xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision 2a598d0b)
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(sizeof(*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_array(qp->rq.wqe_cnt, sizeof(struct mlx4_ib_buf),
476 			      GFP_KERNEL);
477 	if (!qp->sqp_proxy_rcv)
478 		return -ENOMEM;
479 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
480 		qp->sqp_proxy_rcv[i].addr =
481 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
482 				GFP_KERNEL);
483 		if (!qp->sqp_proxy_rcv[i].addr)
484 			goto err;
485 		qp->sqp_proxy_rcv[i].map =
486 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
487 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
488 					  DMA_FROM_DEVICE);
489 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
490 			kfree(qp->sqp_proxy_rcv[i].addr);
491 			goto err;
492 		}
493 	}
494 	return 0;
495 
496 err:
497 	while (i > 0) {
498 		--i;
499 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
500 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
501 				    DMA_FROM_DEVICE);
502 		kfree(qp->sqp_proxy_rcv[i].addr);
503 	}
504 	kfree(qp->sqp_proxy_rcv);
505 	qp->sqp_proxy_rcv = NULL;
506 	return -ENOMEM;
507 }
508 
509 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
510 {
511 	int i;
512 
513 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
514 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
515 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
516 				    DMA_FROM_DEVICE);
517 		kfree(qp->sqp_proxy_rcv[i].addr);
518 	}
519 	kfree(qp->sqp_proxy_rcv);
520 }
521 
522 static bool qp_has_rq(struct ib_qp_init_attr *attr)
523 {
524 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
525 		return false;
526 
527 	return !attr->srq;
528 }
529 
530 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
531 {
532 	int i;
533 	for (i = 0; i < dev->caps.num_ports; i++) {
534 		if (qpn == dev->caps.spec_qps[i].qp0_proxy)
535 			return !!dev->caps.spec_qps[i].qp0_qkey;
536 	}
537 	return 0;
538 }
539 
540 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
541 				    struct mlx4_ib_qp *qp)
542 {
543 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
544 	mlx4_counter_free(dev->dev, qp->counter_index->index);
545 	list_del(&qp->counter_index->list);
546 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
547 
548 	kfree(qp->counter_index);
549 	qp->counter_index = NULL;
550 }
551 
552 static int set_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_rss *rss_ctx,
553 		      struct ib_qp_init_attr *init_attr,
554 		      struct mlx4_ib_create_qp_rss *ucmd)
555 {
556 	rss_ctx->base_qpn_tbl_sz = init_attr->rwq_ind_tbl->ind_tbl[0]->wq_num |
557 		(init_attr->rwq_ind_tbl->log_ind_tbl_size << 24);
558 
559 	if ((ucmd->rx_hash_function == MLX4_IB_RX_HASH_FUNC_TOEPLITZ) &&
560 	    (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS_TOP)) {
561 		memcpy(rss_ctx->rss_key, ucmd->rx_hash_key,
562 		       MLX4_EN_RSS_KEY_SIZE);
563 	} else {
564 		pr_debug("RX Hash function is not supported\n");
565 		return (-EOPNOTSUPP);
566 	}
567 
568 	if (ucmd->rx_hash_fields_mask & ~(MLX4_IB_RX_HASH_SRC_IPV4	|
569 					  MLX4_IB_RX_HASH_DST_IPV4	|
570 					  MLX4_IB_RX_HASH_SRC_IPV6	|
571 					  MLX4_IB_RX_HASH_DST_IPV6	|
572 					  MLX4_IB_RX_HASH_SRC_PORT_TCP	|
573 					  MLX4_IB_RX_HASH_DST_PORT_TCP	|
574 					  MLX4_IB_RX_HASH_SRC_PORT_UDP	|
575 					  MLX4_IB_RX_HASH_DST_PORT_UDP  |
576 					  MLX4_IB_RX_HASH_INNER)) {
577 		pr_debug("RX Hash fields_mask has unsupported mask (0x%llx)\n",
578 			 ucmd->rx_hash_fields_mask);
579 		return (-EOPNOTSUPP);
580 	}
581 
582 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) &&
583 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
584 		rss_ctx->flags = MLX4_RSS_IPV4;
585 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) ||
586 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
587 		pr_debug("RX Hash fields_mask is not supported - both IPv4 SRC and DST must be set\n");
588 		return (-EOPNOTSUPP);
589 	}
590 
591 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) &&
592 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
593 		rss_ctx->flags |= MLX4_RSS_IPV6;
594 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) ||
595 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
596 		pr_debug("RX Hash fields_mask is not supported - both IPv6 SRC and DST must be set\n");
597 		return (-EOPNOTSUPP);
598 	}
599 
600 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) &&
601 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
602 		if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS)) {
603 			pr_debug("RX Hash fields_mask for UDP is not supported\n");
604 			return (-EOPNOTSUPP);
605 		}
606 
607 		if (rss_ctx->flags & MLX4_RSS_IPV4)
608 			rss_ctx->flags |= MLX4_RSS_UDP_IPV4;
609 		if (rss_ctx->flags & MLX4_RSS_IPV6)
610 			rss_ctx->flags |= MLX4_RSS_UDP_IPV6;
611 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
612 			pr_debug("RX Hash fields_mask is not supported - UDP must be set with IPv4 or IPv6\n");
613 			return (-EOPNOTSUPP);
614 		}
615 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) ||
616 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
617 		pr_debug("RX Hash fields_mask is not supported - both UDP SRC and DST must be set\n");
618 		return (-EOPNOTSUPP);
619 	}
620 
621 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) &&
622 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
623 		if (rss_ctx->flags & MLX4_RSS_IPV4)
624 			rss_ctx->flags |= MLX4_RSS_TCP_IPV4;
625 		if (rss_ctx->flags & MLX4_RSS_IPV6)
626 			rss_ctx->flags |= MLX4_RSS_TCP_IPV6;
627 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
628 			pr_debug("RX Hash fields_mask is not supported - TCP must be set with IPv4 or IPv6\n");
629 			return (-EOPNOTSUPP);
630 		}
631 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) ||
632 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
633 		pr_debug("RX Hash fields_mask is not supported - both TCP SRC and DST must be set\n");
634 		return (-EOPNOTSUPP);
635 	}
636 
637 	if (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_INNER) {
638 		if (dev->dev->caps.tunnel_offload_mode ==
639 		    MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
640 			/*
641 			 * Hash according to inner headers if exist, otherwise
642 			 * according to outer headers.
643 			 */
644 			rss_ctx->flags |= MLX4_RSS_BY_INNER_HEADERS_IPONLY;
645 		} else {
646 			pr_debug("RSS Hash for inner headers isn't supported\n");
647 			return (-EOPNOTSUPP);
648 		}
649 	}
650 
651 	return 0;
652 }
653 
654 static int create_qp_rss(struct mlx4_ib_dev *dev,
655 			 struct ib_qp_init_attr *init_attr,
656 			 struct mlx4_ib_create_qp_rss *ucmd,
657 			 struct mlx4_ib_qp *qp)
658 {
659 	int qpn;
660 	int err;
661 
662 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
663 
664 	err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 0, qp->mqp.usage);
665 	if (err)
666 		return err;
667 
668 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
669 	if (err)
670 		goto err_qpn;
671 
672 	INIT_LIST_HEAD(&qp->gid_list);
673 	INIT_LIST_HEAD(&qp->steering_rules);
674 
675 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
676 	qp->state = IB_QPS_RESET;
677 
678 	/* Set dummy send resources to be compatible with HV and PRM */
679 	qp->sq_no_prefetch = 1;
680 	qp->sq.wqe_cnt = 1;
681 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
682 	qp->buf_size = qp->sq.wqe_cnt << MLX4_IB_MIN_SQ_STRIDE;
683 	qp->mtt = (to_mqp(
684 		   (struct ib_qp *)init_attr->rwq_ind_tbl->ind_tbl[0]))->mtt;
685 
686 	qp->rss_ctx = kzalloc(sizeof(*qp->rss_ctx), GFP_KERNEL);
687 	if (!qp->rss_ctx) {
688 		err = -ENOMEM;
689 		goto err_qp_alloc;
690 	}
691 
692 	err = set_qp_rss(dev, qp->rss_ctx, init_attr, ucmd);
693 	if (err)
694 		goto err;
695 
696 	return 0;
697 
698 err:
699 	kfree(qp->rss_ctx);
700 
701 err_qp_alloc:
702 	mlx4_qp_remove(dev->dev, &qp->mqp);
703 	mlx4_qp_free(dev->dev, &qp->mqp);
704 
705 err_qpn:
706 	mlx4_qp_release_range(dev->dev, qpn, 1);
707 	return err;
708 }
709 
710 static int _mlx4_ib_create_qp_rss(struct ib_pd *pd, struct mlx4_ib_qp *qp,
711 				  struct ib_qp_init_attr *init_attr,
712 				  struct ib_udata *udata)
713 {
714 	struct mlx4_ib_create_qp_rss ucmd = {};
715 	size_t required_cmd_sz;
716 	int err;
717 
718 	if (!udata) {
719 		pr_debug("RSS QP with NULL udata\n");
720 		return -EINVAL;
721 	}
722 
723 	if (udata->outlen)
724 		return -EOPNOTSUPP;
725 
726 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) +
727 					sizeof(ucmd.reserved1);
728 	if (udata->inlen < required_cmd_sz) {
729 		pr_debug("invalid inlen\n");
730 		return -EINVAL;
731 	}
732 
733 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
734 		pr_debug("copy failed\n");
735 		return -EFAULT;
736 	}
737 
738 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)))
739 		return -EOPNOTSUPP;
740 
741 	if (ucmd.comp_mask || ucmd.reserved1)
742 		return -EOPNOTSUPP;
743 
744 	if (udata->inlen > sizeof(ucmd) &&
745 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
746 				 udata->inlen - sizeof(ucmd))) {
747 		pr_debug("inlen is not supported\n");
748 		return -EOPNOTSUPP;
749 	}
750 
751 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
752 		pr_debug("RSS QP with unsupported QP type %d\n",
753 			 init_attr->qp_type);
754 		return -EOPNOTSUPP;
755 	}
756 
757 	if (init_attr->create_flags) {
758 		pr_debug("RSS QP doesn't support create flags\n");
759 		return -EOPNOTSUPP;
760 	}
761 
762 	if (init_attr->send_cq || init_attr->cap.max_send_wr) {
763 		pr_debug("RSS QP with unsupported send attributes\n");
764 		return -EOPNOTSUPP;
765 	}
766 
767 	qp->pri.vid = 0xFFFF;
768 	qp->alt.vid = 0xFFFF;
769 
770 	err = create_qp_rss(to_mdev(pd->device), init_attr, &ucmd, qp);
771 	if (err)
772 		return err;
773 
774 	qp->ibqp.qp_num = qp->mqp.qpn;
775 	return 0;
776 }
777 
778 /*
779  * This function allocates a WQN from a range which is consecutive and aligned
780  * to its size. In case the range is full, then it creates a new range and
781  * allocates WQN from it. The new range will be used for following allocations.
782  */
783 static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context,
784 			     struct mlx4_ib_qp *qp, int range_size, int *wqn)
785 {
786 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
787 	struct mlx4_wqn_range *range;
788 	int err = 0;
789 
790 	mutex_lock(&context->wqn_ranges_mutex);
791 
792 	range = list_first_entry_or_null(&context->wqn_ranges_list,
793 					 struct mlx4_wqn_range, list);
794 
795 	if (!range || (range->refcount == range->size) || range->dirty) {
796 		range = kzalloc(sizeof(*range), GFP_KERNEL);
797 		if (!range) {
798 			err = -ENOMEM;
799 			goto out;
800 		}
801 
802 		err = mlx4_qp_reserve_range(dev->dev, range_size,
803 					    range_size, &range->base_wqn, 0,
804 					    qp->mqp.usage);
805 		if (err) {
806 			kfree(range);
807 			goto out;
808 		}
809 
810 		range->size = range_size;
811 		list_add(&range->list, &context->wqn_ranges_list);
812 	} else if (range_size != 1) {
813 		/*
814 		 * Requesting a new range (>1) when last range is still open, is
815 		 * not valid.
816 		 */
817 		err = -EINVAL;
818 		goto out;
819 	}
820 
821 	qp->wqn_range = range;
822 
823 	*wqn = range->base_wqn + range->refcount;
824 
825 	range->refcount++;
826 
827 out:
828 	mutex_unlock(&context->wqn_ranges_mutex);
829 
830 	return err;
831 }
832 
833 static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context,
834 				struct mlx4_ib_qp *qp, bool dirty_release)
835 {
836 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
837 	struct mlx4_wqn_range *range;
838 
839 	mutex_lock(&context->wqn_ranges_mutex);
840 
841 	range = qp->wqn_range;
842 
843 	range->refcount--;
844 	if (!range->refcount) {
845 		mlx4_qp_release_range(dev->dev, range->base_wqn,
846 				      range->size);
847 		list_del(&range->list);
848 		kfree(range);
849 	} else if (dirty_release) {
850 	/*
851 	 * A range which one of its WQNs is destroyed, won't be able to be
852 	 * reused for further WQN allocations.
