }
751 
752 static int adjust_bfregn(struct mlx5_ib_dev *dev,
753 			 struct mlx5_bfreg_info *bfregi, int bfregn)
754 {
755 	return bfregn / MLX5_NON_FP_BFREGS_PER_UAR * MLX5_BFREGS_PER_UAR +
756 				bfregn % MLX5_NON_FP_BFREGS_PER_UAR;
757 }
758 
759 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
760 			  struct mlx5_ib_qp *qp, struct ib_udata *udata,
761 			  struct ib_qp_init_attr *attr,
762 			  u32 **in,
763 			  struct mlx5_ib_create_qp_resp *resp, int *inlen,
764 			  struct mlx5_ib_qp_base *base)
765 {
766 	struct mlx5_ib_ucontext *context;
767 	struct mlx5_ib_create_qp ucmd;
768 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
769 	int page_shift = 0;
770 	int uar_index = 0;
771 	int npages;
772 	u32 offset = 0;
773 	int bfregn;
774 	int ncont = 0;
775 	__be64 *pas;
776 	void *qpc;
777 	int err;
778 	u16 uid;
779 	u32 uar_flags;
780 
781 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
782 	if (err) {
783 		mlx5_ib_dbg(dev, "copy failed\n");
784 		return err;
785 	}
786 
787 	context = to_mucontext(pd->uobject->context);
788 	uar_flags = ucmd.flags & (MLX5_QP_FLAG_UAR_PAGE_INDEX |
789 				  MLX5_QP_FLAG_BFREG_INDEX);
790 	switch (uar_flags) {
791 	case MLX5_QP_FLAG_UAR_PAGE_INDEX:
792 		uar_index = ucmd.bfreg_index;
793 		bfregn = MLX5_IB_INVALID_BFREG;
794 		break;
795 	case MLX5_QP_FLAG_BFREG_INDEX:
796 		uar_index = bfregn_to_uar_index(dev, &context->bfregi,
797 						ucmd.bfreg_index, true);
798 		if (uar_index < 0)
799 			return uar_index;
800 		bfregn = MLX5_IB_INVALID_BFREG;
801 		break;
802 	case 0:
803 		if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
804 			return -EINVAL;
805 		bfregn = alloc_bfreg(dev, &context->bfregi);
806 		if (bfregn < 0)
807 			return bfregn;
808 		break;
809 	default:
810 		return -EINVAL;
811 	}
812 
813 	mlx5_ib_dbg(dev, "bfregn 0x%x, uar_index 0x%x\n", bfregn, uar_index);
814 	if (bfregn != MLX5_IB_INVALID_BFREG)
815 		uar_index = bfregn_to_uar_index(dev, &context->bfregi, bfregn,
816 						false);
817 
818 	qp->rq.offset = 0;
819 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
820 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
821 
822 	err = set_user_buf_size(dev, qp, &ucmd, base, attr);
823 	if (err)
824 		goto err_bfreg;
825 
826 	if (ucmd.buf_addr && ubuffer->buf_size) {
827 		ubuffer->buf_addr = ucmd.buf_addr;
828 		err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr,
829 				       ubuffer->buf_size,
830 				       &ubuffer->umem, &npages, &page_shift,
831 				       &ncont, &offset);
832 		if (err)
833 			goto err_bfreg;
834 	} else {
835 		ubuffer->umem = NULL;
836 	}
837 
838 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
839 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * ncont;
840 	*in = mlx5_vzalloc(*inlen);
841 	if (!*in) {
842 		err = -ENOMEM;
843 		goto err_umem;
844 	}
845 
846 	uid = (attr->qp_type != IB_QPT_XRC_TGT &&
847 	       attr->qp_type != IB_QPT_XRC_INI) ? to_mpd(pd)->uid : 0;
848 	MLX5_SET(create_qp_in, *in, uid, uid);
849 	pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas);
850 	if (ubuffer->umem)
851 		mlx5_ib_populate_pas(dev, ubuffer->umem, page_shift, pas, 0);
852 
853 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
854 
855 	MLX5_SET(qpc, qpc, log_page_size, page_shift - MLX5_ADAPTER_PAGE_SHIFT);
856 	MLX5_SET(qpc, qpc, page_offset, offset);
857 
858 	MLX5_SET(qpc, qpc, uar_page, uar_index);
859 	if (bfregn != MLX5_IB_INVALID_BFREG)
860 		resp->bfreg_index = adjust_bfregn(dev, &context->bfregi, bfregn);
861 	else
862 		resp->bfreg_index = MLX5_IB_INVALID_BFREG;
863 	qp->bfregn = bfregn;
864 
865 	err = mlx5_ib_db_map_user(context, ucmd.db_addr, &qp->db);
866 	if (err) {
867 		mlx5_ib_dbg(dev, "map failed\n");
868 		goto err_free;
869 	}
870 
871 	err = ib_copy_to_udata(udata, resp, sizeof(*resp));
872 	if (err) {
873 		mlx5_ib_dbg(dev, "copy failed\n");
874 		goto err_unmap;
875 	}
876 	qp->create_type = MLX5_QP_USER;
877 
878 	return 0;
879 
880 err_unmap:
881 	mlx5_ib_db_unmap_user(context, &qp->db);
882 
883 err_free:
884 	kvfree(*in);
885 
886 err_umem:
887 	if (ubuffer->umem)
888 		ib_umem_release(ubuffer->umem);
889 
890 err_bfreg:
891 	if (bfregn != MLX5_IB_INVALID_BFREG)
892 		mlx5_ib_free_bfreg(dev, &context->bfregi, bfregn);
893 	return err;
894 }
895 
896 static void destroy_qp_user(struct mlx5_ib_dev *dev, struct ib_pd *pd, struct mlx5_ib_qp *qp,
897 			    struct mlx5_ib_qp_base *base,
898 			    struct ib_udata *udata)
899 {
900 	struct mlx5_ib_ucontext *context =
901 		rdma_udata_to_drv_context(
902 			udata,
903 			struct mlx5_ib_ucontext,
904 			ibucontext);
905 
906 	mlx5_ib_db_unmap_user(context, &qp->db);
907 	if (base->ubuffer.umem)
908 		ib_umem_release(base->ubuffer.umem);
909 
910 	/*
911 	 * Free only the BFREGs which are handled by the kernel.
912 	 * BFREGs of UARs allocated dynamically are handled by user.
913 	 */
914 	if (qp->bfregn != MLX5_IB_INVALID_BFREG)
915 		mlx5_ib_free_bfreg(dev, &context->bfregi, qp->bfregn);
916 }
917 
918 static int create_kernel_qp(struct mlx5_ib_dev *dev,
919 			    struct ib_qp_init_attr *init_attr,
920 			    struct mlx5_ib_qp *qp,
921 			    u32 **in, int *inlen,
922 			    struct mlx5_ib_qp_base *base)
923 {
924 	int uar_index;
925 	void *qpc;
926 	int err;
927 
928 	if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN |
929 					IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
930 					IB_QP_CREATE_IPOIB_UD_LSO |
931 					MLX5_IB_QP_CREATE_SQPN_QP1 |
932 					MLX5_IB_QP_CREATE_WC_TEST))
933 		return -EINVAL;
934 
935 	spin_lock_init(&qp->bf.lock32);
936 
937 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
938 		qp->bf.bfreg = &dev->fp_bfreg;
939 	else if (init_attr->create_flags & MLX5_IB_QP_CREATE_WC_TEST)
940 		qp->bf.bfreg = &dev->wc_bfreg;
941 	else
942 		qp->bf.bfreg = &dev->bfreg;
943 
944 	/* We need to divide by two since each register is comprised of
945 	 * two buffers of identical size, namely odd and even
946 	 */
947 	qp->bf.buf_size = (1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size)) / 2;
948 	uar_index = qp->bf.bfreg->index;
949 
950 	err = calc_sq_size(dev, init_attr, qp);
951 	if (err < 0) {
952 		mlx5_ib_dbg(dev, "err %d\n", err);
953 		return err;
954 	}
955 
956 	qp->rq.offset = 0;
957 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
958 	base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
959 
960 	err = mlx5_buf_alloc(dev->mdev, base->ubuffer.buf_size,
961 	    2 * PAGE_SIZE, &qp->buf);
962 	if (err) {
963 		mlx5_ib_dbg(dev, "err %d\n", err);
964 		return err;
965 	}
966 
967 	qp->sq.qend = mlx5_get_send_wqe(qp, qp->sq.wqe_cnt);
968 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
969 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * qp->buf.npages;
970 	*in = mlx5_vzalloc(*inlen);
971 	if (!*in) {
972 		err = -ENOMEM;
973 		goto err_buf;
974 	}
975 
976 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
977 	MLX5_SET(qpc, qpc, uar_page, uar_index);
978 	MLX5_SET(qpc, qpc, ts_format, mlx5_get_qp_default_ts(dev->mdev));
979 	MLX5_SET(qpc, qpc, log_page_size, qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
980 
981 	/* Set "fast registration enabled" for all kernel QPs */
982 	MLX5_SET(qpc, qpc, fre, 1);
983 	MLX5_SET(qpc, qpc, rlky, 1);
984 
985 	if (init_attr->create_flags & MLX5_IB_QP_CREATE_SQPN_QP1) {
986 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
987 		qp->flags |= MLX5_IB_QP_SQPN_QP1;
988 	}
989 
990 	mlx5_fill_page_array(&qp->buf,
991 			     (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas));
992 
993 	err = mlx5_db_alloc(dev->mdev, &qp->db);
994 	if (err) {
995 		mlx5_ib_dbg(dev, "err %d\n", err);
996 		goto err_free;
997 	}
998 
999 	qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
1000 	qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
1001 	qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
1002 	qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
1003 	qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
1004 
1005 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
1006 	    !qp->sq.w_list || !qp->sq.wqe_head) {
1007 		err = -ENOMEM;
1008 		goto err_wrid;
1009 	}
1010 	qp->create_type = MLX5_QP_KERNEL;
1011 
1012 	return 0;
1013 
1014 err_wrid:
1015 	kfree(qp->sq.wqe_head);
1016 	kfree(qp->sq.w_list);
1017 	kfree(qp->sq.wrid);
1018 	kfree(qp->sq.wr_data);
1019 	kfree(qp->rq.wrid);
1020 	mlx5_db_free(dev->mdev, &qp->db);
1021 
1022 err_free:
1023 	kvfree(*in);
1024 
1025 err_buf:
1026 	mlx5_buf_free(dev->mdev, &qp->buf);
1027 	return err;
1028 }
1029 
1030 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1031 {
1032 	kfree(qp->sq.wqe_head);
1033 	kfree(qp->sq.w_list);
1034 	kfree(qp->sq.wrid);
1035 	kfree(qp->sq.wr_data);
1036 	kfree(qp->rq.wrid);
1037 	mlx5_db_free(dev->mdev, &qp->db);
1038 	mlx5_buf_free(dev->mdev, &qp->buf);
1039 }
1040 
1041 static u32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
1042 {
1043 	if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
1044 	    (attr->qp_type == IB_QPT_XRC_INI))
1045 		return MLX5_SRQ_RQ;
1046 	else if (!qp->has_rq)
1047 		return MLX5_ZERO_LEN_RQ;
1048 	else
1049 		return MLX5_NON_ZERO_RQ;
1050 }
1051 
1052 static int is_connected(enum ib_qp_type qp_type)
1053 {
1054 	if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC)
1055 		return 1;
1056 
1057 	return 0;
1058 }
1059 
1060 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1061 				    struct mlx5_ib_sq *sq, u32 tdn,
1062 				    struct ib_pd *pd)
1063 {
1064 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {0};
1065 	void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1066 
1067 	MLX5_SET(create_tis_in, in, uid, to_mpd(pd)->uid);
1068 	MLX5_SET(tisc, tisc, transport_domain, tdn);
1069 	return mlx5_core_create_tis(dev->mdev, in, sizeof(in), &sq->tisn);
1070 }
1071 
1072 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1073 				      struct mlx5_ib_sq *sq, struct ib_pd *pd)
1074 {
1075 	mlx5_core_destroy_tis(dev->mdev, sq->tisn, to_mpd(pd)->uid);
1076 }
1077 
1078 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1079 				   struct mlx5_ib_sq *sq, void *qpin,
1080 				   struct ib_pd *pd)
1081 {
1082 	struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1083 	__be64 *pas;
1084 	void *in;
1085 	void *sqc;
1086 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1087 	void *wq;
1088 	int inlen;
1089 	int err;
1090 	int page_shift = 0;
1091 	int npages;
1092 	int ncont = 0;
1093 	u32 offset = 0;
1094 	u8 ts_format;
1095 
1096 	ts_format = mlx5_get_sq_default_ts(dev->mdev);
1097 
1098 	err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr, ubuffer->buf_size,
1099 			       &sq->ubuffer.umem, &npages, &page_shift,
1100 			       &ncont, &offset);
1101 	if (err)
1102 		return err;
1103 
1104 	inlen = MLX5_ST_SZ_BYTES(create_sq_in) + sizeof(u64) * ncont;
1105 	in = mlx5_vzalloc(inlen);
1106 	if (!in) {
1107 		err = -ENOMEM;
1108 		goto err_umem;
1109 	}
1110 
1111 	MLX5_SET(create_sq_in, in, uid, to_mpd(pd)->uid);
1112 	sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1113 	MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1114 	MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1115 	MLX5_SET(sqc, sqc, ts_format, ts_format);
1116 	MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1117 	MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1118 	MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1119 	MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1120 
1121 	wq = MLX5_ADDR_OF(sqc, sqc, wq);
1122 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1123 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1124 	MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1125 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1126 	MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1127 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1128 	MLX5_SET(wq, wq, log_wq_pg_sz,  page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1129 	MLX5_SET(wq, wq, page_offset, offset);
1130 
1131 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1132 	mlx5_ib_populate_pas(dev, sq->ubuffer.umem, page_shift, pas, 0);
1133 
1134 	err = mlx5_core_create_sq_tracked(dev->mdev, in, inlen, &sq->base.mqp);
1135 
1136 	kvfree(in);
1137 
1138 	if (err)
1139 		goto err_umem;
1140 
1141 	return 0;
1142 
1143 err_umem:
1144 	ib_umem_release(sq->ubuffer.umem);
1145 	sq->ubuffer.umem = NULL;
1146 
1147 	return err;
1148 }
1149 
1150 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1151 				     struct mlx5_ib_sq *sq)
1152 {
1153 	mlx5_core_destroy_sq_tracked(dev->mdev, &sq->base.mqp);
1154 	ib_umem_release(sq->ubuffer.umem);
1155 }
1156 
1157 static int get_rq_pas_size(void *qpc)
1158 {
1159 	u32 log_page_size = MLX5_GET(qpc, qpc, log_page_size) + 12;
1160 	u32 log_rq_stride = MLX5_GET(qpc, qpc, log_rq_stride);
1161 	u32 log_rq_size   = MLX5_GET(qpc, qpc, log_rq_size);
1162 	u32 page_offset   = MLX5_GET(qpc, qpc, page_offset);
1163 	u32 po_quanta	  = 1 << (log_page_size - 6);
1164 	u32 rq_sz	  = 1 << (log_rq_size + 4 + log_rq_stride);
1165 	u32 page_size	  = 1 << log_page_size;
1166 	u32 rq_sz_po      = rq_sz + (page_offset * po_quanta);
1167 	u32 rq_num_pas	  = (rq_sz_po + page_size - 1) / page_size;
1168 
1169 	return rq_num_pas * sizeof(u64);
1170 }
1171 
1172 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1173 				   struct mlx5_ib_rq *rq, void *qpin,
1174 				   struct ib_pd *pd)
1175 {
1176 	struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1177 	__be64 *pas;
1178 	__be64 *qp_pas;
1179 	void *in;
1180 	void *rqc;
1181 	void *wq;
1182 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1183 	int inlen;
1184 	int err;
1185 	u32 rq_pas_size = get_rq_pas_size(qpc);
1186 	u8 ts_format;
1187 
1188 	ts_format = mlx5_get_rq_default_ts(dev->mdev);
1189 
1190 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + rq_pas_size;
1191 	in = mlx5_vzalloc(inlen);
1192 	if (!in)
1193 		return -ENOMEM;
1194 
1195 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
1196 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1197 	MLX5_SET(rqc, rqc, vlan_strip_disable, 1);
1198 	MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_RQ_TYPE_MEMORY_RQ_INLINE);
1199 	MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1200 	MLX5_SET(rqc, rqc, ts_format, ts_format);
1201 	MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1202 	MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1203 	MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1204 
1205 	if (mqp->flags & MLX5_IB_QP_CAP_SCATTER_FCS)
1206 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
1207 
1208 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
1209 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1210 	MLX5_SET(wq, wq, end_padding_mode,
1211 		 MLX5_GET(qpc, qpc, end_padding_mode));
1212 	MLX5_SET(wq, wq, page_offset, MLX5_GET(qpc, qpc, page_offset));
1213 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1214 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1215 	MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1216 	MLX5_SET(wq, wq, log_wq_pg_sz, MLX5_GET(qpc, qpc, log_page_size));
1217 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1218 
1219 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1220 	qp_pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, qpin, pas);
1221 	memcpy(pas, qp_pas, rq_pas_size);
1222 
1223 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rq->base.mqp);
1224 
1225 	kvfree(in);
1226 
1227 	return err;
1228 }
1229 
1230 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1231 				     struct mlx5_ib_rq *rq)
1232 {
1233 	mlx5_core_destroy_rq_tracked(dev->mdev, &rq->base.mqp);
1234 }
1235 
1236 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1237 				    struct mlx5_ib_rq *rq, u32 tdn,
1238 				    struct ib_pd *pd)
1239 {
1240 	u32 *in;
1241 	void *tirc;
1242 	int inlen;
1243 	int err;
1244 
1245 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1246 	in = mlx5_vzalloc(inlen);
1247 	if (!in)
1248 		return -ENOMEM;
1249 
1250 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1251 	tirc = MLX5_ADDR_OF(create_tir_in, in, tir_context);
1252 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1253 	MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1254 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1255 
1256 	err = mlx5_core_create_tir(dev->mdev, in, inlen, &rq->tirn);
1257 
1258 	kvfree(in);
1259 
1260 	return err;
1261 }
1262 
1263 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1264 				      struct mlx5_ib_rq *rq,
1265 				      struct ib_pd *pd)
1266 {
1267 	mlx5_core_destroy_tir(dev->mdev, rq->tirn, to_mpd(pd)->uid);
1268 }
1269 
1270 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1271 				u32 *in,
1272 				struct ib_pd *pd)
1273 {
1274 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1275 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1276 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1277 	struct ib_uobject *uobj = pd->uobject;
1278 	struct ib_ucontext *ucontext = uobj->context;
1279 	struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1280 	int err;
1281 	u32 tdn = mucontext->tdn;
1282 
1283 	if (qp->sq.wqe_cnt) {
1284 		err = create_raw_packet_qp_tis(dev, sq, tdn, pd);
1285 		if (err)
1286 			return err;
1287 
1288 		err = create_raw_packet_qp_sq(dev, sq, in, pd);
1289 		if (err)
1290 			goto err_destroy_tis;
1291 
1292 		sq->base.container_mibqp = qp;
1293 	}
1294 
1295 	if (qp->rq.wqe_cnt) {
1296 		rq->base.container_mibqp = qp;
1297 
1298 		err = create_raw_packet_qp_rq(dev, rq, in, pd);
1299 		if (err)
1300 			goto err_destroy_sq;
1301 
1302 
1303 		err = create_raw_packet_qp_tir(dev, rq, tdn, pd);
1304 		if (err)
1305 			goto err_destroy_rq;
1306 	}
1307 
1308 	qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1309 						     rq->base.mqp.qpn;
1310 
1311 	return 0;
1312 
1313 err_destroy_rq:
1314 	destroy_raw_packet_qp_rq(dev, rq);
1315 err_destroy_sq:
1316 	if (!qp->sq.wqe_cnt)
1317 		return err;
1318 	destroy_raw_packet_qp_sq(dev, sq);
1319 err_destroy_tis:
1320 	destroy_raw_packet_qp_tis(dev, sq, pd);
1321 
1322 	return err;
1323 }
1324 
1325 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1326 				  struct mlx5_ib_qp *qp)
1327 {
1328 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1329 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1330 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1331 
1332 	if (qp->rq.wqe_cnt) {
1333 		destroy_raw_packet_qp_tir(dev, rq, qp->ibqp.pd);
1334 		destroy_raw_packet_qp_rq(dev, rq);
1335 	}
1336 
1337 	if (qp->sq.wqe_cnt) {
1338 		destroy_raw_packet_qp_sq(dev, sq);
1339 		destroy_raw_packet_qp_tis(dev, sq, qp->ibqp.pd);
1340 	}
1341 }
1342 
1343 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1344 				    struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1345 {
1346 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1347 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1348 
1349 	sq->sq = &qp->sq;
1350 	rq->rq = &qp->rq;
1351 	sq->doorbell = &qp->db;
1352 	rq->doorbell = &qp->db;
1353 }
1354 
1355 static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1356 {
1357 	mlx5_core_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1358 			      to_mpd(qp->ibqp.pd)->uid);
1359 }
1360 
1361 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1362 				 struct ib_pd *pd,
1363 				 struct ib_qp_init_attr *init_attr,
1364 				 struct ib_udata *udata)
1365 {
1366 	struct ib_uobject *uobj = pd->uobject;
1367 	struct ib_ucontext *ucontext = uobj->context;
1368 	struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1369 	struct mlx5_ib_create_qp_resp resp = {};
1370 	int inlen;
1371 	int err;
1372 	u32 *in;
1373 	void *tirc;
1374 	void *hfso;
1375 	u32 selected_fields = 0;
1376 	size_t min_resp_len;
1377 	u32 tdn = mucontext->tdn;
1378 	struct mlx5_ib_create_qp_rss ucmd = {};
1379 	size_t required_cmd_sz;
1380 
1381 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1382 		return -EOPNOTSUPP;
1383 
1384 	if (init_attr->create_flags || init_attr->send_cq)
1385 		return -EINVAL;
1386 
1387 	min_resp_len = offsetof(typeof(resp), bfreg_index) + sizeof(resp.bfreg_index);
1388 	if (udata->outlen < min_resp_len)
1389 		return -EINVAL;
1390 
1391 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) + sizeof(ucmd.reserved1);
1392 	if (udata->inlen < required_cmd_sz) {
1393 		mlx5_ib_dbg(dev, "invalid inlen\n");
