xref: /linux/drivers/infiniband/hw/mlx4/qp.c (revision bf4afc53b77aeaa48b5409da5c8da6bb4eff7f43)
1 /*
2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/log2.h>
35 #include <linux/etherdevice.h>
36 #include <net/ip.h>
37 #include <linux/slab.h>
38 #include <linux/netdevice.h>
39 
40 #include <rdma/ib_cache.h>
41 #include <rdma/ib_pack.h>
42 #include <rdma/ib_addr.h>
43 #include <rdma/ib_mad.h>
44 #include <rdma/uverbs_ioctl.h>
45 
46 #include <linux/mlx4/driver.h>
47 #include <linux/mlx4/qp.h>
48 
49 #include "mlx4_ib.h"
50 #include <rdma/mlx4-abi.h>
51 
52 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
53 			     struct mlx4_ib_cq *recv_cq);
54 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
55 			       struct mlx4_ib_cq *recv_cq);
56 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state,
57 			      struct ib_udata *udata);
58 
59 enum {
60 	MLX4_IB_ACK_REQ_FREQ	= 8,
61 };
62 
63 enum {
64 	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
65 	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
66 	MLX4_IB_LINK_TYPE_IB		= 0,
67 	MLX4_IB_LINK_TYPE_ETH		= 1
68 };
69 
70 enum {
71 	MLX4_IB_MIN_SQ_STRIDE	= 6,
72 	MLX4_IB_CACHE_LINE_SIZE	= 64,
73 };
74 
75 enum {
76 	MLX4_RAW_QP_MTU		= 7,
77 	MLX4_RAW_QP_MSGMAX	= 31,
78 };
79 
80 #ifndef ETH_ALEN
81 #define ETH_ALEN        6
82 #endif
83 
84 static const __be32 mlx4_ib_opcode[] = {
85 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
86 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
87 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
88 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
89 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
90 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
91 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
92 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
93 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
94 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
95 	[IB_WR_REG_MR]				= cpu_to_be32(MLX4_OPCODE_FMR),
96 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
97 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
98 };
99 
100 enum mlx4_ib_source_type {
101 	MLX4_IB_QP_SRC	= 0,
102 	MLX4_IB_RWQ_SRC	= 1,
103 };
104 
105 struct mlx4_ib_qp_event_work {
106 	struct work_struct work;
107 	struct mlx4_qp *qp;
108 	enum mlx4_event type;
109 };
110 
111 static struct workqueue_struct *mlx4_ib_qp_event_wq;
112 
113 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
114 {
115 	if (!mlx4_is_master(dev->dev))
116 		return 0;
117 
118 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
119 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
120 		8 * MLX4_MFUNC_MAX;
121 }
122 
123 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
124 {
125 	int proxy_sqp = 0;
126 	int real_sqp = 0;
127 	int i;
128 	/* PPF or Native -- real SQP */
129 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
130 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
131 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
132 	if (real_sqp)
133 		return 1;
134 	/* VF or PF -- proxy SQP */
135 	if (mlx4_is_mfunc(dev->dev)) {
136 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
137 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy ||
138 			    qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp1_proxy) {
139 				proxy_sqp = 1;
140 				break;
141 			}
142 		}
143 	}
144 	if (proxy_sqp)
145 		return 1;
146 
147 	return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
148 }
149 
150 /* used for INIT/CLOSE port logic */
151 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
152 {
153 	int proxy_qp0 = 0;
154 	int real_qp0 = 0;
155 	int i;
156 	/* PPF or Native -- real QP0 */
157 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
158 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
159 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
160 	if (real_qp0)
161 		return 1;
162 	/* VF or PF -- proxy QP0 */
163 	if (mlx4_is_mfunc(dev->dev)) {
164 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
165 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy) {
166 				proxy_qp0 = 1;
167 				break;
168 			}
169 		}
170 	}
171 	return proxy_qp0;
172 }
173 
174 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
175 {
176 	return mlx4_buf_offset(&qp->buf, offset);
177 }
178 
179 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
180 {
181 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
182 }
183 
184 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
185 {
186 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
187 }
188 
189 /*
190  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
191  * first four bytes of every 64 byte chunk with 0xffffffff, except for
192  * the very first chunk of the WQE.
193  */
194 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n)
195 {
196 	__be32 *wqe;
197 	int i;
198 	int s;
199 	void *buf;
200 	struct mlx4_wqe_ctrl_seg *ctrl;
201 
202 	buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
203 	ctrl = (struct mlx4_wqe_ctrl_seg *)buf;
204 	s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4;
205 	for (i = 64; i < s; i += 64) {
206 		wqe = buf + i;
207 		*wqe = cpu_to_be32(0xffffffff);
208 	}
209 }
210 
211 static void mlx4_ib_handle_qp_event(struct work_struct *_work)
212 {
213 	struct mlx4_ib_qp_event_work *qpe_work =
214 		container_of(_work, struct mlx4_ib_qp_event_work, work);
215 	struct ib_qp *ibqp = &to_mibqp(qpe_work->qp)->ibqp;
216 	struct ib_event event = {};
217 
218 	event.device = ibqp->device;
219 	event.element.qp = ibqp;
220 
221 	switch (qpe_work->type) {
222 	case MLX4_EVENT_TYPE_PATH_MIG:
223 		event.event = IB_EVENT_PATH_MIG;
224 		break;
225 	case MLX4_EVENT_TYPE_COMM_EST:
226 		event.event = IB_EVENT_COMM_EST;
227 		break;
228 	case MLX4_EVENT_TYPE_SQ_DRAINED:
229 		event.event = IB_EVENT_SQ_DRAINED;
230 		break;
231 	case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
232 		event.event = IB_EVENT_QP_LAST_WQE_REACHED;
233 		break;
234 	case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
235 		event.event = IB_EVENT_QP_FATAL;
236 		break;
237 	case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
238 		event.event = IB_EVENT_PATH_MIG_ERR;
239 		break;
240 	case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
241 		event.event = IB_EVENT_QP_REQ_ERR;
242 		break;
243 	case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
244 		event.event = IB_EVENT_QP_ACCESS_ERR;
245 		break;
246 	default:
247 		pr_warn("Unexpected event type %d on QP %06x\n",
248 			qpe_work->type, qpe_work->qp->qpn);
249 		goto out;
250 	}
251 
252 	ibqp->event_handler(&event, ibqp->qp_context);
253 
254 out:
255 	mlx4_put_qp(qpe_work->qp);
256 	kfree(qpe_work);
257 }
258 
259 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
260 {
261 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
262 	struct mlx4_ib_qp_event_work *qpe_work;
263 
264 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
265 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
266 
267 	if (!ibqp->event_handler)
268 		goto out_no_handler;
269 
270 	qpe_work = kzalloc_obj(*qpe_work, GFP_ATOMIC);
271 	if (!qpe_work)
272 		goto out_no_handler;
273 
274 	qpe_work->qp = qp;
275 	qpe_work->type = type;
276 	INIT_WORK(&qpe_work->work, mlx4_ib_handle_qp_event);
277 	queue_work(mlx4_ib_qp_event_wq, &qpe_work->work);
278 	return;
279 
280 out_no_handler:
281 	mlx4_put_qp(qp);
282 }
283 
284 static void mlx4_ib_wq_event(struct mlx4_qp *qp, enum mlx4_event type)
285 {
286 	pr_warn_ratelimited("Unexpected event type %d on WQ 0x%06x. Events are not supported for WQs\n",
287 			    type, qp->qpn);
288 }
289 
290 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
291 {
292 	/*
293 	 * UD WQEs must have a datagram segment.
294 	 * RC and UC WQEs might have a remote address segment.
295 	 * MLX WQEs need two extra inline data segments (for the UD
296 	 * header and space for the ICRC).
297 	 */
298 	switch (type) {
299 	case MLX4_IB_QPT_UD:
300 		return sizeof (struct mlx4_wqe_ctrl_seg) +
301 			sizeof (struct mlx4_wqe_datagram_seg) +
302 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
303 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
304 	case MLX4_IB_QPT_PROXY_SMI:
305 	case MLX4_IB_QPT_PROXY_GSI:
306 		return sizeof (struct mlx4_wqe_ctrl_seg) +
307 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
308 	case MLX4_IB_QPT_TUN_SMI_OWNER:
309 	case MLX4_IB_QPT_TUN_GSI:
310 		return sizeof (struct mlx4_wqe_ctrl_seg) +
311 			sizeof (struct mlx4_wqe_datagram_seg);
312 
313 	case MLX4_IB_QPT_UC:
314 		return sizeof (struct mlx4_wqe_ctrl_seg) +
315 			sizeof (struct mlx4_wqe_raddr_seg);
316 	case MLX4_IB_QPT_RC:
317 		return sizeof (struct mlx4_wqe_ctrl_seg) +
318 			sizeof (struct mlx4_wqe_masked_atomic_seg) +
319 			sizeof (struct mlx4_wqe_raddr_seg);
320 	case MLX4_IB_QPT_SMI:
321 	case MLX4_IB_QPT_GSI:
322 		return sizeof (struct mlx4_wqe_ctrl_seg) +
323 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
324 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
325 					   MLX4_INLINE_ALIGN) *
326 			      sizeof (struct mlx4_wqe_inline_seg),
327 			      sizeof (struct mlx4_wqe_data_seg)) +
328 			ALIGN(4 +
329 			      sizeof (struct mlx4_wqe_inline_seg),
330 			      sizeof (struct mlx4_wqe_data_seg));
331 	default:
332 		return sizeof (struct mlx4_wqe_ctrl_seg);
333 	}
334 }
335 
336 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
337 		       bool is_user, bool has_rq, struct mlx4_ib_qp *qp,
338 		       u32 inl_recv_sz)
339 {
340 	/* Sanity check RQ size before proceeding */
341 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
342 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
343 		return -EINVAL;
344 
345 	if (!has_rq) {
346 		if (cap->max_recv_wr || inl_recv_sz)
347 			return -EINVAL;
348 
349 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
350 	} else {
351 		u32 max_inl_recv_sz = dev->dev->caps.max_rq_sg *
352 			sizeof(struct mlx4_wqe_data_seg);
353 		u32 wqe_size;
354 
355 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
356 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge ||
357 				inl_recv_sz > max_inl_recv_sz))
358 			return -EINVAL;
359 
360 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
361 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
362 		wqe_size = qp->rq.max_gs * sizeof(struct mlx4_wqe_data_seg);
363 		qp->rq.wqe_shift = ilog2(max_t(u32, wqe_size, inl_recv_sz));
364 	}
365 
366 	/* leave userspace return values as they were, so as not to break ABI */
367 	if (is_user) {
368 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
369 		cap->max_recv_sge = qp->rq.max_gs;
370 	} else {
371 		cap->max_recv_wr  = qp->rq.max_post =
372 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
373 		cap->max_recv_sge = min(qp->rq.max_gs,
374 					min(dev->dev->caps.max_sq_sg,
375 					    dev->dev->caps.max_rq_sg));
376 	}
377 
378 	return 0;
379 }
380 
381 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
382 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
383 {
384 	int s;
385 
386 	/* Sanity check SQ size before proceeding */
387 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
388 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
389 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
390 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
391 		return -EINVAL;
392 
393 	/*
394 	 * For MLX transport we need 2 extra S/G entries:
395 	 * one for the header and one for the checksum at the end
396 	 */
397 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
398 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
399 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
400 		return -EINVAL;
401 
402 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
403 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
404 		send_wqe_overhead(type, qp->flags);
405 
406 	if (s > dev->dev->caps.max_sq_desc_sz)
407 		return -EINVAL;
408 
409 	qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
410 
411 	/*
412 	 * We need to leave 2 KB + 1 WR of headroom in the SQ to
413 	 * allow HW to prefetch.
414 	 */
415 	qp->sq_spare_wqes = MLX4_IB_SQ_HEADROOM(qp->sq.wqe_shift);
416 	qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr +
417 					    qp->sq_spare_wqes);
418 
419 	qp->sq.max_gs =
420 		(min(dev->dev->caps.max_sq_desc_sz,
421 		     (1 << qp->sq.wqe_shift)) -
422 		 send_wqe_overhead(type, qp->flags)) /
423 		sizeof (struct mlx4_wqe_data_seg);
424 
425 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
426 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
427 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
428 		qp->rq.offset = 0;
429 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
430 	} else {
431 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
432 		qp->sq.offset = 0;
433 	}
434 
435 	cap->max_send_wr  = qp->sq.max_post =
436 		qp->sq.wqe_cnt - qp->sq_spare_wqes;
437 	cap->max_send_sge = min(qp->sq.max_gs,
438 				min(dev->dev->caps.max_sq_sg,
439 				    dev->dev->caps.max_rq_sg));
440 	/* We don't support inline sends for kernel QPs (yet) */
441 	cap->max_inline_data = 0;
442 
443 	return 0;
444 }
445 
446 static int set_user_sq_size(struct mlx4_ib_dev *dev,
447 			    struct mlx4_ib_qp *qp,
448 			    struct mlx4_ib_create_qp *ucmd)
449 {
450 	u32 cnt;
451 
452 	/* Sanity check SQ size before proceeding */
453 	if (check_shl_overflow(1, ucmd->log_sq_bb_count, &cnt) ||
454 	    cnt > dev->dev->caps.max_wqes)
455 		return -EINVAL;
456 	if (ucmd->log_sq_stride >
457 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
458 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
459 		return -EINVAL;
460 
461 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
462 	qp->sq.wqe_shift = ucmd->log_sq_stride;
463 
464 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
465 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
466 
467 	return 0;
468 }
469 
470 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
471 {
472 	int i;
473 
474 	qp->sqp_proxy_rcv =
475 		kmalloc_objs(struct mlx4_ib_buf, qp->rq.wqe_cnt);
476 	if (!qp->sqp_proxy_rcv)
477 		return -ENOMEM;
478 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
479 		qp->sqp_proxy_rcv[i].addr =
480 			kmalloc_obj(struct mlx4_ib_proxy_sqp_hdr);
481 		if (!qp->sqp_proxy_rcv[i].addr)
482 			goto err;
483 		qp->sqp_proxy_rcv[i].map =
484 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
485 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
486 					  DMA_FROM_DEVICE);
487 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
488 			kfree(qp->sqp_proxy_rcv[i].addr);
489 			goto err;
490 		}
491 	}
492 	return 0;
493 
494 err:
495 	while (i > 0) {
496 		--i;
497 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
498 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
499 				    DMA_FROM_DEVICE);
500 		kfree(qp->sqp_proxy_rcv[i].addr);
501 	}
502 	kfree(qp->sqp_proxy_rcv);
503 	qp->sqp_proxy_rcv = NULL;
504 	return -ENOMEM;
505 }
506 
507 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
508 {
509 	int i;
510 
511 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
512 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
513 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
514 				    DMA_FROM_DEVICE);
515 		kfree(qp->sqp_proxy_rcv[i].addr);
516 	}
517 	kfree(qp->sqp_proxy_rcv);
518 }
519 
520 static bool qp_has_rq(struct ib_qp_init_attr *attr)
521 {
522 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
523 		return false;
524 
525 	return !attr->srq;
526 }
527 
528 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
529 {
530 	int i;
531 	for (i = 0; i < dev->caps.num_ports; i++) {
532 		if (qpn == dev->caps.spec_qps[i].qp0_proxy)
533 			return !!dev->caps.spec_qps[i].qp0_qkey;
534 	}
535 	return 0;
536 }
537 
538 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
539 				    struct mlx4_ib_qp *qp)
540 {
541 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
542 	mlx4_counter_free(dev->dev, qp->counter_index->index);
543 	list_del(&qp->counter_index->list);
544 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
545 
546 	kfree(qp->counter_index);
547 	qp->counter_index = NULL;
548 }
549 
550 static int set_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_rss *rss_ctx,
551 		      struct ib_qp_init_attr *init_attr,
552 		      struct mlx4_ib_create_qp_rss *ucmd)
553 {
554 	rss_ctx->base_qpn_tbl_sz = init_attr->rwq_ind_tbl->ind_tbl[0]->wq_num |
555 		(init_attr->rwq_ind_tbl->log_ind_tbl_size << 24);
556 
557 	if ((ucmd->rx_hash_function == MLX4_IB_RX_HASH_FUNC_TOEPLITZ) &&
558 	    (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS_TOP)) {
559 		memcpy(rss_ctx->rss_key, ucmd->rx_hash_key,
560 		       MLX4_EN_RSS_KEY_SIZE);
561 	} else {
562 		pr_debug("RX Hash function is not supported\n");
563 		return (-EOPNOTSUPP);
564 	}
565 
566 	if (ucmd->rx_hash_fields_mask & ~(u64)(MLX4_IB_RX_HASH_SRC_IPV4	|
567 					       MLX4_IB_RX_HASH_DST_IPV4	|
568 					       MLX4_IB_RX_HASH_SRC_IPV6	|
569 					       MLX4_IB_RX_HASH_DST_IPV6	|
570 					       MLX4_IB_RX_HASH_SRC_PORT_TCP |
571 					       MLX4_IB_RX_HASH_DST_PORT_TCP |
572 					       MLX4_IB_RX_HASH_SRC_PORT_UDP |
573 					       MLX4_IB_RX_HASH_DST_PORT_UDP |
574 					       MLX4_IB_RX_HASH_INNER)) {
575 		pr_debug("RX Hash fields_mask has unsupported mask (0x%llx)\n",
576 			 ucmd->rx_hash_fields_mask);
577 		return (-EOPNOTSUPP);
578 	}
579 
580 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) &&
581 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
582 		rss_ctx->flags = MLX4_RSS_IPV4;
583 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) ||
584 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
585 		pr_debug("RX Hash fields_mask is not supported - both IPv4 SRC and DST must be set\n");
586 		return (-EOPNOTSUPP);
587 	}
588 
589 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) &&
590 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
591 		rss_ctx->flags |= MLX4_RSS_IPV6;
592 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) ||
593 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
594 		pr_debug("RX Hash fields_mask is not supported - both IPv6 SRC and DST must be set\n");
595 		return (-EOPNOTSUPP);
596 	}
597 
598 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) &&
599 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
600 		if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS)) {
601 			pr_debug("RX Hash fields_mask for UDP is not supported\n");
602 			return (-EOPNOTSUPP);
603 		}
604 
605 		if (rss_ctx->flags & MLX4_RSS_IPV4)
606 			rss_ctx->flags |= MLX4_RSS_UDP_IPV4;
607 		if (rss_ctx->flags & MLX4_RSS_IPV6)
608 			rss_ctx->flags |= MLX4_RSS_UDP_IPV6;
609 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
610 			pr_debug("RX Hash fields_mask is not supported - UDP must be set with IPv4 or IPv6\n");
611 			return (-EOPNOTSUPP);
612 		}
613 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) ||
614 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
615 		pr_debug("RX Hash fields_mask is not supported - both UDP SRC and DST must be set\n");
616 		return (-EOPNOTSUPP);
617 	}
618 
619 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) &&
620 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
621 		if (rss_ctx->flags & MLX4_RSS_IPV4)
622 			rss_ctx->flags |= MLX4_RSS_TCP_IPV4;
623 		if (rss_ctx->flags & MLX4_RSS_IPV6)
624 			rss_ctx->flags |= MLX4_RSS_TCP_IPV6;
625 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
626 			pr_debug("RX Hash fields_mask is not supported - TCP must be set with IPv4 or IPv6\n");
627 			return (-EOPNOTSUPP);
628 		}
629 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) ||
630 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
631 		pr_debug("RX Hash fields_mask is not supported - both TCP SRC and DST must be set\n");
632 		return (-EOPNOTSUPP);
633 	}
634 
635 	if (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_INNER) {
636 		if (dev->dev->caps.tunnel_offload_mode ==
637 		    MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
638 			/*
639 			 * Hash according to inner headers if exist, otherwise
640 			 * according to outer headers.