853 	 * The next created WQ will allocate a new range.
854 	 */
855 		range->dirty = true;
856 	}
857 
858 	mutex_unlock(&context->wqn_ranges_mutex);
859 }
860 
861 static int create_rq(struct ib_pd *pd, struct ib_qp_init_attr *init_attr,
862 		     struct ib_udata *udata, struct mlx4_ib_qp *qp)
863 {
864 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
865 	int qpn;
866 	int err;
867 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
868 		udata, struct mlx4_ib_ucontext, ibucontext);
869 	struct mlx4_ib_cq *mcq;
870 	unsigned long flags;
871 	int range_size;
872 	struct mlx4_ib_create_wq wq;
873 	size_t copy_len;
874 	int shift;
875 	int n;
876 
877 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
878 
879 	spin_lock_init(&qp->sq.lock);
880 	spin_lock_init(&qp->rq.lock);
881 	INIT_LIST_HEAD(&qp->gid_list);
882 	INIT_LIST_HEAD(&qp->steering_rules);
883 
884 	qp->state = IB_QPS_RESET;
885 
886 	copy_len = min(sizeof(struct mlx4_ib_create_wq), udata->inlen);
887 
888 	if (ib_copy_from_udata(&wq, udata, copy_len)) {
889 		err = -EFAULT;
890 		goto err;
891 	}
892 
893 	if (wq.comp_mask || wq.reserved[0] || wq.reserved[1] ||
894 	    wq.reserved[2]) {
895 		pr_debug("user command isn't supported\n");
896 		err = -EOPNOTSUPP;
897 		goto err;
898 	}
899 
900 	if (wq.log_range_size > ilog2(dev->dev->caps.max_rss_tbl_sz)) {
901 		pr_debug("WQN range size must be equal or smaller than %d\n",
902 			 dev->dev->caps.max_rss_tbl_sz);
903 		err = -EOPNOTSUPP;
904 		goto err;
905 	}
906 	range_size = 1 << wq.log_range_size;
907 
908 	if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS)
909 		qp->flags |= MLX4_IB_QP_SCATTER_FCS;
910 
911 	err = set_rq_size(dev, &init_attr->cap, true, true, qp, qp->inl_recv_sz);
912 	if (err)
913 		goto err;
914 
915 	qp->sq_no_prefetch = 1;
916 	qp->sq.wqe_cnt = 1;
917 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
918 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
919 		       (qp->sq.wqe_cnt << qp->sq.wqe_shift);
920 
921 	qp->umem = ib_umem_get(pd->device, wq.buf_addr, qp->buf_size, 0);
922 	if (IS_ERR(qp->umem)) {
923 		err = PTR_ERR(qp->umem);
924 		goto err;
925 	}
926 
927 	shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
928 	err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
929 
930 	if (err)
931 		goto err_buf;
932 
933 	err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
934 	if (err)
935 		goto err_mtt;
936 
937 	err = mlx4_ib_db_map_user(udata, wq.db_addr, &qp->db);
938 	if (err)
939 		goto err_mtt;
940 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
941 
942 	err = mlx4_ib_alloc_wqn(context, qp, range_size, &qpn);
943 	if (err)
944 		goto err_wrid;
945 
946 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
947 	if (err)
948 		goto err_qpn;
949 
950 	/*
951 	 * Hardware wants QPN written in big-endian order (after
952 	 * shifting) for send doorbell.  Precompute this value to save
953 	 * a little bit when posting sends.
954 	 */
955 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
956 
957 	qp->mqp.event = mlx4_ib_wq_event;
958 
959 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
960 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
961 			 to_mcq(init_attr->recv_cq));
962 	/* Maintain device to QPs access, needed for further handling
963 	 * via reset flow
964 	 */
965 	list_add_tail(&qp->qps_list, &dev->qp_list);
966 	/* Maintain CQ to QPs access, needed for further handling
967 	 * via reset flow
968 	 */
969 	mcq = to_mcq(init_attr->send_cq);
970 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
971 	mcq = to_mcq(init_attr->recv_cq);
972 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
973 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
974 			   to_mcq(init_attr->recv_cq));
975 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
976 	return 0;
977 
978 err_qpn:
979 	mlx4_ib_release_wqn(context, qp, 0);
980 err_wrid:
981 	mlx4_ib_db_unmap_user(context, &qp->db);
982 
983 err_mtt:
984 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
985 err_buf:
986 	ib_umem_release(qp->umem);
987 err:
988 	return err;
989 }
990 
991 static int create_qp_common(struct ib_pd *pd, struct ib_qp_init_attr *init_attr,
992 			    struct ib_udata *udata, int sqpn,
993 			    struct mlx4_ib_qp *qp)
994 {
995 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
996 	int qpn;
997 	int err;
998 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
999 		udata, struct mlx4_ib_ucontext, ibucontext);
1000 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
1001 	struct mlx4_ib_cq *mcq;
1002 	unsigned long flags;
1003 
1004 	/* When tunneling special qps, we use a plain UD qp */
1005 	if (sqpn) {
1006 		if (mlx4_is_mfunc(dev->dev) &&
1007 		    (!mlx4_is_master(dev->dev) ||
1008 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
1009 			if (init_attr->qp_type == IB_QPT_GSI)
1010 				qp_type = MLX4_IB_QPT_PROXY_GSI;
1011 			else {
1012 				if (mlx4_is_master(dev->dev) ||
1013 				    qp0_enabled_vf(dev->dev, sqpn))
1014 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
1015 				else
1016 					qp_type = MLX4_IB_QPT_PROXY_SMI;
1017 			}
1018 		}
1019 		qpn = sqpn;
1020 		/* add extra sg entry for tunneling */
1021 		init_attr->cap.max_recv_sge++;
1022 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
1023 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
1024 			container_of(init_attr,
1025 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
1026 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
1027 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
1028 		    !mlx4_is_master(dev->dev))
1029 			return -EINVAL;
1030 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
1031 			qp_type = MLX4_IB_QPT_TUN_GSI;
1032 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
1033 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
1034 					     tnl_init->port))
1035 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
1036 		else
1037 			qp_type = MLX4_IB_QPT_TUN_SMI;
1038 		/* we are definitely in the PPF here, since we are creating
1039 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
1040 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
1041 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
1042 		sqpn = qpn;
1043 	}
1044 
1045 	if (init_attr->qp_type == IB_QPT_SMI ||
1046 	    init_attr->qp_type == IB_QPT_GSI || qp_type == MLX4_IB_QPT_SMI ||
1047 	    qp_type == MLX4_IB_QPT_GSI ||
1048 	    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
1049 			MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
1050 		qp->sqp = kzalloc(sizeof(struct mlx4_ib_sqp), GFP_KERNEL);
1051 		if (!qp->sqp)
1052 			return -ENOMEM;
1053 	}
1054 
1055 	qp->mlx4_ib_qp_type = qp_type;
1056 
1057 	spin_lock_init(&qp->sq.lock);
1058 	spin_lock_init(&qp->rq.lock);
1059 	INIT_LIST_HEAD(&qp->gid_list);
1060 	INIT_LIST_HEAD(&qp->steering_rules);
1061 
1062 	qp->state = IB_QPS_RESET;
1063 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1064 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1065 
1066 	if (udata) {
1067 		struct mlx4_ib_create_qp ucmd;
1068 		size_t copy_len;
1069 		int shift;
1070 		int n;
1071 
1072 		copy_len = sizeof(struct mlx4_ib_create_qp);
1073 
1074 		if (ib_copy_from_udata(&ucmd, udata, copy_len)) {
1075 			err = -EFAULT;
1076 			goto err;
1077 		}
1078 
1079 		qp->inl_recv_sz = ucmd.inl_recv_sz;
1080 
1081 		if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1082 			if (!(dev->dev->caps.flags &
1083 			      MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
1084 				pr_debug("scatter FCS is unsupported\n");
1085 				err = -EOPNOTSUPP;
1086 				goto err;
1087 			}
1088 
1089 			qp->flags |= MLX4_IB_QP_SCATTER_FCS;
1090 		}
1091 
1092 		err = set_rq_size(dev, &init_attr->cap, udata,
1093 				  qp_has_rq(init_attr), qp, qp->inl_recv_sz);
1094 		if (err)
1095 			goto err;
1096 
1097 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
1098 
1099 		err = set_user_sq_size(dev, qp, &ucmd);
1100 		if (err)
1101 			goto err;
1102 
1103 		qp->umem =
1104 			ib_umem_get(pd->device, ucmd.buf_addr, qp->buf_size, 0);
1105 		if (IS_ERR(qp->umem)) {
1106 			err = PTR_ERR(qp->umem);
1107 			goto err;
1108 		}
1109 
1110 		shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
1111 		err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
1112 
1113 		if (err)
1114 			goto err_buf;
1115 
1116 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
1117 		if (err)
1118 			goto err_mtt;
1119 
1120 		if (qp_has_rq(init_attr)) {
1121 			err = mlx4_ib_db_map_user(udata, ucmd.db_addr, &qp->db);
1122 			if (err)
1123 				goto err_mtt;
1124 		}
1125 		qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
1126 	} else {
1127 		err = set_rq_size(dev, &init_attr->cap, udata,
1128 				  qp_has_rq(init_attr), qp, 0);
1129 		if (err)
1130 			goto err;
1131 
1132 		qp->sq_no_prefetch = 0;
1133 
1134 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
1135 			qp->flags |= MLX4_IB_QP_LSO;
1136 
1137 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1138 			if (dev->steering_support ==
1139 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
1140 				qp->flags |= MLX4_IB_QP_NETIF;
1141 			else {
1142 				err = -EINVAL;
1143 				goto err;
1144 			}
1145 		}
1146 
1147 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
1148 		if (err)
1149 			goto err;
1150 
1151 		if (qp_has_rq(init_attr)) {
1152 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
1153 			if (err)
1154 				goto err;
1155 
1156 			*qp->db.db = 0;
1157 		}
1158 
1159 		if (mlx4_buf_alloc(dev->dev, qp->buf_size,  PAGE_SIZE * 2,
1160 				   &qp->buf)) {
1161 			err = -ENOMEM;
1162 			goto err_db;
1163 		}
1164 
1165 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
1166 				    &qp->mtt);
1167 		if (err)
1168 			goto err_buf;
1169 
1170 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
1171 		if (err)
1172 			goto err_mtt;
1173 
1174 		qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1175 					     sizeof(u64), GFP_KERNEL);
1176 		qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1177 					     sizeof(u64), GFP_KERNEL);
1178 		if (!qp->sq.wrid || !qp->rq.wrid) {
1179 			err = -ENOMEM;
1180 			goto err_wrid;
1181 		}
1182 		qp->mqp.usage = MLX4_RES_USAGE_DRIVER;
1183 	}
1184 
1185 	if (sqpn) {
1186 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1187 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
1188 			if (alloc_proxy_bufs(pd->device, qp)) {
1189 				err = -ENOMEM;
1190 				goto err_wrid;
1191 			}
1192 		}
1193 	} else {
1194 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
1195 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
1196 		 * VLAN insertion. */
1197 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
1198 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
1199 						    (init_attr->cap.max_send_wr ?
1200 						     MLX4_RESERVE_ETH_BF_QP : 0) |
1201 						    (init_attr->cap.max_recv_wr ?
1202 						     MLX4_RESERVE_A0_QP : 0),
1203 						    qp->mqp.usage);
1204 		else
1205 			if (qp->flags & MLX4_IB_QP_NETIF)
1206 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
1207 			else
1208 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
1209 							    &qpn, 0, qp->mqp.usage);
1210 		if (err)
1211 			goto err_proxy;
1212 	}
1213 
1214 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1215 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1216 
1217 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
1218 	if (err)
1219 		goto err_qpn;
1220 
1221 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1222 		qp->mqp.qpn |= (1 << 23);
1223 
1224 	/*
1225 	 * Hardware wants QPN written in big-endian order (after
1226 	 * shifting) for send doorbell.  Precompute this value to save
1227 	 * a little bit when posting sends.