1394 		return -EINVAL;
1395 	}
1396 
1397 	if (udata->inlen > sizeof(ucmd) &&
1398 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
1399 				 udata->inlen - sizeof(ucmd))) {
1400 		mlx5_ib_dbg(dev, "inlen is not supported\n");
1401 		return -EOPNOTSUPP;
1402 	}
1403 
1404 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
1405 		mlx5_ib_dbg(dev, "copy failed\n");
1406 		return -EFAULT;
1407 	}
1408 
1409 	if (ucmd.comp_mask) {
1410 		mlx5_ib_dbg(dev, "invalid comp mask\n");
1411 		return -EOPNOTSUPP;
1412 	}
1413 
1414 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)) || ucmd.reserved1) {
1415 		mlx5_ib_dbg(dev, "invalid reserved\n");
1416 		return -EOPNOTSUPP;
1417 	}
1418 
1419 	err = ib_copy_to_udata(udata, &resp, min_resp_len);
1420 	if (err) {
1421 		mlx5_ib_dbg(dev, "copy failed\n");
1422 		return -EINVAL;
1423 	}
1424 
1425 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1426 	in = mlx5_vzalloc(inlen);
1427 	if (!in)
1428 		return -ENOMEM;
1429 
1430 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1431 	tirc = MLX5_ADDR_OF(create_tir_in, in, tir_context);
1432 	MLX5_SET(tirc, tirc, disp_type,
1433 		 MLX5_TIRC_DISP_TYPE_INDIRECT);
1434 	MLX5_SET(tirc, tirc, indirect_table,
1435 		 init_attr->rwq_ind_tbl->ind_tbl_num);
1436 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1437 
1438 	hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1439 	switch (ucmd.rx_hash_function) {
1440 	case MLX5_RX_HASH_FUNC_TOEPLITZ:
1441 	{
1442 		void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1443 		size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1444 
1445 		if (len != ucmd.rx_key_len) {
1446 			err = -EINVAL;
1447 			goto err;
1448 		}
1449 
1450 		MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FUNC_TOEPLITZ);
1451 		MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
1452 		memcpy(rss_key, ucmd.rx_hash_key, len);
1453 		break;
1454 	}
1455 	default:
1456 		err = -EOPNOTSUPP;
1457 		goto err;
1458 	}
1459 
1460 	if (!ucmd.rx_hash_fields_mask) {
1461 		/* special case when this TIR serves as steering entry without hashing */
1462 		if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1463 			goto create_tir;
1464 		err = -EINVAL;
1465 		goto err;
1466 	}
1467 
1468 	if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1469 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1470 	     ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1471 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1472 		err = -EINVAL;
1473 		goto err;
1474 	}
1475 
1476 	/* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1477 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1478 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1479 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1480 			 MLX5_L3_PROT_TYPE_IPV4);
1481 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1482 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1483 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1484 			 MLX5_L3_PROT_TYPE_IPV6);
1485 
1486 	if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1487 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP)) &&
1488 	     ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1489 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))) {
1490 		err = -EINVAL;
1491 		goto err;
1492 	}
1493 
1494 	/* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1495 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1496 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1497 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1498 			 MLX5_L4_PROT_TYPE_TCP);
1499 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1500 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1501 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1502 			 MLX5_L4_PROT_TYPE_UDP);
1503 
1504 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1505 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1506 		selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1507 
1508 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1509 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1510 		selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1511 
1512 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1513 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1514 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1515 
1516 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1517 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1518 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1519 
1520 	MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1521 
1522 create_tir:
1523 	err = mlx5_core_create_tir(dev->mdev, in, inlen, &qp->rss_qp.tirn);
1524 
1525 	if (err)
1526 		goto err;
1527 
1528 	kvfree(in);
1529 	/* qpn is reserved for that QP */
1530 	qp->trans_qp.base.mqp.qpn = 0;
1531 	qp->flags |= MLX5_IB_QP_RSS;
1532 	return 0;
1533 
1534 err:
1535 	kvfree(in);
1536 	return err;
1537 }
1538 
1539 static int atomic_size_to_mode(int size_mask)
1540 {
1541 	/* driver does not support atomic_size > 256B
1542 	 * and does not know how to translate bigger sizes
1543 	 */
1544 	int supported_size_mask = size_mask & 0x1ff;
1545 	int log_max_size;
1546 
1547 	if (!supported_size_mask)
1548 		return -EOPNOTSUPP;
1549 
1550 	log_max_size = __fls(supported_size_mask);
1551 
1552 	if (log_max_size > 3)
1553 		return log_max_size;
1554 
1555 	return MLX5_ATOMIC_MODE_8B;
1556 }
1557 
1558 static int get_atomic_mode(struct mlx5_ib_dev *dev,
1559 			   enum ib_qp_type qp_type)
1560 {
1561 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
1562 	u8 atomic = MLX5_CAP_GEN(dev->mdev, atomic);
1563 	int atomic_mode = -EOPNOTSUPP;
1564 	int atomic_size_mask;
1565 
1566 	if (!atomic)
1567 		return -EOPNOTSUPP;
1568 
1569 	if (qp_type == MLX5_IB_QPT_DCT)
1570 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
1571 	else
1572 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
1573 
1574 	if ((atomic_operations & MLX5_ATOMIC_OPS_MASKED_CMP_SWAP) ||
1575 	    (atomic_operations & MLX5_ATOMIC_OPS_MASKED_FETCH_ADD))
1576 		atomic_mode = atomic_size_to_mode(atomic_size_mask);
1577 
1578 	if (atomic_mode <= 0 &&
1579 	    (atomic_operations & MLX5_ATOMIC_OPS_CMP_SWAP &&
1580 	     atomic_operations & MLX5_ATOMIC_OPS_FETCH_ADD))
1581 		atomic_mode = MLX5_ATOMIC_MODE_IB_COMP;
1582 
1583 	return atomic_mode;
1584 }
1585 
1586 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1587 			    struct ib_qp_init_attr *init_attr,
1588 			    struct ib_udata *udata, struct mlx5_ib_qp *qp)
1589 {
1590 	struct mlx5_ib_resources *devr = &dev->devr;
1591 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1592 	struct mlx5_core_dev *mdev = dev->mdev;
1593 	struct mlx5_ib_create_qp_resp resp;
1594 	struct mlx5_ib_cq *send_cq;
1595 	struct mlx5_ib_cq *recv_cq;
1596 	unsigned long flags;
1597 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
1598 	struct mlx5_ib_create_qp ucmd;
1599 	struct mlx5_ib_qp_base *base;
1600 	void *qpc;
1601 	u32 *in;
1602 	int err;
1603 
1604 	base = init_attr->qp_type == IB_QPT_RAW_PACKET ?
1605 	       &qp->raw_packet_qp.rq.base :
1606 	       &qp->trans_qp.base;
1607 
1608 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1609 		mlx5_ib_odp_create_qp(qp);
1610 
1611 	mutex_init(&qp->mutex);
1612 	spin_lock_init(&qp->sq.lock);
1613 	spin_lock_init(&qp->rq.lock);
1614 
1615 	if (init_attr->rwq_ind_tbl) {
1616 		if (!udata)
1617 			return -ENOSYS;
1618 
1619 		err = create_rss_raw_qp_tir(dev, qp, pd, init_attr, udata);
1620 		return err;
1621 	}
1622 
1623 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
1624 		if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
1625 			mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
1626 			return -EINVAL;
1627 		} else {
1628 			qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1629 		}
1630 	}
1631 
1632 	if (init_attr->create_flags &
1633 			(IB_QP_CREATE_CROSS_CHANNEL |
1634 			 IB_QP_CREATE_MANAGED_SEND |
1635 			 IB_QP_CREATE_MANAGED_RECV)) {
1636 		if (!MLX5_CAP_GEN(mdev, cd)) {
1637 			mlx5_ib_dbg(dev, "cross-channel isn't supported\n");
1638 			return -EINVAL;
1639 		}
1640 		if (init_attr->create_flags & IB_QP_CREATE_CROSS_CHANNEL)
1641 			qp->flags |= MLX5_IB_QP_CROSS_CHANNEL;
1642 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_SEND)
1643 			qp->flags |= MLX5_IB_QP_MANAGED_SEND;
1644 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_RECV)
1645 			qp->flags |= MLX5_IB_QP_MANAGED_RECV;
1646 	}
1647 
1648 	if (init_attr->qp_type == IB_QPT_UD &&
1649 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO))
1650 		if (!MLX5_CAP_GEN(mdev, ipoib_ipoib_offloads)) {
1651 			mlx5_ib_dbg(dev, "ipoib UD lso qp isn't supported\n");
1652 			return -EOPNOTSUPP;
1653 		}
1654 
1655 	if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1656 		if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
1657 			mlx5_ib_dbg(dev, "Scatter FCS is supported only for Raw Packet QPs");
1658 			return -EOPNOTSUPP;
1659 		}
1660 		if (!MLX5_CAP_GEN(dev->mdev, eth_net_offloads) ||
1661 		    !MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
1662 			mlx5_ib_dbg(dev, "Scatter FCS isn't supported\n");
1663 			return -EOPNOTSUPP;
1664 		}
1665 		qp->flags |= MLX5_IB_QP_CAP_SCATTER_FCS;
1666 	}
1667 
1668 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1669 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
1670 
1671 	if (pd && pd->uobject) {
1672 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
1673 			mlx5_ib_dbg(dev, "copy failed\n");
1674 			return -EFAULT;
1675 		}
1676 
1677 		err = get_qp_user_index(to_mucontext(pd->uobject->context),
1678 					&ucmd, udata->inlen, &uidx);
1679 		if (err)
1680 			return err;
1681 
1682 		qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
1683 		qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
1684 	} else {
1685 		qp->wq_sig = !!wq_signature;
1686 	}
1687 
1688 	qp->has_rq = qp_has_rq(init_attr);
1689 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
1690 			  qp, (pd && pd->uobject) ? &ucmd : NULL);
1691 	if (err) {
1692 		mlx5_ib_dbg(dev, "err %d\n", err);
1693 		return err;
1694 	}
1695 
1696 	if (pd) {
1697 		if (pd->uobject) {
1698 			__u32 max_wqes =
1699 				1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1700 			mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
1701 			if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
1702 			    ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
1703 				mlx5_ib_dbg(dev, "invalid rq params\n");
1704 				return -EINVAL;
1705 			}
1706 			if (ucmd.sq_wqe_count > max_wqes) {
1707 				mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
1708 					    ucmd.sq_wqe_count, max_wqes);
1709 				return -EINVAL;
1710 			}
1711 			if (init_attr->create_flags &
1712 			    MLX5_IB_QP_CREATE_SQPN_QP1) {
1713 				mlx5_ib_dbg(dev, "user-space is not allowed to create UD QPs spoofing as QP1\n");
1714 				return -EINVAL;
1715 			}
1716 			err = create_user_qp(dev, pd, qp, udata, init_attr, &in,
1717 					     &resp, &inlen, base);
1718 			if (err)
1719 				mlx5_ib_dbg(dev, "err %d\n", err);
1720 		} else {
1721 			err = create_kernel_qp(dev, init_attr, qp, &in, &inlen,
1722 					       base);
1723 			if (err)
1724 				mlx5_ib_dbg(dev, "err %d\n", err);
1725 		}
1726 
1727 		if (err)
1728 			return err;
1729 	} else {
1730 		in = mlx5_vzalloc(inlen);
1731 		if (!in)
1732 			return -ENOMEM;
1733 
1734 		qp->create_type = MLX5_QP_EMPTY;
1735 	}
1736 
1737 	if (is_sqp(init_attr->qp_type))
1738 		qp->port = init_attr->port_num;
1739 
1740 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1741 
1742 	MLX5_SET(qpc, qpc, st, to_mlx5_st(init_attr->qp_type));
1743 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
1744 
1745 	if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
1746 		MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
1747 	else
1748 		MLX5_SET(qpc, qpc, latency_sensitive, 1);
1749 
1750 
1751 	if (qp->wq_sig)
1752 		MLX5_SET(qpc, qpc, wq_signature, 1);
1753 
1754 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
1755 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
1756 
1757 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
1758 		MLX5_SET(qpc, qpc, cd_master, 1);
1759 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
1760 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
1761 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
1762 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
1763 
1764 	if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
1765 		int rcqe_sz;
1766 		int scqe_sz;
1767 
1768 		rcqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->recv_cq);
1769 		scqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->send_cq);
1770 
1771 		if (rcqe_sz == 128)
1772 			MLX5_SET(qpc, qpc, cs_res, MLX5_RES_SCAT_DATA64_CQE);
1773 		else
1774 			MLX5_SET(qpc, qpc, cs_res, MLX5_RES_SCAT_DATA32_CQE);
1775 
1776 		if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) {
1777 			if (scqe_sz == 128)
1778 				MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA64_CQE);
1779 			else
1780 				MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA32_CQE);
1781 		}
1782 	}
1783 
1784 	if (qp->rq.wqe_cnt) {
1785 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
1786 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
1787 	}
1788 
1789 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1790 		MLX5_SET(qpc, qpc, ts_format, mlx5_get_qp_default_ts(dev->mdev));
1791 
1792 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
1793 
1794 	if (qp->sq.wqe_cnt)
1795 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
1796 	else
1797 		MLX5_SET(qpc, qpc, no_sq, 1);
1798 
1799 	/* Set default resources */
1800 	switch (init_attr->qp_type) {
1801 	case IB_QPT_XRC_TGT:
1802 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1803 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(devr->c0)->mcq.cqn);
1804 		MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(devr->s0)->msrq.srqn);
1805 		MLX5_SET(qpc, qpc, xrcd, to_mxrcd(init_attr->xrcd)->xrcdn);
1806 		break;
1807 	case IB_QPT_XRC_INI:
1808 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1809 		MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
1810 		MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(devr->s0)->msrq.srqn);
1811 		break;
1812 	default:
1813 		if (init_attr->srq) {
1814 			MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x0)->xrcdn);
1815 			MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(init_attr->srq)->msrq.srqn);
1816 		} else {
1817 			MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
1818 			MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(devr->s1)->msrq.srqn);
1819 		}
1820 	}
1821 
1822 	if (init_attr->send_cq)
1823 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(init_attr->send_cq)->mcq.cqn);
1824 
1825 	if (init_attr->recv_cq)
1826 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(init_attr->recv_cq)->mcq.cqn);
1827 
1828 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
1829 
1830 	/* 0xffffff means we ask to work with cqe version 0 */
1831 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
1832 		MLX5_SET(qpc, qpc, user_index, uidx);
1833 
1834 	/* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
1835 	if (init_attr->qp_type == IB_QPT_UD &&
1836 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)) {
1837 		MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
1838 		qp->flags |= MLX5_IB_QP_LSO;
1839 	}
1840 
1841 	if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
1842 		qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd.sq_buf_addr;
1843 		raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
1844 		err = create_raw_packet_qp(dev, qp, in, pd);
1845 	} else {
1846 		err = mlx5_core_create_qp(dev->mdev, &base->mqp, in, inlen);
1847 	}
1848 
1849 	if (err) {
1850 		mlx5_ib_dbg(dev, "create qp failed\n");
1851 		goto err_create;
1852 	}
1853 
1854 	kvfree(in);
1855 
1856 	base->container_mibqp = qp;
1857 	base->mqp.event = mlx5_ib_qp_event;
1858 
1859 	get_cqs(init_attr->qp_type, init_attr->send_cq, init_attr->recv_cq,
1860 		&send_cq, &recv_cq);
1861 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1862 	mlx5_ib_lock_cqs(send_cq, recv_cq);
1863 	/* Maintain device to QPs access, needed for further handling via reset
1864 	 * flow
1865 	 */
1866 	list_add_tail(&qp->qps_list, &dev->qp_list);
1867 	/* Maintain CQ to QPs access, needed for further handling via reset flow
1868 	 */
1869 	if (send_cq)
1870 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
1871 	if (recv_cq)
1872 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
1873 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
1874 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1875 
1876 	return 0;
1877 
1878 err_create:
1879 	if (qp->create_type == MLX5_QP_USER)
1880 		destroy_qp_user(dev, pd, qp, base, udata);
1881 	else if (qp->create_type == MLX5_QP_KERNEL)
1882 		destroy_qp_kernel(dev, qp);
1883 
1884 	kvfree(in);
1885 	return err;
1886 }
1887 
1888 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1889 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1890 {
1891 	if (send_cq) {
1892 		if (recv_cq) {
1893 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1894 				spin_lock(&send_cq->lock);
1895 				spin_lock_nested(&recv_cq->lock,
1896 						 SINGLE_DEPTH_NESTING);
1897 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1898 				spin_lock(&send_cq->lock);
1899 				__acquire(&recv_cq->lock);
1900 			} else {
1901 				spin_lock(&recv_cq->lock);
1902 				spin_lock_nested(&send_cq->lock,
1903 						 SINGLE_DEPTH_NESTING);
1904 			}
1905 		} else {
1906 			spin_lock(&send_cq->lock);
1907 			__acquire(&recv_cq->lock);
1908 		}
1909 	} else if (recv_cq) {
1910 		spin_lock(&recv_cq->lock);
1911 		__acquire(&send_cq->lock);
1912 	} else {
1913 		__acquire(&send_cq->lock);
1914 		__acquire(&recv_cq->lock);
1915 	}
1916 }
1917 
1918 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1919 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1920 {
1921 	if (send_cq) {
1922 		if (recv_cq) {
1923 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
1924 				spin_unlock(&recv_cq->lock);
1925 				spin_unlock(&send_cq->lock);
1926 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1927 				__release(&recv_cq->lock);
1928 				spin_unlock(&send_cq->lock);
1929 			} else {
1930 				spin_unlock(&send_cq->lock);
1931 				spin_unlock(&recv_cq->lock);
1932 			}
1933 		} else {
1934 			__release(&recv_cq->lock);
1935 			spin_unlock(&send_cq->lock);
1936 		}
1937 	} else if (recv_cq) {
1938 		__release(&send_cq->lock);
1939 		spin_unlock(&recv_cq->lock);
1940 	} else {
1941 		__release(&recv_cq->lock);
1942 		__release(&send_cq->lock);
1943 	}
1944 }
1945 
1946 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
1947 {
1948 	return to_mpd(qp->ibqp.pd);
1949 }
1950 
1951 static void get_cqs(enum ib_qp_type qp_type,
1952 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
1953 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
1954 {
1955 	switch (qp_type) {
1956 	case IB_QPT_XRC_TGT:
1957 		*send_cq = NULL;
1958 		*recv_cq = NULL;
1959 		break;
1960 	case MLX5_IB_QPT_REG_UMR:
1961 	case IB_QPT_XRC_INI:
1962 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
1963 		*recv_cq = NULL;
1964 		break;
1965 
1966 	case IB_QPT_SMI:
1967 	case MLX5_IB_QPT_HW_GSI:
1968 	case IB_QPT_RC:
1969 	case IB_QPT_UC:
1970 	case IB_QPT_UD:
1971 	case IB_QPT_RAW_IPV6:
1972 	case IB_QPT_RAW_ETHERTYPE:
1973 	case IB_QPT_RAW_PACKET:
1974 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
1975 		*recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
1976 		break;
1977 
1978 	case IB_QPT_MAX:
1979 	default:
1980 		*send_cq = NULL;
1981 		*recv_cq = NULL;
1982 		break;
1983 	}
1984 }
1985 
1986 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1987 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
1988 				u8 lag_tx_affinity);
1989 
1990 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1991 			      struct ib_udata *udata)
1992 {
1993 	struct mlx5_ib_cq *send_cq, *recv_cq;
1994 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
1995 	unsigned long flags;
1996 	int err;
1997 
1998 	if (qp->ibqp.rwq_ind_tbl) {
1999 		destroy_rss_raw_qp_tir(dev, qp);
2000 		return;
2001 	}
2002 
2003 	base = qp->ibqp.qp_type == IB_QPT_RAW_PACKET ?