641 			 */
642 			rss_ctx->flags |= MLX4_RSS_BY_INNER_HEADERS_IPONLY;
643 		} else {
644 			pr_debug("RSS Hash for inner headers isn't supported\n");
645 			return (-EOPNOTSUPP);
646 		}
647 	}
648 
649 	return 0;
650 }
651 
652 static int create_qp_rss(struct mlx4_ib_dev *dev,
653 			 struct ib_qp_init_attr *init_attr,
654 			 struct mlx4_ib_create_qp_rss *ucmd,
655 			 struct mlx4_ib_qp *qp)
656 {
657 	int qpn;
658 	int err;
659 
660 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
661 
662 	err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 0, qp->mqp.usage);
663 	if (err)
664 		return err;
665 
666 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
667 	if (err)
668 		goto err_qpn;
669 
670 	INIT_LIST_HEAD(&qp->gid_list);
671 	INIT_LIST_HEAD(&qp->steering_rules);
672 
673 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
674 	qp->state = IB_QPS_RESET;
675 
676 	/* Set dummy send resources to be compatible with HV and PRM */
677 	qp->sq_no_prefetch = 1;
678 	qp->sq.wqe_cnt = 1;
679 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
680 	qp->buf_size = qp->sq.wqe_cnt << MLX4_IB_MIN_SQ_STRIDE;
681 	qp->mtt = (to_mqp(
682 		   (struct ib_qp *)init_attr->rwq_ind_tbl->ind_tbl[0]))->mtt;
683 
684 	qp->rss_ctx = kzalloc_obj(*qp->rss_ctx);
685 	if (!qp->rss_ctx) {
686 		err = -ENOMEM;
687 		goto err_qp_alloc;
688 	}
689 
690 	err = set_qp_rss(dev, qp->rss_ctx, init_attr, ucmd);
691 	if (err)
692 		goto err;
693 
694 	return 0;
695 
696 err:
697 	kfree(qp->rss_ctx);
698 
699 err_qp_alloc:
700 	mlx4_qp_remove(dev->dev, &qp->mqp);
701 	mlx4_qp_free(dev->dev, &qp->mqp);
702 
703 err_qpn:
704 	mlx4_qp_release_range(dev->dev, qpn, 1);
705 	return err;
706 }
707 
708 static int _mlx4_ib_create_qp_rss(struct ib_pd *pd, struct mlx4_ib_qp *qp,
709 				  struct ib_qp_init_attr *init_attr,
710 				  struct ib_udata *udata)
711 {
712 	struct mlx4_ib_create_qp_rss ucmd = {};
713 	size_t required_cmd_sz;
714 	int err;
715 
716 	if (!udata) {
717 		pr_debug("RSS QP with NULL udata\n");
718 		return -EINVAL;
719 	}
720 
721 	if (udata->outlen)
722 		return -EOPNOTSUPP;
723 
724 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) +
725 					sizeof(ucmd.reserved1);
726 	if (udata->inlen < required_cmd_sz) {
727 		pr_debug("invalid inlen\n");
728 		return -EINVAL;
729 	}
730 
731 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
732 		pr_debug("copy failed\n");
733 		return -EFAULT;
734 	}
735 
736 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)))
737 		return -EOPNOTSUPP;
738 
739 	if (ucmd.comp_mask || ucmd.reserved1)
740 		return -EOPNOTSUPP;
741 
742 	if (udata->inlen > sizeof(ucmd) &&
743 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
744 				 udata->inlen - sizeof(ucmd))) {
745 		pr_debug("inlen is not supported\n");
746 		return -EOPNOTSUPP;
747 	}
748 
749 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
750 		pr_debug("RSS QP with unsupported QP type %d\n",
751 			 init_attr->qp_type);
752 		return -EOPNOTSUPP;
753 	}
754 
755 	if (init_attr->create_flags) {
756 		pr_debug("RSS QP doesn't support create flags\n");
757 		return -EOPNOTSUPP;
758 	}
759 
760 	if (init_attr->send_cq || init_attr->cap.max_send_wr) {
761 		pr_debug("RSS QP with unsupported send attributes\n");
762 		return -EOPNOTSUPP;
763 	}
764 
765 	qp->pri.vid = 0xFFFF;
766 	qp->alt.vid = 0xFFFF;
767 
768 	err = create_qp_rss(to_mdev(pd->device), init_attr, &ucmd, qp);
769 	if (err)
770 		return err;
771 
772 	qp->ibqp.qp_num = qp->mqp.qpn;
773 	return 0;
774 }
775 
776 /*
777  * This function allocates a WQN from a range which is consecutive and aligned
778  * to its size. In case the range is full, then it creates a new range and
779  * allocates WQN from it. The new range will be used for following allocations.
780  */
781 static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context,
782 			     struct mlx4_ib_qp *qp, int range_size, int *wqn)
783 {
784 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
785 	struct mlx4_wqn_range *range;
786 	int err = 0;
787 
788 	mutex_lock(&context->wqn_ranges_mutex);
789 
790 	range = list_first_entry_or_null(&context->wqn_ranges_list,
791 					 struct mlx4_wqn_range, list);
792 
793 	if (!range || (range->refcount == range->size) || range->dirty) {
794 		range = kzalloc_obj(*range);
795 		if (!range) {
796 			err = -ENOMEM;
797 			goto out;
798 		}
799 
800 		err = mlx4_qp_reserve_range(dev->dev, range_size,
801 					    range_size, &range->base_wqn, 0,
802 					    qp->mqp.usage);
803 		if (err) {
804 			kfree(range);
805 			goto out;
806 		}
807 
808 		range->size = range_size;
809 		list_add(&range->list, &context->wqn_ranges_list);
810 	} else if (range_size != 1) {
811 		/*
812 		 * Requesting a new range (>1) when last range is still open, is
813 		 * not valid.
814 		 */
815 		err = -EINVAL;
816 		goto out;
817 	}
818 
819 	qp->wqn_range = range;
820 
821 	*wqn = range->base_wqn + range->refcount;
822 
823 	range->refcount++;
824 
825 out:
826 	mutex_unlock(&context->wqn_ranges_mutex);
827 
828 	return err;
829 }
830 
831 static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context,
832 				struct mlx4_ib_qp *qp, bool dirty_release)
833 {
834 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
835 	struct mlx4_wqn_range *range;
836 
837 	mutex_lock(&context->wqn_ranges_mutex);
838 
839 	range = qp->wqn_range;
840 
841 	range->refcount--;
842 	if (!range->refcount) {
843 		mlx4_qp_release_range(dev->dev, range->base_wqn,
844 				      range->size);
845 		list_del(&range->list);
846 		kfree(range);
847 	} else if (dirty_release) {
848 	/*
849 	 * A range which one of its WQNs is destroyed, won't be able to be
850 	 * reused for further WQN allocations.
851 	 * The next created WQ will allocate a new range.
852 	 */
853 		range->dirty = true;
854 	}
855 
856 	mutex_unlock(&context->wqn_ranges_mutex);
857 }
858 
859 static int create_rq(struct ib_pd *pd, struct ib_qp_init_attr *init_attr,
860 		     struct ib_udata *udata, struct mlx4_ib_qp *qp)
861 {
862 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
863 	int qpn;
864 	int err;
865 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
866 		udata, struct mlx4_ib_ucontext, ibucontext);
867 	struct mlx4_ib_cq *mcq;
868 	unsigned long flags;
869 	int range_size;
870 	struct mlx4_ib_create_wq wq;
871 	size_t copy_len;
872 	int shift;
873 	int n;
874 
875 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
876 
877 	spin_lock_init(&qp->sq.lock);
878 	spin_lock_init(&qp->rq.lock);
879 	INIT_LIST_HEAD(&qp->gid_list);
880 	INIT_LIST_HEAD(&qp->steering_rules);
881 
882 	qp->state = IB_QPS_RESET;
883 
884 	copy_len = min(sizeof(struct mlx4_ib_create_wq), udata->inlen);
885 
886 	if (ib_copy_from_udata(&wq, udata, copy_len)) {
887 		err = -EFAULT;
888 		goto err;
889 	}
890 
891 	if (wq.comp_mask || wq.reserved[0] || wq.reserved[1] ||
892 	    wq.reserved[2]) {
893 		pr_debug("user command isn't supported\n");
894 		err = -EOPNOTSUPP;
895 		goto err;
896 	}
897 
898 	if (wq.log_range_size > ilog2(dev->dev->caps.max_rss_tbl_sz)) {
899 		pr_debug("WQN range size must be equal or smaller than %d\n",
900 			 dev->dev->caps.max_rss_tbl_sz);
901 		err = -EOPNOTSUPP;
902 		goto err;
903 	}
904 	range_size = 1 << wq.log_range_size;
905 
906 	if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS)
907 		qp->flags |= MLX4_IB_QP_SCATTER_FCS;
908 
909 	err = set_rq_size(dev, &init_attr->cap, true, true, qp, qp->inl_recv_sz);
910 	if (err)
911 		goto err;
912 
913 	qp->sq_no_prefetch = 1;
914 	qp->sq.wqe_cnt = 1;
915 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
916 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
917 		       (qp->sq.wqe_cnt << qp->sq.wqe_shift);
918 
919 	qp->umem = ib_umem_get(pd->device, wq.buf_addr, qp->buf_size, 0);
920 	if (IS_ERR(qp->umem)) {
921 		err = PTR_ERR(qp->umem);
922 		goto err;
923 	}
924 
925 	shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
926 	if (shift < 0) {
927 		err = shift;
928 		goto err_buf;
929 	}
930 
931 	err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
932 	if (err)
933 		goto err_buf;
934 
935 	err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
936 	if (err)
937 		goto err_mtt;
938 
939 	err = mlx4_ib_db_map_user(udata, wq.db_addr, &qp->db);
940 	if (err)
941 		goto err_mtt;
942 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
943 
944 	err = mlx4_ib_alloc_wqn(context, qp, range_size, &qpn);
945 	if (err)
946 		goto err_wrid;
947 
948 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
949 	if (err)
950 		goto err_qpn;
951 
952 	/*
953 	 * Hardware wants QPN written in big-endian order (after
954 	 * shifting) for send doorbell.  Precompute this value to save
955 	 * a little bit when posting sends.
956 	 */
957 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
958 
959 	qp->mqp.event = mlx4_ib_wq_event;
960 
961 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
962 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
963 			 to_mcq(init_attr->recv_cq));
964 	/* Maintain device to QPs access, needed for further handling
965 	 * via reset flow
966 	 */
967 	list_add_tail(&qp->qps_list, &dev->qp_list);
968 	/* Maintain CQ to QPs access, needed for further handling
969 	 * via reset flow
970 	 */
971 	mcq = to_mcq(init_attr->send_cq);
972 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
973 	mcq = to_mcq(init_attr->recv_cq);
974 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
975 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
976 			   to_mcq(init_attr->recv_cq));
977 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
978 	return 0;
979 
980 err_qpn:
981 	mlx4_ib_release_wqn(context, qp, 0);
982 err_wrid:
983 	mlx4_ib_db_unmap_user(context, &qp->db);
984 
985 err_mtt:
986 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
987 err_buf:
988 	ib_umem_release(qp->umem);
989 err:
990 	return err;
991 }
992 
993 static int create_qp_common(struct ib_pd *pd, struct ib_qp_init_attr *init_attr,
994 			    struct ib_udata *udata, int sqpn,
995 			    struct mlx4_ib_qp *qp)
996 {
997 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
998 	int qpn;
999 	int err;
1000 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
1001 		udata, struct mlx4_ib_ucontext, ibucontext);
1002 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
1003 	struct mlx4_ib_cq *mcq;
1004 	unsigned long flags;
1005 
1006 	/* When tunneling special qps, we use a plain UD qp */
1007 	if (sqpn) {
1008 		if (mlx4_is_mfunc(dev->dev) &&
1009 		    (!mlx4_is_master(dev->dev) ||
1010 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
1011 			if (init_attr->qp_type == IB_QPT_GSI)
1012 				qp_type = MLX4_IB_QPT_PROXY_GSI;
1013 			else {
1014 				if (mlx4_is_master(dev->dev) ||
1015 				    qp0_enabled_vf(dev->dev, sqpn))
1016 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
1017 				else
1018 					qp_type = MLX4_IB_QPT_PROXY_SMI;
1019 			}
1020 		}
1021 		qpn = sqpn;
1022 		/* add extra sg entry for tunneling */
1023 		init_attr->cap.max_recv_sge++;
1024 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
1025 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
1026 			container_of(init_attr,
1027 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
1028 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
1029 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
1030 		    !mlx4_is_master(dev->dev))
1031 			return -EINVAL;
1032 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
1033 			qp_type = MLX4_IB_QPT_TUN_GSI;
1034 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
1035 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
1036 					     tnl_init->port))
1037 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
1038 		else
1039 			qp_type = MLX4_IB_QPT_TUN_SMI;
1040 		/* we are definitely in the PPF here, since we are creating
1041 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
1042 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
1043 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
1044 		sqpn = qpn;
1045 	}
1046 
1047 	if (init_attr->qp_type == IB_QPT_SMI ||
1048 	    init_attr->qp_type == IB_QPT_GSI || qp_type == MLX4_IB_QPT_SMI ||
1049 	    qp_type == MLX4_IB_QPT_GSI ||
1050 	    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
1051 			MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
1052 		qp->sqp = kzalloc_obj(struct mlx4_ib_sqp);
1053 		if (!qp->sqp)
1054 			return -ENOMEM;
1055 	}
1056 
1057 	qp->mlx4_ib_qp_type = qp_type;
1058 
1059 	spin_lock_init(&qp->sq.lock);
1060 	spin_lock_init(&qp->rq.lock);
1061 	INIT_LIST_HEAD(&qp->gid_list);
1062 	INIT_LIST_HEAD(&qp->steering_rules);
1063 
1064 	qp->state = IB_QPS_RESET;
1065 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1066 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1067 
1068 	if (udata) {
1069 		struct mlx4_ib_create_qp ucmd;
1070 		size_t copy_len;
1071 		int shift;
1072 		int n;
1073 
1074 		copy_len = sizeof(struct mlx4_ib_create_qp);
1075 
1076 		if (ib_copy_from_udata(&ucmd, udata, copy_len)) {
1077 			err = -EFAULT;
1078 			goto err;
1079 		}
1080 
1081 		qp->inl_recv_sz = ucmd.inl_recv_sz;
1082 
1083 		if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1084 			if (!(dev->dev->caps.flags &
1085 			      MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
1086 				pr_debug("scatter FCS is unsupported\n");
1087 				err = -EOPNOTSUPP;
1088 				goto err;
1089 			}
1090 
1091 			qp->flags |= MLX4_IB_QP_SCATTER_FCS;
1092 		}
1093 
1094 		err = set_rq_size(dev, &init_attr->cap, udata,
1095 				  qp_has_rq(init_attr), qp, qp->inl_recv_sz);
1096 		if (err)
1097 			goto err;
1098 
1099 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
1100 
1101 		err = set_user_sq_size(dev, qp, &ucmd);
1102 		if (err)
1103 			goto err;
1104 
1105 		qp->umem =
1106 			ib_umem_get(pd->device, ucmd.buf_addr, qp->buf_size, 0);
1107 		if (IS_ERR(qp->umem)) {
1108 			err = PTR_ERR(qp->umem);
1109 			goto err;
1110 		}
1111 
1112 		shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
1113 		if (shift < 0) {
1114 			err = shift;
1115 			goto err_buf;
1116 		}
1117 
1118 		err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
1119 		if (err)
1120 			goto err_buf;
1121 
1122 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
1123 		if (err)
1124 			goto err_mtt;
1125 
1126 		if (qp_has_rq(init_attr)) {
1127 			err = mlx4_ib_db_map_user(udata, ucmd.db_addr, &qp->db);
1128 			if (err)
1129 				goto err_mtt;
1130 		}
1131 		qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
1132 	} else {
1133 		err = set_rq_size(dev, &init_attr->cap, udata,
1134 				  qp_has_rq(init_attr), qp, 0);
1135 		if (err)
1136 			goto err;
1137 
1138 		qp->sq_no_prefetch = 0;
1139 
1140 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
1141 			qp->flags |= MLX4_IB_QP_LSO;
1142 
1143 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1144 			if (dev->steering_support ==
1145 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
1146 				qp->flags |= MLX4_IB_QP_NETIF;
1147 			else {
1148 				err = -EINVAL;
1149 				goto err;
1150 			}
1151 		}
1152 
1153 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
1154 		if (err)
1155 			goto err;
1156 
1157 		if (qp_has_rq(init_attr)) {
1158 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
1159 			if (err)
1160 				goto err;
1161 
1162 			*qp->db.db = 0;
1163 		}
1164 
1165 		if (mlx4_buf_alloc(dev->dev, qp->buf_size,  PAGE_SIZE * 2,
1166 				   &qp->buf)) {
1167 			err = -ENOMEM;
1168 			goto err_db;
1169 		}
1170 
1171 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
1172 				    &qp->mtt);
1173 		if (err)
1174 			goto err_buf;
1175 
1176 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
1177 		if (err)
1178 			goto err_mtt;
1179 
1180 		qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1181 					     sizeof(u64), GFP_KERNEL);
1182 		qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1183 					     sizeof(u64), GFP_KERNEL);
1184 		if (!qp->sq.wrid || !qp->rq.wrid) {
1185 			err = -ENOMEM;
1186 			goto err_wrid;
1187 		}
1188 		qp->mqp.usage = MLX4_RES_USAGE_DRIVER;
1189 	}
1190 
1191 	if (sqpn) {
1192 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1193 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
1194 			if (alloc_proxy_bufs(pd->device, qp)) {
1195 				err = -ENOMEM;
1196 				goto err_wrid;
1197 			}
1198 		}
1199 	} else {
1200 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
1201 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
1202 		 * VLAN insertion. */
1203 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
1204 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
1205 						    (init_attr->cap.max_send_wr ?