1228 	 */
1229 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1230 
1231 	qp->mqp.event = mlx4_ib_qp_event;
1232 
1233 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1234 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
1235 			 to_mcq(init_attr->recv_cq));
1236 	/* Maintain device to QPs access, needed for further handling
1237 	 * via reset flow
1238 	 */
1239 	list_add_tail(&qp->qps_list, &dev->qp_list);
1240 	/* Maintain CQ to QPs access, needed for further handling
1241 	 * via reset flow
1242 	 */
1243 	mcq = to_mcq(init_attr->send_cq);
1244 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
1245 	mcq = to_mcq(init_attr->recv_cq);
1246 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
1247 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
1248 			   to_mcq(init_attr->recv_cq));
1249 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1250 	return 0;
1251 
1252 err_qpn:
1253 	if (!sqpn) {
1254 		if (qp->flags & MLX4_IB_QP_NETIF)
1255 			mlx4_ib_steer_qp_free(dev, qpn, 1);
1256 		else
1257 			mlx4_qp_release_range(dev->dev, qpn, 1);
1258 	}
1259 err_proxy:
1260 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1261 		free_proxy_bufs(pd->device, qp);
1262 err_wrid:
1263 	if (udata) {
1264 		if (qp_has_rq(init_attr))
1265 			mlx4_ib_db_unmap_user(context, &qp->db);
1266 	} else {
1267 		kvfree(qp->sq.wrid);
1268 		kvfree(qp->rq.wrid);
1269 	}
1270 
1271 err_mtt:
1272 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1273 
1274 err_buf:
1275 	if (!qp->umem)
1276 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1277 	ib_umem_release(qp->umem);
1278 
1279 err_db:
1280 	if (!udata && qp_has_rq(init_attr))
1281 		mlx4_db_free(dev->dev, &qp->db);
1282 
1283 err:
1284 	kfree(qp->sqp);
1285 	return err;
1286 }
1287 
1288 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
1289 {
1290 	switch (state) {
1291 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
1292 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
1293 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
1294 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
1295 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
1296 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
1297 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
1298 	default:		return -1;
1299 	}
1300 }
1301 
1302 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1303 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1304 {
1305 	if (send_cq == recv_cq) {
1306 		spin_lock(&send_cq->lock);
1307 		__acquire(&recv_cq->lock);
1308 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1309 		spin_lock(&send_cq->lock);
1310 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1311 	} else {
1312 		spin_lock(&recv_cq->lock);
1313 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1314 	}
1315 }
1316 
1317 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1318 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1319 {
1320 	if (send_cq == recv_cq) {
1321 		__release(&recv_cq->lock);
1322 		spin_unlock(&send_cq->lock);
1323 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1324 		spin_unlock(&recv_cq->lock);
1325 		spin_unlock(&send_cq->lock);
1326 	} else {
1327 		spin_unlock(&send_cq->lock);
1328 		spin_unlock(&recv_cq->lock);
1329 	}
1330 }
1331 
1332 static void del_gid_entries(struct mlx4_ib_qp *qp)
1333 {
1334 	struct mlx4_ib_gid_entry *ge, *tmp;
1335 
1336 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1337 		list_del(&ge->list);
1338 		kfree(ge);
1339 	}
1340 }
1341 
1342 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
1343 {
1344 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
1345 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
1346 	else
1347 		return to_mpd(qp->ibqp.pd);
1348 }
1349 
1350 static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src,
1351 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1352 {
1353 	switch (qp->ibqp.qp_type) {
1354 	case IB_QPT_XRC_TGT:
1355 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1356 		*recv_cq = *send_cq;
1357 		break;
1358 	case IB_QPT_XRC_INI:
1359 		*send_cq = to_mcq(qp->ibqp.send_cq);
1360 		*recv_cq = *send_cq;
1361 		break;
1362 	default:
1363 		*recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) :
1364 						     to_mcq(qp->ibwq.cq);
1365 		*send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) :
1366 						     *recv_cq;
1367 		break;
1368 	}
1369 }
1370 
1371 static void destroy_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1372 {
1373 	if (qp->state != IB_QPS_RESET) {
1374 		int i;
1375 
1376 		for (i = 0; i < (1 << qp->ibqp.rwq_ind_tbl->log_ind_tbl_size);
1377 		     i++) {
1378 			struct ib_wq *ibwq = qp->ibqp.rwq_ind_tbl->ind_tbl[i];
1379 			struct mlx4_ib_qp *wq =	to_mqp((struct ib_qp *)ibwq);
1380 
1381 			mutex_lock(&wq->mutex);
1382 
1383 			wq->rss_usecnt--;
1384 
1385 			mutex_unlock(&wq->mutex);
1386 		}
1387 
1388 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1389 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1390 			pr_warn("modify QP %06x to RESET failed.\n",
1391 				qp->mqp.qpn);
1392 	}
1393 
1394 	mlx4_qp_remove(dev->dev, &qp->mqp);
1395 	mlx4_qp_free(dev->dev, &qp->mqp);
1396 	mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1397 	del_gid_entries(qp);
1398 }
1399 
1400 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1401 			      enum mlx4_ib_source_type src,
1402 			      struct ib_udata *udata)
1403 {
1404 	struct mlx4_ib_cq *send_cq, *recv_cq;
1405 	unsigned long flags;
1406 
1407 	if (qp->state != IB_QPS_RESET) {
1408 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1409 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1410 			pr_warn("modify QP %06x to RESET failed.\n",
1411 			       qp->mqp.qpn);
1412 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1413 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1414 			qp->pri.smac = 0;
1415 			qp->pri.smac_port = 0;
1416 		}
1417 		if (qp->alt.smac) {
1418 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1419 			qp->alt.smac = 0;
1420 		}
1421 		if (qp->pri.vid < 0x1000) {
1422 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1423 			qp->pri.vid = 0xFFFF;
1424 			qp->pri.candidate_vid = 0xFFFF;
1425 			qp->pri.update_vid = 0;
1426 		}
1427 		if (qp->alt.vid < 0x1000) {
1428 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1429 			qp->alt.vid = 0xFFFF;
1430 			qp->alt.candidate_vid = 0xFFFF;
1431 			qp->alt.update_vid = 0;
1432 		}
1433 	}
1434 
1435 	get_cqs(qp, src, &send_cq, &recv_cq);
1436 
1437 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1438 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1439 
1440 	/* del from lists under both locks above to protect reset flow paths */
1441 	list_del(&qp->qps_list);
1442 	list_del(&qp->cq_send_list);
1443 	list_del(&qp->cq_recv_list);
1444 	if (!udata) {
1445 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1446 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1447 		if (send_cq != recv_cq)
1448 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1449 	}
1450 
1451 	mlx4_qp_remove(dev->dev, &qp->mqp);
1452 
1453 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1454 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1455 
1456 	mlx4_qp_free(dev->dev, &qp->mqp);
1457 
1458 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1459 		if (qp->flags & MLX4_IB_QP_NETIF)
1460 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1461 		else if (src == MLX4_IB_RWQ_SRC)
1462 			mlx4_ib_release_wqn(
1463 				rdma_udata_to_drv_context(
1464 					udata,
1465 					struct mlx4_ib_ucontext,
1466 					ibucontext),
1467 				qp, 1);
1468 		else
1469 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1470 	}
1471 
1472 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1473 
1474 	if (udata) {
1475 		if (qp->rq.wqe_cnt) {
1476 			struct mlx4_ib_ucontext *mcontext =
1477 				rdma_udata_to_drv_context(
1478 					udata,
1479 					struct mlx4_ib_ucontext,
1480 					ibucontext);
1481 
1482 			mlx4_ib_db_unmap_user(mcontext, &qp->db);
1483 		}
1484 	} else {
1485 		kvfree(qp->sq.wrid);
1486 		kvfree(qp->rq.wrid);
1487 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1488 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1489 			free_proxy_bufs(&dev->ib_dev, qp);
1490 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1491 		if (qp->rq.wqe_cnt)
1492 			mlx4_db_free(dev->dev, &qp->db);
1493 	}
1494 	ib_umem_release(qp->umem);
1495 
1496 	del_gid_entries(qp);
1497 }
1498 
1499 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1500 {
1501 	/* Native or PPF */
1502 	if (!mlx4_is_mfunc(dev->dev) ||
1503 	    (mlx4_is_master(dev->dev) &&
1504 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1505 		return  dev->dev->phys_caps.base_sqpn +
1506 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1507 			attr->port_num - 1;
1508 	}
1509 	/* PF or VF -- creating proxies */
1510 	if (attr->qp_type == IB_QPT_SMI)
1511 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp0_proxy;
1512 	else
1513 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp1_proxy;
1514 }
1515 
1516 static int _mlx4_ib_create_qp(struct ib_pd *pd, struct mlx4_ib_qp *qp,
1517 			      struct ib_qp_init_attr *init_attr,
1518 			      struct ib_udata *udata)
1519 {
1520 	int err;
1521 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1522 	u16 xrcdn = 0;
1523 
1524 	if (init_attr->rwq_ind_tbl)
1525 		return _mlx4_ib_create_qp_rss(pd, qp, init_attr, udata);
1526 
1527 	/*
1528 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1529 	 * and only for kernel UD QPs.
1530 	 */
1531 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1532 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1533 					MLX4_IB_SRIOV_TUNNEL_QP |
1534 					MLX4_IB_SRIOV_SQP |
1535 					MLX4_IB_QP_NETIF |
1536 					MLX4_IB_QP_CREATE_ROCE_V2_GSI))
1537 		return -EOPNOTSUPP;
1538 
1539 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1540 		if (init_attr->qp_type != IB_QPT_UD)
1541 			return -EINVAL;
1542 	}
1543 
1544 	if (init_attr->create_flags) {
1545 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1546 			return -EINVAL;
1547 
1548 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1549 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1550 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1551 		     init_attr->qp_type != IB_QPT_UD) ||
1552 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1553 		     init_attr->qp_type > IB_QPT_GSI) ||
1554 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1555 		     init_attr->qp_type != IB_QPT_GSI))
1556 			return -EINVAL;
1557 	}
1558 
1559 	switch (init_attr->qp_type) {
1560 	case IB_QPT_XRC_TGT:
1561 		pd = to_mxrcd(init_attr->xrcd)->pd;
1562 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1563 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1564 		fallthrough;
1565 	case IB_QPT_XRC_INI:
1566 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1567 			return -ENOSYS;
1568 		init_attr->recv_cq = init_attr->send_cq;
1569 		fallthrough;
1570 	case IB_QPT_RC:
1571 	case IB_QPT_UC:
1572 	case IB_QPT_RAW_PACKET:
1573 	case IB_QPT_UD:
1574 		qp->pri.vid = 0xFFFF;
1575 		qp->alt.vid = 0xFFFF;
1576 		err = create_qp_common(pd, init_attr, udata, 0, qp);
1577 		if (err)
1578 			return err;
1579 
1580 		qp->ibqp.qp_num = qp->mqp.qpn;
1581 		qp->xrcdn = xrcdn;
1582 		break;
1583 	case IB_QPT_SMI:
1584 	case IB_QPT_GSI:
1585 	{
1586 		int sqpn;
1587 
1588 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1589 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev,
1590 							1, 1, &sqpn, 0,
1591 							MLX4_RES_USAGE_DRIVER);
1592 
1593 			if (res)
1594 				return res;
1595 		} else {
1596 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1597 		}
1598 
1599 		qp->pri.vid = 0xFFFF;
1600 		qp->alt.vid = 0xFFFF;
1601 		err = create_qp_common(pd, init_attr, udata, sqpn, qp);
1602 		if (err)
1603 			return err;
1604 
1605 		if (init_attr->create_flags &
1606 		    (MLX4_IB_SRIOV_SQP | MLX4_IB_SRIOV_TUNNEL_QP))
1607 			/* Internal QP created with ib_create_qp */
1608 			rdma_restrack_no_track(&qp->ibqp.res);
1609 
1610 		qp->port	= init_attr->port_num;
1611 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1612 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1613 		break;
1614 	}
1615 	default:
1616 		/* Don't support raw QPs */
1617 		return -EOPNOTSUPP;
1618 	}
1619 	return 0;
1620 }
1621 
1622 int mlx4_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr,
1623 		      struct ib_udata *udata)
1624 {
1625 	struct ib_device *device = ibqp->device;
1626 	struct mlx4_ib_dev *dev = to_mdev(device);
1627 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1628 	struct ib_pd *pd = ibqp->pd;
1629 	int ret;
1630 
1631 	mutex_init(&qp->mutex);
1632 	ret = _mlx4_ib_create_qp(pd, qp, init_attr, udata);