2004 	       &qp->raw_packet_qp.rq.base :
2005 	       &qp->trans_qp.base;
2006 
2007 	if (qp->state != IB_QPS_RESET) {
2008 		if (qp->ibqp.qp_type != IB_QPT_RAW_PACKET) {
2009 			mlx5_ib_qp_disable_pagefaults(qp);
2010 			err = mlx5_core_qp_modify(dev->mdev,
2011 						  MLX5_CMD_OP_2RST_QP, 0,
2012 						  NULL, &base->mqp);
2013 		} else {
2014 			struct mlx5_modify_raw_qp_param raw_qp_param = {
2015 				.operation = MLX5_CMD_OP_2RST_QP
2016 			};
2017 
2018 			err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
2019 		}
2020 		if (err)
2021 			mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
2022 				     base->mqp.qpn);
2023 	}
2024 
2025 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
2026 		&send_cq, &recv_cq);
2027 
2028 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2029 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2030 	/* del from lists under both locks above to protect reset flow paths */
2031 	list_del(&qp->qps_list);
2032 	if (send_cq)
2033 		list_del(&qp->cq_send_list);
2034 
2035 	if (recv_cq)
2036 		list_del(&qp->cq_recv_list);
2037 
2038 	if (qp->create_type == MLX5_QP_KERNEL) {
2039 		__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2040 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
2041 		if (send_cq != recv_cq)
2042 			__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
2043 					   NULL);
2044 	}
2045 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2046 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2047 
2048 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
2049 		destroy_raw_packet_qp(dev, qp);
2050 	} else {
2051 		err = mlx5_core_destroy_qp(dev->mdev, &base->mqp);
2052 		if (err)
2053 			mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
2054 				     base->mqp.qpn);
2055 	}
2056 
2057 	if (qp->create_type == MLX5_QP_KERNEL)
2058 		destroy_qp_kernel(dev, qp);
2059 	else if (qp->create_type == MLX5_QP_USER)
2060 		destroy_qp_user(dev, &get_pd(qp)->ibpd, qp, base, udata);
2061 }
2062 
2063 static const char *ib_qp_type_str(enum ib_qp_type type)
2064 {
2065 	switch (type) {
2066 	case IB_QPT_SMI:
2067 		return "IB_QPT_SMI";
2068 	case IB_QPT_GSI:
2069 		return "IB_QPT_GSI";
2070 	case IB_QPT_RC:
2071 		return "IB_QPT_RC";
2072 	case IB_QPT_UC:
2073 		return "IB_QPT_UC";
2074 	case IB_QPT_UD:
2075 		return "IB_QPT_UD";
2076 	case IB_QPT_RAW_IPV6:
2077 		return "IB_QPT_RAW_IPV6";
2078 	case IB_QPT_RAW_ETHERTYPE:
2079 		return "IB_QPT_RAW_ETHERTYPE";
2080 	case IB_QPT_XRC_INI:
2081 		return "IB_QPT_XRC_INI";
2082 	case IB_QPT_XRC_TGT:
2083 		return "IB_QPT_XRC_TGT";
2084 	case IB_QPT_RAW_PACKET:
2085 		return "IB_QPT_RAW_PACKET";
2086 	case MLX5_IB_QPT_REG_UMR:
2087 		return "MLX5_IB_QPT_REG_UMR";
2088 	case IB_QPT_MAX:
2089 	default:
2090 		return "Invalid QP type";
2091 	}
2092 }
2093 
2094 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
2095 				struct ib_qp_init_attr *init_attr,
2096 				struct ib_udata *udata)
2097 {
2098 	struct mlx5_ib_dev *dev;
2099 	struct mlx5_ib_qp *qp;
2100 	u16 xrcdn = 0;
2101 	int err;
2102 
2103 	if (pd) {
2104 		dev = to_mdev(pd->device);
2105 
2106 		if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
2107 			if (!pd->uobject) {
2108 				mlx5_ib_dbg(dev, "Raw Packet QP is not supported for kernel consumers\n");
2109 				return ERR_PTR(-EINVAL);
2110 			} else if (!to_mucontext(pd->uobject->context)->cqe_version) {
2111 				mlx5_ib_dbg(dev, "Raw Packet QP is only supported for CQE version > 0\n");
2112 				return ERR_PTR(-EINVAL);
2113 			}
2114 		}
2115 	} else {
2116 		/* being cautious here */
2117 		if (init_attr->qp_type != IB_QPT_XRC_TGT &&
2118 		    init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
2119 			pr_warn("%s: no PD for transport %s\n", __func__,
2120 				ib_qp_type_str(init_attr->qp_type));
2121 			return ERR_PTR(-EINVAL);
2122 		}
2123 		dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
2124 	}
2125 
2126 	switch (init_attr->qp_type) {
2127 	case IB_QPT_XRC_TGT:
2128 	case IB_QPT_XRC_INI:
2129 		if (!MLX5_CAP_GEN(dev->mdev, xrc)) {
2130 			mlx5_ib_dbg(dev, "XRC not supported\n");
2131 			return ERR_PTR(-ENOSYS);
2132 		}
2133 		init_attr->recv_cq = NULL;
2134 		if (init_attr->qp_type == IB_QPT_XRC_TGT) {
2135 			xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
2136 			init_attr->send_cq = NULL;
2137 		}
2138 
2139 		/* fall through */
2140 	case IB_QPT_RAW_PACKET:
2141 	case IB_QPT_RC:
2142 	case IB_QPT_UC:
2143 	case IB_QPT_UD:
2144 	case IB_QPT_SMI:
2145 	case MLX5_IB_QPT_HW_GSI:
2146 	case MLX5_IB_QPT_REG_UMR:
2147 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2148 		if (!qp)
2149 			return ERR_PTR(-ENOMEM);
2150 
2151 		err = create_qp_common(dev, pd, init_attr, udata, qp);
2152 		if (err) {
2153 			mlx5_ib_dbg(dev, "create_qp_common failed\n");
2154 			kfree(qp);
2155 			return ERR_PTR(err);
2156 		}
2157 
2158 		if (is_qp0(init_attr->qp_type))
2159 			qp->ibqp.qp_num = 0;
2160 		else if (is_qp1(init_attr->qp_type))
2161 			qp->ibqp.qp_num = 1;
2162 		else
2163 			qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
2164 
2165 		mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
2166 			    qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
2167 			    init_attr->recv_cq ? to_mcq(init_attr->recv_cq)->mcq.cqn : -1,
2168 			    init_attr->send_cq ? to_mcq(init_attr->send_cq)->mcq.cqn : -1);
2169 
2170 		qp->trans_qp.xrcdn = xrcdn;
2171 
2172 		break;
2173 
2174 	case IB_QPT_GSI:
2175 		return mlx5_ib_gsi_create_qp(pd, init_attr);
2176 
2177 	case IB_QPT_RAW_IPV6:
2178 	case IB_QPT_RAW_ETHERTYPE:
2179 	case IB_QPT_MAX:
2180 	default:
2181 		mlx5_ib_dbg(dev, "unsupported qp type %d\n",
2182 			    init_attr->qp_type);
2183 		/* Don't support raw QPs */
2184 		return ERR_PTR(-EINVAL);
2185 	}
2186 
2187 	return &qp->ibqp;
2188 }
2189 
2190 int mlx5_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
2191 {
2192 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2193 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2194 
2195 	if (unlikely(qp->qp_type == IB_QPT_GSI))
2196 		return mlx5_ib_gsi_destroy_qp(qp);
2197 
2198 	destroy_qp_common(dev, mqp, udata);
2199 
2200 	kfree(mqp);
2201 
2202 	return 0;
2203 }
2204 
2205 static int to_mlx5_access_flags(struct mlx5_ib_qp *qp,
2206 				const struct ib_qp_attr *attr,
2207 				int attr_mask, __be32 *hw_access_flags_be)
2208 {
2209 	u8 dest_rd_atomic;
2210 	u32 access_flags, hw_access_flags = 0;
2211 
2212 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
2213 
2214 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2215 		dest_rd_atomic = attr->max_dest_rd_atomic;
2216 	else
2217 		dest_rd_atomic = qp->trans_qp.resp_depth;
2218 
2219 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2220 		access_flags = attr->qp_access_flags;
2221 	else
2222 		access_flags = qp->trans_qp.atomic_rd_en;
2223 
2224 	if (!dest_rd_atomic)
2225 		access_flags &= IB_ACCESS_REMOTE_WRITE;
2226 
2227 	if (access_flags & IB_ACCESS_REMOTE_READ)
2228 		hw_access_flags |= MLX5_QP_BIT_RRE;
2229 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC) {
2230 		int atomic_mode;
2231 
2232 		atomic_mode = get_atomic_mode(dev, qp->ibqp.qp_type);
2233 		if (atomic_mode < 0)
2234 			return -EOPNOTSUPP;
2235 
2236 		hw_access_flags |= MLX5_QP_BIT_RAE;
2237 		hw_access_flags |= atomic_mode << MLX5_ATOMIC_MODE_OFF;
2238 	}
2239 
2240 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
2241 		hw_access_flags |= MLX5_QP_BIT_RWE;
2242 
2243 	*hw_access_flags_be = cpu_to_be32(hw_access_flags);
2244 
2245 	return 0;
2246 }
2247 
2248 enum {
2249 	MLX5_PATH_FLAG_FL	= 1 << 0,
2250 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
2251 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
2252 };
2253 
2254 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
2255 {
2256 	if (rate == IB_RATE_PORT_CURRENT) {
2257 		return 0;
2258 	} else if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_600_GBPS) {
2259 		return -EINVAL;
2260 	} else {
2261 		while (rate != IB_RATE_2_5_GBPS &&
2262 		       !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
2263 			 MLX5_CAP_GEN(dev->mdev, stat_rate_support)))
2264 			--rate;
2265 	}
2266 
2267 	return rate + MLX5_STAT_RATE_OFFSET;
2268 }
2269 
2270 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
2271 				      struct mlx5_ib_sq *sq, u8 sl,
2272 				      struct ib_pd *pd)
2273 {
2274 	void *in;
2275 	void *tisc;
2276 	int inlen;
2277 	int err;
2278 
2279 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2280 	in = mlx5_vzalloc(inlen);
2281 	if (!in)
2282 		return -ENOMEM;
2283 
2284 	MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
2285 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
2286 
2287 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2288 	MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
2289 
2290 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2291 
2292 	kvfree(in);
2293 
2294 	return err;
2295 }
2296 
2297 static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,
2298 					 struct mlx5_ib_sq *sq, u8 tx_affinity,
2299 					 struct ib_pd *pd)
2300 {
2301 	void *in;
2302 	void *tisc;
2303 	int inlen;
2304 	int err;
2305 
2306 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2307 	in = mlx5_vzalloc(inlen);
2308 	if (!in)
2309 		return -ENOMEM;
2310 
2311 	MLX5_SET(modify_tis_in, in, bitmask.lag_tx_port_affinity, 1);
2312 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
2313 
2314 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2315 	MLX5_SET(tisc, tisc, lag_tx_port_affinity, tx_affinity);
2316 
2317 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2318 
2319 	kvfree(in);
2320 
2321 	return err;
2322 }
2323 
2324 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2325 			 const struct ib_ah_attr *ah,
2326 			 struct mlx5_qp_path *path, u8 port, int attr_mask,
2327 			 u32 path_flags, const struct ib_qp_attr *attr,
2328 			 bool alt)
2329 {
2330 	enum rdma_link_layer ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2331 	int err;
2332 	enum ib_gid_type gid_type;
2333 
2334 	if (attr_mask & IB_QP_PKEY_INDEX)
2335 		path->pkey_index = cpu_to_be16(alt ? attr->alt_pkey_index :
2336 						     attr->pkey_index);
2337 
2338 	if (ah->ah_flags & IB_AH_GRH) {
2339 		if (ah->grh.sgid_index >=
2340 		    dev->mdev->port_caps[port - 1].gid_table_len) {
2341 			pr_err("sgid_index (%u) too large. max is %d\n",
2342 			       ah->grh.sgid_index,
2343 			       dev->mdev->port_caps[port - 1].gid_table_len);
2344 			return -EINVAL;
2345 		}
2346 	}
2347 
2348 	if (ll == IB_LINK_LAYER_ETHERNET) {
2349 		if (!(ah->ah_flags & IB_AH_GRH))
2350 			return -EINVAL;
2351 		err = mlx5_get_roce_gid_type(dev, port, ah->grh.sgid_index,
2352 					     &gid_type);
2353 		if (err)
2354 			return err;
2355 		memcpy(path->rmac, ah->dmac, sizeof(ah->dmac));
2356 		path->udp_sport = mlx5_get_roce_udp_sport(dev, port,
2357 							  ah->grh.sgid_index);
2358 		path->dci_cfi_prio_sl = (ah->sl & 0x7) << 4;
2359 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
2360 			path->ecn_dscp = (ah->grh.traffic_class >> 2) & 0x3f;
2361 	} else {
2362 		path->fl_free_ar = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
2363 		path->fl_free_ar |=
2364 			(path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x40 : 0;
2365 		path->rlid = cpu_to_be16(ah->dlid);
2366 		path->grh_mlid = ah->src_path_bits & 0x7f;
2367 		if (ah->ah_flags & IB_AH_GRH)
2368 			path->grh_mlid	|= 1 << 7;
2369 		path->dci_cfi_prio_sl = ah->sl & 0xf;
2370 	}
2371 
2372 	if (ah->ah_flags & IB_AH_GRH) {
2373 		path->mgid_index = ah->grh.sgid_index;
2374 		path->hop_limit  = ah->grh.hop_limit;
2375 		path->tclass_flowlabel =
2376 			cpu_to_be32((ah->grh.traffic_class << 20) |
2377 				    (ah->grh.flow_label));
2378 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
2379 	}
2380 
2381 	err = ib_rate_to_mlx5(dev, ah->static_rate);
2382 	if (err < 0)
2383 		return err;
2384 	path->static_rate = err;
2385 	path->port = port;
2386 
2387 	if (attr_mask & IB_QP_TIMEOUT)
2388 		path->ackto_lt = (alt ? attr->alt_timeout : attr->timeout) << 3;
2389 
2390 	if ((qp->ibqp.qp_type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
2391 		return modify_raw_packet_eth_prio(dev->mdev,
2392 						  &qp->raw_packet_qp.sq,
2393 						  ah->sl & 0xf, qp->ibqp.pd);
2394 
2395 	return 0;
2396 }
2397 
2398 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
2399 	[MLX5_QP_STATE_INIT] = {
2400 		[MLX5_QP_STATE_INIT] = {
2401 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
2402 					  MLX5_QP_OPTPAR_RAE		|
2403 					  MLX5_QP_OPTPAR_RWE		|
2404 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
2405 					  MLX5_QP_OPTPAR_PRI_PORT,
2406 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
2407 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
2408 					  MLX5_QP_OPTPAR_PRI_PORT,
2409 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
2410 					  MLX5_QP_OPTPAR_Q_KEY		|
2411 					  MLX5_QP_OPTPAR_PRI_PORT,
2412 		},
2413 		[MLX5_QP_STATE_RTR] = {
2414 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
2415 					  MLX5_QP_OPTPAR_RRE            |
2416 					  MLX5_QP_OPTPAR_RAE            |
2417 					  MLX5_QP_OPTPAR_RWE            |
2418 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2419 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
2420 					  MLX5_QP_OPTPAR_RWE            |
2421 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2422 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
2423 					  MLX5_QP_OPTPAR_Q_KEY,
2424 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
2425 					   MLX5_QP_OPTPAR_Q_KEY,
2426 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2427 					  MLX5_QP_OPTPAR_RRE            |
2428 					  MLX5_QP_OPTPAR_RAE            |
2429 					  MLX5_QP_OPTPAR_RWE            |
2430 					  MLX5_QP_OPTPAR_PKEY_INDEX,
2431 		},
2432 	},
2433 	[MLX5_QP_STATE_RTR] = {
2434 		[MLX5_QP_STATE_RTS] = {
2435 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
2436 					  MLX5_QP_OPTPAR_RRE		|
2437 					  MLX5_QP_OPTPAR_RAE		|
2438 					  MLX5_QP_OPTPAR_RWE		|
2439 					  MLX5_QP_OPTPAR_PM_STATE	|
2440 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
2441 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
2442 					  MLX5_QP_OPTPAR_RWE		|
2443 					  MLX5_QP_OPTPAR_PM_STATE,
2444 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
2445 		},
2446 	},
2447 	[MLX5_QP_STATE_RTS] = {
2448 		[MLX5_QP_STATE_RTS] = {
2449 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
2450 					  MLX5_QP_OPTPAR_RAE		|
2451 					  MLX5_QP_OPTPAR_RWE		|
2452 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
2453 					  MLX5_QP_OPTPAR_PM_STATE	|
2454 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
2455 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
2456 					  MLX5_QP_OPTPAR_PM_STATE	|
2457 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
2458 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
2459 					  MLX5_QP_OPTPAR_SRQN		|
2460 					  MLX5_QP_OPTPAR_CQN_RCV,
2461 		},
2462 	},
2463 	[MLX5_QP_STATE_SQER] = {
2464 		[MLX5_QP_STATE_RTS] = {
2465 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
2466 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
2467 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
2468 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
2469 					   MLX5_QP_OPTPAR_RWE		|
2470 					   MLX5_QP_OPTPAR_RAE		|
2471 					   MLX5_QP_OPTPAR_RRE,
2472 		},
2473 	},
2474 };
2475 