1206 						     MLX4_RESERVE_ETH_BF_QP : 0) |
1207 						    (init_attr->cap.max_recv_wr ?
1208 						     MLX4_RESERVE_A0_QP : 0),
1209 						    qp->mqp.usage);
1210 		else
1211 			if (qp->flags & MLX4_IB_QP_NETIF)
1212 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
1213 			else
1214 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
1215 							    &qpn, 0, qp->mqp.usage);
1216 		if (err)
1217 			goto err_proxy;
1218 	}
1219 
1220 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1221 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1222 
1223 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
1224 	if (err)
1225 		goto err_qpn;
1226 
1227 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1228 		qp->mqp.qpn |= (1 << 23);
1229 
1230 	/*
1231 	 * Hardware wants QPN written in big-endian order (after
1232 	 * shifting) for send doorbell.  Precompute this value to save
1233 	 * a little bit when posting sends.
1234 	 */
1235 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1236 
1237 	qp->mqp.event = mlx4_ib_qp_event;
1238 
1239 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1240 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
1241 			 to_mcq(init_attr->recv_cq));
1242 	/* Maintain device to QPs access, needed for further handling
1243 	 * via reset flow
1244 	 */
1245 	list_add_tail(&qp->qps_list, &dev->qp_list);
1246 	/* Maintain CQ to QPs access, needed for further handling
1247 	 * via reset flow
1248 	 */
1249 	mcq = to_mcq(init_attr->send_cq);
1250 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
1251 	mcq = to_mcq(init_attr->recv_cq);
1252 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
1253 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
1254 			   to_mcq(init_attr->recv_cq));
1255 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1256 	return 0;
1257 
1258 err_qpn:
1259 	if (!sqpn) {
1260 		if (qp->flags & MLX4_IB_QP_NETIF)
1261 			mlx4_ib_steer_qp_free(dev, qpn, 1);
1262 		else
1263 			mlx4_qp_release_range(dev->dev, qpn, 1);
1264 	}
1265 err_proxy:
1266 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1267 		free_proxy_bufs(pd->device, qp);
1268 err_wrid:
1269 	if (udata) {
1270 		if (qp_has_rq(init_attr))
1271 			mlx4_ib_db_unmap_user(context, &qp->db);
1272 	} else {
1273 		kvfree(qp->sq.wrid);
1274 		kvfree(qp->rq.wrid);
1275 	}
1276 
1277 err_mtt:
1278 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1279 
1280 err_buf:
1281 	if (!qp->umem)
1282 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1283 	ib_umem_release(qp->umem);
1284 
1285 err_db:
1286 	if (!udata && qp_has_rq(init_attr))
1287 		mlx4_db_free(dev->dev, &qp->db);
1288 
1289 err:
1290 	kfree(qp->sqp);
1291 	return err;
1292 }
1293 
1294 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
1295 {
1296 	switch (state) {
1297 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
1298 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
1299 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
1300 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
1301 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
1302 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
1303 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
1304 	default:		return -1;
1305 	}
1306 }
1307 
1308 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1309 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1310 {
1311 	if (send_cq == recv_cq) {
1312 		spin_lock(&send_cq->lock);
1313 		__acquire(&recv_cq->lock);
1314 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1315 		spin_lock(&send_cq->lock);
1316 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1317 	} else {
1318 		spin_lock(&recv_cq->lock);
1319 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1320 	}
1321 }
1322 
1323 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1324 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1325 {
1326 	if (send_cq == recv_cq) {
1327 		__release(&recv_cq->lock);
1328 		spin_unlock(&send_cq->lock);
1329 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1330 		spin_unlock(&recv_cq->lock);
1331 		spin_unlock(&send_cq->lock);
1332 	} else {
1333 		spin_unlock(&send_cq->lock);
1334 		spin_unlock(&recv_cq->lock);
1335 	}
1336 }
1337 
1338 static void del_gid_entries(struct mlx4_ib_qp *qp)
1339 {
1340 	struct mlx4_ib_gid_entry *ge, *tmp;
1341 
1342 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1343 		list_del(&ge->list);
1344 		kfree(ge);
1345 	}
1346 }
1347 
1348 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
1349 {
1350 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
1351 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
1352 	else
1353 		return to_mpd(qp->ibqp.pd);
1354 }
1355 
1356 static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src,
1357 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1358 {
1359 	switch (qp->ibqp.qp_type) {
1360 	case IB_QPT_XRC_TGT:
1361 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1362 		*recv_cq = *send_cq;
1363 		break;
1364 	case IB_QPT_XRC_INI:
1365 		*send_cq = to_mcq(qp->ibqp.send_cq);
1366 		*recv_cq = *send_cq;
1367 		break;
1368 	default:
1369 		*recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) :
1370 						     to_mcq(qp->ibwq.cq);
1371 		*send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) :
1372 						     *recv_cq;
1373 		break;
1374 	}
1375 }
1376 
1377 static void destroy_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1378 {
1379 	if (qp->state != IB_QPS_RESET) {
1380 		int i;
1381 
1382 		for (i = 0; i < (1 << qp->ibqp.rwq_ind_tbl->log_ind_tbl_size);
1383 		     i++) {
1384 			struct ib_wq *ibwq = qp->ibqp.rwq_ind_tbl->ind_tbl[i];
1385 			struct mlx4_ib_qp *wq =	to_mqp((struct ib_qp *)ibwq);
1386 
1387 			mutex_lock(&wq->mutex);
1388 
1389 			wq->rss_usecnt--;
1390 
1391 			mutex_unlock(&wq->mutex);
1392 		}
1393 
1394 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1395 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1396 			pr_warn("modify QP %06x to RESET failed.\n",
1397 				qp->mqp.qpn);
1398 	}
1399 
1400 	mlx4_qp_remove(dev->dev, &qp->mqp);
1401 	mlx4_qp_free(dev->dev, &qp->mqp);
1402 	mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1403 	del_gid_entries(qp);
1404 }
1405 
1406 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1407 			      enum mlx4_ib_source_type src,
1408 			      struct ib_udata *udata)
1409 {
1410 	struct mlx4_ib_cq *send_cq, *recv_cq;
1411 	unsigned long flags;
1412 
1413 	if (qp->state != IB_QPS_RESET) {
1414 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1415 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1416 			pr_warn("modify QP %06x to RESET failed.\n",
1417 			       qp->mqp.qpn);
1418 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1419 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1420 			qp->pri.smac = 0;
1421 			qp->pri.smac_port = 0;
1422 		}
1423 		if (qp->alt.smac) {
1424 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1425 			qp->alt.smac = 0;
1426 		}
1427 		if (qp->pri.vid < 0x1000) {
1428 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1429 			qp->pri.vid = 0xFFFF;
1430 			qp->pri.candidate_vid = 0xFFFF;
1431 			qp->pri.update_vid = 0;
1432 		}
1433 		if (qp->alt.vid < 0x1000) {
1434 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1435 			qp->alt.vid = 0xFFFF;
1436 			qp->alt.candidate_vid = 0xFFFF;
1437 			qp->alt.update_vid = 0;
1438 		}
1439 	}
1440 
1441 	get_cqs(qp, src, &send_cq, &recv_cq);
1442 
1443 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1444 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1445 
1446 	/* del from lists under both locks above to protect reset flow paths */
1447 	list_del(&qp->qps_list);
1448 	list_del(&qp->cq_send_list);
1449 	list_del(&qp->cq_recv_list);
1450 	if (!udata) {
1451 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1452 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1453 		if (send_cq != recv_cq)
1454 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1455 	}
1456 
1457 	mlx4_qp_remove(dev->dev, &qp->mqp);
1458 
1459 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1460 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1461 
1462 	mlx4_qp_free(dev->dev, &qp->mqp);
1463 
1464 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1465 		if (qp->flags & MLX4_IB_QP_NETIF)
1466 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1467 		else if (src == MLX4_IB_RWQ_SRC)
1468 			mlx4_ib_release_wqn(
1469 				rdma_udata_to_drv_context(
1470 					udata,
1471 					struct mlx4_ib_ucontext,
1472 					ibucontext),
1473 				qp, 1);
1474 		else
1475 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1476 	}
1477 
1478 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1479 
1480 	if (udata) {
1481 		if (qp->rq.wqe_cnt) {
1482 			struct mlx4_ib_ucontext *mcontext =
1483 				rdma_udata_to_drv_context(
1484 					udata,
1485 					struct mlx4_ib_ucontext,
1486 					ibucontext);
1487 
1488 			mlx4_ib_db_unmap_user(mcontext, &qp->db);
1489 		}
1490 	} else {
1491 		kvfree(qp->sq.wrid);
1492 		kvfree(qp->rq.wrid);
1493 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1494 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1495 			free_proxy_bufs(&dev->ib_dev, qp);
1496 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1497 		if (qp->rq.wqe_cnt)
1498 			mlx4_db_free(dev->dev, &qp->db);
1499 	}
1500 	ib_umem_release(qp->umem);
1501 
1502 	del_gid_entries(qp);
1503 }
1504 
1505 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1506 {
1507 	/* Native or PPF */
1508 	if (!mlx4_is_mfunc(dev->dev) ||
1509 	    (mlx4_is_master(dev->dev) &&
1510 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1511 		return  dev->dev->phys_caps.base_sqpn +
1512 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1513 			attr->port_num - 1;
1514 	}
1515 	/* PF or VF -- creating proxies */
1516 	if (attr->qp_type == IB_QPT_SMI)
1517 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp0_proxy;
1518 	else
1519 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp1_proxy;
1520 }
1521 
1522 static int _mlx4_ib_create_qp(struct ib_pd *pd, struct mlx4_ib_qp *qp,
1523 			      struct ib_qp_init_attr *init_attr,
1524 			      struct ib_udata *udata)
1525 {
1526 	int err;
1527 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1528 	u16 xrcdn = 0;
1529 
1530 	if (init_attr->rwq_ind_tbl)
1531 		return _mlx4_ib_create_qp_rss(pd, qp, init_attr, udata);
1532 
1533 	/*
1534 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1535 	 * and only for kernel UD QPs.
1536 	 */
1537 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1538 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1539 					MLX4_IB_SRIOV_TUNNEL_QP |
1540 					MLX4_IB_SRIOV_SQP |
1541 					MLX4_IB_QP_NETIF |
1542 					MLX4_IB_QP_CREATE_ROCE_V2_GSI))
1543 		return -EOPNOTSUPP;
1544 
1545 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1546 		if (init_attr->qp_type != IB_QPT_UD)
1547 			return -EINVAL;
1548 	}
1549 
1550 	if (init_attr->create_flags) {
1551 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1552 			return -EINVAL;
1553 
1554 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1555 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1556 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1557 		     init_attr->qp_type != IB_QPT_UD) ||
1558 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1559 		     init_attr->qp_type > IB_QPT_GSI) ||
1560 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1561 		     init_attr->qp_type != IB_QPT_GSI))
1562 			return -EINVAL;
1563 	}
1564 
1565 	switch (init_attr->qp_type) {
1566 	case IB_QPT_XRC_TGT:
1567 		pd = to_mxrcd(init_attr->xrcd)->pd;
1568 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1569 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1570 		fallthrough;
1571 	case IB_QPT_XRC_INI:
1572 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1573 			return -ENOSYS;
1574 		init_attr->recv_cq = init_attr->send_cq;
1575 		fallthrough;
1576 	case IB_QPT_RC:
1577 	case IB_QPT_UC:
1578 	case IB_QPT_RAW_PACKET:
1579 	case IB_QPT_UD:
1580 		qp->pri.vid = 0xFFFF;
1581 		qp->alt.vid = 0xFFFF;
1582 		err = create_qp_common(pd, init_attr, udata, 0, qp);
1583 		if (err)
1584 			return err;
1585 
1586 		qp->ibqp.qp_num = qp->mqp.qpn;
1587 		qp->xrcdn = xrcdn;
1588 		break;
1589 	case IB_QPT_SMI:
1590 	case IB_QPT_GSI:
1591 	{
1592 		int sqpn;
1593 
1594 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1595 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev,
1596 							1, 1, &sqpn, 0,
1597 							MLX4_RES_USAGE_DRIVER);
1598 
1599 			if (res)
1600 				return res;
1601 		} else {
1602 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1603 		}
1604 
1605 		qp->pri.vid = 0xFFFF;
1606 		qp->alt.vid = 0xFFFF;
1607 		err = create_qp_common(pd, init_attr, udata, sqpn, qp);
1608 		if (err)
1609 			return err;
1610 
1611 		if (init_attr->create_flags &
1612 		    (MLX4_IB_SRIOV_SQP | MLX4_IB_SRIOV_TUNNEL_QP))
1613 			/* Internal QP created with ib_create_qp */
1614 			rdma_restrack_no_track(&qp->ibqp.res);
1615 
1616 		qp->port	= init_attr->port_num;
1617 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1618 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1619 		break;
1620 	}
1621 	default:
1622 		/* Don't support raw QPs */
1623 		return -EOPNOTSUPP;
1624 	}
1625 	return 0;
1626 }
1627 
1628 int mlx4_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr,
1629 		      struct ib_udata *udata)
1630 {
1631 	struct ib_device *device = ibqp->device;
1632 	struct mlx4_ib_dev *dev = to_mdev(device);
1633 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1634 	struct ib_pd *pd = ibqp->pd;
1635 	int ret;
1636 
1637 	mutex_init(&qp->mutex);
1638 	ret = _mlx4_ib_create_qp(pd, qp, init_attr, udata);
1639 	if (ret)
1640 		return ret;
1641 
1642 	if (init_attr->qp_type == IB_QPT_GSI &&
1643 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1644 		struct mlx4_ib_sqp *sqp = qp->sqp;
1645 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1646 
1647 		if (is_eth &&
1648 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1649 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1650 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1651 
1652 			if (IS_ERR(sqp->roce_v2_gsi)) {
1653 				pr_err("Failed to create GSI QP for RoCEv2 (%pe)\n",
1654 				       sqp->roce_v2_gsi);
1655 				sqp->roce_v2_gsi = NULL;
1656 			} else {
1657 				to_mqp(sqp->roce_v2_gsi)->flags |=
1658 					MLX4_IB_ROCE_V2_GSI_QP;
1659 			}
1660 
1661 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1662 		}
1663 	}
1664 	return 0;
1665 }
1666 
1667 static int _mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
1668 {
1669 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1670 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1671 
1672 	if (is_qp0(dev, mqp))
1673 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1674 
1675 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1676 	    dev->qp1_proxy[mqp->port - 1] == mqp) {
1677 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1678 		dev->qp1_proxy[mqp->port - 1] = NULL;
1679 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1680 	}
1681 
1682 	if (mqp->counter_index)
1683 		mlx4_ib_free_qp_counter(dev, mqp);
1684 
1685 	if (qp->rwq_ind_tbl) {
1686 		destroy_qp_rss(dev, mqp);
1687 	} else {
1688 		destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, udata);
1689 	}
1690 
1691 	kfree(mqp->sqp);
1692 	return 0;
1693 }
1694 
1695 int mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
1696 {
1697 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1698 
1699 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1700 		struct mlx4_ib_sqp *sqp = mqp->sqp;
1701 
1702 		if (sqp->roce_v2_gsi)
1703 			ib_destroy_qp(sqp->roce_v2_gsi);
1704 	}
1705 
1706 	return _mlx4_ib_destroy_qp(qp, udata);
1707 }
1708 
1709 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1710 {
1711 	switch (type) {
1712 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1713 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1714 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1715 	case MLX4_IB_QPT_XRC_INI:
1716 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1717 	case MLX4_IB_QPT_SMI:
1718 	case MLX4_IB_QPT_GSI:
1719 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1720 
1721 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1722 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1723 						MLX4_QP_ST_MLX : -1);
1724 	case MLX4_IB_QPT_PROXY_SMI:
1725 	case MLX4_IB_QPT_TUN_SMI:
1726 	case MLX4_IB_QPT_PROXY_GSI:
1727 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1728 						MLX4_QP_ST_UD : -1);
1729 	default:			return -1;
1730 	}
1731 }
1732 
1733 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1734 				   int attr_mask)
1735 {
1736 	u8 dest_rd_atomic;
1737 	u32 access_flags;
1738 	u32 hw_access_flags = 0;
1739 
1740 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1741 		dest_rd_atomic = attr->max_dest_rd_atomic;
1742 	else
1743 		dest_rd_atomic = qp->resp_depth;
1744 
1745 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1746 		access_flags = attr->qp_access_flags;
1747 	else
1748 		access_flags = qp->atomic_rd_en;
1749 
1750 	if (!dest_rd_atomic)
1751 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1752 
1753 	if (access_flags & IB_ACCESS_REMOTE_READ)
1754 		hw_access_flags |= MLX4_QP_BIT_RRE;
1755 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1756 		hw_access_flags |= MLX4_QP_BIT_RAE;
1757 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1758 		hw_access_flags |= MLX4_QP_BIT_RWE;
1759 
1760 	return cpu_to_be32(hw_access_flags);
1761 }
1762 
1763 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1764 			    int attr_mask)
1765 {
1766 	if (attr_mask & IB_QP_PKEY_INDEX)
1767 		sqp->pkey_index = attr->pkey_index;
1768 	if (attr_mask & IB_QP_QKEY)
1769 		sqp->qkey = attr->qkey;
1770 	if (attr_mask & IB_QP_SQ_PSN)
1771 		sqp->send_psn = attr->sq_psn;
1772 }
1773 
1774 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1775 {
1776 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1777 }
1778 
1779 static int _mlx4_set_path(struct mlx4_ib_dev *dev,
1780 			  const struct rdma_ah_attr *ah,
1781 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1782 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1783 {
1784 	int vidx;
1785 	int smac_index;
1786 	int err;
1787 
1788 	path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
1789 	path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
1790 	if (rdma_ah_get_static_rate(ah)) {
1791 		path->static_rate = rdma_ah_get_static_rate(ah) +
1792 				    MLX4_STAT_RATE_OFFSET;
1793 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1794 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1795 			--path->static_rate;
1796 	} else
1797 		path->static_rate = 0;
1798 
1799 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
1800 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
1801 		int real_sgid_index =
1802 			mlx4_ib_gid_index_to_real_index(dev, grh->sgid_attr);
1803 
1804 		if (real_sgid_index < 0)
1805 			return real_sgid_index;
1806 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1807 			pr_err("sgid_index (%u) too large. max is %d\n",
1808 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1809 			return -1;
1810 		}
1811 
1812 		path->grh_mylmc |= 1 << 7;
1813 		path->mgid_index = real_sgid_index;
1814 		path->hop_limit  = grh->hop_limit;
1815 		path->tclass_flowlabel =
1816 			cpu_to_be32((grh->traffic_class << 20) |
1817 				    (grh->flow_label));
1818 		memcpy(path->rgid, grh->dgid.raw, 16);
1819 	}
1820 
1821 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
1822 		if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
1823 			return -1;
1824 
1825 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1826 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
1827 
1828 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1829 		if (vlan_tag < 0x1000) {
1830 			if (smac_info->vid < 0x1000) {
1831 				/* both valid vlan ids */
1832 				if (smac_info->vid != vlan_tag) {
1833 					/* different VIDs.  unreg old and reg new */
1834 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1835 					if (err)
1836 						return err;
1837 					smac_info->candidate_vid = vlan_tag;
1838 					smac_info->candidate_vlan_index = vidx;
1839 					smac_info->candidate_vlan_port = port;
1840 					smac_info->update_vid = 1;
1841 					path->vlan_index = vidx;
1842 				} else {
1843 					path->vlan_index = smac_info->vlan_index;
1844 				}
1845 			} else {
1846 				/* no current vlan tag in qp */
1847 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1848 				if (err)
1849 					return err;
1850 				smac_info->candidate_vid = vlan_tag;
1851 				smac_info->candidate_vlan_index = vidx;
1852 				smac_info->candidate_vlan_port = port;
1853 				smac_info->update_vid = 1;
1854 				path->vlan_index = vidx;
1855 			}
1856 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1857 			path->fl = 1 << 6;
1858 		} else {
1859 			/* have current vlan tag. unregister it at modify-qp success */
1860 			if (smac_info->vid < 0x1000) {
1861 				smac_info->candidate_vid = 0xFFFF;
1862 				smac_info->update_vid = 1;
1863 			}
1864 		}
1865 
1866 		/* get smac_index for RoCE use.