1633 	if (ret)
1634 		return ret;
1635 
1636 	if (init_attr->qp_type == IB_QPT_GSI &&
1637 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1638 		struct mlx4_ib_sqp *sqp = qp->sqp;
1639 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1640 
1641 		if (is_eth &&
1642 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1643 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1644 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1645 
1646 			if (IS_ERR(sqp->roce_v2_gsi)) {
1647 				pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1648 				sqp->roce_v2_gsi = NULL;
1649 			} else {
1650 				to_mqp(sqp->roce_v2_gsi)->flags |=
1651 					MLX4_IB_ROCE_V2_GSI_QP;
1652 			}
1653 
1654 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1655 		}
1656 	}
1657 	return 0;
1658 }
1659 
1660 static int _mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
1661 {
1662 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1663 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1664 
1665 	if (is_qp0(dev, mqp))
1666 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1667 
1668 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1669 	    dev->qp1_proxy[mqp->port - 1] == mqp) {
1670 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1671 		dev->qp1_proxy[mqp->port - 1] = NULL;
1672 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1673 	}
1674 
1675 	if (mqp->counter_index)
1676 		mlx4_ib_free_qp_counter(dev, mqp);
1677 
1678 	if (qp->rwq_ind_tbl) {
1679 		destroy_qp_rss(dev, mqp);
1680 	} else {
1681 		destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, udata);
1682 	}
1683 
1684 	kfree(mqp->sqp);
1685 	return 0;
1686 }
1687 
1688 int mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
1689 {
1690 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1691 
1692 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1693 		struct mlx4_ib_sqp *sqp = mqp->sqp;
1694 
1695 		if (sqp->roce_v2_gsi)
1696 			ib_destroy_qp(sqp->roce_v2_gsi);
1697 	}
1698 
1699 	return _mlx4_ib_destroy_qp(qp, udata);
1700 }
1701 
1702 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1703 {
1704 	switch (type) {
1705 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1706 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1707 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1708 	case MLX4_IB_QPT_XRC_INI:
1709 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1710 	case MLX4_IB_QPT_SMI:
1711 	case MLX4_IB_QPT_GSI:
1712 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1713 
1714 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1715 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1716 						MLX4_QP_ST_MLX : -1);
1717 	case MLX4_IB_QPT_PROXY_SMI:
1718 	case MLX4_IB_QPT_TUN_SMI:
1719 	case MLX4_IB_QPT_PROXY_GSI:
1720 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1721 						MLX4_QP_ST_UD : -1);
1722 	default:			return -1;
1723 	}
1724 }
1725 
1726 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1727 				   int attr_mask)
1728 {
1729 	u8 dest_rd_atomic;
1730 	u32 access_flags;
1731 	u32 hw_access_flags = 0;
1732 
1733 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1734 		dest_rd_atomic = attr->max_dest_rd_atomic;
1735 	else
1736 		dest_rd_atomic = qp->resp_depth;
1737 
1738 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1739 		access_flags = attr->qp_access_flags;
1740 	else
1741 		access_flags = qp->atomic_rd_en;
1742 
1743 	if (!dest_rd_atomic)
1744 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1745 
1746 	if (access_flags & IB_ACCESS_REMOTE_READ)
1747 		hw_access_flags |= MLX4_QP_BIT_RRE;
1748 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1749 		hw_access_flags |= MLX4_QP_BIT_RAE;
1750 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1751 		hw_access_flags |= MLX4_QP_BIT_RWE;
1752 
1753 	return cpu_to_be32(hw_access_flags);
1754 }
1755 
1756 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1757 			    int attr_mask)
1758 {
1759 	if (attr_mask & IB_QP_PKEY_INDEX)
1760 		sqp->pkey_index = attr->pkey_index;
1761 	if (attr_mask & IB_QP_QKEY)
1762 		sqp->qkey = attr->qkey;
1763 	if (attr_mask & IB_QP_SQ_PSN)
1764 		sqp->send_psn = attr->sq_psn;
1765 }
1766 
1767 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1768 {
1769 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1770 }
1771 
1772 static int _mlx4_set_path(struct mlx4_ib_dev *dev,
1773 			  const struct rdma_ah_attr *ah,
1774 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1775 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1776 {
1777 	int vidx;
1778 	int smac_index;
1779 	int err;
1780 
1781 	path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
1782 	path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
1783 	if (rdma_ah_get_static_rate(ah)) {
1784 		path->static_rate = rdma_ah_get_static_rate(ah) +
1785 				    MLX4_STAT_RATE_OFFSET;
1786 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1787 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1788 			--path->static_rate;
1789 	} else
1790 		path->static_rate = 0;
1791 
1792 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
1793 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
1794 		int real_sgid_index =
1795 			mlx4_ib_gid_index_to_real_index(dev, grh->sgid_attr);
1796 
1797 		if (real_sgid_index < 0)
1798 			return real_sgid_index;
1799 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1800 			pr_err("sgid_index (%u) too large. max is %d\n",
1801 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1802 			return -1;
1803 		}
1804 
1805 		path->grh_mylmc |= 1 << 7;
1806 		path->mgid_index = real_sgid_index;
1807 		path->hop_limit  = grh->hop_limit;
1808 		path->tclass_flowlabel =
1809 			cpu_to_be32((grh->traffic_class << 20) |
1810 				    (grh->flow_label));
1811 		memcpy(path->rgid, grh->dgid.raw, 16);
1812 	}
1813 
1814 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
1815 		if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
1816 			return -1;
1817 
1818 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1819 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
1820 
1821 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1822 		if (vlan_tag < 0x1000) {
1823 			if (smac_info->vid < 0x1000) {
1824 				/* both valid vlan ids */
1825 				if (smac_info->vid != vlan_tag) {
1826 					/* different VIDs.  unreg old and reg new */
1827 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1828 					if (err)
1829 						return err;
1830 					smac_info->candidate_vid = vlan_tag;
1831 					smac_info->candidate_vlan_index = vidx;
1832 					smac_info->candidate_vlan_port = port;
1833 					smac_info->update_vid = 1;
1834 					path->vlan_index = vidx;
1835 				} else {
1836 					path->vlan_index = smac_info->vlan_index;
1837 				}
1838 			} else {
1839 				/* no current vlan tag in qp */
1840 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1841 				if (err)
1842 					return err;
1843 				smac_info->candidate_vid = vlan_tag;
1844 				smac_info->candidate_vlan_index = vidx;
1845 				smac_info->candidate_vlan_port = port;
1846 				smac_info->update_vid = 1;
1847 				path->vlan_index = vidx;
1848 			}
1849 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1850 			path->fl = 1 << 6;
1851 		} else {
1852 			/* have current vlan tag. unregister it at modify-qp success */
1853 			if (smac_info->vid < 0x1000) {
1854 				smac_info->candidate_vid = 0xFFFF;
1855 				smac_info->update_vid = 1;
1856 			}
1857 		}
1858 
1859 		/* get smac_index for RoCE use.
1860 		 * If no smac was yet assigned, register one.
1861 		 * If one was already assigned, but the new mac differs,
1862 		 * unregister the old one and register the new one.
1863 		*/
1864 		if ((!smac_info->smac && !smac_info->smac_port) ||
1865 		    smac_info->smac != smac) {
1866 			/* register candidate now, unreg if needed, after success */
1867 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1868 			if (smac_index >= 0) {
1869 				smac_info->candidate_smac_index = smac_index;
1870 				smac_info->candidate_smac = smac;
1871 				smac_info->candidate_smac_port = port;
1872 			} else {
1873 				return -EINVAL;
1874 			}
1875 		} else {
1876 			smac_index = smac_info->smac_index;
1877 		}
1878 		memcpy(path->dmac, ah->roce.dmac, 6);
1879 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1880 		/* put MAC table smac index for IBoE */
1881 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1882 	} else {
1883 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1884 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
1885 	}
1886 
1887 	return 0;
1888 }
1889 
1890 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1891 			 enum ib_qp_attr_mask qp_attr_mask,
1892 			 struct mlx4_ib_qp *mqp,
1893 			 struct mlx4_qp_path *path, u8 port,
1894 			 u16 vlan_id, u8 *smac)
1895 {
1896 	return _mlx4_set_path(dev, &qp->ah_attr,
1897 			      ether_addr_to_u64(smac),
1898 			      vlan_id,
1899 			      path, &mqp->pri, port);
1900 }
1901 
1902 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1903 			     const struct ib_qp_attr *qp,
1904 			     enum ib_qp_attr_mask qp_attr_mask,
1905 			     struct mlx4_ib_qp *mqp,
1906 			     struct mlx4_qp_path *path, u8 port)
1907 {
1908 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1909 			      0,
1910 			      0xffff,
1911 			      path, &mqp->alt, port);
1912 }
1913 
1914 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1915 {
1916 	struct mlx4_ib_gid_entry *ge, *tmp;
1917 
1918 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1919 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1920 			ge->added = 1;
1921 			ge->port = qp->port;
1922 		}
1923 	}
1924 }
1925 
1926 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1927 				    struct mlx4_ib_qp *qp,
1928 				    struct mlx4_qp_context *context)
1929 {
1930 	u64 u64_mac;
1931 	int smac_index;
1932 
1933 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1934 
1935 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1936 	if (!qp->pri.smac && !qp->pri.smac_port) {
1937 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1938 		if (smac_index >= 0) {
1939 			qp->pri.candidate_smac_index = smac_index;
1940 			qp->pri.candidate_smac = u64_mac;
1941 			qp->pri.candidate_smac_port = qp->port;
1942 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1943 		} else {
1944 			return -ENOENT;
1945 		}
1946 	}
1947 	return 0;
1948 }
1949 
1950 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1951 {
1952 	struct counter_index *new_counter_index;
1953 	int err;
1954 	u32 tmp_idx;
1955 
1956 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1957 	    IB_LINK_LAYER_ETHERNET ||
1958 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1959 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1960 		return 0;
1961 
1962 	err = mlx4_counter_alloc(dev->dev, &tmp_idx, MLX4_RES_USAGE_DRIVER);
1963 	if (err)
1964 		return err;
1965 
1966 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1967 	if (!new_counter_index) {
1968 		mlx4_counter_free(dev->dev, tmp_idx);
1969 		return -ENOMEM;
1970 	}
1971 
1972 	new_counter_index->index = tmp_idx;
1973 	new_counter_index->allocated = 1;
1974 	qp->counter_index = new_counter_index;
1975 
1976 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
1977 	list_add_tail(&new_counter_index->list,
1978 		      &dev->counters_table[qp->port - 1].counters_list);
1979 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
1980 
1981 	return 0;
1982 }
1983 
1984 enum {
1985 	MLX4_QPC_ROCE_MODE_1 = 0,
1986 	MLX4_QPC_ROCE_MODE_2 = 2,
1987 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
1988 };
1989 
1990 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
1991 {
1992 	switch (gid_type) {
1993 	case IB_GID_TYPE_ROCE:
1994 		return MLX4_QPC_ROCE_MODE_1;
1995 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
1996 		return MLX4_QPC_ROCE_MODE_2;
1997 	default:
1998 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
1999 	}
2000 }
2001 
2002 /*
2003  * Go over all RSS QP's childes (WQs) and apply their HW state according to
2004  * their logic state if the RSS QP is the first RSS QP associated for the WQ.
2005  */
2006 static int bringup_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, u8 port_num,
2007 			    struct ib_udata *udata)
2008 {
2009 	int err = 0;
2010 	int i;
2011 
2012 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2013 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2014 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2015 
2016 		mutex_lock(&wq->mutex);
2017 
2018 		/* Mlx4_ib restrictions:
2019 		 * WQ's is associated to a port according to the RSS QP it is
2020 		 * associates to.
2021 		 * In case the WQ is associated to a different port by another
2022 		 * RSS QP, return a failure.