2476 static int ib_nr_to_mlx5_nr(int ib_mask)
2477 {
2478 	switch (ib_mask) {
2479 	case IB_QP_STATE:
2480 		return 0;
2481 	case IB_QP_CUR_STATE:
2482 		return 0;
2483 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
2484 		return 0;
2485 	case IB_QP_ACCESS_FLAGS:
2486 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
2487 			MLX5_QP_OPTPAR_RAE;
2488 	case IB_QP_PKEY_INDEX:
2489 		return MLX5_QP_OPTPAR_PKEY_INDEX;
2490 	case IB_QP_PORT:
2491 		return MLX5_QP_OPTPAR_PRI_PORT;
2492 	case IB_QP_QKEY:
2493 		return MLX5_QP_OPTPAR_Q_KEY;
2494 	case IB_QP_AV:
2495 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
2496 			MLX5_QP_OPTPAR_PRI_PORT;
2497 	case IB_QP_PATH_MTU:
2498 		return 0;
2499 	case IB_QP_TIMEOUT:
2500 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
2501 	case IB_QP_RETRY_CNT:
2502 		return MLX5_QP_OPTPAR_RETRY_COUNT;
2503 	case IB_QP_RNR_RETRY:
2504 		return MLX5_QP_OPTPAR_RNR_RETRY;
2505 	case IB_QP_RQ_PSN:
2506 		return 0;
2507 	case IB_QP_MAX_QP_RD_ATOMIC:
2508 		return MLX5_QP_OPTPAR_SRA_MAX;
2509 	case IB_QP_ALT_PATH:
2510 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
2511 	case IB_QP_MIN_RNR_TIMER:
2512 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
2513 	case IB_QP_SQ_PSN:
2514 		return 0;
2515 	case IB_QP_MAX_DEST_RD_ATOMIC:
2516 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
2517 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
2518 	case IB_QP_PATH_MIG_STATE:
2519 		return MLX5_QP_OPTPAR_PM_STATE;
2520 	case IB_QP_CAP:
2521 		return 0;
2522 	case IB_QP_DEST_QPN:
2523 		return 0;
2524 	}
2525 	return 0;
2526 }
2527 
2528 static int ib_mask_to_mlx5_opt(int ib_mask)
2529 {
2530 	int result = 0;
2531 	int i;
2532 
2533 	for (i = 0; i < 8 * sizeof(int); i++) {
2534 		if ((1 << i) & ib_mask)
2535 			result |= ib_nr_to_mlx5_nr(1 << i);
2536 	}
2537 
2538 	return result;
2539 }
2540 
2541 static int modify_raw_packet_qp_rq(
2542 	struct mlx5_ib_dev *dev, struct mlx5_ib_rq *rq, int new_state,
2543 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
2544 {
2545 	void *in;
2546 	void *rqc;
2547 	int inlen;
2548 	int err;
2549 
2550 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
2551 	in = mlx5_vzalloc(inlen);
2552 	if (!in)
2553 		return -ENOMEM;
2554 
2555 	MLX5_SET(modify_rq_in, in, rqn, rq->base.mqp.qpn);
2556 	MLX5_SET(modify_rq_in, in, rq_state, rq->state);
2557 	MLX5_SET(modify_rq_in, in, uid, to_mpd(pd)->uid);
2558 
2559 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
2560 	MLX5_SET(rqc, rqc, state, new_state);
2561 
2562 	if (raw_qp_param->set_mask & MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID) {
2563 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counters_set_id)) {
2564 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
2565 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_MODIFY_RQ_COUNTER_SET_ID);
2566 			MLX5_SET(rqc, rqc, counter_set_id, raw_qp_param->rq_q_ctr_id);
2567 		} else
2568 			pr_info_once("%s: RAW PACKET QP counters are not supported on current FW\n",
2569 				     dev->ib_dev.name);
2570 	}
2571 
2572 	err = mlx5_core_modify_rq(dev->mdev, in, inlen);
2573 	if (err)
2574 		goto out;
2575 
2576 	rq->state = new_state;
2577 
2578 out:
2579 	kvfree(in);
2580 	return err;
2581 }
2582 
2583 static int modify_raw_packet_qp_sq(struct mlx5_core_dev *dev,
2584 				   struct mlx5_ib_sq *sq, int new_state,
2585 				   struct ib_pd *pd)
2586 {
2587 	void *in;
2588 	void *sqc;
2589 	int inlen;
2590 	int err;
2591 
2592 	inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
2593 	in = mlx5_vzalloc(inlen);
2594 	if (!in)
2595 		return -ENOMEM;
2596 
2597 	MLX5_SET(modify_sq_in, in, sqn, sq->base.mqp.qpn);
2598 	MLX5_SET(modify_sq_in, in, uid, to_mpd(pd)->uid);
2599 	MLX5_SET(modify_sq_in, in, sq_state, sq->state);
2600 
2601 	sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
2602 	MLX5_SET(sqc, sqc, state, new_state);
2603 
2604 	err = mlx5_core_modify_sq(dev, in, inlen);
2605 	if (err)
2606 		goto out;
2607 
2608 	sq->state = new_state;
2609 
2610 out:
2611 	kvfree(in);
2612 	return err;
2613 }
2614 
2615 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2616 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
2617 				u8 tx_affinity)
2618 {
2619 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
2620 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
2621 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
2622 	int modify_rq = !!qp->rq.wqe_cnt;
2623 	int modify_sq = !!qp->sq.wqe_cnt;
2624 	int rq_state;
2625 	int sq_state;
2626 	int err;
2627 
2628 	switch (raw_qp_param->operation) {
2629 	case MLX5_CMD_OP_RST2INIT_QP:
2630 		rq_state = MLX5_RQC_STATE_RDY;
2631 		sq_state = MLX5_SQC_STATE_RDY;
2632 		break;
2633 	case MLX5_CMD_OP_2ERR_QP:
2634 		rq_state = MLX5_RQC_STATE_ERR;
2635 		sq_state = MLX5_SQC_STATE_ERR;
2636 		break;
2637 	case MLX5_CMD_OP_2RST_QP:
2638 		rq_state = MLX5_RQC_STATE_RST;
2639 		sq_state = MLX5_SQC_STATE_RST;
2640 		break;
2641 	case MLX5_CMD_OP_RTR2RTS_QP:
2642 	case MLX5_CMD_OP_RTS2RTS_QP:
2643 		return raw_qp_param->set_mask ? -EINVAL : 0;
2644 	case MLX5_CMD_OP_INIT2INIT_QP:
2645 	case MLX5_CMD_OP_INIT2RTR_QP:
2646 		if (raw_qp_param->set_mask)
2647 			return -EINVAL;
2648 		else
2649 			return 0;
2650 	default:
2651 		WARN_ON(1);
2652 		return -EINVAL;
2653 	}
2654 
2655 	if (modify_rq) {
2656 		err =  modify_raw_packet_qp_rq(dev, rq, rq_state, raw_qp_param,
2657 					       qp->ibqp.pd);
2658 		if (err)
2659 			return err;
2660 	}
2661 
2662 	if (modify_sq) {
2663 		if (tx_affinity) {
2664 			err = modify_raw_packet_tx_affinity(dev->mdev, sq,
2665 							    tx_affinity,
2666 							    qp->ibqp.pd);
2667 			if (err)
2668 				return err;
2669 		}
2670 
2671 		return modify_raw_packet_qp_sq(dev->mdev, sq, sq_state, qp->ibqp.pd);
2672 	}
2673 
2674 	return 0;
2675 }
2676 
2677 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
2678 			       const struct ib_qp_attr *attr, int attr_mask,
2679 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
2680 {
2681 	static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
2682 		[MLX5_QP_STATE_RST] = {
2683 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2684 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2685 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_RST2INIT_QP,
2686 		},
2687 		[MLX5_QP_STATE_INIT]  = {
2688 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2689 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2690 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_INIT2INIT_QP,
2691 			[MLX5_QP_STATE_RTR]	= MLX5_CMD_OP_INIT2RTR_QP,
2692 		},
2693 		[MLX5_QP_STATE_RTR]   = {
2694 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2695 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2696 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTR2RTS_QP,
2697 		},
2698 		[MLX5_QP_STATE_RTS]   = {
2699 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2700 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2701 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTS2RTS_QP,
2702 		},
2703 		[MLX5_QP_STATE_SQD] = {
2704 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2705 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2706 		},
2707 		[MLX5_QP_STATE_SQER] = {
2708 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2709 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2710 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQERR2RTS_QP,
2711 		},
2712 		[MLX5_QP_STATE_ERR] = {
2713 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
2714 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
2715 		}
2716 	};
2717 
2718 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2719 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2720 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
2721 	struct mlx5_ib_cq *send_cq, *recv_cq;
2722 	struct mlx5_qp_context *context;
2723 	struct mlx5_ib_pd *pd;
2724 	struct mlx5_ib_port *mibport = NULL;
2725 	enum mlx5_qp_state mlx5_cur, mlx5_new;
2726 	enum mlx5_qp_optpar optpar;
2727 	int sqd_event;
2728 	int mlx5_st;
2729 	int err;
2730 	u16 op;
2731 
2732 	context = kzalloc(sizeof(*context), GFP_KERNEL);
2733 	if (!context)
2734 		return -ENOMEM;
2735 
2736 	err = to_mlx5_st(ibqp->qp_type);
2737 	if (err < 0) {
2738 		mlx5_ib_dbg(dev, "unsupported qp type %d\n", ibqp->qp_type);
2739 		goto out;
2740 	}
2741 
2742 	context->flags = cpu_to_be32(err << 16);
2743 
2744 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
2745 		context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2746 	} else {
2747 		switch (attr->path_mig_state) {
2748 		case IB_MIG_MIGRATED:
2749 			context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2750 			break;
2751 		case IB_MIG_REARM:
2752 			context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
2753 			break;
2754 		case IB_MIG_ARMED:
2755 			context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
2756 			break;
2757 		}
2758 	}
2759 
2760 	if (is_sqp(ibqp->qp_type)) {
2761 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
2762 	} else if (ibqp->qp_type == IB_QPT_UD ||
2763 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
2764 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
2765 	} else if (attr_mask & IB_QP_PATH_MTU) {
2766 		if (attr->path_mtu < IB_MTU_256 ||
2767 		    attr->path_mtu > IB_MTU_4096) {
2768 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
2769 			err = -EINVAL;
2770 			goto out;
2771 		}
2772 		context->mtu_msgmax = (attr->path_mtu << 5) |
2773 				      (u8)MLX5_CAP_GEN(dev->mdev, log_max_msg);
2774 	}
2775 
2776 	if (attr_mask & IB_QP_DEST_QPN)
2777 		context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
2778 
2779 	if (attr_mask & IB_QP_PKEY_INDEX)
2780 		context->pri_path.pkey_index = cpu_to_be16(attr->pkey_index);
2781 
2782 	/* todo implement counter_index functionality */
2783 
2784 	if (is_sqp(ibqp->qp_type))
2785 		context->pri_path.port = qp->port;
2786 
2787 	if (attr_mask & IB_QP_PORT)
2788 		context->pri_path.port = attr->port_num;
2789 
2790 	if (attr_mask & IB_QP_AV) {
2791 		err = mlx5_set_path(dev, qp, &attr->ah_attr, &context->pri_path,
2792 				    attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
2793 				    attr_mask, 0, attr, false);
2794 		if (err)
2795 			goto out;
2796 	}
2797 
2798 	if (attr_mask & IB_QP_TIMEOUT)
2799 		context->pri_path.ackto_lt |= attr->timeout << 3;
2800 
2801 	if (attr_mask & IB_QP_ALT_PATH) {
2802 		err = mlx5_set_path(dev, qp, &attr->alt_ah_attr,
2803 				    &context->alt_path,
2804 				    attr->alt_port_num,
2805 				    attr_mask | IB_QP_PKEY_INDEX | IB_QP_TIMEOUT,
2806 				    0, attr, true);
2807 		if (err)
2808 			goto out;
2809 	}
2810 
2811 	pd = get_pd(qp);
2812 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
2813 		&send_cq, &recv_cq);
2814 
2815 	context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
2816 	context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
2817 	context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
2818 	context->params1  = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
2819 
2820 	if (attr_mask & IB_QP_RNR_RETRY)
2821 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2822 
2823 	if (attr_mask & IB_QP_RETRY_CNT)
2824 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2825 
2826 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2827 		if (attr->max_rd_atomic)
2828 			context->params1 |=
2829 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2830 	}
2831 
2832 	if (attr_mask & IB_QP_SQ_PSN)
2833 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2834 
2835 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2836 		if (attr->max_dest_rd_atomic)
2837 			context->params2 |=
2838 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2839 	}
2840 
2841 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
2842 		__be32 access_flags;
2843 
2844 		err = to_mlx5_access_flags(qp, attr, attr_mask, &access_flags);
2845 		if (err)
2846 			goto out;
2847 
2848 		context->params2 |= access_flags;
2849 	}
2850 
2851 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
2852 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2853 
2854 	if (attr_mask & IB_QP_RQ_PSN)
2855 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2856 
2857 	if (attr_mask & IB_QP_QKEY)
2858 		context->qkey = cpu_to_be32(attr->qkey);
2859 
2860 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2861 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2862 
2863 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2864 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2865 		sqd_event = 1;
2866 	else
2867 		sqd_event = 0;
2868 
2869 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2870 		u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
2871 			       qp->port) - 1;
2872 		mibport = &dev->port[port_num];
2873 		context->qp_counter_set_usr_page |=
2874 			cpu_to_be32((u32)(mibport->q_cnt_id) << 24);
2875 	}
2876 
2877 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2878 		context->sq_crq_size |= cpu_to_be16(1 << 4);
2879 
2880 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
2881 		context->deth_sqpn = cpu_to_be32(1);
2882 
2883 	mlx5_cur = to_mlx5_state(cur_state);
2884 	mlx5_new = to_mlx5_state(new_state);
2885 	mlx5_st = to_mlx5_st(ibqp->qp_type);
2886 	if (mlx5_st < 0)
2887 		goto out;
2888 
2889 	/* If moving to a reset or error state, we must disable page faults on
2890 	 * this QP and flush all current page faults. Otherwise a stale page
2891 	 * fault may attempt to work on this QP after it is reset and moved
2892 	 * again to RTS, and may cause the driver and the device to get out of
2893 	 * sync. */
2894 	if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2895 	    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR) &&
2896 	    (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
2897 		mlx5_ib_qp_disable_pagefaults(qp);
2898 
2899 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
2900 	    !optab[mlx5_cur][mlx5_new])
2901 		goto out;
2902 
2903 	op = optab[mlx5_cur][mlx5_new];
2904 	optpar = ib_mask_to_mlx5_opt(attr_mask);
2905 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
2906 
2907 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
2908 		struct mlx5_modify_raw_qp_param raw_qp_param = {};
2909 
2910 		raw_qp_param.operation = op;
2911 		if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2912 			raw_qp_param.rq_q_ctr_id = mibport->q_cnt_id;
2913 			raw_qp_param.set_mask |= MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID;
2914 		}
2915 		err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
2916 	} else {
2917 		err = mlx5_core_qp_modify(dev->mdev, op, optpar, context,
2918 					  &base->mqp);
2919 	}
2920 
2921 	if (err)
2922 		goto out;
2923 
2924 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT &&
2925 	    (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
2926 		mlx5_ib_qp_enable_pagefaults(qp);
2927 
2928 	qp->state = new_state;
2929 
2930 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2931 		qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
2932 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2933 		qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
2934 	if (attr_mask & IB_QP_PORT)
2935 		qp->port = attr->port_num;
2936 	if (attr_mask & IB_QP_ALT_PATH)
2937 		qp->trans_qp.alt_port = attr->alt_port_num;
2938 
2939 	/*
2940 	 * If we moved a kernel QP to RESET, clean up all old CQ
2941 	 * entries and reinitialize the QP.