1867 		 * If no smac was yet assigned, register one.
1868 		 * If one was already assigned, but the new mac differs,
1869 		 * unregister the old one and register the new one.
1870 		*/
1871 		if ((!smac_info->smac && !smac_info->smac_port) ||
1872 		    smac_info->smac != smac) {
1873 			/* register candidate now, unreg if needed, after success */
1874 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1875 			if (smac_index >= 0) {
1876 				smac_info->candidate_smac_index = smac_index;
1877 				smac_info->candidate_smac = smac;
1878 				smac_info->candidate_smac_port = port;
1879 			} else {
1880 				return -EINVAL;
1881 			}
1882 		} else {
1883 			smac_index = smac_info->smac_index;
1884 		}
1885 		memcpy(path->dmac, ah->roce.dmac, 6);
1886 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1887 		/* put MAC table smac index for IBoE */
1888 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1889 	} else {
1890 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1891 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
1892 	}
1893 
1894 	return 0;
1895 }
1896 
1897 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1898 			 enum ib_qp_attr_mask qp_attr_mask,
1899 			 struct mlx4_ib_qp *mqp,
1900 			 struct mlx4_qp_path *path, u8 port,
1901 			 u16 vlan_id, u8 *smac)
1902 {
1903 	return _mlx4_set_path(dev, &qp->ah_attr,
1904 			      ether_addr_to_u64(smac),
1905 			      vlan_id,
1906 			      path, &mqp->pri, port);
1907 }
1908 
1909 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1910 			     const struct ib_qp_attr *qp,
1911 			     enum ib_qp_attr_mask qp_attr_mask,
1912 			     struct mlx4_ib_qp *mqp,
1913 			     struct mlx4_qp_path *path, u8 port)
1914 {
1915 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1916 			      0,
1917 			      0xffff,
1918 			      path, &mqp->alt, port);
1919 }
1920 
1921 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1922 {
1923 	struct mlx4_ib_gid_entry *ge, *tmp;
1924 
1925 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1926 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1927 			ge->added = 1;
1928 			ge->port = qp->port;
1929 		}
1930 	}
1931 }
1932 
1933 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1934 				    struct mlx4_ib_qp *qp,
1935 				    struct mlx4_qp_context *context)
1936 {
1937 	u64 u64_mac;
1938 	int smac_index;
1939 
1940 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1941 
1942 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1943 	if (!qp->pri.smac && !qp->pri.smac_port) {
1944 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1945 		if (smac_index >= 0) {
1946 			qp->pri.candidate_smac_index = smac_index;
1947 			qp->pri.candidate_smac = u64_mac;
1948 			qp->pri.candidate_smac_port = qp->port;
1949 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1950 		} else {
1951 			return -ENOENT;
1952 		}
1953 	}
1954 	return 0;
1955 }
1956 
1957 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1958 {
1959 	struct counter_index *new_counter_index;
1960 	int err;
1961 	u32 tmp_idx;
1962 
1963 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1964 	    IB_LINK_LAYER_ETHERNET ||
1965 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1966 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1967 		return 0;
1968 
1969 	err = mlx4_counter_alloc(dev->dev, &tmp_idx, MLX4_RES_USAGE_DRIVER);
1970 	if (err)
1971 		return err;
1972 
1973 	new_counter_index = kmalloc_obj(*new_counter_index);
1974 	if (!new_counter_index) {
1975 		mlx4_counter_free(dev->dev, tmp_idx);
1976 		return -ENOMEM;
1977 	}
1978 
1979 	new_counter_index->index = tmp_idx;
1980 	new_counter_index->allocated = 1;
1981 	qp->counter_index = new_counter_index;
1982 
1983 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
1984 	list_add_tail(&new_counter_index->list,
1985 		      &dev->counters_table[qp->port - 1].counters_list);
1986 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
1987 
1988 	return 0;
1989 }
1990 
1991 enum {
1992 	MLX4_QPC_ROCE_MODE_1 = 0,
1993 	MLX4_QPC_ROCE_MODE_2 = 2,
1994 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
1995 };
1996 
1997 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
1998 {
1999 	switch (gid_type) {
2000 	case IB_GID_TYPE_ROCE:
2001 		return MLX4_QPC_ROCE_MODE_1;
2002 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
2003 		return MLX4_QPC_ROCE_MODE_2;
2004 	default:
2005 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
2006 	}
2007 }
2008 
2009 /*
2010  * Go over all RSS QP's childes (WQs) and apply their HW state according to
2011  * their logic state if the RSS QP is the first RSS QP associated for the WQ.
2012  */
2013 static int bringup_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, u8 port_num,
2014 			    struct ib_udata *udata)
2015 {
2016 	int err = 0;
2017 	int i;
2018 
2019 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2020 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2021 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2022 
2023 		mutex_lock(&wq->mutex);
2024 
2025 		/* Mlx4_ib restrictions:
2026 		 * WQ's is associated to a port according to the RSS QP it is
2027 		 * associates to.
2028 		 * In case the WQ is associated to a different port by another
2029 		 * RSS QP, return a failure.
2030 		 */
2031 		if ((wq->rss_usecnt > 0) && (wq->port != port_num)) {
2032 			err = -EINVAL;
2033 			mutex_unlock(&wq->mutex);
2034 			break;
2035 		}
2036 		wq->port = port_num;
2037 		if ((wq->rss_usecnt == 0) && (ibwq->state == IB_WQS_RDY)) {
2038 			err = _mlx4_ib_modify_wq(ibwq, IB_WQS_RDY, udata);
2039 			if (err) {
2040 				mutex_unlock(&wq->mutex);
2041 				break;
2042 			}
2043 		}
2044 		wq->rss_usecnt++;
2045 
2046 		mutex_unlock(&wq->mutex);
2047 	}
2048 
2049 	if (i && err) {
2050 		int j;
2051 
2052 		for (j = (i - 1); j >= 0; j--) {
2053 			struct ib_wq *ibwq = ind_tbl->ind_tbl[j];
2054 			struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2055 
2056 			mutex_lock(&wq->mutex);
2057 
2058 			if ((wq->rss_usecnt == 1) &&
2059 			    (ibwq->state == IB_WQS_RDY))
2060 				if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET,
2061 						       udata))
2062 					pr_warn("failed to reverse WQN=0x%06x\n",
2063 						ibwq->wq_num);
2064 			wq->rss_usecnt--;
2065 
2066 			mutex_unlock(&wq->mutex);
2067 		}
2068 	}
2069 
2070 	return err;
2071 }
2072 
2073 static void bring_down_rss_rwqs(struct ib_rwq_ind_table *ind_tbl,
2074 				struct ib_udata *udata)
2075 {
2076 	int i;
2077 
2078 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2079 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2080 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2081 
2082 		mutex_lock(&wq->mutex);
2083 
2084 		if ((wq->rss_usecnt == 1) && (ibwq->state == IB_WQS_RDY))
2085 			if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET, udata))
2086 				pr_warn("failed to reverse WQN=%x\n",
2087 					ibwq->wq_num);
2088 		wq->rss_usecnt--;
2089 
2090 		mutex_unlock(&wq->mutex);
2091 	}
2092 }
2093 
2094 static void fill_qp_rss_context(struct mlx4_qp_context *context,
2095 				struct mlx4_ib_qp *qp)
2096 {
2097 	struct mlx4_rss_context *rss_context;
2098 
2099 	rss_context = (void *)context + offsetof(struct mlx4_qp_context,
2100 			pri_path) + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
2101 
2102 	rss_context->base_qpn = cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz);
2103 	rss_context->default_qpn =
2104 		cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz & 0xffffff);
2105 	if (qp->rss_ctx->flags & (MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6))
2106 		rss_context->base_qpn_udp = rss_context->default_qpn;
2107 	rss_context->flags = qp->rss_ctx->flags;
2108 	/* Currently support just toeplitz */
2109 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
2110 
2111 	memcpy(rss_context->rss_key, qp->rss_ctx->rss_key,
2112 	       MLX4_EN_RSS_KEY_SIZE);
2113 }
2114 
2115 static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type,
2116 			       const struct ib_qp_attr *attr, int attr_mask,
2117 			       enum ib_qp_state cur_state,
2118 			       enum ib_qp_state new_state,
2119 			       struct ib_udata *udata)
2120 {
2121 	struct ib_srq  *ibsrq;
2122 	const struct ib_gid_attr *gid_attr = NULL;
2123 	struct ib_rwq_ind_table *rwq_ind_tbl;
2124 	enum ib_qp_type qp_type;
2125 	struct mlx4_ib_dev *dev;
2126 	struct mlx4_ib_qp *qp;
2127 	struct mlx4_ib_pd *pd;
2128 	struct mlx4_ib_cq *send_cq, *recv_cq;
2129 	struct mlx4_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2130 		udata, struct mlx4_ib_ucontext, ibucontext);
2131 	struct mlx4_qp_context *context;
2132 	enum mlx4_qp_optpar optpar = 0;
2133 	int sqd_event;
2134 	int steer_qp = 0;
2135 	int err = -EINVAL;
2136 	int counter_index;
2137 
2138 	if (src_type == MLX4_IB_RWQ_SRC) {
2139 		struct ib_wq *ibwq;
2140 
2141 		ibwq	    = (struct ib_wq *)src;
2142 		ibsrq	    = NULL;
2143 		rwq_ind_tbl = NULL;
2144 		qp_type     = IB_QPT_RAW_PACKET;
2145 		qp	    = to_mqp((struct ib_qp *)ibwq);
2146 		dev	    = to_mdev(ibwq->device);
2147 		pd	    = to_mpd(ibwq->pd);
2148 	} else {
2149 		struct ib_qp *ibqp;
2150 
2151 		ibqp	    = (struct ib_qp *)src;
2152 		ibsrq	    = ibqp->srq;
2153 		rwq_ind_tbl = ibqp->rwq_ind_tbl;
2154 		qp_type     = ibqp->qp_type;
2155 		qp	    = to_mqp(ibqp);
2156 		dev	    = to_mdev(ibqp->device);
2157 		pd	    = get_pd(qp);
2158 	}
2159 
2160 	/* APM is not supported under RoCE */
2161 	if (attr_mask & IB_QP_ALT_PATH &&
2162 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2163 	    IB_LINK_LAYER_ETHERNET)
2164 		return -ENOTSUPP;
2165 
2166 	context = kzalloc_obj(*context);
2167 	if (!context)
2168 		return -ENOMEM;
2169 
2170 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
2171 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
2172 
2173 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
2174 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2175 	else {
2176 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
2177 		switch (attr->path_mig_state) {
2178 		case IB_MIG_MIGRATED:
2179 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2180 			break;
2181 		case IB_MIG_REARM:
2182 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
2183 			break;
2184 		case IB_MIG_ARMED:
2185 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
2186 			break;
2187 		}
2188 	}
2189 
2190 	if (qp->inl_recv_sz)
2191 		context->param3 |= cpu_to_be32(1 << 25);
2192 
2193 	if (qp->flags & MLX4_IB_QP_SCATTER_FCS)
2194 		context->param3 |= cpu_to_be32(1 << 29);
2195 
2196 	if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI)
2197 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
2198 	else if (qp_type == IB_QPT_RAW_PACKET)
2199 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
2200 	else if (qp_type == IB_QPT_UD) {
2201 		if (qp->flags & MLX4_IB_QP_LSO)
2202 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
2203 					      ilog2(dev->dev->caps.max_gso_sz);
2204 		else
2205 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 13;
2206 	} else if (attr_mask & IB_QP_PATH_MTU) {
2207 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
2208 			pr_err("path MTU (%u) is invalid\n",
2209 			       attr->path_mtu);
2210 			goto out;
2211 		}
2212 		context->mtu_msgmax = (attr->path_mtu << 5) |
2213 			ilog2(dev->dev->caps.max_msg_sz);
2214 	}
2215 
2216 	if (!rwq_ind_tbl) { /* PRM RSS receive side should be left zeros */
2217 		if (qp->rq.wqe_cnt)
2218 			context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
2219 		context->rq_size_stride |= qp->rq.wqe_shift - 4;
2220 	}
2221 
2222 	if (qp->sq.wqe_cnt)
2223 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
2224 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
2225 
2226 	if (new_state == IB_QPS_RESET && qp->counter_index)
2227 		mlx4_ib_free_qp_counter(dev, qp);
2228 
2229 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2230 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
2231 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
2232 		if (qp_type == IB_QPT_RAW_PACKET)
2233 			context->param3 |= cpu_to_be32(1 << 30);
2234 	}
2235 
2236 	if (ucontext)
2237 		context->usr_page = cpu_to_be32(
2238 			mlx4_to_hw_uar_index(dev->dev, ucontext->uar.index));
2239 	else
2240 		context->usr_page = cpu_to_be32(
2241 			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
2242 
2243 	if (attr_mask & IB_QP_DEST_QPN)
2244 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
2245 
2246 	if (attr_mask & IB_QP_PORT) {
2247 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
2248 		    !(attr_mask & IB_QP_AV)) {
2249 			mlx4_set_sched(&context->pri_path, attr->port_num);
2250 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
2251 		}
2252 	}
2253 
2254 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
2255 		err = create_qp_lb_counter(dev, qp);
2256 		if (err)
2257 			goto out;
2258 
2259 		counter_index =
2260 			dev->counters_table[qp->port - 1].default_counter;
2261 		if (qp->counter_index)
2262 			counter_index = qp->counter_index->index;
2263 
2264 		if (counter_index != -1) {
2265 			context->pri_path.counter_index = counter_index;
2266 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
2267 			if (qp->counter_index) {
2268 				context->pri_path.fl |=
2269 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
2270 				context->pri_path.vlan_control |=
2271 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
2272 			}
2273 		} else
2274 			context->pri_path.counter_index =
2275 				MLX4_SINK_COUNTER_INDEX(dev->dev);
2276 
2277 		if (qp->flags & MLX4_IB_QP_NETIF) {
2278 			mlx4_ib_steer_qp_reg(dev, qp, 1);
2279 			steer_qp = 1;
2280 		}
2281 
2282 		if (qp_type == IB_QPT_GSI) {
2283 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
2284 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
2285 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
2286 
2287 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2288 		}
2289 	}
2290 
2291 	if (attr_mask & IB_QP_PKEY_INDEX) {
2292 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2293 			context->pri_path.disable_pkey_check = 0x40;
2294 		context->pri_path.pkey_index = attr->pkey_index;
2295 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
2296 	}
2297 
2298 	if (attr_mask & IB_QP_AV) {
2299 		u8 port_num = mlx4_is_bonded(dev->dev) ? 1 :
2300 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2301 		u16 vlan = 0xffff;
2302 		u8 smac[ETH_ALEN];
2303 		int is_eth =
2304 			rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
2305 			rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
2306 
2307 		if (is_eth) {
2308 			gid_attr = attr->ah_attr.grh.sgid_attr;
2309 			err = rdma_read_gid_l2_fields(gid_attr, &vlan,
2310 						      &smac[0]);
2311 			if (err)
2312 				goto out;
2313 		}
2314 
2315 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
2316 				  port_num, vlan, smac))
2317 			goto out;
2318 
2319 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
2320 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
2321 
2322 		if (is_eth &&
2323 		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
2324 			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr->gid_type);
2325 
2326 			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
2327 				err = -EINVAL;
2328 				goto out;
2329 			}
2330 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2331 		}
2332 
2333 	}
2334 
2335 	if (attr_mask & IB_QP_TIMEOUT) {
2336 		context->pri_path.ackto |= attr->timeout << 3;
2337 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
2338 	}
2339 
2340 	if (attr_mask & IB_QP_ALT_PATH) {
2341 		if (attr->alt_port_num == 0 ||
2342 		    attr->alt_port_num > dev->dev->caps.num_ports)
2343 			goto out;
2344 
2345 		if (attr->alt_pkey_index >=
2346 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
2347 			goto out;
2348 
2349 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
2350 				      &context->alt_path,
2351 				      attr->alt_port_num))
2352 			goto out;
2353 
2354 		context->alt_path.pkey_index = attr->alt_pkey_index;
2355 		context->alt_path.ackto = attr->alt_timeout << 3;
2356 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
2357 	}
2358 
2359 	context->pd = cpu_to_be32(pd->pdn);
2360 
2361 	if (!rwq_ind_tbl) {
2362 		context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
2363 		get_cqs(qp, src_type, &send_cq, &recv_cq);
2364 	} else { /* Set dummy CQs to be compatible with HV and PRM */
2365 		send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq);
2366 		recv_cq = send_cq;
2367 	}
2368 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
2369 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
2370 
2371 	/* Set "fast registration enabled" for all kernel QPs */
2372 	if (!ucontext)
2373 		context->params1 |= cpu_to_be32(1 << 11);
2374 
2375 	if (attr_mask & IB_QP_RNR_RETRY) {
2376 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2377 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
2378 	}
2379 
2380 	if (attr_mask & IB_QP_RETRY_CNT) {
2381 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2382 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
2383 	}
2384 
2385 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2386 		if (attr->max_rd_atomic)
2387 			context->params1 |=
2388 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2389 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
2390 	}
2391 
2392 	if (attr_mask & IB_QP_SQ_PSN)
2393 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2394 
2395 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2396 		if (attr->max_dest_rd_atomic)
2397 			context->params2 |=
2398 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2399 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
2400 	}
2401 
2402 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
2403 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
2404 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
2405 	}
2406 
2407 	if (ibsrq)
2408 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
2409 
2410 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
2411 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2412 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
2413 	}
2414 	if (attr_mask & IB_QP_RQ_PSN)