2023 		 */
2024 		if ((wq->rss_usecnt > 0) && (wq->port != port_num)) {
2025 			err = -EINVAL;
2026 			mutex_unlock(&wq->mutex);
2027 			break;
2028 		}
2029 		wq->port = port_num;
2030 		if ((wq->rss_usecnt == 0) && (ibwq->state == IB_WQS_RDY)) {
2031 			err = _mlx4_ib_modify_wq(ibwq, IB_WQS_RDY, udata);
2032 			if (err) {
2033 				mutex_unlock(&wq->mutex);
2034 				break;
2035 			}
2036 		}
2037 		wq->rss_usecnt++;
2038 
2039 		mutex_unlock(&wq->mutex);
2040 	}
2041 
2042 	if (i && err) {
2043 		int j;
2044 
2045 		for (j = (i - 1); j >= 0; j--) {
2046 			struct ib_wq *ibwq = ind_tbl->ind_tbl[j];
2047 			struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2048 
2049 			mutex_lock(&wq->mutex);
2050 
2051 			if ((wq->rss_usecnt == 1) &&
2052 			    (ibwq->state == IB_WQS_RDY))
2053 				if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET,
2054 						       udata))
2055 					pr_warn("failed to reverse WQN=0x%06x\n",
2056 						ibwq->wq_num);
2057 			wq->rss_usecnt--;
2058 
2059 			mutex_unlock(&wq->mutex);
2060 		}
2061 	}
2062 
2063 	return err;
2064 }
2065 
2066 static void bring_down_rss_rwqs(struct ib_rwq_ind_table *ind_tbl,
2067 				struct ib_udata *udata)
2068 {
2069 	int i;
2070 
2071 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2072 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2073 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2074 
2075 		mutex_lock(&wq->mutex);
2076 
2077 		if ((wq->rss_usecnt == 1) && (ibwq->state == IB_WQS_RDY))
2078 			if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET, udata))
2079 				pr_warn("failed to reverse WQN=%x\n",
2080 					ibwq->wq_num);
2081 		wq->rss_usecnt--;
2082 
2083 		mutex_unlock(&wq->mutex);
2084 	}
2085 }
2086 
2087 static void fill_qp_rss_context(struct mlx4_qp_context *context,
2088 				struct mlx4_ib_qp *qp)
2089 {
2090 	struct mlx4_rss_context *rss_context;
2091 
2092 	rss_context = (void *)context + offsetof(struct mlx4_qp_context,
2093 			pri_path) + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
2094 
2095 	rss_context->base_qpn = cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz);
2096 	rss_context->default_qpn =
2097 		cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz & 0xffffff);
2098 	if (qp->rss_ctx->flags & (MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6))
2099 		rss_context->base_qpn_udp = rss_context->default_qpn;
2100 	rss_context->flags = qp->rss_ctx->flags;
2101 	/* Currently support just toeplitz */
2102 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
2103 
2104 	memcpy(rss_context->rss_key, qp->rss_ctx->rss_key,
2105 	       MLX4_EN_RSS_KEY_SIZE);
2106 }
2107 
2108 static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type,
2109 			       const struct ib_qp_attr *attr, int attr_mask,
2110 			       enum ib_qp_state cur_state,
2111 			       enum ib_qp_state new_state,
2112 			       struct ib_udata *udata)
2113 {
2114 	struct ib_srq  *ibsrq;
2115 	const struct ib_gid_attr *gid_attr = NULL;
2116 	struct ib_rwq_ind_table *rwq_ind_tbl;
2117 	enum ib_qp_type qp_type;
2118 	struct mlx4_ib_dev *dev;
2119 	struct mlx4_ib_qp *qp;
2120 	struct mlx4_ib_pd *pd;
2121 	struct mlx4_ib_cq *send_cq, *recv_cq;
2122 	struct mlx4_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2123 		udata, struct mlx4_ib_ucontext, ibucontext);
2124 	struct mlx4_qp_context *context;
2125 	enum mlx4_qp_optpar optpar = 0;
2126 	int sqd_event;
2127 	int steer_qp = 0;
2128 	int err = -EINVAL;
2129 	int counter_index;
2130 
2131 	if (src_type == MLX4_IB_RWQ_SRC) {
2132 		struct ib_wq *ibwq;
2133 
2134 		ibwq	    = (struct ib_wq *)src;
2135 		ibsrq	    = NULL;
2136 		rwq_ind_tbl = NULL;
2137 		qp_type     = IB_QPT_RAW_PACKET;
2138 		qp	    = to_mqp((struct ib_qp *)ibwq);
2139 		dev	    = to_mdev(ibwq->device);
2140 		pd	    = to_mpd(ibwq->pd);
2141 	} else {
2142 		struct ib_qp *ibqp;
2143 
2144 		ibqp	    = (struct ib_qp *)src;
2145 		ibsrq	    = ibqp->srq;
2146 		rwq_ind_tbl = ibqp->rwq_ind_tbl;
2147 		qp_type     = ibqp->qp_type;
2148 		qp	    = to_mqp(ibqp);
2149 		dev	    = to_mdev(ibqp->device);
2150 		pd	    = get_pd(qp);
2151 	}
2152 
2153 	/* APM is not supported under RoCE */
2154 	if (attr_mask & IB_QP_ALT_PATH &&
2155 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2156 	    IB_LINK_LAYER_ETHERNET)
2157 		return -ENOTSUPP;
2158 
2159 	context = kzalloc(sizeof *context, GFP_KERNEL);
2160 	if (!context)
2161 		return -ENOMEM;
2162 
2163 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
2164 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
2165 
2166 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
2167 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2168 	else {
2169 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
2170 		switch (attr->path_mig_state) {
2171 		case IB_MIG_MIGRATED:
2172 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2173 			break;
2174 		case IB_MIG_REARM:
2175 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
2176 			break;
2177 		case IB_MIG_ARMED:
2178 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
2179 			break;
2180 		}
2181 	}
2182 
2183 	if (qp->inl_recv_sz)
2184 		context->param3 |= cpu_to_be32(1 << 25);
2185 
2186 	if (qp->flags & MLX4_IB_QP_SCATTER_FCS)
2187 		context->param3 |= cpu_to_be32(1 << 29);
2188 
2189 	if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI)
2190 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
2191 	else if (qp_type == IB_QPT_RAW_PACKET)
2192 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
2193 	else if (qp_type == IB_QPT_UD) {
2194 		if (qp->flags & MLX4_IB_QP_LSO)
2195 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
2196 					      ilog2(dev->dev->caps.max_gso_sz);
2197 		else
2198 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 13;
2199 	} else if (attr_mask & IB_QP_PATH_MTU) {
2200 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
2201 			pr_err("path MTU (%u) is invalid\n",
2202 			       attr->path_mtu);
2203 			goto out;
2204 		}
2205 		context->mtu_msgmax = (attr->path_mtu << 5) |
2206 			ilog2(dev->dev->caps.max_msg_sz);
2207 	}
2208 
2209 	if (!rwq_ind_tbl) { /* PRM RSS receive side should be left zeros */
2210 		if (qp->rq.wqe_cnt)
2211 			context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
2212 		context->rq_size_stride |= qp->rq.wqe_shift - 4;
2213 	}
2214 
2215 	if (qp->sq.wqe_cnt)
2216 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
2217 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
2218 
2219 	if (new_state == IB_QPS_RESET && qp->counter_index)
2220 		mlx4_ib_free_qp_counter(dev, qp);
2221 
2222 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2223 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
2224 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
2225 		if (qp_type == IB_QPT_RAW_PACKET)
2226 			context->param3 |= cpu_to_be32(1 << 30);
2227 	}
2228 
2229 	if (ucontext)
2230 		context->usr_page = cpu_to_be32(
2231 			mlx4_to_hw_uar_index(dev->dev, ucontext->uar.index));
2232 	else
2233 		context->usr_page = cpu_to_be32(
2234 			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
2235 
2236 	if (attr_mask & IB_QP_DEST_QPN)
2237 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
2238 
2239 	if (attr_mask & IB_QP_PORT) {
2240 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
2241 		    !(attr_mask & IB_QP_AV)) {
2242 			mlx4_set_sched(&context->pri_path, attr->port_num);
2243 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
2244 		}
2245 	}
2246 
2247 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
2248 		err = create_qp_lb_counter(dev, qp);
2249 		if (err)
2250 			goto out;
2251 
2252 		counter_index =
2253 			dev->counters_table[qp->port - 1].default_counter;
2254 		if (qp->counter_index)
2255 			counter_index = qp->counter_index->index;
2256 
2257 		if (counter_index != -1) {
2258 			context->pri_path.counter_index = counter_index;
2259 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
2260 			if (qp->counter_index) {
2261 				context->pri_path.fl |=
2262 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
2263 				context->pri_path.vlan_control |=
2264 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
2265 			}
2266 		} else
2267 			context->pri_path.counter_index =
2268 				MLX4_SINK_COUNTER_INDEX(dev->dev);
2269 
2270 		if (qp->flags & MLX4_IB_QP_NETIF) {
2271 			mlx4_ib_steer_qp_reg(dev, qp, 1);
2272 			steer_qp = 1;
2273 		}
2274 
2275 		if (qp_type == IB_QPT_GSI) {
2276 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
2277 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
2278 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
2279 
2280 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2281 		}
2282 	}
2283 
2284 	if (attr_mask & IB_QP_PKEY_INDEX) {
2285 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2286 			context->pri_path.disable_pkey_check = 0x40;
2287 		context->pri_path.pkey_index = attr->pkey_index;
2288 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
2289 	}
2290 
2291 	if (attr_mask & IB_QP_AV) {
2292 		u8 port_num = mlx4_is_bonded(dev->dev) ? 1 :
2293 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2294 		u16 vlan = 0xffff;
2295 		u8 smac[ETH_ALEN];
2296 		int is_eth =
2297 			rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
2298 			rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
2299 
2300 		if (is_eth) {
2301 			gid_attr = attr->ah_attr.grh.sgid_attr;
2302 			err = rdma_read_gid_l2_fields(gid_attr, &vlan,
2303 						      &smac[0]);
2304 			if (err)
2305 				goto out;
2306 		}
2307 
2308 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
2309 				  port_num, vlan, smac))
2310 			goto out;
2311 
2312 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
2313 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
2314 
2315 		if (is_eth &&
2316 		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
2317 			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr->gid_type);
2318 
2319 			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
2320 				err = -EINVAL;
2321 				goto out;
2322 			}
2323 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2324 		}
2325 
2326 	}
2327 
2328 	if (attr_mask & IB_QP_TIMEOUT) {
2329 		context->pri_path.ackto |= attr->timeout << 3;
2330 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
2331 	}
2332 
2333 	if (attr_mask & IB_QP_ALT_PATH) {
2334 		if (attr->alt_port_num == 0 ||
2335 		    attr->alt_port_num > dev->dev->caps.num_ports)
2336 			goto out;
2337 
2338 		if (attr->alt_pkey_index >=
2339 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
2340 			goto out;
2341 
2342 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
2343 				      &context->alt_path,
2344 				      attr->alt_port_num))
2345 			goto out;
2346 
2347 		context->alt_path.pkey_index = attr->alt_pkey_index;
2348 		context->alt_path.ackto = attr->alt_timeout << 3;
2349 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
2350 	}
2351 
2352 	context->pd = cpu_to_be32(pd->pdn);
2353 
2354 	if (!rwq_ind_tbl) {
2355 		context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
2356 		get_cqs(qp, src_type, &send_cq, &recv_cq);
2357 	} else { /* Set dummy CQs to be compatible with HV and PRM */
2358 		send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq);
2359 		recv_cq = send_cq;
2360 	}
2361 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
2362 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
2363 
2364 	/* Set "fast registration enabled" for all kernel QPs */
2365 	if (!ucontext)
2366 		context->params1 |= cpu_to_be32(1 << 11);
2367 
2368 	if (attr_mask & IB_QP_RNR_RETRY) {
2369 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2370 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
2371 	}
2372 
2373 	if (attr_mask & IB_QP_RETRY_CNT) {
2374 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2375 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
2376 	}
2377 
2378 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2379 		if (attr->max_rd_atomic)
2380 			context->params1 |=
2381 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2382 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
2383 	}
2384 
2385 	if (attr_mask & IB_QP_SQ_PSN)
2386 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2387 
2388 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2389 		if (attr->max_dest_rd_atomic)
2390 			context->params2 |=
2391 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2392 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
2393 	}
2394 
2395 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
2396 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
2397 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
2398 	}
2399 
2400 	if (ibsrq)
2401 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
2402 
2403 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
2404 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2405 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
2406 	}
2407 	if (attr_mask & IB_QP_RQ_PSN)
2408 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2409 
2410 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
2411 	if (attr_mask & IB_QP_QKEY) {
2412 		if (qp->mlx4_ib_qp_type &
2413 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
2414 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2415 		else {
2416 			if (mlx4_is_mfunc(dev->dev) &&
2417 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
2418 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
2419 			    MLX4_RESERVED_QKEY_BASE) {
2420 				pr_err("Cannot use reserved QKEY"
2421 				       " 0x%x (range 0xffff0000..0xffffffff"
2422 				       " is reserved)\n", attr->qkey);
2423 				err = -EINVAL;
2424 				goto out;
2425 			}
2426 			context->qkey = cpu_to_be32(attr->qkey);
2427 		}
2428 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
2429 	}
2430 
2431 	if (ibsrq)
2432 		context->srqn = cpu_to_be32(1 << 24 |
2433 					    to_msrq(ibsrq)->msrq.srqn);
2434 
2435 	if (qp->rq.wqe_cnt &&
2436 	    cur_state == IB_QPS_RESET &&
2437 	    new_state == IB_QPS_INIT)
2438 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2439 
2440 	if (cur_state == IB_QPS_INIT &&
2441 	    new_state == IB_QPS_RTR  &&
2442 	    (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI ||
2443 	     qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) {
2444 		context->pri_path.sched_queue = (qp->port - 1) << 6;
2445 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2446 		    qp->mlx4_ib_qp_type &
2447 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
2448 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
2449 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
2450 				context->pri_path.fl = 0x80;
2451 		} else {
2452 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2453 				context->pri_path.fl = 0x80;
2454 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
2455 		}
2456 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2457 		    IB_LINK_LAYER_ETHERNET) {
2458 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
2459 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
2460 				context->pri_path.feup = 1 << 7; /* don't fsm */
2461 			/* handle smac_index */
2462 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
2463 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
2464 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
2465 				err = handle_eth_ud_smac_index(dev, qp, context);
2466 				if (err) {
2467 					err = -EINVAL;
2468 					goto out;
2469 				}
2470 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
2471 					dev->qp1_proxy[qp->port - 1] = qp;
2472 			}
2473 		}
2474 	}
2475 
2476 	if (qp_type == IB_QPT_RAW_PACKET) {
2477 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
2478 					MLX4_IB_LINK_TYPE_ETH;
2479 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
2480 			/* set QP to receive both tunneled & non-tunneled packets */
2481 			if (!rwq_ind_tbl)
2482 				context->srqn = cpu_to_be32(7 << 28);
2483 		}
2484 	}
2485 
2486 	if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
2487 		int is_eth = rdma_port_get_link_layer(
2488 				&dev->ib_dev, qp->port) ==
2489 				IB_LINK_LAYER_ETHERNET;
2490 		if (is_eth) {
2491 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
2492 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
2493 		}
2494 	}
2495 
2496 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2497 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2498 		sqd_event = 1;
2499 	else
2500 		sqd_event = 0;
2501 
2502 	if (!ucontext &&
2503 	    cur_state == IB_QPS_RESET &&
2504 	    new_state == IB_QPS_INIT)
2505 		context->rlkey_roce_mode |= (1 << 4);
2506 
2507 	/*
2508 	 * Before passing a kernel QP to the HW, make sure that the
2509 	 * ownership bits of the send queue are set and the SQ
2510 	 * headroom is stamped so that the hardware doesn't start
2511 	 * processing stale work requests.