2942 	 */
2943 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
2944 		mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2945 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
2946 		if (send_cq != recv_cq)
2947 			mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
2948 
2949 		qp->rq.head = 0;
2950 		qp->rq.tail = 0;
2951 		qp->sq.head = 0;
2952 		qp->sq.tail = 0;
2953 		qp->sq.cur_post = 0;
2954 		qp->sq.last_poll = 0;
2955 		qp->db.db[MLX5_RCV_DBR] = 0;
2956 		qp->db.db[MLX5_SND_DBR] = 0;
2957 	}
2958 
2959 out:
2960 	kfree(context);
2961 	return err;
2962 }
2963 
2964 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2965 		      int attr_mask, struct ib_udata *udata)
2966 {
2967 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2968 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
2969 	enum ib_qp_type qp_type;
2970 	enum ib_qp_state cur_state, new_state;
2971 	int err = -EINVAL;
2972 	int port;
2973 
2974 	if (ibqp->rwq_ind_tbl)
2975 		return -ENOSYS;
2976 
2977 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
2978 		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
2979 
2980 	qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ?
2981 		IB_QPT_GSI : ibqp->qp_type;
2982 
2983 	mutex_lock(&qp->mutex);
2984 
2985 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2986 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2987 
2988 	if (qp_type != MLX5_IB_QPT_REG_UMR &&
2989 	    !ib_modify_qp_is_ok(cur_state, new_state, qp_type, attr_mask)) {
2990 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
2991 			    cur_state, new_state, ibqp->qp_type, attr_mask);
2992 		goto out;
2993 	}
2994 
2995 	if ((attr_mask & IB_QP_PORT) &&
2996 	    (attr->port_num == 0 ||
2997 	     attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports))) {
2998 		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
2999 			    attr->port_num, dev->num_ports);
3000 		goto out;
3001 	}
3002 
3003 	if (attr_mask & IB_QP_PKEY_INDEX) {
3004 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
3005 		if (attr->pkey_index >=
3006 		    dev->mdev->port_caps[port - 1].pkey_table_len) {
3007 			mlx5_ib_dbg(dev, "invalid pkey index %d\n",
3008 				    attr->pkey_index);
3009 			goto out;
3010 		}
3011 	}
3012 
3013 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
3014 	    attr->max_rd_atomic >
3015 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
3016 		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
3017 			    attr->max_rd_atomic);
3018 		goto out;
3019 	}
3020 
3021 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
3022 	    attr->max_dest_rd_atomic >
3023 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
3024 		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
3025 			    attr->max_dest_rd_atomic);
3026 		goto out;
3027 	}
3028 
3029 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
3030 		err = 0;
3031 		goto out;
3032 	}
3033 
3034 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
3035 
3036 out:
3037 	mutex_unlock(&qp->mutex);
3038 	return err;
3039 }
3040 
3041 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3042 {
3043 	struct mlx5_ib_cq *cq;
3044 	unsigned cur;
3045 
3046 	cur = wq->head - wq->tail;
3047 	if (likely(cur + nreq < wq->max_post))
3048 		return 0;
3049 
3050 	cq = to_mcq(ib_cq);
3051 	spin_lock(&cq->lock);
3052 	cur = wq->head - wq->tail;
3053 	spin_unlock(&cq->lock);
3054 
3055 	return cur + nreq >= wq->max_post;
3056 }
3057 
3058 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
3059 					  u64 remote_addr, u32 rkey)
3060 {
3061 	rseg->raddr    = cpu_to_be64(remote_addr);
3062 	rseg->rkey     = cpu_to_be32(rkey);
3063 	rseg->reserved = 0;
3064 }
3065 
3066 static void *set_eth_seg(struct mlx5_wqe_eth_seg *eseg,
3067 			 const struct ib_send_wr *wr, void *qend,
3068 			 struct mlx5_ib_qp *qp, int *size)
3069 {
3070 	void *seg = eseg;
3071 
3072 	memset(eseg, 0, sizeof(struct mlx5_wqe_eth_seg));
3073 
3074 	if (wr->send_flags & IB_SEND_IP_CSUM)
3075 		eseg->cs_flags = MLX5_ETH_WQE_L3_CSUM |
3076 				 MLX5_ETH_WQE_L4_CSUM;
3077 
3078 	seg += sizeof(struct mlx5_wqe_eth_seg);
3079 	*size += sizeof(struct mlx5_wqe_eth_seg) / 16;
3080 
3081 	if (wr->opcode == IB_WR_LSO) {
3082 		struct ib_ud_wr *ud_wr = container_of(wr, struct ib_ud_wr, wr);
3083 		int size_of_inl_hdr_start = sizeof(eseg->inline_hdr_start);
3084 		u64 left, leftlen, copysz;
3085 		void *pdata = ud_wr->header;
3086 
3087 		left = ud_wr->hlen;
3088 		eseg->mss = cpu_to_be16(ud_wr->mss);
3089 		eseg->inline_hdr_sz = cpu_to_be16(left);
3090 
3091 		/*
3092 		 * check if there is space till the end of queue, if yes,
3093 		 * copy all in one shot, otherwise copy till the end of queue,
3094 		 * rollback and than the copy the left
3095 		 */
3096 		leftlen = qend - (void *)eseg->inline_hdr_start;
3097 		copysz = min_t(u64, leftlen, left);
3098 
3099 		memcpy(seg - size_of_inl_hdr_start, pdata, copysz);
3100 
3101 		if (likely(copysz > size_of_inl_hdr_start)) {
3102 			seg += ALIGN(copysz - size_of_inl_hdr_start, 16);
3103 			*size += ALIGN(copysz - size_of_inl_hdr_start, 16) / 16;
3104 		}
3105 
3106 		if (unlikely(copysz < left)) { /* the last wqe in the queue */
3107 			seg = mlx5_get_send_wqe(qp, 0);
3108 			left -= copysz;
3109 			pdata += copysz;
3110 			memcpy(seg, pdata, left);
3111 			seg += ALIGN(left, 16);
3112 			*size += ALIGN(left, 16) / 16;
3113 		}
3114 	}
3115 
3116 	return seg;
3117 }
3118 
3119 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
3120 			     const struct ib_send_wr *wr)
3121 {
3122 	memcpy(&dseg->av, &to_mah(ud_wr(wr)->ah)->av, sizeof(struct mlx5_av));
3123 	dseg->av.dqp_dct = cpu_to_be32(ud_wr(wr)->remote_qpn | MLX5_EXTENDED_UD_AV);
3124 	dseg->av.key.qkey.qkey = cpu_to_be32(ud_wr(wr)->remote_qkey);
3125 }
3126 
3127 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
3128 {
3129 	dseg->byte_count = cpu_to_be32(sg->length);
3130 	dseg->lkey       = cpu_to_be32(sg->lkey);
3131 	dseg->addr       = cpu_to_be64(sg->addr);
3132 }
3133 
3134 static __be16 get_klm_octo(int npages)
3135 {
3136 	return cpu_to_be16(ALIGN(npages, 8) / 2);
3137 }
3138 
3139 static __be64 frwr_mkey_mask(void)
3140 {
3141 	u64 result;
3142 
3143 	result = MLX5_MKEY_MASK_LEN		|
3144 		MLX5_MKEY_MASK_PAGE_SIZE	|
3145 		MLX5_MKEY_MASK_START_ADDR	|
3146 		MLX5_MKEY_MASK_EN_RINVAL	|
3147 		MLX5_MKEY_MASK_KEY		|
3148 		MLX5_MKEY_MASK_LR		|
3149 		MLX5_MKEY_MASK_LW		|
3150 		MLX5_MKEY_MASK_RR		|
3151 		MLX5_MKEY_MASK_RW		|
3152 		MLX5_MKEY_MASK_A		|
3153 		MLX5_MKEY_MASK_SMALL_FENCE	|
3154 		MLX5_MKEY_MASK_FREE;
3155 
3156 	return cpu_to_be64(result);
3157 }
3158 
3159 static __be64 sig_mkey_mask(void)
3160 {
3161 	u64 result;
3162 
3163 	result = MLX5_MKEY_MASK_LEN		|
3164 		MLX5_MKEY_MASK_PAGE_SIZE	|
3165 		MLX5_MKEY_MASK_START_ADDR	|
3166 		MLX5_MKEY_MASK_EN_SIGERR	|
3167 		MLX5_MKEY_MASK_EN_RINVAL	|
3168 		MLX5_MKEY_MASK_KEY		|
3169 		MLX5_MKEY_MASK_LR		|
3170 		MLX5_MKEY_MASK_LW		|
3171 		MLX5_MKEY_MASK_RR		|
3172 		MLX5_MKEY_MASK_RW		|
3173 		MLX5_MKEY_MASK_SMALL_FENCE	|
3174 		MLX5_MKEY_MASK_FREE		|
3175 		MLX5_MKEY_MASK_BSF_EN;
3176 
3177 	return cpu_to_be64(result);
3178 }
3179 
3180 static void set_reg_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr,
3181 				struct mlx5_ib_mr *mr)
3182 {
3183 	int ndescs = mr->ndescs;
3184 
3185 	memset(umr, 0, sizeof(*umr));
3186 
3187 	if (mr->access_mode == MLX5_ACCESS_MODE_KLM)
3188 		/* KLMs take twice the size of MTTs */
3189 		ndescs *= 2;
3190 
3191 	umr->flags = MLX5_UMR_CHECK_NOT_FREE;
3192 	umr->klm_octowords = get_klm_octo(ndescs);
3193 	umr->mkey_mask = frwr_mkey_mask();
3194 }
3195 
3196 static void set_linv_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr)
3197 {
3198 	memset(umr, 0, sizeof(*umr));
3199 	umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
3200 	umr->flags = 1 << 7;
3201 }
3202 
3203 static __be64 get_umr_reg_mr_mask(void)
3204 {
3205 	u64 result;
3206 
3207 	result = MLX5_MKEY_MASK_LEN		|
3208 		 MLX5_MKEY_MASK_PAGE_SIZE	|
3209 		 MLX5_MKEY_MASK_START_ADDR	|
3210 		 MLX5_MKEY_MASK_PD		|
3211 		 MLX5_MKEY_MASK_LR		|
3212 		 MLX5_MKEY_MASK_LW		|
3213 		 MLX5_MKEY_MASK_KEY		|
3214 		 MLX5_MKEY_MASK_RR		|
3215 		 MLX5_MKEY_MASK_RW		|
3216 		 MLX5_MKEY_MASK_A		|
3217 		 MLX5_MKEY_MASK_FREE;
3218 
3219 	return cpu_to_be64(result);
3220 }
3221 
3222 static __be64 get_umr_unreg_mr_mask(void)
3223 {
3224 	u64 result;
3225 
3226 	result = MLX5_MKEY_MASK_FREE;
3227 
3228 	return cpu_to_be64(result);
3229 }
3230 
3231 static __be64 get_umr_update_mtt_mask(void)
3232 {
3233 	u64 result;
3234 
3235 	result = MLX5_MKEY_MASK_FREE;
3236 
3237 	return cpu_to_be64(result);
3238 }
3239 
3240 static __be64 get_umr_update_translation_mask(void)
3241 {
3242 	u64 result;
3243 
3244 	result = MLX5_MKEY_MASK_LEN |
3245 		 MLX5_MKEY_MASK_PAGE_SIZE |
3246 		 MLX5_MKEY_MASK_START_ADDR |
3247 		 MLX5_MKEY_MASK_KEY |
3248 		 MLX5_MKEY_MASK_FREE;
3249 
3250 	return cpu_to_be64(result);
3251 }
3252 
3253 static __be64 get_umr_update_access_mask(void)
3254 {
3255 	u64 result;
3256 
3257 	result = MLX5_MKEY_MASK_LW |
3258 		 MLX5_MKEY_MASK_RR |
3259 		 MLX5_MKEY_MASK_RW |
3260 		 MLX5_MKEY_MASK_A |
3261 		 MLX5_MKEY_MASK_KEY |
3262 		 MLX5_MKEY_MASK_FREE;
3263 
3264 	return cpu_to_be64(result);
3265 }
3266 
3267 static __be64 get_umr_update_pd_mask(void)
3268 {
3269 	u64 result;
3270 
3271 	result = MLX5_MKEY_MASK_PD |
3272 		 MLX5_MKEY_MASK_KEY |
3273 		 MLX5_MKEY_MASK_FREE;
3274 
3275 	return cpu_to_be64(result);
3276 }
3277 
3278 static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
3279 				const struct ib_send_wr *wr)
3280 {
3281 	const struct mlx5_umr_wr *umrwr = umr_wr(wr);
3282 
3283 	memset(umr, 0, sizeof(*umr));
3284 
3285 	if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
3286 		umr->flags = MLX5_UMR_CHECK_FREE; /* fail if free */
3287 	else
3288 		umr->flags = MLX5_UMR_CHECK_NOT_FREE; /* fail if not free */
3289 
3290 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UNREG)) {
3291 		umr->klm_octowords = get_klm_octo(umrwr->npages);
3292 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT) {
3293 			umr->mkey_mask = get_umr_update_mtt_mask();
3294 			umr->bsf_octowords = get_klm_octo(umrwr->target.offset);
3295 			umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
3296 		}
3297 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION)
3298 			umr->mkey_mask |= get_umr_update_translation_mask();
3299 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_ACCESS)
3300 			umr->mkey_mask |= get_umr_update_access_mask();
3301 		if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_PD)
3302 			umr->mkey_mask |= get_umr_update_pd_mask();
3303 		if (!umr->mkey_mask)
3304 			umr->mkey_mask = get_umr_reg_mr_mask();
3305 	} else {
3306 		umr->mkey_mask = get_umr_unreg_mr_mask();
3307 	}
3308 
3309 	if (!wr->num_sge)
3310 		umr->flags |= MLX5_UMR_INLINE;
3311 }
3312 
3313 static u8 get_umr_flags(int acc)
3314 {
3315 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
3316 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
3317 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
3318 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
3319 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
3320 }
3321 
3322 static void set_reg_mkey_seg(struct mlx5_mkey_seg *seg,
3323 			     struct mlx5_ib_mr *mr,
3324 			     u32 key, int access)
3325 {
3326 	int ndescs = ALIGN(mr->ndescs, 8) >> 1;
3327 
3328 	memset(seg, 0, sizeof(*seg));
3329 
3330 	if (mr->access_mode == MLX5_ACCESS_MODE_MTT)
3331 		seg->log2_page_size = ilog2(mr->ibmr.page_size);
3332 	else if (mr->access_mode == MLX5_ACCESS_MODE_KLM)
3333 		/* KLMs take twice the size of MTTs */
3334 		ndescs *= 2;
3335 
3336 	seg->flags = get_umr_flags(access) | mr->access_mode;
3337 	seg->qpn_mkey7_0 = cpu_to_be32((key & 0xff) | 0xffffff00);
3338 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
3339 	seg->start_addr = cpu_to_be64(mr->ibmr.iova);
3340 	seg->len = cpu_to_be64(mr->ibmr.length);
3341 	seg->xlt_oct_size = cpu_to_be32(ndescs);
3342 }
3343 
3344 static void set_linv_mkey_seg(struct mlx5_mkey_seg *seg)
3345 {
3346 	memset(seg, 0, sizeof(*seg));
3347 	seg->status = MLX5_MKEY_STATUS_FREE;
3348 }
3349 
3350 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, const struct ib_send_wr *wr)
3351 {
3352 	const struct mlx5_umr_wr *umrwr = umr_wr(wr);
3353 
3354 	memset(seg, 0, sizeof(*seg));
3355 	if (wr->send_flags & MLX5_IB_SEND_UMR_UNREG) {
3356 		seg->status = MLX5_MKEY_STATUS_FREE;
3357 		return;
3358 	}
3359 
3360 	seg->flags = convert_access(umrwr->access_flags);
3361 	if (!(wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT)) {
3362 		if (umrwr->pd)
3363 			seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
3364 		seg->start_addr = cpu_to_be64(umrwr->target.virt_addr);
3365 	}
3366 	seg->len = cpu_to_be64(umrwr->length);
3367 	seg->log2_page_size = umrwr->page_shift;
3368 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
3369 				       mlx5_mkey_variant(umrwr->mkey));
3370 }
3371 
3372 static void set_reg_data_seg(struct mlx5_wqe_data_seg *dseg,
3373 			     struct mlx5_ib_mr *mr,
3374 			     struct mlx5_ib_pd *pd)
3375 {
3376 	int bcount = mr->desc_size * mr->ndescs;
3377 
3378 	dseg->addr = cpu_to_be64(mr->desc_map);
3379 	dseg->byte_count = cpu_to_be32(ALIGN(bcount, 64));
3380 	dseg->lkey = cpu_to_be32(pd->ibpd.local_dma_lkey);
3381 }
3382 
3383 static __be32 send_ieth(const struct ib_send_wr *wr)
3384 {
3385 	switch (wr->opcode) {
3386 	case IB_WR_SEND_WITH_IMM:
3387 	case IB_WR_RDMA_WRITE_WITH_IMM:
3388 		return wr->ex.imm_data;
3389 
3390 	case IB_WR_SEND_WITH_INV:
3391 		return cpu_to_be32(wr->ex.invalidate_rkey);
3392 
3393 	default:
3394 		return 0;
3395 	}
3396 }
3397 
3398 static u8 calc_sig(void *wqe, int size)
3399 {
3400 	u8 *p = wqe;
3401 	u8 res = 0;
3402 	int i;
3403 
3404 	for (i = 0; i < size; i++)
3405 		res ^= p[i];
3406 
3407 	return ~res;
3408 }
3409 
3410 static u8 wq_sig(void *wqe)
3411 {
3412 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
3413 }
3414 
3415 static int set_data_inl_seg(struct mlx5_ib_qp *qp, const struct ib_send_wr *wr,
3416 			    void *wqe, int *sz)
3417 {
3418 	struct mlx5_wqe_inline_seg *seg;
3419 	void *qend = qp->sq.qend;
3420 	void *addr;
3421 	int inl = 0;
3422 	int copy;
3423 	int len;
3424 	int i;
3425 
3426 	seg = wqe;
3427 	wqe += sizeof(*seg);
3428 	for (i = 0; i < wr->num_sge; i++) {
3429 		addr = (void *)(unsigned long)(wr->sg_list[i].addr);
3430 		len  = wr->sg_list[i].length;
3431 		inl += len;
3432 
3433 		if (unlikely(inl > qp->max_inline_data))
3434 			return -ENOMEM;
3435 
3436 		if (unlikely(wqe + len > qend)) {
3437 			copy = qend - wqe;
3438 			memcpy(wqe, addr, copy);
3439 			addr += copy;
3440 			len -= copy;
3441 			wqe = mlx5_get_send_wqe(qp, 0);
3442 		}
3443 		memcpy(wqe, addr, len);
3444 		wqe += len;
3445 	}
3446 
3447 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
3448 
3449 	*sz = ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
3450 
3451 	return 0;
3452 }
3453 
3454 static u16 prot_field_size(enum ib_signature_type type)
3455 {
3456 	switch (type) {
3457 	case IB_SIG_TYPE_T10_DIF:
3458 		return MLX5_DIF_SIZE;
3459 	default:
3460 		return 0;
3461 	}
3462 }
3463 
3464 static u8 bs_selector(int block_size)
3465 {
3466 	switch (block_size) {
3467 	case 512:	    return 0x1;
3468 	case 520:	    return 0x2;
3469 	case 4096:	    return 0x3;
3470 	case 4160:	    return 0x4;
3471 	case 1073741824:    return 0x5;
3472 	default:	    return 0;
3473 	}
3474 }
3475 
3476 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
3477 			      struct mlx5_bsf_inl *inl)
3478 {
3479 	/* Valid inline section and allow BSF refresh */
3480 	inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
3481 				       MLX5_BSF_REFRESH_DIF);
3482 	inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
3483 	inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
3484 	/* repeating block */
3485 	inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
3486 	inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
3487 			MLX5_DIF_CRC : MLX5_DIF_IPCS;
3488 
3489 	if (domain->sig.dif.ref_remap)
3490 		inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
3491 
3492 	if (domain->sig.dif.app_escape) {
3493 		if (domain->sig.dif.ref_escape)
3494 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
3495 		else
3496 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
3497 	}
3498 
3499 	inl->dif_app_bitmask_check =
3500 		cpu_to_be16(domain->sig.dif.apptag_check_mask);
3501 }
3502 
3503 static int mlx5_set_bsf(struct ib_mr *sig_mr,
3504 			struct ib_sig_attrs *sig_attrs,
3505 			struct mlx5_bsf *bsf, u32 data_size)
3506 {
3507 	struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
3508 	struct mlx5_bsf_basic *basic = &bsf->basic;
3509 	struct ib_sig_domain *mem = &sig_attrs->mem;
3510 	struct ib_sig_domain *wire = &sig_attrs->wire;
3511 
3512 	memset(bsf, 0, sizeof(*bsf));
3513 
3514 	/* Basic + Extended + Inline */
3515 	basic->bsf_size_sbs = 1 << 7;
3516 	/* Input domain check byte mask */
3517 	basic->check_byte_mask = sig_attrs->check_mask;
3518 	basic->raw_data_size = cpu_to_be32(data_size);
3519 
3520 	/* Memory domain */
3521 	switch (sig_attrs->mem.sig_type) {
3522 	case IB_SIG_TYPE_NONE:
3523 		break;
3524 	case IB_SIG_TYPE_T10_DIF:
3525 		basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
3526 		basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
3527 		mlx5_fill_inl_bsf(mem, &bsf->m_inl);
3528 		break;
3529 	default:
3530 		return -EINVAL;
3531 	}
3532 
3533 	/* Wire domain */
3534 	switch (sig_attrs->wire.sig_type) {
3535 	case IB_SIG_TYPE_NONE:
3536 		break;
3537 	case IB_SIG_TYPE_T10_DIF:
3538 		if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
3539 		    mem->sig_type == wire->sig_type) {
3540 			/* Same block structure */
3541 			basic->bsf_size_sbs |= 1 << 4;
3542 			if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
3543 				basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
3544 			if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
3545 				basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
3546 			if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
3547 				basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
3548 		} else
3549 			basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
3550 
3551 		basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
3552 		mlx5_fill_inl_bsf(wire, &bsf->w_inl);
3553 		break;
3554 	default:
3555 		return -EINVAL;
3556 	}
3557 
3558 	return 0;
3559 }
3560 
3561 static int set_sig_data_segment(const struct ib_sig_handover_wr *wr,
3562 				struct mlx5_ib_qp *qp, void **seg, int *size)
3563 {
3564 	struct ib_sig_attrs *sig_attrs = wr->sig_attrs;
3565 	struct ib_mr *sig_mr = wr->sig_mr;
3566 	struct mlx5_bsf *bsf;
3567 	u32 data_len = wr->wr.sg_list->length;
3568 	u32 data_key = wr->wr.sg_list->lkey;
3569 	u64 data_va = wr->wr.sg_list->addr;
3570 	int ret;
3571 	int wqe_size;
3572 
3573 	if (!wr->prot ||
3574 	    (data_key == wr->prot->lkey &&
3575 	     data_va == wr->prot->addr &&
3576 	     data_len == wr->prot->length)) {
3577 		/**
3578 		 * Source domain doesn't contain signature information
3579 		 * or data and protection are interleaved in memory.