2415 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2416 
2417 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
2418 	if (attr_mask & IB_QP_QKEY) {
2419 		if (qp->mlx4_ib_qp_type &
2420 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
2421 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2422 		else {
2423 			if (mlx4_is_mfunc(dev->dev) &&
2424 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
2425 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
2426 			    MLX4_RESERVED_QKEY_BASE) {
2427 				pr_err("Cannot use reserved QKEY"
2428 				       " 0x%x (range 0xffff0000..0xffffffff"
2429 				       " is reserved)\n", attr->qkey);
2430 				err = -EINVAL;
2431 				goto out;
2432 			}
2433 			context->qkey = cpu_to_be32(attr->qkey);
2434 		}
2435 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
2436 	}
2437 
2438 	if (ibsrq)
2439 		context->srqn = cpu_to_be32(1 << 24 |
2440 					    to_msrq(ibsrq)->msrq.srqn);
2441 
2442 	if (qp->rq.wqe_cnt &&
2443 	    cur_state == IB_QPS_RESET &&
2444 	    new_state == IB_QPS_INIT)
2445 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2446 
2447 	if (cur_state == IB_QPS_INIT &&
2448 	    new_state == IB_QPS_RTR  &&
2449 	    (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI ||
2450 	     qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) {
2451 		context->pri_path.sched_queue = (qp->port - 1) << 6;
2452 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2453 		    qp->mlx4_ib_qp_type &
2454 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
2455 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
2456 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
2457 				context->pri_path.fl = 0x80;
2458 		} else {
2459 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2460 				context->pri_path.fl = 0x80;
2461 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
2462 		}
2463 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2464 		    IB_LINK_LAYER_ETHERNET) {
2465 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
2466 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
2467 				context->pri_path.feup = 1 << 7; /* don't fsm */
2468 			/* handle smac_index */
2469 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
2470 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
2471 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
2472 				err = handle_eth_ud_smac_index(dev, qp, context);
2473 				if (err) {
2474 					err = -EINVAL;
2475 					goto out;
2476 				}
2477 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
2478 					dev->qp1_proxy[qp->port - 1] = qp;
2479 			}
2480 		}
2481 	}
2482 
2483 	if (qp_type == IB_QPT_RAW_PACKET) {
2484 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
2485 					MLX4_IB_LINK_TYPE_ETH;
2486 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
2487 			/* set QP to receive both tunneled & non-tunneled packets */
2488 			if (!rwq_ind_tbl)
2489 				context->srqn = cpu_to_be32(7 << 28);
2490 		}
2491 	}
2492 
2493 	if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
2494 		int is_eth = rdma_port_get_link_layer(
2495 				&dev->ib_dev, qp->port) ==
2496 				IB_LINK_LAYER_ETHERNET;
2497 		if (is_eth) {
2498 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
2499 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
2500 		}
2501 	}
2502 
2503 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2504 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2505 		sqd_event = 1;
2506 	else
2507 		sqd_event = 0;
2508 
2509 	if (!ucontext &&
2510 	    cur_state == IB_QPS_RESET &&
2511 	    new_state == IB_QPS_INIT)
2512 		context->rlkey_roce_mode |= (1 << 4);
2513 
2514 	/*
2515 	 * Before passing a kernel QP to the HW, make sure that the
2516 	 * ownership bits of the send queue are set and the SQ
2517 	 * headroom is stamped so that the hardware doesn't start
2518 	 * processing stale work requests.
2519 	 */
2520 	if (!ucontext &&
2521 	    cur_state == IB_QPS_RESET &&
2522 	    new_state == IB_QPS_INIT) {
2523 		struct mlx4_wqe_ctrl_seg *ctrl;
2524 		int i;
2525 
2526 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
2527 			ctrl = get_send_wqe(qp, i);
2528 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
2529 			ctrl->qpn_vlan.fence_size =
2530 				1 << (qp->sq.wqe_shift - 4);
2531 			stamp_send_wqe(qp, i);
2532 		}
2533 	}
2534 
2535 	if (rwq_ind_tbl	&&
2536 	    cur_state == IB_QPS_RESET &&
2537 	    new_state == IB_QPS_INIT) {
2538 		fill_qp_rss_context(context, qp);
2539 		context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET);
2540 	}
2541 
2542 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2543 			     to_mlx4_state(new_state), context, optpar,
2544 			     sqd_event, &qp->mqp);
2545 	if (err)
2546 		goto out;
2547 
2548 	qp->state = new_state;
2549 
2550 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2551 		qp->atomic_rd_en = attr->qp_access_flags;
2552 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2553 		qp->resp_depth = attr->max_dest_rd_atomic;
2554 	if (attr_mask & IB_QP_PORT) {
2555 		qp->port = attr->port_num;
2556 		update_mcg_macs(dev, qp);
2557 	}
2558 	if (attr_mask & IB_QP_ALT_PATH)
2559 		qp->alt_port = attr->alt_port_num;
2560 
2561 	if (is_sqp(dev, qp))
2562 		store_sqp_attrs(qp->sqp, attr, attr_mask);
2563 
2564 	/*
2565 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
2566 	 * QP0 to RESET or ERROR, bring the link back down.
2567 	 */
2568 	if (is_qp0(dev, qp)) {
2569 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2570 			if (mlx4_INIT_PORT(dev->dev, qp->port))
2571 				pr_warn("INIT_PORT failed for port %d\n",
2572 				       qp->port);
2573 
2574 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2575 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2576 			mlx4_CLOSE_PORT(dev->dev, qp->port);
2577 	}
2578 
2579 	/*
2580 	 * If we moved a kernel QP to RESET, clean up all old CQ
2581 	 * entries and reinitialize the QP.
2582 	 */
2583 	if (new_state == IB_QPS_RESET) {
2584 		if (!ucontext) {
2585 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2586 					 ibsrq ? to_msrq(ibsrq) : NULL);
2587 			if (send_cq != recv_cq)
2588 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2589 
2590 			qp->rq.head = 0;
2591 			qp->rq.tail = 0;
2592 			qp->sq.head = 0;
2593 			qp->sq.tail = 0;
2594 			qp->sq_next_wqe = 0;
2595 			if (qp->rq.wqe_cnt)
2596 				*qp->db.db  = 0;
2597 
2598 			if (qp->flags & MLX4_IB_QP_NETIF)
2599 				mlx4_ib_steer_qp_reg(dev, qp, 0);
2600 		}
2601 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2602 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2603 			qp->pri.smac = 0;
2604 			qp->pri.smac_port = 0;
2605 		}
2606 		if (qp->alt.smac) {
2607 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2608 			qp->alt.smac = 0;
2609 		}
2610 		if (qp->pri.vid < 0x1000) {
2611 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2612 			qp->pri.vid = 0xFFFF;
2613 			qp->pri.candidate_vid = 0xFFFF;
2614 			qp->pri.update_vid = 0;
2615 		}
2616 
2617 		if (qp->alt.vid < 0x1000) {
2618 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2619 			qp->alt.vid = 0xFFFF;
2620 			qp->alt.candidate_vid = 0xFFFF;
2621 			qp->alt.update_vid = 0;
2622 		}
2623 	}
2624 out:
2625 	if (err && qp->counter_index)
2626 		mlx4_ib_free_qp_counter(dev, qp);
2627 	if (err && steer_qp)
2628 		mlx4_ib_steer_qp_reg(dev, qp, 0);
2629 	kfree(context);
2630 	if (qp->pri.candidate_smac ||
2631 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2632 		if (err) {
2633 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2634 		} else {
2635 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2636 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2637 			qp->pri.smac = qp->pri.candidate_smac;
2638 			qp->pri.smac_index = qp->pri.candidate_smac_index;
2639 			qp->pri.smac_port = qp->pri.candidate_smac_port;
2640 		}
2641 		qp->pri.candidate_smac = 0;
2642 		qp->pri.candidate_smac_index = 0;
2643 		qp->pri.candidate_smac_port = 0;
2644 	}
2645 	if (qp->alt.candidate_smac) {
2646 		if (err) {
2647 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2648 		} else {
2649 			if (qp->alt.smac)
2650 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2651 			qp->alt.smac = qp->alt.candidate_smac;
2652 			qp->alt.smac_index = qp->alt.candidate_smac_index;
2653 			qp->alt.smac_port = qp->alt.candidate_smac_port;
2654 		}
2655 		qp->alt.candidate_smac = 0;
2656 		qp->alt.candidate_smac_index = 0;
2657 		qp->alt.candidate_smac_port = 0;
2658 	}
2659 
2660 	if (qp->pri.update_vid) {
2661 		if (err) {
2662 			if (qp->pri.candidate_vid < 0x1000)
2663 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2664 						     qp->pri.candidate_vid);
2665 		} else {
2666 			if (qp->pri.vid < 0x1000)
2667 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2668 						     qp->pri.vid);
2669 			qp->pri.vid = qp->pri.candidate_vid;
2670 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2671 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
2672 		}
2673 		qp->pri.candidate_vid = 0xFFFF;
2674 		qp->pri.update_vid = 0;
2675 	}
2676 
2677 	if (qp->alt.update_vid) {
2678 		if (err) {
2679 			if (qp->alt.candidate_vid < 0x1000)
2680 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2681 						     qp->alt.candidate_vid);
2682 		} else {
2683 			if (qp->alt.vid < 0x1000)
2684 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2685 						     qp->alt.vid);
2686 			qp->alt.vid = qp->alt.candidate_vid;
2687 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2688 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2689 		}
2690 		qp->alt.candidate_vid = 0xFFFF;
2691 		qp->alt.update_vid = 0;
2692 	}
2693 
2694 	return err;
2695 }
2696 
2697 enum {
2698 	MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK = (IB_QP_STATE	|
2699 					      IB_QP_PORT),
2700 };
2701 
2702 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2703 			      int attr_mask, struct ib_udata *udata)
2704 {
2705 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2706 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2707 	enum ib_qp_state cur_state, new_state;
2708 	int err = -EINVAL;
2709 	mutex_lock(&qp->mutex);
2710 
2711 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2712 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2713 
2714 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2715 				attr_mask)) {
2716 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2717 			 "for transition %d to %d. qp_type %d,"
2718 			 " attr_mask 0x%x\n",
2719 			 ibqp->qp_num, cur_state, new_state,
2720 			 ibqp->qp_type, attr_mask);
2721 		goto out;
2722 	}
2723 
2724 	if (ibqp->rwq_ind_tbl) {
2725 		if (!(((cur_state == IB_QPS_RESET) &&
2726 		       (new_state == IB_QPS_INIT)) ||
2727 		      ((cur_state == IB_QPS_INIT)  &&
2728 		       (new_state == IB_QPS_RTR)))) {
2729 			pr_debug("qpn 0x%x: RSS QP unsupported transition %d to %d\n",
2730 				 ibqp->qp_num, cur_state, new_state);
2731 
2732 			err = -EOPNOTSUPP;
2733 			goto out;
2734 		}
2735 
2736 		if (attr_mask & ~MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK) {
2737 			pr_debug("qpn 0x%x: RSS QP unsupported attribute mask 0x%x for transition %d to %d\n",
2738 				 ibqp->qp_num, attr_mask, cur_state, new_state);
2739 
2740 			err = -EOPNOTSUPP;
2741 			goto out;
2742 		}
2743 	}
2744 
2745 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2746 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2747 			if ((ibqp->qp_type == IB_QPT_RC) ||
2748 			    (ibqp->qp_type == IB_QPT_UD) ||
2749 			    (ibqp->qp_type == IB_QPT_UC) ||
2750 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2751 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2752 				attr->port_num = mlx4_ib_bond_next_port(dev);
2753 			}
2754 		} else {
2755 			/* no sense in changing port_num
2756 			 * when ports are bonded */
2757 			attr_mask &= ~IB_QP_PORT;
2758 		}
2759 	}
2760 
2761 	if ((attr_mask & IB_QP_PORT) &&
2762 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2763 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2764 			 "for transition %d to %d. qp_type %d\n",
2765 			 ibqp->qp_num, attr->port_num, cur_state,
2766 			 new_state, ibqp->qp_type);
2767 		goto out;
2768 	}
2769 
2770 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2771 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2772 	     IB_LINK_LAYER_ETHERNET))
2773 		goto out;
2774 
2775 	if (attr_mask & IB_QP_PKEY_INDEX) {
2776 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2777 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2778 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2779 				 "for transition %d to %d. qp_type %d\n",
2780 				 ibqp->qp_num, attr->pkey_index, cur_state,
2781 				 new_state, ibqp->qp_type);
2782 			goto out;
2783 		}
2784 	}
2785 
2786 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2787 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2788 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2789 			 "Transition %d to %d. qp_type %d\n",
2790 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2791 			 new_state, ibqp->qp_type);
2792 		goto out;
2793 	}
2794 
2795 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2796 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2797 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2798 			 "Transition %d to %d. qp_type %d\n",
2799 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2800 			 new_state, ibqp->qp_type);
2801 		goto out;
2802 	}
2803 
2804 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2805 		err = 0;
2806 		goto out;
2807 	}
2808 
2809 	if (ibqp->rwq_ind_tbl && (new_state == IB_QPS_INIT)) {
2810 		err = bringup_rss_rwqs(ibqp->rwq_ind_tbl, attr->port_num,
2811 				       udata);
2812 		if (err)
2813 			goto out;
2814 	}
2815 
2816 	err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask,
2817 				  cur_state, new_state, udata);
2818 
2819 	if (ibqp->rwq_ind_tbl && err)
2820 		bring_down_rss_rwqs(ibqp->rwq_ind_tbl, udata);
2821 
2822 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2823 		attr->port_num = 1;
2824 
2825 out:
2826 	mutex_unlock(&qp->mutex);
2827 	return err;
2828 }
2829 
2830 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2831 		      int attr_mask, struct ib_udata *udata)
2832 {
2833 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2834 	int ret;
2835 
2836 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
2837 		return -EOPNOTSUPP;
2838 
2839 	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2840 
2841 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2842 		struct mlx4_ib_sqp *sqp = mqp->sqp;
2843 		int err = 0;
2844 
2845 		if (sqp->roce_v2_gsi)
2846 			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2847 		if (err)
2848 			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2849 			       err);
2850 	}
2851 	return ret;
2852 }
2853 
2854 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2855 {
2856 	int i;
2857 	for (i = 0; i < dev->caps.num_ports; i++) {
2858 		if (qpn == dev->caps.spec_qps[i].qp0_proxy ||
2859 		    qpn == dev->caps.spec_qps[i].qp0_tunnel) {
2860 			*qkey = dev->caps.spec_qps[i].qp0_qkey;
2861 			return 0;
2862 		}
2863 	}
2864 	return -EINVAL;
2865 }
2866 
2867 static int build_sriov_qp0_header(struct mlx4_ib_qp *qp,
2868 				  const struct ib_ud_wr *wr,
2869 				  void *wqe, unsigned *mlx_seg_len)
2870 {
2871 	struct mlx4_ib_dev *mdev = to_mdev(qp->ibqp.device);
2872 	struct mlx4_ib_sqp *sqp = qp->sqp;
2873 	struct ib_device *ib_dev = qp->ibqp.device;
2874 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2875 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2876 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2877 	u16 pkey;
2878 	u32 qkey;
2879 	int send_size;
2880 	int header_size;
2881 	int spc;
2882 	int err;
2883 	int i;
2884 
2885 	if (wr->wr.opcode != IB_WR_SEND)
2886 		return -EINVAL;
2887 
2888 	send_size = 0;
2889 
2890 	for (i = 0; i < wr->wr.num_sge; ++i)
2891 		send_size += wr->wr.sg_list[i].length;
2892 
2893 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2894 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2895 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2896 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2897 
2898 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2899 
2900 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2901 		sqp->ud_header.lrh.service_level =
2902 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2903 		sqp->ud_header.lrh.destination_lid =
2904 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2905 		sqp->ud_header.lrh.source_lid =
2906 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2907 	}
2908 
2909 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2910 
2911 	/* force loopback */
2912 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2913 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2914 
2915 	sqp->ud_header.lrh.virtual_lane    = 0;
2916 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2917 	err = ib_get_cached_pkey(ib_dev, qp->port, 0, &pkey);
2918 	if (err)
2919 		return err;
2920 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2921 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2922 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2923 	else
2924 		sqp->ud_header.bth.destination_qpn =
2925 			cpu_to_be32(mdev->dev->caps.spec_qps[qp->port - 1].qp0_tunnel);
2926 
2927 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2928 	if (mlx4_is_master(mdev->dev)) {
2929 		if (mlx4_get_parav_qkey(mdev->dev, qp->mqp.qpn, &qkey))
2930 			return -EINVAL;
2931 	} else {
2932 		if (vf_get_qp0_qkey(mdev->dev, qp->mqp.qpn, &qkey))
2933 			return -EINVAL;
2934 	}
2935 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2936 	sqp->ud_header.deth.source_qpn = cpu_to_be32(qp->mqp.qpn);
2937 
2938 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2939 	sqp->ud_header.immediate_present = 0;
2940 
2941 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2942 
2943 	/*
2944 	 * Inline data segments may not cross a 64 byte boundary.  If
2945 	 * our UD header is bigger than the space available up to the
2946 	 * next 64 byte boundary in the WQE, use two inline data
2947 	 * segments to hold the UD header.