2512 	 */
2513 	if (!ucontext &&
2514 	    cur_state == IB_QPS_RESET &&
2515 	    new_state == IB_QPS_INIT) {
2516 		struct mlx4_wqe_ctrl_seg *ctrl;
2517 		int i;
2518 
2519 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
2520 			ctrl = get_send_wqe(qp, i);
2521 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
2522 			ctrl->qpn_vlan.fence_size =
2523 				1 << (qp->sq.wqe_shift - 4);
2524 			stamp_send_wqe(qp, i);
2525 		}
2526 	}
2527 
2528 	if (rwq_ind_tbl	&&
2529 	    cur_state == IB_QPS_RESET &&
2530 	    new_state == IB_QPS_INIT) {
2531 		fill_qp_rss_context(context, qp);
2532 		context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET);
2533 	}
2534 
2535 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2536 			     to_mlx4_state(new_state), context, optpar,
2537 			     sqd_event, &qp->mqp);
2538 	if (err)
2539 		goto out;
2540 
2541 	qp->state = new_state;
2542 
2543 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2544 		qp->atomic_rd_en = attr->qp_access_flags;
2545 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2546 		qp->resp_depth = attr->max_dest_rd_atomic;
2547 	if (attr_mask & IB_QP_PORT) {
2548 		qp->port = attr->port_num;
2549 		update_mcg_macs(dev, qp);
2550 	}
2551 	if (attr_mask & IB_QP_ALT_PATH)
2552 		qp->alt_port = attr->alt_port_num;
2553 
2554 	if (is_sqp(dev, qp))
2555 		store_sqp_attrs(qp->sqp, attr, attr_mask);
2556 
2557 	/*
2558 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
2559 	 * QP0 to RESET or ERROR, bring the link back down.
2560 	 */
2561 	if (is_qp0(dev, qp)) {
2562 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2563 			if (mlx4_INIT_PORT(dev->dev, qp->port))
2564 				pr_warn("INIT_PORT failed for port %d\n",
2565 				       qp->port);
2566 
2567 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2568 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2569 			mlx4_CLOSE_PORT(dev->dev, qp->port);
2570 	}
2571 
2572 	/*
2573 	 * If we moved a kernel QP to RESET, clean up all old CQ
2574 	 * entries and reinitialize the QP.
2575 	 */
2576 	if (new_state == IB_QPS_RESET) {
2577 		if (!ucontext) {
2578 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2579 					 ibsrq ? to_msrq(ibsrq) : NULL);
2580 			if (send_cq != recv_cq)
2581 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2582 
2583 			qp->rq.head = 0;
2584 			qp->rq.tail = 0;
2585 			qp->sq.head = 0;
2586 			qp->sq.tail = 0;
2587 			qp->sq_next_wqe = 0;
2588 			if (qp->rq.wqe_cnt)
2589 				*qp->db.db  = 0;
2590 
2591 			if (qp->flags & MLX4_IB_QP_NETIF)
2592 				mlx4_ib_steer_qp_reg(dev, qp, 0);
2593 		}
2594 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2595 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2596 			qp->pri.smac = 0;
2597 			qp->pri.smac_port = 0;
2598 		}
2599 		if (qp->alt.smac) {
2600 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2601 			qp->alt.smac = 0;
2602 		}
2603 		if (qp->pri.vid < 0x1000) {
2604 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2605 			qp->pri.vid = 0xFFFF;
2606 			qp->pri.candidate_vid = 0xFFFF;
2607 			qp->pri.update_vid = 0;
2608 		}
2609 
2610 		if (qp->alt.vid < 0x1000) {
2611 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2612 			qp->alt.vid = 0xFFFF;
2613 			qp->alt.candidate_vid = 0xFFFF;
2614 			qp->alt.update_vid = 0;
2615 		}
2616 	}
2617 out:
2618 	if (err && qp->counter_index)
2619 		mlx4_ib_free_qp_counter(dev, qp);
2620 	if (err && steer_qp)
2621 		mlx4_ib_steer_qp_reg(dev, qp, 0);
2622 	kfree(context);
2623 	if (qp->pri.candidate_smac ||
2624 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2625 		if (err) {
2626 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2627 		} else {
2628 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2629 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2630 			qp->pri.smac = qp->pri.candidate_smac;
2631 			qp->pri.smac_index = qp->pri.candidate_smac_index;
2632 			qp->pri.smac_port = qp->pri.candidate_smac_port;
2633 		}
2634 		qp->pri.candidate_smac = 0;
2635 		qp->pri.candidate_smac_index = 0;
2636 		qp->pri.candidate_smac_port = 0;
2637 	}
2638 	if (qp->alt.candidate_smac) {
2639 		if (err) {
2640 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2641 		} else {
2642 			if (qp->alt.smac)
2643 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2644 			qp->alt.smac = qp->alt.candidate_smac;
2645 			qp->alt.smac_index = qp->alt.candidate_smac_index;
2646 			qp->alt.smac_port = qp->alt.candidate_smac_port;
2647 		}
2648 		qp->alt.candidate_smac = 0;
2649 		qp->alt.candidate_smac_index = 0;
2650 		qp->alt.candidate_smac_port = 0;
2651 	}
2652 
2653 	if (qp->pri.update_vid) {
2654 		if (err) {
2655 			if (qp->pri.candidate_vid < 0x1000)
2656 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2657 						     qp->pri.candidate_vid);
2658 		} else {
2659 			if (qp->pri.vid < 0x1000)
2660 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2661 						     qp->pri.vid);
2662 			qp->pri.vid = qp->pri.candidate_vid;
2663 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2664 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
2665 		}
2666 		qp->pri.candidate_vid = 0xFFFF;
2667 		qp->pri.update_vid = 0;
2668 	}
2669 
2670 	if (qp->alt.update_vid) {
2671 		if (err) {
2672 			if (qp->alt.candidate_vid < 0x1000)
2673 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2674 						     qp->alt.candidate_vid);
2675 		} else {
2676 			if (qp->alt.vid < 0x1000)
2677 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2678 						     qp->alt.vid);
2679 			qp->alt.vid = qp->alt.candidate_vid;
2680 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2681 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2682 		}
2683 		qp->alt.candidate_vid = 0xFFFF;
2684 		qp->alt.update_vid = 0;
2685 	}
2686 
2687 	return err;
2688 }
2689 
2690 enum {
2691 	MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK = (IB_QP_STATE	|
2692 					      IB_QP_PORT),
2693 };
2694 
2695 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2696 			      int attr_mask, struct ib_udata *udata)
2697 {
2698 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2699 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2700 	enum ib_qp_state cur_state, new_state;
2701 	int err = -EINVAL;
2702 	mutex_lock(&qp->mutex);
2703 
2704 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2705 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2706 
2707 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2708 				attr_mask)) {
2709 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2710 			 "for transition %d to %d. qp_type %d,"
2711 			 " attr_mask 0x%x\n",
2712 			 ibqp->qp_num, cur_state, new_state,
2713 			 ibqp->qp_type, attr_mask);
2714 		goto out;
2715 	}
2716 
2717 	if (ibqp->rwq_ind_tbl) {
2718 		if (!(((cur_state == IB_QPS_RESET) &&
2719 		       (new_state == IB_QPS_INIT)) ||
2720 		      ((cur_state == IB_QPS_INIT)  &&
2721 		       (new_state == IB_QPS_RTR)))) {
2722 			pr_debug("qpn 0x%x: RSS QP unsupported transition %d to %d\n",
2723 				 ibqp->qp_num, cur_state, new_state);
2724 
2725 			err = -EOPNOTSUPP;
2726 			goto out;
2727 		}
2728 
2729 		if (attr_mask & ~MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK) {
2730 			pr_debug("qpn 0x%x: RSS QP unsupported attribute mask 0x%x for transition %d to %d\n",
2731 				 ibqp->qp_num, attr_mask, cur_state, new_state);
2732 
2733 			err = -EOPNOTSUPP;
2734 			goto out;
2735 		}
2736 	}
2737 
2738 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2739 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2740 			if ((ibqp->qp_type == IB_QPT_RC) ||
2741 			    (ibqp->qp_type == IB_QPT_UD) ||
2742 			    (ibqp->qp_type == IB_QPT_UC) ||
2743 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2744 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2745 				attr->port_num = mlx4_ib_bond_next_port(dev);
2746 			}
2747 		} else {
2748 			/* no sense in changing port_num
2749 			 * when ports are bonded */
2750 			attr_mask &= ~IB_QP_PORT;
2751 		}
2752 	}
2753 
2754 	if ((attr_mask & IB_QP_PORT) &&
2755 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2756 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2757 			 "for transition %d to %d. qp_type %d\n",
2758 			 ibqp->qp_num, attr->port_num, cur_state,
2759 			 new_state, ibqp->qp_type);
2760 		goto out;
2761 	}
2762 
2763 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2764 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2765 	     IB_LINK_LAYER_ETHERNET))
2766 		goto out;
2767 
2768 	if (attr_mask & IB_QP_PKEY_INDEX) {
2769 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2770 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2771 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2772 				 "for transition %d to %d. qp_type %d\n",
2773 				 ibqp->qp_num, attr->pkey_index, cur_state,
2774 				 new_state, ibqp->qp_type);
2775 			goto out;
2776 		}
2777 	}
2778 
2779 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2780 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2781 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2782 			 "Transition %d to %d. qp_type %d\n",
2783 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2784 			 new_state, ibqp->qp_type);
2785 		goto out;
2786 	}
2787 
2788 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2789 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2790 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2791 			 "Transition %d to %d. qp_type %d\n",
2792 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2793 			 new_state, ibqp->qp_type);
2794 		goto out;
2795 	}
2796 
2797 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2798 		err = 0;
2799 		goto out;
2800 	}
2801 
2802 	if (ibqp->rwq_ind_tbl && (new_state == IB_QPS_INIT)) {
2803 		err = bringup_rss_rwqs(ibqp->rwq_ind_tbl, attr->port_num,
2804 				       udata);
2805 		if (err)
2806 			goto out;
2807 	}
2808 
2809 	err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask,
2810 				  cur_state, new_state, udata);
2811 
2812 	if (ibqp->rwq_ind_tbl && err)
2813 		bring_down_rss_rwqs(ibqp->rwq_ind_tbl, udata);
2814 
2815 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2816 		attr->port_num = 1;
2817 
2818 out:
2819 	mutex_unlock(&qp->mutex);
2820 	return err;
2821 }
2822 
2823 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2824 		      int attr_mask, struct ib_udata *udata)
2825 {
2826 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2827 	int ret;
2828 
2829 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
2830 		return -EOPNOTSUPP;
2831 
2832 	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2833 
2834 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2835 		struct mlx4_ib_sqp *sqp = mqp->sqp;
2836 		int err = 0;
2837 
2838 		if (sqp->roce_v2_gsi)
2839 			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2840 		if (err)
2841 			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2842 			       err);
2843 	}
2844 	return ret;
2845 }
2846 
2847 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2848 {
2849 	int i;
2850 	for (i = 0; i < dev->caps.num_ports; i++) {
2851 		if (qpn == dev->caps.spec_qps[i].qp0_proxy ||
2852 		    qpn == dev->caps.spec_qps[i].qp0_tunnel) {
2853 			*qkey = dev->caps.spec_qps[i].qp0_qkey;
2854 			return 0;
2855 		}
2856 	}
2857 	return -EINVAL;
2858 }
2859 
2860 static int build_sriov_qp0_header(struct mlx4_ib_qp *qp,
2861 				  const struct ib_ud_wr *wr,
2862 				  void *wqe, unsigned *mlx_seg_len)
2863 {
2864 	struct mlx4_ib_dev *mdev = to_mdev(qp->ibqp.device);
2865 	struct mlx4_ib_sqp *sqp = qp->sqp;
2866 	struct ib_device *ib_dev = qp->ibqp.device;
2867 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2868 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2869 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2870 	u16 pkey;
2871 	u32 qkey;
2872 	int send_size;
2873 	int header_size;
2874 	int spc;
2875 	int err;
2876 	int i;
2877 
2878 	if (wr->wr.opcode != IB_WR_SEND)
2879 		return -EINVAL;
2880 
2881 	send_size = 0;
2882 
2883 	for (i = 0; i < wr->wr.num_sge; ++i)
2884 		send_size += wr->wr.sg_list[i].length;
2885 
2886 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2887 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2888 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2889 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2890 
2891 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2892 
2893 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2894 		sqp->ud_header.lrh.service_level =
2895 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2896 		sqp->ud_header.lrh.destination_lid =
2897 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2898 		sqp->ud_header.lrh.source_lid =
2899 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2900 	}
2901 
2902 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2903 
2904 	/* force loopback */
2905 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2906 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2907 
2908 	sqp->ud_header.lrh.virtual_lane    = 0;
2909 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2910 	err = ib_get_cached_pkey(ib_dev, qp->port, 0, &pkey);
2911 	if (err)
2912 		return err;
2913 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2914 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2915 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2916 	else
2917 		sqp->ud_header.bth.destination_qpn =
2918 			cpu_to_be32(mdev->dev->caps.spec_qps[qp->port - 1].qp0_tunnel);
2919 
2920 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2921 	if (mlx4_is_master(mdev->dev)) {
2922 		if (mlx4_get_parav_qkey(mdev->dev, qp->mqp.qpn, &qkey))
2923 			return -EINVAL;
2924 	} else {
2925 		if (vf_get_qp0_qkey(mdev->dev, qp->mqp.qpn, &qkey))
2926 			return -EINVAL;
2927 	}
2928 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2929 	sqp->ud_header.deth.source_qpn = cpu_to_be32(qp->mqp.qpn);
2930 
2931 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2932 	sqp->ud_header.immediate_present = 0;
2933 
2934 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2935 
2936 	/*
2937 	 * Inline data segments may not cross a 64 byte boundary.  If
2938 	 * our UD header is bigger than the space available up to the
2939 	 * next 64 byte boundary in the WQE, use two inline data
2940 	 * segments to hold the UD header.