3580 		 * So need construct:
3581 		 *                  ------------------
3582 		 *                 |     data_klm     |
3583 		 *                  ------------------
3584 		 *                 |       BSF        |
3585 		 *                  ------------------
3586 		 **/
3587 		struct mlx5_klm *data_klm = *seg;
3588 
3589 		data_klm->bcount = cpu_to_be32(data_len);
3590 		data_klm->key = cpu_to_be32(data_key);
3591 		data_klm->va = cpu_to_be64(data_va);
3592 		wqe_size = ALIGN(sizeof(*data_klm), 64);
3593 	} else {
3594 		/**
3595 		 * Source domain contains signature information
3596 		 * So need construct a strided block format:
3597 		 *               ---------------------------
3598 		 *              |     stride_block_ctrl     |
3599 		 *               ---------------------------
3600 		 *              |          data_klm         |
3601 		 *               ---------------------------
3602 		 *              |          prot_klm         |
3603 		 *               ---------------------------
3604 		 *              |             BSF           |
3605 		 *               ---------------------------
3606 		 **/
3607 		struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
3608 		struct mlx5_stride_block_entry *data_sentry;
3609 		struct mlx5_stride_block_entry *prot_sentry;
3610 		u32 prot_key = wr->prot->lkey;
3611 		u64 prot_va = wr->prot->addr;
3612 		u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
3613 		int prot_size;
3614 
3615 		sblock_ctrl = *seg;
3616 		data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
3617 		prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
3618 
3619 		prot_size = prot_field_size(sig_attrs->mem.sig_type);
3620 		if (!prot_size) {
3621 			pr_err("Bad block size given: %u\n", block_size);
3622 			return -EINVAL;
3623 		}
3624 		sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
3625 							    prot_size);
3626 		sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
3627 		sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
3628 		sblock_ctrl->num_entries = cpu_to_be16(2);
3629 
3630 		data_sentry->bcount = cpu_to_be16(block_size);
3631 		data_sentry->key = cpu_to_be32(data_key);
3632 		data_sentry->va = cpu_to_be64(data_va);
3633 		data_sentry->stride = cpu_to_be16(block_size);
3634 
3635 		prot_sentry->bcount = cpu_to_be16(prot_size);
3636 		prot_sentry->key = cpu_to_be32(prot_key);
3637 		prot_sentry->va = cpu_to_be64(prot_va);
3638 		prot_sentry->stride = cpu_to_be16(prot_size);
3639 
3640 		wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
3641 				 sizeof(*prot_sentry), 64);
3642 	}
3643 
3644 	*seg += wqe_size;
3645 	*size += wqe_size / 16;
3646 	if (unlikely((*seg == qp->sq.qend)))
3647 		*seg = mlx5_get_send_wqe(qp, 0);
3648 
3649 	bsf = *seg;
3650 	ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
3651 	if (ret)
3652 		return -EINVAL;
3653 
3654 	*seg += sizeof(*bsf);
3655 	*size += sizeof(*bsf) / 16;
3656 	if (unlikely((*seg == qp->sq.qend)))
3657 		*seg = mlx5_get_send_wqe(qp, 0);
3658 
3659 	return 0;
3660 }
3661 
3662 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
3663 				 const struct ib_sig_handover_wr *wr, u32 nelements,
3664 				 u32 length, u32 pdn)
3665 {
3666 	struct ib_mr *sig_mr = wr->sig_mr;
3667 	u32 sig_key = sig_mr->rkey;
3668 	u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
3669 
3670 	memset(seg, 0, sizeof(*seg));
3671 
3672 	seg->flags = get_umr_flags(wr->access_flags) |
3673 				   MLX5_ACCESS_MODE_KLM;
3674 	seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
3675 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
3676 				    MLX5_MKEY_BSF_EN | pdn);
3677 	seg->len = cpu_to_be64(length);
3678 	seg->xlt_oct_size = cpu_to_be32(be16_to_cpu(get_klm_octo(nelements)));
3679 	seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
3680 }
3681 
3682 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
3683 				u32 nelements)
3684 {
3685 	memset(umr, 0, sizeof(*umr));
3686 
3687 	umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
3688 	umr->klm_octowords = get_klm_octo(nelements);
3689 	umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
3690 	umr->mkey_mask = sig_mkey_mask();
3691 }
3692 
3693 
3694 static int set_sig_umr_wr(const struct ib_send_wr *send_wr, struct mlx5_ib_qp *qp,
3695 			  void **seg, int *size)
3696 {
3697 	const struct ib_sig_handover_wr *wr = sig_handover_wr(send_wr);
3698 	struct mlx5_ib_mr *sig_mr = to_mmr(wr->sig_mr);
3699 	u32 pdn = get_pd(qp)->pdn;
3700 	u32 klm_oct_size;
3701 	int region_len, ret;
3702 
3703 	if (unlikely(wr->wr.num_sge != 1) ||
3704 	    unlikely(wr->access_flags & IB_ACCESS_REMOTE_ATOMIC) ||
3705 	    unlikely(!sig_mr->sig) || unlikely(!qp->signature_en) ||
3706 	    unlikely(!sig_mr->sig->sig_status_checked))
3707 		return -EINVAL;
3708 
3709 	/* length of the protected region, data + protection */
3710 	region_len = wr->wr.sg_list->length;
3711 	if (wr->prot &&
3712 	    (wr->prot->lkey != wr->wr.sg_list->lkey  ||
3713 	     wr->prot->addr != wr->wr.sg_list->addr  ||
3714 	     wr->prot->length != wr->wr.sg_list->length))
3715 		region_len += wr->prot->length;
3716 
3717 	/**
3718 	 * KLM octoword size - if protection was provided
3719 	 * then we use strided block format (3 octowords),
3720 	 * else we use single KLM (1 octoword)
3721 	 **/
3722 	klm_oct_size = wr->prot ? 3 : 1;
3723 
3724 	set_sig_umr_segment(*seg, klm_oct_size);
3725 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3726 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3727 	if (unlikely((*seg == qp->sq.qend)))
3728 		*seg = mlx5_get_send_wqe(qp, 0);
3729 
3730 	set_sig_mkey_segment(*seg, wr, klm_oct_size, region_len, pdn);
3731 	*seg += sizeof(struct mlx5_mkey_seg);
3732 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3733 	if (unlikely((*seg == qp->sq.qend)))
3734 		*seg = mlx5_get_send_wqe(qp, 0);
3735 
3736 	ret = set_sig_data_segment(wr, qp, seg, size);
3737 	if (ret)
3738 		return ret;
3739 
3740 	sig_mr->sig->sig_status_checked = false;
3741 	return 0;
3742 }
3743 
3744 static int set_psv_wr(struct ib_sig_domain *domain,
3745 		      u32 psv_idx, void **seg, int *size)
3746 {
3747 	struct mlx5_seg_set_psv *psv_seg = *seg;
3748 
3749 	memset(psv_seg, 0, sizeof(*psv_seg));
3750 	psv_seg->psv_num = cpu_to_be32(psv_idx);
3751 	switch (domain->sig_type) {
3752 	case IB_SIG_TYPE_NONE:
3753 		break;
3754 	case IB_SIG_TYPE_T10_DIF:
3755 		psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
3756 						     domain->sig.dif.app_tag);
3757 		psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
3758 		break;
3759 	default:
3760 		pr_err("Bad signature type given.\n");
3761 		return 1;
3762 	}
3763 
3764 	*seg += sizeof(*psv_seg);
3765 	*size += sizeof(*psv_seg) / 16;
3766 
3767 	return 0;
3768 }
3769 
3770 static int set_reg_wr(struct mlx5_ib_qp *qp,
3771 		      const struct ib_reg_wr *wr,
3772 		      void **seg, int *size)
3773 {
3774 	struct mlx5_ib_mr *mr = to_mmr(wr->mr);
3775 	struct mlx5_ib_pd *pd = to_mpd(qp->ibqp.pd);
3776 
3777 	if (unlikely(wr->wr.send_flags & IB_SEND_INLINE)) {
3778 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
3779 			     "Invalid IB_SEND_INLINE send flag\n");
3780 		return -EINVAL;
3781 	}
3782 
3783 	set_reg_umr_seg(*seg, mr);
3784 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3785 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3786 	if (unlikely((*seg == qp->sq.qend)))
3787 		*seg = mlx5_get_send_wqe(qp, 0);
3788 
3789 	set_reg_mkey_seg(*seg, mr, wr->key, wr->access);
3790 	*seg += sizeof(struct mlx5_mkey_seg);
3791 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3792 	if (unlikely((*seg == qp->sq.qend)))
3793 		*seg = mlx5_get_send_wqe(qp, 0);
3794 
3795 	set_reg_data_seg(*seg, mr, pd);
3796 	*seg += sizeof(struct mlx5_wqe_data_seg);
3797 	*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
3798 
3799 	return 0;
3800 }
3801 
3802 static void set_linv_wr(struct mlx5_ib_qp *qp, void **seg, int *size)
3803 {
3804 	set_linv_umr_seg(*seg);
3805 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3806 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3807 	if (unlikely((*seg == qp->sq.qend)))
3808 		*seg = mlx5_get_send_wqe(qp, 0);
3809 	set_linv_mkey_seg(*seg);
3810 	*seg += sizeof(struct mlx5_mkey_seg);
3811 	*size += sizeof(struct mlx5_mkey_seg) / 16;
3812 	if (unlikely((*seg == qp->sq.qend)))
3813 		*seg = mlx5_get_send_wqe(qp, 0);
3814 }
3815 
3816 static void dump_wqe(struct mlx5_ib_qp *qp, int idx, int size_16)
3817 {
3818 	__be32 *p = NULL;
3819 	int tidx = idx;
3820 	int i, j;
3821 
3822 	pr_debug("dump wqe at %p\n", mlx5_get_send_wqe(qp, tidx));
3823 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
3824 		if ((i & 0xf) == 0) {
3825 			void *buf = mlx5_get_send_wqe(qp, tidx);
3826 			tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
3827 			p = buf;
3828 			j = 0;
3829 		}
3830 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
3831 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
3832 			 be32_to_cpu(p[j + 3]));
3833 	}
3834 }
3835 
3836 static u8 get_fence(u8 fence, const struct ib_send_wr *wr)
3837 {
3838 	if (unlikely(wr->opcode == IB_WR_LOCAL_INV &&
3839 		     wr->send_flags & IB_SEND_FENCE))
3840 		return MLX5_FENCE_MODE_STRONG_ORDERING;
3841 
3842 	if (unlikely(fence)) {
3843 		if (wr->send_flags & IB_SEND_FENCE)
3844 			return MLX5_FENCE_MODE_SMALL_AND_FENCE;
3845 		else
3846 			return fence;
3847 	} else if (unlikely(wr->send_flags & IB_SEND_FENCE)) {
3848 		return MLX5_FENCE_MODE_FENCE;
3849 	}
3850 
3851 	return 0;
3852 }
3853 
3854 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
3855 		     struct mlx5_wqe_ctrl_seg **ctrl,
3856 		     const struct ib_send_wr *wr, unsigned *idx,
3857 		     int *size, int nreq, int send_flags)
3858 {
3859 	if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)))
3860 		return -ENOMEM;
3861 
3862 	*idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
3863 	*seg = mlx5_get_send_wqe(qp, *idx);
3864 	*ctrl = *seg;
3865 	*(uint32_t *)(*seg + 8) = 0;
3866 	(*ctrl)->imm = send_ieth(wr);
3867 	(*ctrl)->fm_ce_se = qp->sq_signal_bits |
3868 		(send_flags & IB_SEND_SIGNALED ?
3869 		 MLX5_WQE_CTRL_CQ_UPDATE : 0) |
3870 		(send_flags & IB_SEND_SOLICITED ?