2948 	 */
2949 	spc = MLX4_INLINE_ALIGN -
2950 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2951 	if (header_size <= spc) {
2952 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2953 		memcpy(inl + 1, sqp->header_buf, header_size);
2954 		i = 1;
2955 	} else {
2956 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2957 		memcpy(inl + 1, sqp->header_buf, spc);
2958 
2959 		inl = (void *) (inl + 1) + spc;
2960 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2961 		/*
2962 		 * Need a barrier here to make sure all the data is
2963 		 * visible before the byte_count field is set.
2964 		 * Otherwise the HCA prefetcher could grab the 64-byte
2965 		 * chunk with this inline segment and get a valid (!=
2966 		 * 0xffffffff) byte count but stale data, and end up
2967 		 * generating a packet with bad headers.
2968 		 *
2969 		 * The first inline segment's byte_count field doesn't
2970 		 * need a barrier, because it comes after a
2971 		 * control/MLX segment and therefore is at an offset
2972 		 * of 16 mod 64.
2973 		 */
2974 		wmb();
2975 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2976 		i = 2;
2977 	}
2978 
2979 	*mlx_seg_len =
2980 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2981 	return 0;
2982 }
2983 
2984 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
2985 {
2986 	union sl2vl_tbl_to_u64 tmp_vltab;
2987 	u8 vl;
2988 
2989 	if (sl > 15)
2990 		return 0xf;
2991 	tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
2992 	vl = tmp_vltab.sl8[sl >> 1];
2993 	if (sl & 1)
2994 		vl &= 0x0f;
2995 	else
2996 		vl >>= 4;
2997 	return vl;
2998 }
2999 
3000 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
3001 				int index, union ib_gid *gid,
3002 				enum ib_gid_type *gid_type)
3003 {
3004 	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
3005 	struct mlx4_port_gid_table *port_gid_table;
3006 	unsigned long flags;
3007 
3008 	port_gid_table = &iboe->gids[port_num - 1];
3009 	spin_lock_irqsave(&iboe->lock, flags);
3010 	memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
3011 	*gid_type = port_gid_table->gids[index].gid_type;
3012 	spin_unlock_irqrestore(&iboe->lock, flags);
3013 	if (rdma_is_zero_gid(gid))
3014 		return -ENOENT;
3015 
3016 	return 0;
3017 }
3018 
3019 #define MLX4_ROCEV2_QP1_SPORT 0xC000
3020 static int build_mlx_header(struct mlx4_ib_qp *qp, const struct ib_ud_wr *wr,
3021 			    void *wqe, unsigned *mlx_seg_len)
3022 {
3023 	struct mlx4_ib_sqp *sqp = qp->sqp;
3024 	struct ib_device *ib_dev = qp->ibqp.device;
3025 	struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
3026 	struct mlx4_wqe_mlx_seg *mlx = wqe;
3027 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
3028 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
3029 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3030 	union ib_gid sgid;
3031 	u16 pkey;
3032 	int send_size;
3033 	int header_size;
3034 	int spc;
3035 	int i;
3036 	int err = 0;
3037 	u16 vlan = 0xffff;
3038 	bool is_eth;
3039 	bool is_vlan = false;
3040 	bool is_grh;
3041 	bool is_udp = false;
3042 	int ip_version = 0;
3043 
3044 	send_size = 0;
3045 	for (i = 0; i < wr->wr.num_sge; ++i)
3046 		send_size += wr->wr.sg_list[i].length;
3047 
3048 	is_eth = rdma_port_get_link_layer(qp->ibqp.device, qp->port) == IB_LINK_LAYER_ETHERNET;
3049 	is_grh = mlx4_ib_ah_grh_present(ah);
3050 	if (is_eth) {
3051 		enum ib_gid_type gid_type;
3052 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3053 			/* When multi-function is enabled, the ib_core gid
3054 			 * indexes don't necessarily match the hw ones, so
3055 			 * we must use our own cache */
3056 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
3057 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
3058 							   ah->av.ib.gid_index, &sgid.raw[0]);
3059 			if (err)
3060 				return err;
3061 		} else  {
3062 			err = fill_gid_by_hw_index(ibdev, qp->port,
3063 						   ah->av.ib.gid_index, &sgid,
3064 						   &gid_type);
3065 			if (!err) {
3066 				is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
3067 				if (is_udp) {
3068 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
3069 						ip_version = 4;
3070 					else
3071 						ip_version = 6;
3072 					is_grh = false;
3073 				}
3074 			} else {
3075 				return err;
3076 			}
3077 		}
3078 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
3079 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
3080 			is_vlan = true;
3081 		}
3082 	}
3083 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
3084 			  ip_version, is_udp, 0, &sqp->ud_header);
3085 	if (err)
3086 		return err;
3087 
3088 	if (!is_eth) {
3089 		sqp->ud_header.lrh.service_level =
3090 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
3091 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
3092 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
3093 	}
3094 
3095 	if (is_grh || (ip_version == 6)) {
3096 		sqp->ud_header.grh.traffic_class =
3097 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3098 		sqp->ud_header.grh.flow_label    =
3099 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
3100 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
3101 		if (is_eth) {
3102 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
3103 		} else {
3104 			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3105 				/* When multi-function is enabled, the ib_core gid
3106 				 * indexes don't necessarily match the hw ones, so
3107 				 * we must use our own cache
3108 				 */
3109 				sqp->ud_header.grh.source_gid.global
3110 					.subnet_prefix =
3111 					cpu_to_be64(atomic64_read(
3112 						&(to_mdev(ib_dev)
3113 							  ->sriov
3114 							  .demux[qp->port - 1]
3115 							  .subnet_prefix)));
3116 				sqp->ud_header.grh.source_gid.global
3117 					.interface_id =
3118 					to_mdev(ib_dev)
3119 						->sriov.demux[qp->port - 1]
3120 						.guid_cache[ah->av.ib.gid_index];
3121 			} else {
3122 				sqp->ud_header.grh.source_gid =
3123 					ah->ibah.sgid_attr->gid;
3124 			}
3125 		}
3126 		memcpy(sqp->ud_header.grh.destination_gid.raw,
3127 		       ah->av.ib.dgid, 16);
3128 	}
3129 
3130 	if (ip_version == 4) {
3131 		sqp->ud_header.ip4.tos =
3132 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3133 		sqp->ud_header.ip4.id = 0;
3134 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
3135 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
3136 
3137 		memcpy(&sqp->ud_header.ip4.saddr,
3138 		       sgid.raw + 12, 4);
3139 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
3140 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
3141 	}
3142 
3143 	if (is_udp) {
3144 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
3145 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
3146 		sqp->ud_header.udp.csum = 0;
3147 	}
3148 
3149 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
3150 
3151 	if (!is_eth) {
3152 		mlx->flags |=
3153 			cpu_to_be32((!qp->ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
3154 				    (sqp->ud_header.lrh.destination_lid ==
3155 						     IB_LID_PERMISSIVE ?
3156 					     MLX4_WQE_MLX_SLR :
3157 					     0) |
3158 				    (sqp->ud_header.lrh.service_level << 8));
3159 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
3160 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
3161 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
3162 	}
3163 
3164 	switch (wr->wr.opcode) {
3165 	case IB_WR_SEND:
3166 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
3167 		sqp->ud_header.immediate_present = 0;
3168 		break;
3169 	case IB_WR_SEND_WITH_IMM:
3170 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
3171 		sqp->ud_header.immediate_present = 1;
3172 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
3173 		break;
3174 	default:
3175 		return -EINVAL;
3176 	}
3177 
3178 	if (is_eth) {
3179 		u16 ether_type;
3180 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
3181 
3182 		ether_type = (!is_udp) ? ETH_P_IBOE:
3183 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
3184 
3185 		mlx->sched_prio = cpu_to_be16(pcp);
3186 
3187 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
3188 		ether_addr_copy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac);
3189 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
3190 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
3191 
3192 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
3193 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
3194 		if (!is_vlan) {
3195 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
3196 		} else {
3197 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
3198 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
3199 		}
3200 	} else {
3201 		sqp->ud_header.lrh.virtual_lane =
3202 			!qp->ibqp.qp_num ?
3203 				15 :
3204 				sl_to_vl(to_mdev(ib_dev),
3205 					 sqp->ud_header.lrh.service_level,
3206 					 qp->port);
3207 		if (qp->ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
3208 			return -EINVAL;
3209 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
3210 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
3211 	}
3212 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
3213 	if (!qp->ibqp.qp_num)
3214 		err = ib_get_cached_pkey(ib_dev, qp->port, sqp->pkey_index,
3215 					 &pkey);
3216 	else
3217 		err = ib_get_cached_pkey(ib_dev, qp->port, wr->pkey_index,
3218 					 &pkey);
3219 	if (err)
3220 		return err;
3221 
3222 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
3223 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
3224 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
3225 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
3226 					       sqp->qkey : wr->remote_qkey);
3227 	sqp->ud_header.deth.source_qpn = cpu_to_be32(qp->ibqp.qp_num);
3228 
3229 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3230 
3231 	if (0) {
3232 		pr_err("built UD header of size %d:\n", header_size);
3233 		for (i = 0; i < header_size / 4; ++i) {
3234 			if (i % 8 == 0)
3235 				pr_err("  [%02x] ", i * 4);
3236 			pr_cont(" %08x",
3237 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
3238 			if ((i + 1) % 8 == 0)
3239 				pr_cont("\n");
3240 		}
3241 		pr_err("\n");
3242 	}
3243 
3244 	/*
3245 	 * Inline data segments may not cross a 64 byte boundary.  If
3246 	 * our UD header is bigger than the space available up to the
3247 	 * next 64 byte boundary in the WQE, use two inline data
3248 	 * segments to hold the UD header.
3249 	 */
3250 	spc = MLX4_INLINE_ALIGN -
3251 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3252 	if (header_size <= spc) {
3253 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3254 		memcpy(inl + 1, sqp->header_buf, header_size);
3255 		i = 1;
3256 	} else {
3257 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3258 		memcpy(inl + 1, sqp->header_buf, spc);
3259 
3260 		inl = (void *) (inl + 1) + spc;
3261 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3262 		/*
3263 		 * Need a barrier here to make sure all the data is
3264 		 * visible before the byte_count field is set.
3265 		 * Otherwise the HCA prefetcher could grab the 64-byte
3266 		 * chunk with this inline segment and get a valid (!=
3267 		 * 0xffffffff) byte count but stale data, and end up
3268 		 * generating a packet with bad headers.
3269 		 *
3270 		 * The first inline segment's byte_count field doesn't
3271 		 * need a barrier, because it comes after a
3272 		 * control/MLX segment and therefore is at an offset
3273 		 * of 16 mod 64.
3274 		 */
3275 		wmb();
3276 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3277 		i = 2;
3278 	}
3279 
3280 	*mlx_seg_len =
3281 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3282 	return 0;
3283 }
3284 
3285 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3286 {
3287 	unsigned cur;
3288 	struct mlx4_ib_cq *cq;
3289 
3290 	cur = wq->head - wq->tail;
3291 	if (likely(cur + nreq < wq->max_post))
3292 		return 0;
3293 
3294 	cq = to_mcq(ib_cq);
3295 	spin_lock(&cq->lock);
3296 	cur = wq->head - wq->tail;
3297 	spin_unlock(&cq->lock);
3298 
3299 	return cur + nreq >= wq->max_post;
3300 }
3301 
3302 static __be32 convert_access(int acc)
3303 {
3304 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
3305 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
3306 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
3307 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
3308 	       (acc & IB_ACCESS_REMOTE_READ   ?