2941 	 */
2942 	spc = MLX4_INLINE_ALIGN -
2943 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2944 	if (header_size <= spc) {
2945 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2946 		memcpy(inl + 1, sqp->header_buf, header_size);
2947 		i = 1;
2948 	} else {
2949 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2950 		memcpy(inl + 1, sqp->header_buf, spc);
2951 
2952 		inl = (void *) (inl + 1) + spc;
2953 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2954 		/*
2955 		 * Need a barrier here to make sure all the data is
2956 		 * visible before the byte_count field is set.
2957 		 * Otherwise the HCA prefetcher could grab the 64-byte
2958 		 * chunk with this inline segment and get a valid (!=
2959 		 * 0xffffffff) byte count but stale data, and end up
2960 		 * generating a packet with bad headers.
2961 		 *
2962 		 * The first inline segment's byte_count field doesn't
2963 		 * need a barrier, because it comes after a
2964 		 * control/MLX segment and therefore is at an offset
2965 		 * of 16 mod 64.
2966 		 */
2967 		wmb();
2968 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2969 		i = 2;
2970 	}
2971 
2972 	*mlx_seg_len =
2973 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2974 	return 0;
2975 }
2976 
2977 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
2978 {
2979 	union sl2vl_tbl_to_u64 tmp_vltab;
2980 	u8 vl;
2981 
2982 	if (sl > 15)
2983 		return 0xf;
2984 	tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
2985 	vl = tmp_vltab.sl8[sl >> 1];
2986 	if (sl & 1)
2987 		vl &= 0x0f;
2988 	else
2989 		vl >>= 4;
2990 	return vl;
2991 }
2992 
2993 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
2994 				int index, union ib_gid *gid,
2995 				enum ib_gid_type *gid_type)
2996 {
2997 	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
2998 	struct mlx4_port_gid_table *port_gid_table;
2999 	unsigned long flags;
3000 
3001 	port_gid_table = &iboe->gids[port_num - 1];
3002 	spin_lock_irqsave(&iboe->lock, flags);
3003 	memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
3004 	*gid_type = port_gid_table->gids[index].gid_type;
3005 	spin_unlock_irqrestore(&iboe->lock, flags);
3006 	if (rdma_is_zero_gid(gid))
3007 		return -ENOENT;
3008 
3009 	return 0;
3010 }
3011 
3012 #define MLX4_ROCEV2_QP1_SPORT 0xC000
3013 static int build_mlx_header(struct mlx4_ib_qp *qp, const struct ib_ud_wr *wr,
3014 			    void *wqe, unsigned *mlx_seg_len)
3015 {
3016 	struct mlx4_ib_sqp *sqp = qp->sqp;
3017 	struct ib_device *ib_dev = qp->ibqp.device;
3018 	struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
3019 	struct mlx4_wqe_mlx_seg *mlx = wqe;
3020 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
3021 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
3022 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3023 	union ib_gid sgid;
3024 	u16 pkey;
3025 	int send_size;
3026 	int header_size;
3027 	int spc;
3028 	int i;
3029 	int err = 0;
3030 	u16 vlan = 0xffff;
3031 	bool is_eth;
3032 	bool is_vlan = false;
3033 	bool is_grh;
3034 	bool is_udp = false;
3035 	int ip_version = 0;
3036 
3037 	send_size = 0;
3038 	for (i = 0; i < wr->wr.num_sge; ++i)
3039 		send_size += wr->wr.sg_list[i].length;
3040 
3041 	is_eth = rdma_port_get_link_layer(qp->ibqp.device, qp->port) == IB_LINK_LAYER_ETHERNET;
3042 	is_grh = mlx4_ib_ah_grh_present(ah);
3043 	if (is_eth) {
3044 		enum ib_gid_type gid_type;
3045 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3046 			/* When multi-function is enabled, the ib_core gid
3047 			 * indexes don't necessarily match the hw ones, so
3048 			 * we must use our own cache */
3049 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
3050 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
3051 							   ah->av.ib.gid_index, &sgid.raw[0]);
3052 			if (err)
3053 				return err;
3054 		} else  {
3055 			err = fill_gid_by_hw_index(ibdev, qp->port,
3056 						   ah->av.ib.gid_index, &sgid,
3057 						   &gid_type);
3058 			if (!err) {
3059 				is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
3060 				if (is_udp) {
3061 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
3062 						ip_version = 4;
3063 					else
3064 						ip_version = 6;
3065 					is_grh = false;
3066 				}
3067 			} else {
3068 				return err;
3069 			}
3070 		}
3071 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
3072 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
3073 			is_vlan = true;
3074 		}
3075 	}
3076 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
3077 			  ip_version, is_udp, 0, &sqp->ud_header);
3078 	if (err)
3079 		return err;
3080 
3081 	if (!is_eth) {
3082 		sqp->ud_header.lrh.service_level =
3083 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
3084 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
3085 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
3086 	}
3087 
3088 	if (is_grh || (ip_version == 6)) {
3089 		sqp->ud_header.grh.traffic_class =
3090 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3091 		sqp->ud_header.grh.flow_label    =
3092 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
3093 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
3094 		if (is_eth) {
3095 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
3096 		} else {
3097 			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3098 				/* When multi-function is enabled, the ib_core gid
3099 				 * indexes don't necessarily match the hw ones, so
3100 				 * we must use our own cache
3101 				 */
3102 				sqp->ud_header.grh.source_gid.global
3103 					.subnet_prefix =
3104 					cpu_to_be64(atomic64_read(
3105 						&(to_mdev(ib_dev)
3106 							  ->sriov
3107 							  .demux[qp->port - 1]
3108 							  .subnet_prefix)));
3109 				sqp->ud_header.grh.source_gid.global
3110 					.interface_id =
3111 					to_mdev(ib_dev)
3112 						->sriov.demux[qp->port - 1]
3113 						.guid_cache[ah->av.ib.gid_index];
3114 			} else {
3115 				sqp->ud_header.grh.source_gid =
3116 					ah->ibah.sgid_attr->gid;
3117 			}
3118 		}
3119 		memcpy(sqp->ud_header.grh.destination_gid.raw,
3120 		       ah->av.ib.dgid, 16);
3121 	}
3122 
3123 	if (ip_version == 4) {
3124 		sqp->ud_header.ip4.tos =
3125 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3126 		sqp->ud_header.ip4.id = 0;
3127 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
3128 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
3129 
3130 		memcpy(&sqp->ud_header.ip4.saddr,
3131 		       sgid.raw + 12, 4);
3132 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
3133 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
3134 	}
3135 
3136 	if (is_udp) {
3137 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
3138 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
3139 		sqp->ud_header.udp.csum = 0;
3140 	}
3141 
3142 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
3143 
3144 	if (!is_eth) {
3145 		mlx->flags |=
3146 			cpu_to_be32((!qp->ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
3147 				    (sqp->ud_header.lrh.destination_lid ==
3148 						     IB_LID_PERMISSIVE ?
3149 					     MLX4_WQE_MLX_SLR :
3150 					     0) |
3151 				    (sqp->ud_header.lrh.service_level << 8));
3152 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
3153 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
3154 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
3155 	}
3156 
3157 	switch (wr->wr.opcode) {
3158 	case IB_WR_SEND:
3159 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
3160 		sqp->ud_header.immediate_present = 0;
3161 		break;
3162 	case IB_WR_SEND_WITH_IMM:
3163 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
3164 		sqp->ud_header.immediate_present = 1;
3165 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
3166 		break;
3167 	default:
3168 		return -EINVAL;
3169 	}
3170 
3171 	if (is_eth) {
3172 		u16 ether_type;
3173 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
3174 
3175 		ether_type = (!is_udp) ? ETH_P_IBOE:
3176 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
3177 
3178 		mlx->sched_prio = cpu_to_be16(pcp);
3179 
3180 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
3181 		ether_addr_copy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac);
3182 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
3183 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
3184 
3185 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
3186 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
3187 		if (!is_vlan) {
3188 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
3189 		} else {
3190 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
3191 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
3192 		}
3193 	} else {
3194 		sqp->ud_header.lrh.virtual_lane =
3195 			!qp->ibqp.qp_num ?
3196 				15 :
3197 				sl_to_vl(to_mdev(ib_dev),
3198 					 sqp->ud_header.lrh.service_level,
3199 					 qp->port);
3200 		if (qp->ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
3201 			return -EINVAL;
3202 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
3203 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
3204 	}
3205 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
3206 	if (!qp->ibqp.qp_num)
3207 		err = ib_get_cached_pkey(ib_dev, qp->port, sqp->pkey_index,
3208 					 &pkey);
3209 	else
3210 		err = ib_get_cached_pkey(ib_dev, qp->port, wr->pkey_index,
3211 					 &pkey);
3212 	if (err)
3213 		return err;
3214 
3215 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
3216 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
3217 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
3218 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
3219 					       sqp->qkey : wr->remote_qkey);
3220 	sqp->ud_header.deth.source_qpn = cpu_to_be32(qp->ibqp.qp_num);
3221 
3222 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3223 
3224 	if (0) {
3225 		pr_err("built UD header of size %d:\n", header_size);
3226 		for (i = 0; i < header_size / 4; ++i) {
3227 			if (i % 8 == 0)
3228 				pr_err("  [%02x] ", i * 4);
3229 			pr_cont(" %08x",
3230 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
3231 			if ((i + 1) % 8 == 0)
3232 				pr_cont("\n");
3233 		}
3234 		pr_err("\n");
3235 	}
3236 
3237 	/*
3238 	 * Inline data segments may not cross a 64 byte boundary.  If
3239 	 * our UD header is bigger than the space available up to the
3240 	 * next 64 byte boundary in the WQE, use two inline data
3241 	 * segments to hold the UD header.
3242 	 */
3243 	spc = MLX4_INLINE_ALIGN -
3244 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3245 	if (header_size <= spc) {
3246 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3247 		memcpy(inl + 1, sqp->header_buf, header_size);
3248 		i = 1;
3249 	} else {
3250 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3251 		memcpy(inl + 1, sqp->header_buf, spc);
3252 
3253 		inl = (void *) (inl + 1) + spc;
3254 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3255 		/*
3256 		 * Need a barrier here to make sure all the data is
3257 		 * visible before the byte_count field is set.
3258 		 * Otherwise the HCA prefetcher could grab the 64-byte
3259 		 * chunk with this inline segment and get a valid (!=
3260 		 * 0xffffffff) byte count but stale data, and end up
3261 		 * generating a packet with bad headers.
3262 		 *
3263 		 * The first inline segment's byte_count field doesn't
3264 		 * need a barrier, because it comes after a
3265 		 * control/MLX segment and therefore is at an offset
3266 		 * of 16 mod 64.
3267 		 */
3268 		wmb();
3269 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3270 		i = 2;
3271 	}
3272 
3273 	*mlx_seg_len =
3274 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3275 	return 0;
3276 }
3277 
3278 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3279 {
3280 	unsigned cur;
3281 	struct mlx4_ib_cq *cq;
3282 
3283 	cur = wq->head - wq->tail;
3284 	if (likely(cur + nreq < wq->max_post))
3285 		return 0;
3286 
3287 	cq = to_mcq(ib_cq);
3288 	spin_lock(&cq->lock);
3289 	cur = wq->head - wq->tail;
3290 	spin_unlock(&cq->lock);
3291 
3292 	return cur + nreq >= wq->max_post;
3293 }
3294 
3295 static __be32 convert_access(int acc)
3296 {
3297 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
3298 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
3299 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
3300 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
3301 	       (acc & IB_ACCESS_REMOTE_READ   ?