3871 		 MLX5_WQE_CTRL_SOLICITED : 0);
3872 
3873 	*seg += sizeof(**ctrl);
3874 	*size = sizeof(**ctrl) / 16;
3875 
3876 	return 0;
3877 }
3878 
3879 static void finish_wqe(struct mlx5_ib_qp *qp,
3880 		       struct mlx5_wqe_ctrl_seg *ctrl,
3881 		       u8 size, unsigned idx, u64 wr_id,
3882 		       int nreq, u8 fence, u8 next_fence,
3883 		       u32 mlx5_opcode)
3884 {
3885 	u8 opmod = 0;
3886 
3887 	ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
3888 					     mlx5_opcode | ((u32)opmod << 24));
3889 	ctrl->qpn_ds = cpu_to_be32(size | (qp->trans_qp.base.mqp.qpn << 8));
3890 	ctrl->fm_ce_se |= fence;
3891 	qp->fm_cache = next_fence;
3892 	if (unlikely(qp->wq_sig))
3893 		ctrl->signature = wq_sig(ctrl);
3894 
3895 	qp->sq.wrid[idx] = wr_id;
3896 	qp->sq.w_list[idx].opcode = mlx5_opcode;
3897 	qp->sq.wqe_head[idx] = qp->sq.head + nreq;
3898 	qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
3899 	qp->sq.w_list[idx].next = qp->sq.cur_post;
3900 }
3901 
3902 
3903 int mlx5_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3904 		      const struct ib_send_wr **bad_wr)
3905 {
3906 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
3907 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3908 	struct mlx5_core_dev *mdev = dev->mdev;
3909 	struct mlx5_ib_qp *qp;
3910 	struct mlx5_ib_mr *mr;
3911 	struct mlx5_wqe_data_seg *dpseg;
3912 	struct mlx5_wqe_xrc_seg *xrc;
3913 	struct mlx5_bf *bf;
3914 	int uninitialized_var(size);
3915 	void *qend;
3916 	unsigned long flags;
3917 	unsigned idx;
3918 	int err = 0;
3919 	int inl = 0;
3920 	int num_sge;
3921 	void *seg;
3922 	int nreq;
3923 	int i;
3924 	u8 next_fence = 0;
3925 	u8 fence;
3926 
3927 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
3928 		return mlx5_ib_gsi_post_send(ibqp, wr, bad_wr);
3929 
3930 	qp = to_mqp(ibqp);
3931 	bf = &qp->bf;
3932 	qend = qp->sq.qend;
3933 
3934 	spin_lock_irqsave(&qp->sq.lock, flags);
3935 
3936 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
3937 		err = -EIO;
3938 		*bad_wr = wr;
3939 		nreq = 0;
3940 		goto out;
3941 	}
3942 
3943 	for (nreq = 0; wr; nreq++, wr = wr->next) {
3944 		if (unlikely(wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
3945 			mlx5_ib_warn(dev, "\n");
3946 			err = -EINVAL;
3947 			*bad_wr = wr;
3948 			goto out;
3949 		}
3950 
3951 		fence = qp->fm_cache;
3952 		num_sge = wr->num_sge;
3953 		if (unlikely(num_sge > qp->sq.max_gs)) {
3954 			mlx5_ib_warn(dev, "\n");
3955 			err = -EINVAL;
3956 			*bad_wr = wr;
3957 			goto out;
3958 		}
3959 
3960 		err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, nreq, wr->send_flags);
3961 		if (err) {
3962 			mlx5_ib_warn(dev, "\n");
3963 			err = -ENOMEM;
3964 			*bad_wr = wr;
3965 			goto out;
3966 		}
3967 
3968 		switch (ibqp->qp_type) {
3969 		case IB_QPT_XRC_INI:
3970 			xrc = seg;
3971 			seg += sizeof(*xrc);
3972 			size += sizeof(*xrc) / 16;
3973 			/* fall through */
3974 		case IB_QPT_RC:
3975 			switch (wr->opcode) {
3976 			case IB_WR_RDMA_READ:
3977 			case IB_WR_RDMA_WRITE:
3978 			case IB_WR_RDMA_WRITE_WITH_IMM:
3979 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
3980 					      rdma_wr(wr)->rkey);
3981 				seg += sizeof(struct mlx5_wqe_raddr_seg);
3982 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
3983 				break;
3984 
3985 			case IB_WR_ATOMIC_CMP_AND_SWP:
3986 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3987 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3988 				mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
3989 				err = -ENOSYS;
3990 				*bad_wr = wr;
3991 				goto out;
3992 
3993 			case IB_WR_LOCAL_INV:
3994 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3995 				qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
3996 				ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
3997 				set_linv_wr(qp, &seg, &size);
3998 				num_sge = 0;
3999 				break;
4000 
4001 			case IB_WR_REG_MR:
4002 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
4003 				qp->sq.wr_data[idx] = IB_WR_REG_MR;
4004 				ctrl->imm = cpu_to_be32(reg_wr(wr)->key);
4005 				err = set_reg_wr(qp, reg_wr(wr), &seg, &size);
4006 				if (err) {
4007 					*bad_wr = wr;
4008 					goto out;
4009 				}
4010 				num_sge = 0;
4011 				break;
4012 
4013 			case IB_WR_REG_SIG_MR:
4014 				qp->sq.wr_data[idx] = IB_WR_REG_SIG_MR;
4015 				mr = to_mmr(sig_handover_wr(wr)->sig_mr);
4016 
4017 				ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
4018 				err = set_sig_umr_wr(wr, qp, &seg, &size);
4019 				if (err) {
4020 					mlx5_ib_warn(dev, "\n");
4021 					*bad_wr = wr;
4022 					goto out;
4023 				}
4024 
4025 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
4026 					   nreq, get_fence(fence, wr),
4027 					   next_fence, MLX5_OPCODE_UMR);
4028 				/*
4029 				 * SET_PSV WQEs are not signaled and solicited
4030 				 * on error
4031 				 */
4032 				err = begin_wqe(qp, &seg, &ctrl, wr,
4033 						&idx, &size, nreq, IB_SEND_SOLICITED);
4034 				if (err) {
4035 					mlx5_ib_warn(dev, "\n");
4036 					err = -ENOMEM;
4037 					*bad_wr = wr;
4038 					goto out;
4039 				}
4040 
4041 				err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->mem,
4042 						 mr->sig->psv_memory.psv_idx, &seg,
4043 						 &size);
4044 				if (err) {
4045 					mlx5_ib_warn(dev, "\n");
4046 					*bad_wr = wr;
4047 					goto out;
4048 				}
4049 
4050 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
4051 					   nreq, get_fence(fence, wr),
4052 					   next_fence, MLX5_OPCODE_SET_PSV);
4053 				err = begin_wqe(qp, &seg, &ctrl, wr,
4054 						&idx, &size, nreq, wr->send_flags);
4055 				if (err) {
4056 					mlx5_ib_warn(dev, "\n");
4057 					err = -ENOMEM;
4058 					*bad_wr = wr;
4059 					goto out;
4060 				}
4061 
4062 				next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
4063 				err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->wire,
4064 						 mr->sig->psv_wire.psv_idx, &seg,
4065 						 &size);
4066 				if (err) {
4067 					mlx5_ib_warn(dev, "\n");
4068 					*bad_wr = wr;
4069 					goto out;
4070 				}
4071 
4072 				finish_wqe(qp, ctrl, size, idx, wr->wr_id,
4073 					   nreq, get_fence(fence, wr),
4074 					   next_fence, MLX5_OPCODE_SET_PSV);
4075 				num_sge = 0;
4076 				goto skip_psv;
4077 
4078 			default:
4079 				break;
4080 			}
4081 			break;
4082 
4083 		case IB_QPT_UC:
4084 			switch (wr->opcode) {
4085 			case IB_WR_RDMA_WRITE:
4086 			case IB_WR_RDMA_WRITE_WITH_IMM:
4087 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
4088 					      rdma_wr(wr)->rkey);
4089 				seg  += sizeof(struct mlx5_wqe_raddr_seg);
4090 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
4091 				break;
4092 
4093 			default:
4094 				break;
4095 			}
4096 			break;
4097 
4098 		case IB_QPT_SMI:
4099 		case MLX5_IB_QPT_HW_GSI:
4100 			set_datagram_seg(seg, wr);
4101 			seg += sizeof(struct mlx5_wqe_datagram_seg);
4102 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
4103 			if (unlikely((seg == qend)))
4104 				seg = mlx5_get_send_wqe(qp, 0);
4105 			break;
4106 		case IB_QPT_UD:
4107 			set_datagram_seg(seg, wr);
4108 			seg += sizeof(struct mlx5_wqe_datagram_seg);
4109 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
4110 
4111 			if (unlikely((seg == qend)))
4112 				seg = mlx5_get_send_wqe(qp, 0);
4113 
4114 			/* handle qp that supports ud offload */
4115 			if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO) {
4116 				struct mlx5_wqe_eth_pad *pad;
4117 
4118 				pad = seg;
4119 				memset(pad, 0, sizeof(struct mlx5_wqe_eth_pad));
4120 				seg += sizeof(struct mlx5_wqe_eth_pad);
4121 				size += sizeof(struct mlx5_wqe_eth_pad) / 16;
4122 
4123 				seg = set_eth_seg(seg, wr, qend, qp, &size);
4124 
4125 				if (unlikely((seg == qend)))
4126 					seg = mlx5_get_send_wqe(qp, 0);
4127 			}
4128 			break;
4129 		case MLX5_IB_QPT_REG_UMR:
4130 			if (wr->opcode != MLX5_IB_WR_UMR) {
4131 				err = -EINVAL;
4132 				mlx5_ib_warn(dev, "bad opcode\n");
4133 				goto out;
4134 			}
4135 			qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
4136 			ctrl->imm = cpu_to_be32(umr_wr(wr)->mkey);
4137 			set_reg_umr_segment(seg, wr);
4138 			seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4139 			size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4140 			if (unlikely((seg == qend)))
4141 				seg = mlx5_get_send_wqe(qp, 0);
4142 			set_reg_mkey_segment(seg, wr);
4143 			seg += sizeof(struct mlx5_mkey_seg);
4144 			size += sizeof(struct mlx5_mkey_seg) / 16;
4145 			if (unlikely((seg == qend)))
4146 				seg = mlx5_get_send_wqe(qp, 0);
4147 			break;
4148 
4149 		default:
4150 			break;
4151 		}
4152 
4153 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
4154 			int uninitialized_var(sz);
4155 
4156 			err = set_data_inl_seg(qp, wr, seg, &sz);
4157 			if (unlikely(err)) {
4158 				mlx5_ib_warn(dev, "\n");
4159 				*bad_wr = wr;
4160 				goto out;
4161 			}
4162 			inl = 1;
4163 			size += sz;
4164 		} else {
4165 			dpseg = seg;
4166 			for (i = 0; i < num_sge; i++) {
4167 				if (unlikely(dpseg == qend)) {
4168 					seg = mlx5_get_send_wqe(qp, 0);
4169 					dpseg = seg;
4170 				}
4171 				if (likely(wr->sg_list[i].length)) {
4172 					set_data_ptr_seg(dpseg, wr->sg_list + i);
4173 					size += sizeof(struct mlx5_wqe_data_seg) / 16;
4174 					dpseg++;
4175 				}
4176 			}
4177 		}
4178 
4179 		finish_wqe(qp, ctrl, size, idx, wr->wr_id, nreq,
4180 			   get_fence(fence, wr), next_fence,
4181 			   mlx5_ib_opcode[wr->opcode]);
4182 skip_psv:
4183 		if (0)
4184 			dump_wqe(qp, idx, size);
4185 	}
4186 
4187 out:
4188 	if (likely(nreq)) {
4189 		qp->sq.head += nreq;
4190 
4191 		/* Make sure that descriptors are written before
4192 		 * updating doorbell record and ringing the doorbell
4193 		 */
4194 		wmb();
4195 
4196 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
4197 
4198 		/* Make sure doorbell record is visible to the HCA before
4199 		 * we hit doorbell */
4200 		wmb();
4201 
4202 		mlx5_write64((__be32 *)ctrl, bf->bfreg->map + bf->offset,
4203 			     MLX5_GET_DOORBELL_LOCK(&bf->lock32));
4204 		/* Make sure doorbells don't leak out of SQ spinlock
4205 		 * and reach the HCA out of order.
4206 		 */
4207 		bf->offset ^= bf->buf_size;
4208 	}
4209 
4210 	spin_unlock_irqrestore(&qp->sq.lock, flags);
4211 
4212 	return err;
4213 }
4214 
4215 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
4216 {
4217 	sig->signature = calc_sig(sig, size);
4218 }
4219 
4220 int mlx5_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
4221 		      const struct ib_recv_wr **bad_wr)
4222 {
4223 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4224 	struct mlx5_wqe_data_seg *scat;
4225 	struct mlx5_rwqe_sig *sig;
4226 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4227 	struct mlx5_core_dev *mdev = dev->mdev;
4228 	unsigned long flags;
4229 	int err = 0;
4230 	int nreq;
4231 	int ind;
4232 	int i;
4233 
4234 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4235 		return mlx5_ib_gsi_post_recv(ibqp, wr, bad_wr);
4236 
4237 	spin_lock_irqsave(&qp->rq.lock, flags);
4238 
4239 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
4240 		err = -EIO;
4241 		*bad_wr = wr;
4242 		nreq = 0;
4243 		goto out;
4244 	}
4245 
4246 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
4247 
4248 	for (nreq = 0; wr; nreq++, wr = wr->next) {
4249 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
4250 			err = -ENOMEM;
4251 			*bad_wr = wr;
4252 			goto out;
4253 		}
4254 
4255 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
4256 			err = -EINVAL;
4257 			*bad_wr = wr;
4258 			goto out;
4259 		}
4260 
4261 		scat = get_recv_wqe(qp, ind);
4262 		if (qp->wq_sig)
4263 			scat++;
4264 
4265 		for (i = 0; i < wr->num_sge; i++)
4266 			set_data_ptr_seg(scat + i, wr->sg_list + i);
4267 
4268 		if (i < qp->rq.max_gs) {
4269 			scat[i].byte_count = 0;
4270 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
4271 			scat[i].addr       = 0;
4272 		}
4273 
4274 		if (qp->wq_sig) {
4275 			sig = (struct mlx5_rwqe_sig *)scat;
4276 			set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
4277 		}
4278 
4279 		qp->rq.wrid[ind] = wr->wr_id;
4280 
4281 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
4282 	}
4283 
4284 out:
4285 	if (likely(nreq)) {
4286 		qp->rq.head += nreq;
4287 
4288 		/* Make sure that descriptors are written before
4289 		 * doorbell record.
4290 		 */
4291 		wmb();
4292 
4293 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
4294 	}
4295 
4296 	spin_unlock_irqrestore(&qp->rq.lock, flags);
4297 
4298 	return err;
4299 }
4300 
4301 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
4302 {
4303 	switch (mlx5_state) {
4304 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
4305 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
4306 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
4307 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
4308 	case MLX5_QP_STATE_SQ_DRAINING:
4309 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
4310 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
4311 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
4312 	default:		     return -1;
4313 	}
4314 }
4315 
4316 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
4317 {
4318 	switch (mlx5_mig_state) {
4319 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
4320 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
4321 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
4322 	default: return -1;
4323 	}
4324 }
4325 
4326 static int to_ib_qp_access_flags(int mlx5_flags)
4327 {
4328 	int ib_flags = 0;
4329 
4330 	if (mlx5_flags & MLX5_QP_BIT_RRE)
4331 		ib_flags |= IB_ACCESS_REMOTE_READ;
4332 	if (mlx5_flags & MLX5_QP_BIT_RWE)
4333 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
4334 	if (mlx5_flags & MLX5_QP_BIT_RAE)
4335 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
4336 
4337 	return ib_flags;
4338 }
4339 
4340 static void to_ib_ah_attr(struct mlx5_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
4341 				struct mlx5_qp_path *path)
4342 {
4343 	struct mlx5_core_dev *dev = ibdev->mdev;
4344 
4345 	memset(ib_ah_attr, 0, sizeof(*ib_ah_attr));
4346 	ib_ah_attr->port_num	  = path->port;
4347 
4348 	if (ib_ah_attr->port_num == 0 ||
4349 	    ib_ah_attr->port_num > MLX5_CAP_GEN(dev, num_ports))
4350 		return;
4351 
4352 	ib_ah_attr->sl = path->dci_cfi_prio_sl & 0xf;
4353 
4354 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
4355 	ib_ah_attr->src_path_bits = path->grh_mlid & 0x7f;
4356 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
4357 	ib_ah_attr->ah_flags      = (path->grh_mlid & (1 << 7)) ? IB_AH_GRH : 0;
4358 	if (ib_ah_attr->ah_flags) {
4359 		ib_ah_attr->grh.sgid_index = path->mgid_index;
4360 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
4361 		ib_ah_attr->grh.traffic_class =
4362 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
4363 		ib_ah_attr->grh.flow_label =
4364 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
4365 		memcpy(ib_ah_attr->grh.dgid.raw,
4366 		       path->rgid, sizeof(ib_ah_attr->grh.dgid.raw));
4367 	}
4368 }
4369 
4370 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
4371 					struct mlx5_ib_sq *sq,
4372 					u8 *sq_state)
4373 {
4374 	void *out;
4375 	void *sqc;
4376 	int inlen;
4377 	int err;
4378 
4379 	inlen = MLX5_ST_SZ_BYTES(query_sq_out);
4380 	out = mlx5_vzalloc(inlen);
4381 	if (!out)
4382 		return -ENOMEM;
4383 
4384 	err = mlx5_core_query_sq(dev->mdev, sq->base.mqp.qpn, out);
4385 	if (err)
4386 		goto out;
4387 
4388 	sqc = MLX5_ADDR_OF(query_sq_out, out, sq_context);
4389 	*sq_state = MLX5_GET(sqc, sqc, state);
4390 	sq->state = *sq_state;
4391 
4392 out:
4393 	kvfree(out);
4394 	return err;
4395 }
4396 
4397 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
4398 					struct mlx5_ib_rq *rq,
4399 					u8 *rq_state)
4400 {
4401 	void *out;
4402 	void *rqc;
4403 	int inlen;
4404 	int err;
4405 
4406 	inlen = MLX5_ST_SZ_BYTES(query_rq_out);
4407 	out = mlx5_vzalloc(inlen);
4408 	if (!out)
4409 		return -ENOMEM;
4410 
4411 	err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
4412 	if (err)
4413 		goto out;
4414 
4415 	rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
4416 	*rq_state = MLX5_GET(rqc, rqc, state);
4417 	rq->state = *rq_state;
4418 
4419 out:
4420 	kvfree(out);
4421 	return err;
4422 }
4423 
4424 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
4425 				  struct mlx5_ib_qp *qp, u8 *qp_state)
4426 {
4427 	static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
4428 		[MLX5_RQC_STATE_RST] = {
4429 			[MLX5_SQC_STATE_RST]	= IB_QPS_RESET,
4430 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4431 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE_BAD,
4432 			[MLX5_SQ_STATE_NA]	= IB_QPS_RESET,
4433 		},
4434 		[MLX5_RQC_STATE_RDY] = {
4435 			[MLX5_SQC_STATE_RST]	= MLX5_QP_STATE_BAD,
4436 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4437 			[MLX5_SQC_STATE_ERR]	= IB_QPS_SQE,
4438 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE,
4439 		},
4440 		[MLX5_RQC_STATE_ERR] = {
4441 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE_BAD,
4442 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4443 			[MLX5_SQC_STATE_ERR]	= IB_QPS_ERR,
4444 			[MLX5_SQ_STATE_NA]	= IB_QPS_ERR,
4445 		},
4446 		[MLX5_RQ_STATE_NA] = {
4447 			[MLX5_SQC_STATE_RST]    = IB_QPS_RESET,
4448 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4449 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE,
4450 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE_BAD,
4451 		},
4452 	};
4453 
4454 	*qp_state = sqrq_trans[rq_state][sq_state];
4455 
4456 	if (*qp_state == MLX5_QP_STATE_BAD) {
4457 		WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
4458 		     qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
4459 		     qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
4460 		return -EINVAL;
4461 	}
4462 
4463 	if (*qp_state == MLX5_QP_STATE)
4464 		*qp_state = qp->state;
4465 
4466 	return 0;
4467 }
4468 
4469 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
4470 				     struct mlx5_ib_qp *qp,
4471 				     u8 *raw_packet_qp_state)
4472 {
4473 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
4474 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
4475 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
4476 	int err;
4477 	u8 sq_state = MLX5_SQ_STATE_NA;
4478 	u8 rq_state = MLX5_RQ_STATE_NA;
4479 
4480 	if (qp->sq.wqe_cnt) {
4481 		err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
4482 		if (err)
4483 			return err;
4484 	}
4485 
4486 	if (qp->rq.wqe_cnt) {
4487 		err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
4488 		if (err)
4489 			return err;
4490 	}
4491 
4492 	return sqrq_state_to_qp_state(sq_state, rq_state, qp,
4493 				      raw_packet_qp_state);
4494 }
4495 
4496 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
4497 			 struct ib_qp_attr *qp_attr)
4498 {
4499 	int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
4500 	struct mlx5_qp_context *context;
4501 	int mlx5_state;
4502 	u32 *outb;
4503 	int err = 0;
4504 
4505 	outb = kzalloc(outlen, GFP_KERNEL);
4506 	if (!outb)
4507 		return -ENOMEM;
4508 
4509 	err = mlx5_core_qp_query(dev->mdev, &qp->trans_qp.base.mqp, outb,
4510 				 outlen);
4511 	if (err)
4512 		goto out;
4513 
4514 	/* FIXME: use MLX5_GET rather than mlx5_qp_context manual struct */
4515 	context = (struct mlx5_qp_context *)MLX5_ADDR_OF(query_qp_out, outb, qpc);
4516 
4517 	mlx5_state = be32_to_cpu(context->flags) >> 28;
4518 
4519 	qp->state		     = to_ib_qp_state(mlx5_state);
4520 	qp_attr->path_mtu	     = context->mtu_msgmax >> 5;
4521 	qp_attr->path_mig_state	     =
4522 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
4523 	qp_attr->qkey		     = be32_to_cpu(context->qkey);
4524 	qp_attr->rq_psn		     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
4525 	qp_attr->sq_psn		     = be32_to_cpu(context->next_send_psn) & 0xffffff;
4526 	qp_attr->dest_qp_num	     = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
4527 	qp_attr->qp_access_flags     =
4528 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
4529 
4530 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4531 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
4532 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
4533 		qp_attr->alt_pkey_index =
4534 			be16_to_cpu(context->alt_path.pkey_index);
4535 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
4536 	}
4537 
4538 	qp_attr->pkey_index = be16_to_cpu(context->pri_path.pkey_index);
4539 	qp_attr->port_num = context->pri_path.port;
4540 
4541 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4542 	qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
4543 
4544 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
4545 
4546 	qp_attr->max_dest_rd_atomic =
4547 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
4548 	qp_attr->min_rnr_timer	    =
4549 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
4550 	qp_attr->timeout	    = context->pri_path.ackto_lt >> 3;
4551 	qp_attr->retry_cnt	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
4552 	qp_attr->rnr_retry	    = (be32_to_cpu(context->params1) >> 13) & 0x7;
4553 	qp_attr->alt_timeout	    = context->alt_path.ackto_lt >> 3;
4554 
4555 out:
4556 	kfree(outb);
4557 	return err;
4558 }
4559 
4560 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
4561 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
4562 {
4563 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4564 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4565 	int err = 0;
4566 	u8 raw_packet_qp_state;
4567 
4568 	if (ibqp->rwq_ind_tbl)
4569 		return -ENOSYS;
4570 
4571 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4572 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
4573 					    qp_init_attr);
4574 
4575 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4576 	/*
4577 	 * Wait for any outstanding page faults, in case the user frees memory
4578 	 * based upon this query's result.