3309 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
3310 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
3311 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
3312 }
3313 
3314 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
3315 			const struct ib_reg_wr *wr)
3316 {
3317 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
3318 
3319 	fseg->flags		= convert_access(wr->access);
3320 	fseg->mem_key		= cpu_to_be32(wr->key);
3321 	fseg->buf_list		= cpu_to_be64(mr->page_map);
3322 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
3323 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
3324 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
3325 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
3326 	fseg->reserved[0]	= 0;
3327 	fseg->reserved[1]	= 0;
3328 }
3329 
3330 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
3331 {
3332 	memset(iseg, 0, sizeof(*iseg));
3333 	iseg->mem_key = cpu_to_be32(rkey);
3334 }
3335 
3336 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
3337 					  u64 remote_addr, u32 rkey)
3338 {
3339 	rseg->raddr    = cpu_to_be64(remote_addr);
3340 	rseg->rkey     = cpu_to_be32(rkey);
3341 	rseg->reserved = 0;
3342 }
3343 
3344 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
3345 			   const struct ib_atomic_wr *wr)
3346 {
3347 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
3348 		aseg->swap_add = cpu_to_be64(wr->swap);
3349 		aseg->compare  = cpu_to_be64(wr->compare_add);
3350 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
3351 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3352 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
3353 	} else {
3354 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3355 		aseg->compare  = 0;
3356 	}
3357 
3358 }
3359 
3360 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
3361 				  const struct ib_atomic_wr *wr)
3362 {
3363 	aseg->swap_add		= cpu_to_be64(wr->swap);
3364 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
3365 	aseg->compare		= cpu_to_be64(wr->compare_add);
3366 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
3367 }
3368 
3369 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
3370 			     const struct ib_ud_wr *wr)
3371 {
3372 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
3373 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
3374 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
3375 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
3376 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
3377 }
3378 
3379 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
3380 				    struct mlx4_wqe_datagram_seg *dseg,
3381 				    const struct ib_ud_wr *wr,
3382 				    enum mlx4_ib_qp_type qpt)
3383 {
3384 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
3385 	struct mlx4_av sqp_av = {0};
3386 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
3387 
3388 	/* force loopback */
3389 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
3390 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
3391 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
3392 			cpu_to_be32(0xf0000000);
3393 
3394 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
3395 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
3396 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp1_tunnel);
3397 	else
3398 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp0_tunnel);
3399 	/* Use QKEY from the QP context, which is set by master */
3400 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
3401 }
3402 
3403 static void build_tunnel_header(const struct ib_ud_wr *wr, void *wqe,
3404 				unsigned *mlx_seg_len)
3405 {
3406 	struct mlx4_wqe_inline_seg *inl = wqe;
3407 	struct mlx4_ib_tunnel_header hdr;
3408 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3409 	int spc;
3410 	int i;
3411 
3412 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
3413 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
3414 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
3415 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
3416 	memcpy(hdr.mac, ah->av.eth.mac, 6);
3417 	hdr.vlan = ah->av.eth.vlan;
3418 
3419 	spc = MLX4_INLINE_ALIGN -
3420 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3421 	if (sizeof (hdr) <= spc) {
3422 		memcpy(inl + 1, &hdr, sizeof (hdr));
3423 		wmb();
3424 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
3425 		i = 1;
3426 	} else {
3427 		memcpy(inl + 1, &hdr, spc);
3428 		wmb();
3429 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3430 
3431 		inl = (void *) (inl + 1) + spc;
3432 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
3433 		wmb();
3434 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
3435 		i = 2;
3436 	}
3437 
3438 	*mlx_seg_len =
3439 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
3440 }
3441 
3442 static void set_mlx_icrc_seg(void *dseg)
3443 {
3444 	u32 *t = dseg;
3445 	struct mlx4_wqe_inline_seg *iseg = dseg;
3446 
3447 	t[1] = 0;
3448 
3449 	/*
3450 	 * Need a barrier here before writing the byte_count field to
3451 	 * make sure that all the data is visible before the
3452 	 * byte_count field is set.  Otherwise, if the segment begins
3453 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3454 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3455 	 * stale data, and end up sending the wrong data.
3456 	 */
3457 	wmb();
3458 
3459 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
3460 }
3461 
3462 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3463 {
3464 	dseg->lkey       = cpu_to_be32(sg->lkey);
3465 	dseg->addr       = cpu_to_be64(sg->addr);
3466 
3467 	/*
3468 	 * Need a barrier here before writing the byte_count field to
3469 	 * make sure that all the data is visible before the
3470 	 * byte_count field is set.  Otherwise, if the segment begins
3471 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3472 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3473 	 * stale data, and end up sending the wrong data.
3474 	 */
3475 	wmb();
3476 
3477 	dseg->byte_count = cpu_to_be32(sg->length);
3478 }
3479 
3480 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3481 {
3482 	dseg->byte_count = cpu_to_be32(sg->length);
3483 	dseg->lkey       = cpu_to_be32(sg->lkey);
3484 	dseg->addr       = cpu_to_be64(sg->addr);
3485 }
3486 
3487 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe,
3488 			 const struct ib_ud_wr *wr, struct mlx4_ib_qp *qp,
3489 			 unsigned *lso_seg_len, __be32 *lso_hdr_sz, __be32 *blh)
3490 {
3491 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
3492 
3493 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
3494 		*blh = cpu_to_be32(1 << 6);
3495 
3496 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
3497 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
3498 		return -EINVAL;
3499 
3500 	memcpy(wqe->header, wr->header, wr->hlen);
3501 
3502 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
3503 	*lso_seg_len = halign;
3504 	return 0;
3505 }
3506 
3507 static __be32 send_ieth(const struct ib_send_wr *wr)
3508 {
3509 	switch (wr->opcode) {
3510 	case IB_WR_SEND_WITH_IMM:
3511 	case IB_WR_RDMA_WRITE_WITH_IMM:
3512 		return wr->ex.imm_data;
3513 
3514 	case IB_WR_SEND_WITH_INV:
3515 		return cpu_to_be32(wr->ex.invalidate_rkey);
3516 
3517 	default:
3518 		return 0;
3519 	}
3520 }
3521 
3522 static void add_zero_len_inline(void *wqe)
3523 {
3524 	struct mlx4_wqe_inline_seg *inl = wqe;
3525 	memset(wqe, 0, 16);
3526 	inl->byte_count = cpu_to_be32(1 << 31);
3527 }
3528 
3529 static int _mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3530 			      const struct ib_send_wr **bad_wr, bool drain)
3531 {
3532 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3533 	void *wqe;
3534 	struct mlx4_wqe_ctrl_seg *ctrl;
3535 	struct mlx4_wqe_data_seg *dseg;
3536 	unsigned long flags;
3537 	int nreq;
3538 	int err = 0;
3539 	unsigned ind;
3540 	int size;
3541 	unsigned seglen;
3542 	__be32 dummy;
3543 	__be32 *lso_wqe;
3544 	__be32 lso_hdr_sz;
3545 	__be32 blh;
3546 	int i;
3547 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3548 
3549 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
3550 		struct mlx4_ib_sqp *sqp = qp->sqp;
3551 
3552 		if (sqp->roce_v2_gsi) {
3553 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
3554 			enum ib_gid_type gid_type;
3555 			union ib_gid gid;
3556 
3557 			if (!fill_gid_by_hw_index(mdev, qp->port,
3558 					   ah->av.ib.gid_index,
3559 					   &gid, &gid_type))
3560 				qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
3561 						to_mqp(sqp->roce_v2_gsi) : qp;
3562 			else
3563 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
3564 				       ah->av.ib.gid_index);
3565 		}
3566 	}
3567 
3568 	spin_lock_irqsave(&qp->sq.lock, flags);
3569 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3570 	    !drain) {
3571 		err = -EIO;
3572 		*bad_wr = wr;
3573 		nreq = 0;
3574 		goto out;
3575 	}
3576 
3577 	ind = qp->sq_next_wqe;
3578 
3579 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3580 		lso_wqe = &dummy;
3581 		blh = 0;
3582 
3583 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
3584 			err = -ENOMEM;
3585 			*bad_wr = wr;
3586 			goto out;
3587 		}
3588 
3589 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3590 			err = -EINVAL;
3591 			*bad_wr = wr;
3592 			goto out;
3593 		}
3594 
3595 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3596 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3597 
3598 		ctrl->srcrb_flags =
3599 			(wr->send_flags & IB_SEND_SIGNALED ?
3600 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3601 			(wr->send_flags & IB_SEND_SOLICITED ?
3602 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3603 			((wr->send_flags & IB_SEND_IP_CSUM) ?
3604 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3605 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3606 			qp->sq_signal_bits;
3607 
3608 		ctrl->imm = send_ieth(wr);
3609 
3610 		wqe += sizeof *ctrl;
3611 		size = sizeof *ctrl / 16;
3612 
3613 		switch (qp->mlx4_ib_qp_type) {
3614 		case MLX4_IB_QPT_RC:
3615 		case MLX4_IB_QPT_UC:
3616 			switch (wr->opcode) {
3617 			case IB_WR_ATOMIC_CMP_AND_SWP:
3618 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3619 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3620 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3621 					      atomic_wr(wr)->rkey);
3622 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3623 
3624 				set_atomic_seg(wqe, atomic_wr(wr));
3625 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
3626 
3627 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3628 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3629 
3630 				break;
3631 
3632 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3633 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3634 					      atomic_wr(wr)->rkey);
3635 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3636 
3637 				set_masked_atomic_seg(wqe, atomic_wr(wr));
3638 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
3639 
3640 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3641 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3642 
3643 				break;
3644 
3645 			case IB_WR_RDMA_READ:
3646 			case IB_WR_RDMA_WRITE:
3647 			case IB_WR_RDMA_WRITE_WITH_IMM:
3648 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3649 					      rdma_wr(wr)->rkey);
3650 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3651 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3652 				break;
3653 
3654 			case IB_WR_LOCAL_INV:
3655 				ctrl->srcrb_flags |=
3656 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3657 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3658 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3659 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3660 				break;
3661 
3662 			case IB_WR_REG_MR:
3663 				ctrl->srcrb_flags |=
3664 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3665 				set_reg_seg(wqe, reg_wr(wr));
3666 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3667 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3668 				break;
3669 
3670 			default:
3671 				/* No extra segments required for sends */
3672 				break;
3673 			}
3674 			break;
3675 
3676 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3677 			err = build_sriov_qp0_header(qp, ud_wr(wr), ctrl,
3678 						     &seglen);
3679 			if (unlikely(err)) {
3680 				*bad_wr = wr;
3681 				goto out;
3682 			}
3683 			wqe  += seglen;
3684 			size += seglen / 16;
3685 			break;
3686 		case MLX4_IB_QPT_TUN_SMI:
3687 		case MLX4_IB_QPT_TUN_GSI:
3688 			/* this is a UD qp used in MAD responses to slaves. */
3689 			set_datagram_seg(wqe, ud_wr(wr));
3690 			/* set the forced-loopback bit in the data seg av */
3691 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3692 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3693 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3694 			break;
3695 		case MLX4_IB_QPT_UD:
3696 			set_datagram_seg(wqe, ud_wr(wr));
3697 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3698 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3699 
3700 			if (wr->opcode == IB_WR_LSO) {
3701 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3702 						&lso_hdr_sz, &blh);
3703 				if (unlikely(err)) {
3704 					*bad_wr = wr;
3705 					goto out;
3706 				}
3707 				lso_wqe = (__be32 *) wqe;
3708 				wqe  += seglen;
3709 				size += seglen / 16;
3710 			}
3711 			break;
3712 
3713 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3714 			err = build_sriov_qp0_header(qp, ud_wr(wr), ctrl,
3715 						     &seglen);
3716 			if (unlikely(err)) {
3717 				*bad_wr = wr;
3718 				goto out;
3719 			}
3720 			wqe  += seglen;
3721 			size += seglen / 16;
3722 			/* to start tunnel header on a cache-line boundary */
3723 			add_zero_len_inline(wqe);
3724 			wqe += 16;
3725 			size++;
3726 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3727 			wqe  += seglen;
3728 			size += seglen / 16;
3729 			break;
3730 		case MLX4_IB_QPT_PROXY_SMI:
3731 		case MLX4_IB_QPT_PROXY_GSI:
3732 			/* If we are tunneling special qps, this is a UD qp.
3733 			 * In this case we first add a UD segment targeting
3734 			 * the tunnel qp, and then add a header with address
3735 			 * information */
3736 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3737 						ud_wr(wr),
3738 						qp->mlx4_ib_qp_type);
3739 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3740 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3741 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3742 			wqe  += seglen;
3743 			size += seglen / 16;
3744 			break;
3745 
3746 		case MLX4_IB_QPT_SMI:
3747 		case MLX4_IB_QPT_GSI:
3748 			err = build_mlx_header(qp, ud_wr(wr), ctrl, &seglen);
3749 			if (unlikely(err)) {
3750 				*bad_wr = wr;
3751 				goto out;
3752 			}
3753 			wqe  += seglen;
3754 			size += seglen / 16;
3755 			break;
3756 
3757 		default:
3758 			break;
3759 		}
3760 
3761 		/*
3762 		 * Write data segments in reverse order, so as to
3763 		 * overwrite cacheline stamp last within each
3764 		 * cacheline.  This avoids issues with WQE
3765 		 * prefetching.
3766 		 */
3767 
3768 		dseg = wqe;
3769 		dseg += wr->num_sge - 1;
3770 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3771 
3772 		/* Add one more inline data segment for ICRC for MLX sends */
3773 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3774 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3775 			     qp->mlx4_ib_qp_type &
3776 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3777 			set_mlx_icrc_seg(dseg + 1);
3778 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3779 		}
3780 
3781 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3782 			set_data_seg(dseg, wr->sg_list + i);
3783 
3784 		/*
3785 		 * Possibly overwrite stamping in cacheline with LSO
3786 		 * segment only after making sure all data segments
3787 		 * are written.
3788 		 */
3789 		wmb();
3790 		*lso_wqe = lso_hdr_sz;
3791 
3792 		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3793 					     MLX4_WQE_CTRL_FENCE : 0) | size;
3794 
3795 		/*
3796 		 * Make sure descriptor is fully written before
3797 		 * setting ownership bit (because HW can start
3798 		 * executing as soon as we do).
3799 		 */
3800 		wmb();
3801 
3802 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3803 			*bad_wr = wr;
3804 			err = -EINVAL;
3805 			goto out;
3806 		}
3807 
3808 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3809 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3810 
3811 		/*
3812 		 * We can improve latency by not stamping the last
3813 		 * send queue WQE until after ringing the doorbell, so
3814 		 * only stamp here if there are still more WQEs to post.
3815 		 */
3816 		if (wr->next)
3817 			stamp_send_wqe(qp, ind + qp->sq_spare_wqes);
3818 		ind++;
3819 	}
3820 
3821 out:
3822 	if (likely(nreq)) {
3823 		qp->sq.head += nreq;
3824 
3825 		/*
3826 		 * Make sure that descriptors are written before
3827 		 * doorbell record.
3828 		 */
3829 		wmb();
3830 
3831 		writel_relaxed(qp->doorbell_qpn,
3832 			to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3833 
3834 		stamp_send_wqe(qp, ind + qp->sq_spare_wqes - 1);
3835 
3836 		qp->sq_next_wqe = ind;
3837 	}
3838 
3839 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3840 
3841 	return err;
3842 }
3843 
3844 int mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3845 		      const struct ib_send_wr **bad_wr)
3846 {
3847 	return _mlx4_ib_post_send(ibqp, wr, bad_wr, false);
3848 }
3849 
3850 static int _mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3851 			      const struct ib_recv_wr **bad_wr, bool drain)
3852 {
3853 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3854 	struct mlx4_wqe_data_seg *scat;
3855 	unsigned long flags;
3856 	int err = 0;
3857 	int nreq;
3858 	int ind;
3859 	int max_gs;
3860 	int i;
3861 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3862 
3863 	max_gs = qp->rq.max_gs;
3864 	spin_lock_irqsave(&qp->rq.lock, flags);
3865 
3866 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3867 	    !drain) {
3868 		err = -EIO;
3869 		*bad_wr = wr;
3870 		nreq = 0;
3871 		goto out;
3872 	}
3873 
3874 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3875 
3876 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3877 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3878 			err = -ENOMEM;
3879 			*bad_wr = wr;
3880 			goto out;
3881 		}
3882 
3883 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3884 			err = -EINVAL;
3885 			*bad_wr = wr;
3886 			goto out;
3887 		}
3888 
3889 		scat = get_recv_wqe(qp, ind);
3890 
3891 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3892 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3893 			ib_dma_sync_single_for_device(ibqp->device,
3894 						      qp->sqp_proxy_rcv[ind].map,
3895 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3896 						      DMA_FROM_DEVICE);
3897 			scat->byte_count =
3898 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3899 			/* use dma lkey from upper layer entry */
3900 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3901 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3902 			scat++;
3903 			max_gs--;
3904 		}
3905 
3906 		for (i = 0; i < wr->num_sge; ++i)
3907 			__set_data_seg(scat + i, wr->sg_list + i);
3908 
3909 		if (i < max_gs) {
3910 			scat[i].byte_count = 0;
3911 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3912 			scat[i].addr       = 0;
3913 		}
3914 
3915 		qp->rq.wrid[ind] = wr->wr_id;
3916 
3917 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3918 	}
3919 
3920 out:
3921 	if (likely(nreq)) {
3922 		qp->rq.head += nreq;
3923 
3924 		/*
3925 		 * Make sure that descriptors are written before
3926 		 * doorbell record.
3927 		 */
3928 		wmb();
3929 
3930 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3931 	}
3932 
3933 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3934 
3935 	return err;
3936 }
3937 
3938 int mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3939 		      const struct ib_recv_wr **bad_wr)
3940 {
3941 	return _mlx4_ib_post_recv(ibqp, wr, bad_wr, false);
3942 }
3943 
3944 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3945 {
3946 	switch (mlx4_state) {
3947 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3948 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3949 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3950 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3951 	case MLX4_QP_STATE_SQ_DRAINING:
3952 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3953 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3954 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3955 	default:		     return -1;
3956 	}
3957 }
3958 
3959 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3960 {
3961 	switch (mlx4_mig_state) {
3962 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3963 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3964 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3965 	default: return -1;
3966 	}
3967 }
3968 
3969 static int to_ib_qp_access_flags(int mlx4_flags)
3970 {
3971 	int ib_flags = 0;
3972 
3973 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3974 		ib_flags |= IB_ACCESS_REMOTE_READ;
3975 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3976 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3977 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3978 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3979 
3980 	return ib_flags;
3981 }
3982 
3983 static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
3984 			    struct rdma_ah_attr *ah_attr,
3985 			    struct mlx4_qp_path *path)
3986 {
3987 	struct mlx4_dev *dev = ibdev->dev;
3988 	u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
3989 
3990 	memset(ah_attr, 0, sizeof(*ah_attr));
3991 	if (port_num == 0 || port_num > dev->caps.num_ports)
3992 		return;
3993 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
3994 
3995 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
3996 		rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
3997 			       ((path->sched_queue & 4) << 1));
3998 	else
3999 		rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
4000 	rdma_ah_set_port_num(ah_attr, port_num);
4001 
4002 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
4003 	rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
4004 	rdma_ah_set_static_rate(ah_attr,
4005 				path->static_rate ? path->static_rate - 5 : 0);
4006 	if (path->grh_mylmc & (1 << 7)) {
4007 		rdma_ah_set_grh(ah_attr, NULL,
4008 				be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
4009 				path->mgid_index,
4010 				path->hop_limit,
4011 				(be32_to_cpu(path->tclass_flowlabel)
4012 				 >> 20) & 0xff);
4013 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
4014 	}
4015 }
4016 
4017 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
4018 		     struct ib_qp_init_attr *qp_init_attr)
4019 {
4020 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
4021 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
4022 	struct mlx4_qp_context context;
4023 	int mlx4_state;
4024 	int err = 0;
4025 
4026 	if (ibqp->rwq_ind_tbl)
4027 		return -EOPNOTSUPP;
4028 
4029 	mutex_lock(&qp->mutex);
4030 
4031 	if (qp->state == IB_QPS_RESET) {
4032 		qp_attr->qp_state = IB_QPS_RESET;
4033 		goto done;
4034 	}
4035 
4036 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
4037 	if (err) {
4038 		err = -EINVAL;
4039 		goto out;
4040 	}
4041 
4042 	mlx4_state = be32_to_cpu(context.flags) >> 28;
4043 
4044 	qp->state		     = to_ib_qp_state(mlx4_state);
4045 	qp_attr->qp_state	     = qp->state;
4046 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
4047 	qp_attr->path_mig_state	     =
4048 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
4049 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
4050 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
4051 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
4052 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
4053 	qp_attr->qp_access_flags     =
4054 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
4055 
4056 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC ||
4057 	    qp->ibqp.qp_type == IB_QPT_XRC_INI ||
4058 	    qp->ibqp.qp_type == IB_QPT_XRC_TGT) {
4059 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
4060 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
4061 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
4062 		qp_attr->alt_port_num	=
4063 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
4064 	}
4065 
4066 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
4067 	if (qp_attr->qp_state == IB_QPS_INIT)
4068 		qp_attr->port_num = qp->port;
4069 	else
4070 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
4071 
4072 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4073 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
4074 
4075 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
4076 
4077 	qp_attr->max_dest_rd_atomic =
4078 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
4079 	qp_attr->min_rnr_timer	    =
4080 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
4081 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
4082 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
4083 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
4084 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
4085 
4086 done:
4087 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4088 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4089 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4090 
4091 	if (!ibqp->uobject) {
4092 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
4093 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4094 	} else {
4095 		qp_attr->cap.max_send_wr  = 0;
4096 		qp_attr->cap.max_send_sge = 0;
4097 	}
4098 
4099 	/*
4100 	 * We don't support inline sends for kernel QPs (yet), and we
4101 	 * don't know what userspace's value should be.
4102 	 */
4103 	qp_attr->cap.max_inline_data = 0;
4104 
4105 	qp_init_attr->cap	     = qp_attr->cap;
4106 
4107 	qp_init_attr->create_flags = 0;
4108 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4109 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4110 
4111 	if (qp->flags & MLX4_IB_QP_LSO)
4112 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
4113 
4114 	if (qp->flags & MLX4_IB_QP_NETIF)
4115 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
4116 
4117 	qp_init_attr->sq_sig_type =
4118 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
4119 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4120 
4121 out:
4122 	mutex_unlock(&qp->mutex);
4123 	return err;
4124 }
4125 
4126 struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
4127 				struct ib_wq_init_attr *init_attr,
4128 				struct ib_udata *udata)
4129 {
4130 	struct mlx4_dev *dev = to_mdev(pd->device)->dev;
4131 	struct ib_qp_init_attr ib_qp_init_attr = {};
4132 	struct mlx4_ib_qp *qp;
4133 	struct mlx4_ib_create_wq ucmd;
4134 	int err, required_cmd_sz;
4135 
4136 	if (!udata)
4137 		return ERR_PTR(-EINVAL);
4138 
4139 	required_cmd_sz = offsetof(typeof(ucmd), comp_mask) +
4140 			  sizeof(ucmd.comp_mask);
4141 	if (udata->inlen < required_cmd_sz) {
4142 		pr_debug("invalid inlen\n");
4143 		return ERR_PTR(-EINVAL);
4144 	}
4145 
4146 	if (udata->inlen > sizeof(ucmd) &&
4147 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4148 				 udata->inlen - sizeof(ucmd))) {
4149 		pr_debug("inlen is not supported\n");
4150 		return ERR_PTR(-EOPNOTSUPP);
4151 	}
4152 
4153 	if (udata->outlen)
4154 		return ERR_PTR(-EOPNOTSUPP);
4155 
4156 	if (init_attr->wq_type != IB_WQT_RQ) {
4157 		pr_debug("unsupported wq type %d\n", init_attr->wq_type);
4158 		return ERR_PTR(-EOPNOTSUPP);
4159 	}
4160 
4161 	if (init_attr->create_flags & ~IB_WQ_FLAGS_SCATTER_FCS ||
4162 	    !(dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
4163 		pr_debug("unsupported create_flags %u\n",
4164 			 init_attr->create_flags);
4165 		return ERR_PTR(-EOPNOTSUPP);
4166 	}
4167 
4168 	qp = kzalloc_obj(*qp);
4169 	if (!qp)
4170 		return ERR_PTR(-ENOMEM);
4171 
4172 	mutex_init(&qp->mutex);
4173 	qp->pri.vid = 0xFFFF;
4174 	qp->alt.vid = 0xFFFF;
4175 
4176 	ib_qp_init_attr.qp_context = init_attr->wq_context;
4177 	ib_qp_init_attr.qp_type = IB_QPT_RAW_PACKET;
4178 	ib_qp_init_attr.cap.max_recv_wr = init_attr->max_wr;
4179 	ib_qp_init_attr.cap.max_recv_sge = init_attr->max_sge;
4180 	ib_qp_init_attr.recv_cq = init_attr->cq;
4181 	ib_qp_init_attr.send_cq = ib_qp_init_attr.recv_cq; /* Dummy CQ */
4182 
4183 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS)
4184 		ib_qp_init_attr.create_flags |= IB_QP_CREATE_SCATTER_FCS;
4185 
4186 	err = create_rq(pd, &ib_qp_init_attr, udata, qp);
4187 	if (err) {
4188 		kfree(qp);
4189 		return ERR_PTR(err);
4190 	}
4191 
4192 	qp->ibwq.event_handler = init_attr->event_handler;
4193 	qp->ibwq.wq_num = qp->mqp.qpn;
4194 	qp->ibwq.state = IB_WQS_RESET;
4195 
4196 	return &qp->ibwq;
4197 }
4198 
4199 static int ib_wq2qp_state(enum ib_wq_state state)
4200 {
4201 	switch (state) {
4202 	case IB_WQS_RESET:
4203 		return IB_QPS_RESET;
4204 	case IB_WQS_RDY:
4205 		return IB_QPS_RTR;
4206 	default:
4207 		return IB_QPS_ERR;
4208 	}
4209 }
4210 
4211 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state,
4212 			      struct ib_udata *udata)
4213 {
4214 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4215 	enum ib_qp_state qp_cur_state;
4216 	enum ib_qp_state qp_new_state;
4217 	int attr_mask;
4218 	int err;
4219 
4220 	/* ib_qp.state represents the WQ HW state while ib_wq.state represents
4221 	 * the WQ logic state.
4222 	 */
4223 	qp_cur_state = qp->state;
4224 	qp_new_state = ib_wq2qp_state(new_state);
4225 
4226 	if (ib_wq2qp_state(new_state) == qp_cur_state)
4227 		return 0;
4228 
4229 	if (new_state == IB_WQS_RDY) {
4230 		struct ib_qp_attr attr = {};
4231 
4232 		attr.port_num = qp->port;
4233 		attr_mask = IB_QP_PORT;
4234 
4235 		err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, &attr,
4236 					  attr_mask, IB_QPS_RESET, IB_QPS_INIT,
4237 					  udata);
4238 		if (err) {
4239 			pr_debug("WQN=0x%06x failed to apply RST->INIT on the HW QP\n",
4240 				 ibwq->wq_num);
4241 			return err;
4242 		}
4243 
4244 		qp_cur_state = IB_QPS_INIT;
4245 	}
4246 
4247 	attr_mask = 0;
4248 	err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, attr_mask,
4249 				  qp_cur_state,  qp_new_state, udata);
4250 
4251 	if (err && (qp_cur_state == IB_QPS_INIT)) {
4252 		qp_new_state = IB_QPS_RESET;
4253 		if (__mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL,
4254 					attr_mask, IB_QPS_INIT, IB_QPS_RESET,
4255 					udata)) {
4256 			pr_warn("WQN=0x%06x failed with reverting HW's resources failure\n",
4257 				ibwq->wq_num);
4258 			qp_new_state = IB_QPS_INIT;
4259 		}
4260 	}
4261 
4262 	qp->state = qp_new_state;
4263 
4264 	return err;
4265 }
4266 
4267 int mlx4_ib_modify_wq(struct ib_wq *ibwq, struct ib_wq_attr *wq_attr,
4268 		      u32 wq_attr_mask, struct ib_udata *udata)
4269 {
4270 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4271 	struct mlx4_ib_modify_wq ucmd = {};
4272 	size_t required_cmd_sz;
4273 	enum ib_wq_state cur_state, new_state;
4274 	int err = 0;
4275 
4276 	required_cmd_sz = offsetof(typeof(ucmd), reserved) +
4277 				   sizeof(ucmd.reserved);
4278 	if (udata->inlen < required_cmd_sz)
4279 		return -EINVAL;
4280 
4281 	if (udata->inlen > sizeof(ucmd) &&
4282 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4283 				 udata->inlen - sizeof(ucmd)))
4284 		return -EOPNOTSUPP;
4285 
4286 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4287 		return -EFAULT;
4288 
4289 	if (ucmd.comp_mask || ucmd.reserved)
4290 		return -EOPNOTSUPP;
4291 
4292 	if (wq_attr_mask & IB_WQ_FLAGS)
4293 		return -EOPNOTSUPP;
4294 
4295 	cur_state = wq_attr->curr_wq_state;
4296 	new_state = wq_attr->wq_state;
4297 
4298 	if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR))
4299 		return -EINVAL;
4300 
4301 	if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET))
4302 		return -EINVAL;
4303 
4304 	/* Need to protect against the parent RSS which also may modify WQ
4305 	 * state.
4306 	 */
4307 	mutex_lock(&qp->mutex);
4308 
4309 	/* Can update HW state only if a RSS QP has already associated to this
4310 	 * WQ, so we can apply its port on the WQ.
4311 	 */
4312 	if (qp->rss_usecnt)
4313 		err = _mlx4_ib_modify_wq(ibwq, new_state, udata);
4314 
4315 	if (!err)
4316 		ibwq->state = new_state;
4317 
4318 	mutex_unlock(&qp->mutex);
4319 
4320 	return err;
4321 }
4322 
4323 int mlx4_ib_destroy_wq(struct ib_wq *ibwq, struct ib_udata *udata)
4324 {
4325 	struct mlx4_ib_dev *dev = to_mdev(ibwq->device);
4326 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4327 
4328 	if (qp->counter_index)
4329 		mlx4_ib_free_qp_counter(dev, qp);
4330 
4331 	destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, udata);
4332 
4333 	kfree(qp);
4334 	return 0;
4335 }
4336 
4337 int mlx4_ib_create_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table,
4338 				 struct ib_rwq_ind_table_init_attr *init_attr,
4339 				 struct ib_udata *udata)
4340 {
4341 	struct mlx4_ib_create_rwq_ind_tbl_resp resp = {};
4342 	unsigned int ind_tbl_size = 1 << init_attr->log_ind_tbl_size;
4343 	struct ib_device *device = rwq_ind_table->device;
4344 	unsigned int base_wqn;
4345 	size_t min_resp_len;
4346 	int i, err = 0;
4347 
4348 	if (udata->inlen > 0 &&
4349 	    !ib_is_udata_cleared(udata, 0,
4350 				 udata->inlen))
4351 		return -EOPNOTSUPP;
4352 
4353 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4354 	if (udata->outlen && udata->outlen < min_resp_len)
4355 		return -EINVAL;
4356 
4357 	if (ind_tbl_size >
4358 	    device->attrs.rss_caps.max_rwq_indirection_table_size) {
4359 		pr_debug("log_ind_tbl_size = %d is bigger than supported = %d\n",
4360 			 ind_tbl_size,
4361 			 device->attrs.rss_caps.max_rwq_indirection_table_size);
4362 		return -EINVAL;
4363 	}
4364 
4365 	base_wqn = init_attr->ind_tbl[0]->wq_num;
4366 
4367 	if (base_wqn % ind_tbl_size) {
4368 		pr_debug("WQN=0x%x isn't aligned with indirection table size\n",
4369 			 base_wqn);
4370 		return -EINVAL;
4371 	}
4372 
4373 	for (i = 1; i < ind_tbl_size; i++) {
4374 		if (++base_wqn != init_attr->ind_tbl[i]->wq_num) {
4375 			pr_debug("indirection table's WQNs aren't consecutive\n");
4376 			return -EINVAL;
4377 		}
4378 	}
4379 
4380 	if (udata->outlen) {
4381 		resp.response_length = offsetof(typeof(resp), response_length) +
4382 					sizeof(resp.response_length);
4383 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4384 	}
4385 
4386 	return err;
4387 }
4388 
4389 struct mlx4_ib_drain_cqe {
4390 	struct ib_cqe cqe;
4391 	struct completion done;
4392 };
4393 
4394 static void mlx4_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
4395 {
4396 	struct mlx4_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
4397 						     struct mlx4_ib_drain_cqe,
4398 						     cqe);
4399 
4400 	complete(&cqe->done);
4401 }
4402 
4403 /* This function returns only once the drained WR was completed */
4404 static void handle_drain_completion(struct ib_cq *cq,
4405 				    struct mlx4_ib_drain_cqe *sdrain,
4406 				    struct mlx4_ib_dev *dev)
4407 {
4408 	struct mlx4_dev *mdev = dev->dev;
4409 
4410 	if (cq->poll_ctx == IB_POLL_DIRECT) {
4411 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
4412 			ib_process_cq_direct(cq, -1);
4413 		return;
4414 	}
4415 
4416 	if (mdev->persist->state == MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4417 		struct mlx4_ib_cq *mcq = to_mcq(cq);
4418 		bool triggered = false;
4419 		unsigned long flags;
4420 
4421 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
4422 		/* Make sure that the CQ handler won't run if wasn't run yet */
4423 		if (!mcq->mcq.reset_notify_added)
4424 			mcq->mcq.reset_notify_added = 1;
4425 		else
4426 			triggered = true;
4427 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
4428 
4429 		if (triggered) {
4430 			/* Wait for any scheduled/running task to be ended */
4431 			switch (cq->poll_ctx) {
4432 			case IB_POLL_SOFTIRQ:
4433 				irq_poll_disable(&cq->iop);
4434 				irq_poll_enable(&cq->iop);
4435 				break;
4436 			case IB_POLL_WORKQUEUE:
4437 				cancel_work_sync(&cq->work);
4438 				break;
4439 			default:
4440 				WARN_ON_ONCE(1);
4441 			}
4442 		}
4443 
4444 		/* Run the CQ handler - this makes sure that the drain WR will
4445 		 * be processed if wasn't processed yet.
4446 		 */
4447 		mcq->mcq.comp(&mcq->mcq);
4448 	}
4449 
4450 	wait_for_completion(&sdrain->done);
4451 }
4452 
4453 void mlx4_ib_drain_sq(struct ib_qp *qp)
4454 {
4455 	struct ib_cq *cq = qp->send_cq;
4456 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4457 	struct mlx4_ib_drain_cqe sdrain;
4458 	const struct ib_send_wr *bad_swr;
4459 	struct ib_rdma_wr swr = {
4460 		.wr = {
4461 			.next = NULL,
4462 			{ .wr_cqe	= &sdrain.cqe, },
4463 			.opcode	= IB_WR_RDMA_WRITE,
4464 		},
4465 	};
4466 	int ret;
4467 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
4468 	struct mlx4_dev *mdev = dev->dev;
4469 
4470 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4471 	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4472 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4473 		return;
4474 	}
4475 
4476 	sdrain.cqe.done = mlx4_ib_drain_qp_done;
4477 	init_completion(&sdrain.done);
4478 
4479 	ret = _mlx4_ib_post_send(qp, &swr.wr, &bad_swr, true);
4480 	if (ret) {
4481 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4482 		return;
4483 	}
4484 
4485 	handle_drain_completion(cq, &sdrain, dev);
4486 }
4487 
4488 void mlx4_ib_drain_rq(struct ib_qp *qp)
4489 {
4490 	struct ib_cq *cq = qp->recv_cq;
4491 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4492 	struct mlx4_ib_drain_cqe rdrain;
4493 	struct ib_recv_wr rwr = {};
4494 	const struct ib_recv_wr *bad_rwr;
4495 	int ret;
4496 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
4497 	struct mlx4_dev *mdev = dev->dev;
4498 
4499 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4500 	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4501 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4502 		return;
4503 	}
4504 
4505 	rwr.wr_cqe = &rdrain.cqe;
4506 	rdrain.cqe.done = mlx4_ib_drain_qp_done;
4507 	init_completion(&rdrain.done);
4508 
4509 	ret = _mlx4_ib_post_recv(qp, &rwr, &bad_rwr, true);
4510 	if (ret) {
4511 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4512 		return;
4513 	}
4514 
4515 	handle_drain_completion(cq, &rdrain, dev);
4516 }
4517 
4518 int mlx4_ib_qp_event_init(void)
4519 {
4520 	mlx4_ib_qp_event_wq = alloc_ordered_workqueue("mlx4_ib_qp_event_wq", 0);
4521 	if (!mlx4_ib_qp_event_wq)
4522 		return -ENOMEM;
4523 
4524 	return 0;
4525 }
4526 
4527 void mlx4_ib_qp_event_cleanup(void)
4528 {
4529 	destroy_workqueue(mlx4_ib_qp_event_wq);
4530 }
4531