4579 	 */
4580 	flush_workqueue(mlx5_ib_page_fault_wq);
4581 #endif
4582 
4583 	mutex_lock(&qp->mutex);
4584 
4585 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
4586 		err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
4587 		if (err)
4588 			goto out;
4589 		qp->state = raw_packet_qp_state;
4590 		qp_attr->port_num = 1;
4591 	} else {
4592 		err = query_qp_attr(dev, qp, qp_attr);
4593 		if (err)
4594 			goto out;
4595 	}
4596 
4597 	qp_attr->qp_state	     = qp->state;
4598 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4599 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4600 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4601 
4602 	if (!ibqp->uobject) {
4603 		qp_attr->cap.max_send_wr  = qp->sq.max_post;
4604 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4605 		qp_init_attr->qp_context = ibqp->qp_context;
4606 	} else {
4607 		qp_attr->cap.max_send_wr  = 0;
4608 		qp_attr->cap.max_send_sge = 0;
4609 	}
4610 
4611 	qp_init_attr->qp_type = ibqp->qp_type;
4612 	qp_init_attr->recv_cq = ibqp->recv_cq;
4613 	qp_init_attr->send_cq = ibqp->send_cq;
4614 	qp_init_attr->srq = ibqp->srq;
4615 	qp_attr->cap.max_inline_data = qp->max_inline_data;
4616 
4617 	qp_init_attr->cap	     = qp_attr->cap;
4618 
4619 	qp_init_attr->create_flags = 0;
4620 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4621 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4622 
4623 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
4624 		qp_init_attr->create_flags |= IB_QP_CREATE_CROSS_CHANNEL;
4625 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
4626 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_SEND;
4627 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
4628 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_RECV;
4629 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
4630 		qp_init_attr->create_flags |= MLX5_IB_QP_CREATE_SQPN_QP1;
4631 
4632 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
4633 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4634 
4635 out:
4636 	mutex_unlock(&qp->mutex);
4637 	return err;
4638 }
4639 
4640 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
4641 				   struct ib_udata *udata)
4642 {
4643 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
4644 	struct mlx5_ib_xrcd *xrcd;
4645 	int err;
4646 
4647 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
4648 		return ERR_PTR(-ENOSYS);
4649 
4650 	xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
4651 	if (!xrcd)
4652 		return ERR_PTR(-ENOMEM);
4653 
4654 	err = mlx5_core_xrcd_alloc(dev->mdev, &xrcd->xrcdn);
4655 	if (err) {
4656 		kfree(xrcd);
4657 		return ERR_PTR(-ENOMEM);
4658 	}
4659 
4660 	return &xrcd->ibxrcd;
4661 }
4662 
4663 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata)
4664 {
4665 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
4666 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
4667 	int err;
4668 
4669 	err = mlx5_core_xrcd_dealloc(dev->mdev, xrcdn);
4670 	if (err)
4671 		mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
4672 
4673 	kfree(xrcd);
4674 	return 0;
4675 }
4676 
4677 static void mlx5_ib_wq_event(struct mlx5_core_qp *core_qp, int type)
4678 {
4679 	struct mlx5_ib_rwq *rwq = to_mibrwq(core_qp);
4680 	struct mlx5_ib_dev *dev = to_mdev(rwq->ibwq.device);
4681 	struct ib_event event;
4682 
4683 	if (rwq->ibwq.event_handler) {
4684 		event.device     = rwq->ibwq.device;
4685 		event.element.wq = &rwq->ibwq;
4686 		switch (type) {
4687 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
4688 			event.event = IB_EVENT_WQ_FATAL;
4689 			break;
4690 		default:
4691 			mlx5_ib_warn(dev, "Unexpected event type %d on WQ %06x\n", type, core_qp->qpn);
4692 			return;
4693 		}
4694 
4695 		rwq->ibwq.event_handler(&event, rwq->ibwq.wq_context);
4696 	}
4697 }
4698 
4699 static int  create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
4700 		      struct ib_wq_init_attr *init_attr)
4701 {
4702 	struct mlx5_ib_dev *dev;
4703 	__be64 *rq_pas0;
4704 	void *in;
4705 	void *rqc;
4706 	void *wq;
4707 	int inlen;
4708 	int err;
4709 
4710 	dev = to_mdev(pd->device);
4711 
4712 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
4713 	in = mlx5_vzalloc(inlen);
4714 	if (!in)
4715 		return -ENOMEM;
4716 
4717 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
4718 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
4719 	MLX5_SET(rqc,  rqc, mem_rq_type,
4720 		 MLX5_RQC_RQ_TYPE_MEMORY_RQ_INLINE);
4721 	MLX5_SET(rqc, rqc, user_index, rwq->user_index);
4722 	MLX5_SET(rqc,  rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
4723 	MLX5_SET(rqc,  rqc, state, MLX5_RQC_STATE_RST);
4724 	MLX5_SET(rqc,  rqc, flush_in_error_en, 1);
4725 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
4726 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
4727 	MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
4728 	MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
4729 	MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
4730 	MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
4731 	MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
4732 	MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
4733 	MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
4734 	MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
4735 	rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
4736 	mlx5_ib_populate_pas(dev, rwq->umem, rwq->page_shift, rq_pas0, 0);
4737 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rwq->core_qp);
4738 	kvfree(in);
4739 	return err;
4740 }
4741 
4742 static int set_user_rq_size(struct mlx5_ib_dev *dev,
4743 			    struct ib_wq_init_attr *wq_init_attr,
4744 			    struct mlx5_ib_create_wq *ucmd,
4745 			    struct mlx5_ib_rwq *rwq)
4746 {
4747 	/* Sanity check RQ size before proceeding */
4748 	if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
4749 		return -EINVAL;
4750 
4751 	if (!ucmd->rq_wqe_count)
4752 		return -EINVAL;
4753 
4754 	rwq->wqe_count = ucmd->rq_wqe_count;
4755 	rwq->wqe_shift = ucmd->rq_wqe_shift;
4756 	rwq->buf_size = (rwq->wqe_count << rwq->wqe_shift);
4757 	rwq->log_rq_stride = rwq->wqe_shift;
4758 	rwq->log_rq_size = ilog2(rwq->wqe_count);
4759 	return 0;
4760 }
4761 
4762 static int prepare_user_rq(struct ib_pd *pd,
4763 			   struct ib_wq_init_attr *init_attr,
4764 			   struct ib_udata *udata,
4765 			   struct mlx5_ib_rwq *rwq)
4766 {
4767 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
4768 	struct mlx5_ib_create_wq ucmd = {};
4769 	int err;
4770 	size_t required_cmd_sz;
4771 
4772 	required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
4773 	if (udata->inlen < required_cmd_sz) {
4774 		mlx5_ib_dbg(dev, "invalid inlen\n");
4775 		return -EINVAL;
4776 	}
4777 
4778 	if (udata->inlen > sizeof(ucmd) &&
4779 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4780 				 udata->inlen - sizeof(ucmd))) {
4781 		mlx5_ib_dbg(dev, "inlen is not supported\n");
4782 		return -EOPNOTSUPP;
4783 	}
4784 
4785 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
4786 		mlx5_ib_dbg(dev, "copy failed\n");
4787 		return -EFAULT;
4788 	}
4789 
4790 	if (ucmd.comp_mask) {
4791 		mlx5_ib_dbg(dev, "invalid comp mask\n");
4792 		return -EOPNOTSUPP;
4793 	}
4794 
4795 	if (ucmd.reserved) {
4796 		mlx5_ib_dbg(dev, "invalid reserved\n");
4797 		return -EOPNOTSUPP;
4798 	}
4799 
4800 	err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
4801 	if (err) {
4802 		mlx5_ib_dbg(dev, "err %d\n", err);
4803 		return err;
4804 	}
4805 
4806 	err = create_user_rq(dev, pd, rwq, &ucmd);
4807 	if (err) {
4808 		mlx5_ib_dbg(dev, "err %d\n", err);
4809 		if (err)
4810 			return err;
4811 	}
4812 
4813 	rwq->user_index = ucmd.user_index;
4814 	return 0;
4815 }
4816 
4817 struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
4818 				struct ib_wq_init_attr *init_attr,
4819 				struct ib_udata *udata)
4820 {
4821 	struct mlx5_ib_dev *dev;
4822 	struct mlx5_ib_rwq *rwq;
4823 	struct mlx5_ib_create_wq_resp resp = {};
4824 	size_t min_resp_len;
4825 	int err;
4826 
4827 	if (!udata)
4828 		return ERR_PTR(-ENOSYS);
4829 
4830 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4831 	if (udata->outlen && udata->outlen < min_resp_len)
4832 		return ERR_PTR(-EINVAL);
4833 
4834 	dev = to_mdev(pd->device);
4835 	switch (init_attr->wq_type) {
4836 	case IB_WQT_RQ:
4837 		rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
4838 		if (!rwq)
4839 			return ERR_PTR(-ENOMEM);
4840 		err = prepare_user_rq(pd, init_attr, udata, rwq);
4841 		if (err)
4842 			goto err;
4843 		err = create_rq(rwq, pd, init_attr);
4844 		if (err)
4845 			goto err_user_rq;
4846 		break;
4847 	default:
4848 		mlx5_ib_dbg(dev, "unsupported wq type %d\n",
4849 			    init_attr->wq_type);
4850 		return ERR_PTR(-EINVAL);
4851 	}
4852 
4853 	rwq->ibwq.wq_num = rwq->core_qp.qpn;
4854 	rwq->ibwq.state = IB_WQS_RESET;
4855 	if (udata->outlen) {
4856 		resp.response_length = offsetof(typeof(resp), response_length) +
4857 				sizeof(resp.response_length);
4858 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4859 		if (err)
4860 			goto err_copy;
4861 	}
4862 
4863 	rwq->core_qp.event = mlx5_ib_wq_event;
4864 	rwq->ibwq.event_handler = init_attr->event_handler;
4865 	return &rwq->ibwq;
4866 
4867 err_copy:
4868 	mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
4869 err_user_rq:
4870 	destroy_user_rq(pd, rwq, udata);
4871 err:
4872 	kfree(rwq);
4873 	return ERR_PTR(err);
4874 }
4875 
4876 void mlx5_ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata)
4877 {
4878 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
4879 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
4880 
4881 	mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
4882 	destroy_user_rq(wq->pd, rwq, udata);
4883 	kfree(rwq);
4884 }
4885 
4886 struct ib_rwq_ind_table *mlx5_ib_create_rwq_ind_table(struct ib_device *device,
4887 						      struct ib_rwq_ind_table_init_attr *init_attr,
4888 						      struct ib_udata *udata)
4889 {
4890 	struct mlx5_ib_dev *dev = to_mdev(device);
4891 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl;
4892 	int sz = 1 << init_attr->log_ind_tbl_size;
4893 	struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
4894 	size_t min_resp_len;
4895 	int inlen;
4896 	int err;
4897 	int i;
4898 	u32 *in;
4899 	void *rqtc;
4900 
4901 	if (udata->inlen > 0 &&
4902 	    !ib_is_udata_cleared(udata, 0,
4903 				 udata->inlen))
4904 		return ERR_PTR(-EOPNOTSUPP);
4905 
4906 	if (init_attr->log_ind_tbl_size >
4907 	    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size)) {
4908 		mlx5_ib_dbg(dev, "log_ind_tbl_size = %d is bigger than supported = %d\n",
4909 			    init_attr->log_ind_tbl_size,
4910 			    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size));
4911 		return ERR_PTR(-EINVAL);
4912 	}
4913 
4914 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4915 	if (udata->outlen && udata->outlen < min_resp_len)
4916 		return ERR_PTR(-EINVAL);
4917 
4918 	rwq_ind_tbl = kzalloc(sizeof(*rwq_ind_tbl), GFP_KERNEL);
4919 	if (!rwq_ind_tbl)
4920 		return ERR_PTR(-ENOMEM);
4921 
4922 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
4923 	in = mlx5_vzalloc(inlen);
4924 	if (!in) {
4925 		err = -ENOMEM;
4926 		goto err;
4927 	}
4928 
4929 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
4930 
4931 	MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
4932 	MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
4933 
4934 	for (i = 0; i < sz; i++)
4935 		MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
4936 
4937 	rwq_ind_tbl->uid = to_mpd(init_attr->ind_tbl[0]->pd)->uid;
4938 	MLX5_SET(create_rqt_in, in, uid, rwq_ind_tbl->uid);
4939 
4940 	err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
4941 	kvfree(in);
4942 
4943 	if (err)
4944 		goto err;
4945 
4946 	rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
4947 	if (udata->outlen) {
4948 		resp.response_length = offsetof(typeof(resp), response_length) +
4949 					sizeof(resp.response_length);
4950 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4951 		if (err)
4952 			goto err_copy;
4953 	}
4954 
4955 	return &rwq_ind_tbl->ib_rwq_ind_tbl;
4956 
4957 err_copy:
4958 	mlx5_core_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
4959 err:
4960 	kfree(rwq_ind_tbl);
4961 	return ERR_PTR(err);
4962 }
4963 
4964 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4965 {
4966 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
4967 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
4968 
4969 	mlx5_core_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
4970 
4971 	kfree(rwq_ind_tbl);
4972 	return 0;
4973 }
4974 
4975 int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
4976 		      u32 wq_attr_mask, struct ib_udata *udata)
4977 {
4978 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
4979 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
4980 	struct mlx5_ib_modify_wq ucmd = {};
4981 	size_t required_cmd_sz;
4982 	int curr_wq_state;
4983 	int wq_state;
4984 	int inlen;
4985 	int err;
4986 	void *rqc;
4987 	void *in;
4988 
4989 	required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
4990 	if (udata->inlen < required_cmd_sz)
4991 		return -EINVAL;
4992 
4993 	if (udata->inlen > sizeof(ucmd) &&
4994 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4995 				 udata->inlen - sizeof(ucmd)))
4996 		return -EOPNOTSUPP;
4997 
4998 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4999 		return -EFAULT;
5000 
5001 	if (ucmd.comp_mask || ucmd.reserved)
5002 		return -EOPNOTSUPP;
5003 
5004 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
5005 	in = mlx5_vzalloc(inlen);
5006 	if (!in)
5007 		return -ENOMEM;
5008 
5009 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
5010 
5011 	MLX5_SET(modify_rq_in, in, rqn, rwq->core_qp.qpn);
5012 	curr_wq_state = (wq_attr_mask & IB_WQ_CUR_STATE) ?
5013 		wq_attr->curr_wq_state : wq->state;
5014 	wq_state = (wq_attr_mask & IB_WQ_STATE) ?
5015 		wq_attr->wq_state : curr_wq_state;
5016 	if (curr_wq_state == IB_WQS_ERR)
5017 		curr_wq_state = MLX5_RQC_STATE_ERR;
5018 	if (wq_state == IB_WQS_ERR)
5019 		wq_state = MLX5_RQC_STATE_ERR;
5020 	MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
5021 	MLX5_SET(modify_rq_in, in, uid, to_mpd(wq->pd)->uid);
5022 	MLX5_SET(rqc, rqc, state, wq_state);
5023 
5024 	err = mlx5_core_modify_rq(dev->mdev, in, inlen);
5025 	kvfree(in);
5026 	if (!err)
5027 		rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;
5028 
5029 	return err;
5030 }
5031