3302 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
3303 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
3304 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
3305 }
3306 
3307 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
3308 			const struct ib_reg_wr *wr)
3309 {
3310 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
3311 
3312 	fseg->flags		= convert_access(wr->access);
3313 	fseg->mem_key		= cpu_to_be32(wr->key);
3314 	fseg->buf_list		= cpu_to_be64(mr->page_map);
3315 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
3316 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
3317 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
3318 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
3319 	fseg->reserved[0]	= 0;
3320 	fseg->reserved[1]	= 0;
3321 }
3322 
3323 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
3324 {
3325 	memset(iseg, 0, sizeof(*iseg));
3326 	iseg->mem_key = cpu_to_be32(rkey);
3327 }
3328 
3329 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
3330 					  u64 remote_addr, u32 rkey)
3331 {
3332 	rseg->raddr    = cpu_to_be64(remote_addr);
3333 	rseg->rkey     = cpu_to_be32(rkey);
3334 	rseg->reserved = 0;
3335 }
3336 
3337 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
3338 			   const struct ib_atomic_wr *wr)
3339 {
3340 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
3341 		aseg->swap_add = cpu_to_be64(wr->swap);
3342 		aseg->compare  = cpu_to_be64(wr->compare_add);
3343 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
3344 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3345 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
3346 	} else {
3347 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3348 		aseg->compare  = 0;
3349 	}
3350 
3351 }
3352 
3353 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
3354 				  const struct ib_atomic_wr *wr)
3355 {
3356 	aseg->swap_add		= cpu_to_be64(wr->swap);
3357 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
3358 	aseg->compare		= cpu_to_be64(wr->compare_add);
3359 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
3360 }
3361 
3362 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
3363 			     const struct ib_ud_wr *wr)
3364 {
3365 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
3366 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
3367 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
3368 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
3369 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
3370 }
3371 
3372 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
3373 				    struct mlx4_wqe_datagram_seg *dseg,
3374 				    const struct ib_ud_wr *wr,
3375 				    enum mlx4_ib_qp_type qpt)
3376 {
3377 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
3378 	struct mlx4_av sqp_av = {0};
3379 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
3380 
3381 	/* force loopback */
3382 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
3383 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
3384 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
3385 			cpu_to_be32(0xf0000000);
3386 
3387 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
3388 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
3389 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp1_tunnel);
3390 	else
3391 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp0_tunnel);
3392 	/* Use QKEY from the QP context, which is set by master */
3393 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
3394 }
3395 
3396 static void build_tunnel_header(const struct ib_ud_wr *wr, void *wqe,
3397 				unsigned *mlx_seg_len)
3398 {
3399 	struct mlx4_wqe_inline_seg *inl = wqe;
3400 	struct mlx4_ib_tunnel_header hdr;
3401 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3402 	int spc;
3403 	int i;
3404 
3405 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
3406 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
3407 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
3408 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
3409 	memcpy(hdr.mac, ah->av.eth.mac, 6);
3410 	hdr.vlan = ah->av.eth.vlan;
3411 
3412 	spc = MLX4_INLINE_ALIGN -
3413 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3414 	if (sizeof (hdr) <= spc) {
3415 		memcpy(inl + 1, &hdr, sizeof (hdr));
3416 		wmb();
3417 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
3418 		i = 1;
3419 	} else {
3420 		memcpy(inl + 1, &hdr, spc);
3421 		wmb();
3422 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3423 
3424 		inl = (void *) (inl + 1) + spc;
3425 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
3426 		wmb();
3427 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
3428 		i = 2;
3429 	}
3430 
3431 	*mlx_seg_len =
3432 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
3433 }
3434 
3435 static void set_mlx_icrc_seg(void *dseg)
3436 {
3437 	u32 *t = dseg;
3438 	struct mlx4_wqe_inline_seg *iseg = dseg;
3439 
3440 	t[1] = 0;
3441 
3442 	/*
3443 	 * Need a barrier here before writing the byte_count field to
3444 	 * make sure that all the data is visible before the
3445 	 * byte_count field is set.  Otherwise, if the segment begins
3446 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3447 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3448 	 * stale data, and end up sending the wrong data.
3449 	 */
3450 	wmb();
3451 
3452 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
3453 }
3454 
3455 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3456 {
3457 	dseg->lkey       = cpu_to_be32(sg->lkey);
3458 	dseg->addr       = cpu_to_be64(sg->addr);
3459 
3460 	/*
3461 	 * Need a barrier here before writing the byte_count field to
3462 	 * make sure that all the data is visible before the
3463 	 * byte_count field is set.  Otherwise, if the segment begins
3464 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3465 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3466 	 * stale data, and end up sending the wrong data.
3467 	 */
3468 	wmb();
3469 
3470 	dseg->byte_count = cpu_to_be32(sg->length);
3471 }
3472 
3473 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3474 {
3475 	dseg->byte_count = cpu_to_be32(sg->length);
3476 	dseg->lkey       = cpu_to_be32(sg->lkey);
3477 	dseg->addr       = cpu_to_be64(sg->addr);
3478 }
3479 
3480 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe,
3481 			 const struct ib_ud_wr *wr, struct mlx4_ib_qp *qp,
3482 			 unsigned *lso_seg_len, __be32 *lso_hdr_sz, __be32 *blh)
3483 {
3484 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
3485 
3486 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
3487 		*blh = cpu_to_be32(1 << 6);
3488 
3489 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
3490 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
3491 		return -EINVAL;
3492 
3493 	memcpy(wqe->header, wr->header, wr->hlen);
3494 
3495 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
3496 	*lso_seg_len = halign;
3497 	return 0;
3498 }
3499 
3500 static __be32 send_ieth(const struct ib_send_wr *wr)
3501 {
3502 	switch (wr->opcode) {
3503 	case IB_WR_SEND_WITH_IMM:
3504 	case IB_WR_RDMA_WRITE_WITH_IMM:
3505 		return wr->ex.imm_data;
3506 
3507 	case IB_WR_SEND_WITH_INV:
3508 		return cpu_to_be32(wr->ex.invalidate_rkey);
3509 
3510 	default:
3511 		return 0;
3512 	}
3513 }
3514 
3515 static void add_zero_len_inline(void *wqe)
3516 {
3517 	struct mlx4_wqe_inline_seg *inl = wqe;
3518 	memset(wqe, 0, 16);
3519 	inl->byte_count = cpu_to_be32(1 << 31);
3520 }
3521 
3522 static int _mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3523 			      const struct ib_send_wr **bad_wr, bool drain)
3524 {
3525 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3526 	void *wqe;
3527 	struct mlx4_wqe_ctrl_seg *ctrl;
3528 	struct mlx4_wqe_data_seg *dseg;
3529 	unsigned long flags;
3530 	int nreq;
3531 	int err = 0;
3532 	unsigned ind;
3533 	int size;
3534 	unsigned seglen;
3535 	__be32 dummy;
3536 	__be32 *lso_wqe;
3537 	__be32 lso_hdr_sz;
3538 	__be32 blh;
3539 	int i;
3540 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3541 
3542 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
3543 		struct mlx4_ib_sqp *sqp = qp->sqp;
3544 
3545 		if (sqp->roce_v2_gsi) {
3546 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
3547 			enum ib_gid_type gid_type;
3548 			union ib_gid gid;
3549 
3550 			if (!fill_gid_by_hw_index(mdev, qp->port,
3551 					   ah->av.ib.gid_index,
3552 					   &gid, &gid_type))
3553 				qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
3554 						to_mqp(sqp->roce_v2_gsi) : qp;
3555 			else
3556 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
3557 				       ah->av.ib.gid_index);
3558 		}
3559 	}
3560 
3561 	spin_lock_irqsave(&qp->sq.lock, flags);
3562 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3563 	    !drain) {
3564 		err = -EIO;
3565 		*bad_wr = wr;
3566 		nreq = 0;
3567 		goto out;
3568 	}
3569 
3570 	ind = qp->sq_next_wqe;
3571 
3572 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3573 		lso_wqe = &dummy;
3574 		blh = 0;
3575 
3576 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
3577 			err = -ENOMEM;
3578 			*bad_wr = wr;
3579 			goto out;
3580 		}
3581 
3582 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3583 			err = -EINVAL;
3584 			*bad_wr = wr;
3585 			goto out;
3586 		}
3587 
3588 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3589 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3590 
3591 		ctrl->srcrb_flags =
3592 			(wr->send_flags & IB_SEND_SIGNALED ?
3593 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3594 			(wr->send_flags & IB_SEND_SOLICITED ?
3595 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3596 			((wr->send_flags & IB_SEND_IP_CSUM) ?
3597 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3598 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3599 			qp->sq_signal_bits;
3600 
3601 		ctrl->imm = send_ieth(wr);
3602 
3603 		wqe += sizeof *ctrl;
3604 		size = sizeof *ctrl / 16;
3605 
3606 		switch (qp->mlx4_ib_qp_type) {
3607 		case MLX4_IB_QPT_RC:
3608 		case MLX4_IB_QPT_UC:
3609 			switch (wr->opcode) {
3610 			case IB_WR_ATOMIC_CMP_AND_SWP:
3611 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3612 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3613 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3614 					      atomic_wr(wr)->rkey);
3615 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3616 
3617 				set_atomic_seg(wqe, atomic_wr(wr));
3618 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
3619 
3620 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3621 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3622 
3623 				break;
3624 
3625 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3626 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3627 					      atomic_wr(wr)->rkey);
3628 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3629 
3630 				set_masked_atomic_seg(wqe, atomic_wr(wr));
3631 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
3632 
3633 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3634 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3635 
3636 				break;
3637 
3638 			case IB_WR_RDMA_READ:
3639 			case IB_WR_RDMA_WRITE:
3640 			case IB_WR_RDMA_WRITE_WITH_IMM:
3641 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3642 					      rdma_wr(wr)->rkey);
3643 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3644 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3645 				break;
3646 
3647 			case IB_WR_LOCAL_INV:
3648 				ctrl->srcrb_flags |=
3649 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3650 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3651 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3652 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3653 				break;
3654 
3655 			case IB_WR_REG_MR:
3656 				ctrl->srcrb_flags |=
3657 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3658 				set_reg_seg(wqe, reg_wr(wr));
3659 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3660 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3661 				break;
3662 
3663 			default:
3664 				/* No extra segments required for sends */
3665 				break;
3666 			}
3667 			break;
3668 
3669 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3670 			err = build_sriov_qp0_header(qp, ud_wr(wr), ctrl,
3671 						     &seglen);
3672 			if (unlikely(err)) {
3673 				*bad_wr = wr;
3674 				goto out;
3675 			}
3676 			wqe  += seglen;
3677 			size += seglen / 16;
3678 			break;
3679 		case MLX4_IB_QPT_TUN_SMI:
3680 		case MLX4_IB_QPT_TUN_GSI:
3681 			/* this is a UD qp used in MAD responses to slaves. */
3682 			set_datagram_seg(wqe, ud_wr(wr));
3683 			/* set the forced-loopback bit in the data seg av */
3684 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3685 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3686 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3687 			break;
3688 		case MLX4_IB_QPT_UD:
3689 			set_datagram_seg(wqe, ud_wr(wr));
3690 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3691 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3692 
3693 			if (wr->opcode == IB_WR_LSO) {
3694 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3695 						&lso_hdr_sz, &blh);
3696 				if (unlikely(err)) {
3697 					*bad_wr = wr;
3698 					goto out;
3699 				}
3700 				lso_wqe = (__be32 *) wqe;
3701 				wqe  += seglen;
3702 				size += seglen / 16;
3703 			}
3704 			break;
3705 
3706 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3707 			err = build_sriov_qp0_header(qp, ud_wr(wr), ctrl,
3708 						     &seglen);
3709 			if (unlikely(err)) {
3710 				*bad_wr = wr;
3711 				goto out;
3712 			}
3713 			wqe  += seglen;
3714 			size += seglen / 16;
3715 			/* to start tunnel header on a cache-line boundary */
3716 			add_zero_len_inline(wqe);
3717 			wqe += 16;
3718 			size++;
3719 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3720 			wqe  += seglen;
3721 			size += seglen / 16;
3722 			break;
3723 		case MLX4_IB_QPT_PROXY_SMI:
3724 		case MLX4_IB_QPT_PROXY_GSI:
3725 			/* If we are tunneling special qps, this is a UD qp.
3726 			 * In this case we first add a UD segment targeting
3727 			 * the tunnel qp, and then add a header with address
3728 			 * information */
3729 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3730 						ud_wr(wr),
3731 						qp->mlx4_ib_qp_type);
3732 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3733 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3734 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3735 			wqe  += seglen;
3736 			size += seglen / 16;
3737 			break;
3738 
3739 		case MLX4_IB_QPT_SMI:
3740 		case MLX4_IB_QPT_GSI:
3741 			err = build_mlx_header(qp, ud_wr(wr), ctrl, &seglen);
3742 			if (unlikely(err)) {
3743 				*bad_wr = wr;
3744 				goto out;
3745 			}
3746 			wqe  += seglen;
3747 			size += seglen / 16;
3748 			break;
3749 
3750 		default:
3751 			break;
3752 		}
3753 
3754 		/*
3755 		 * Write data segments in reverse order, so as to
3756 		 * overwrite cacheline stamp last within each
3757 		 * cacheline.  This avoids issues with WQE
3758 		 * prefetching.
3759 		 */
3760 
3761 		dseg = wqe;
3762 		dseg += wr->num_sge - 1;
3763 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3764 
3765 		/* Add one more inline data segment for ICRC for MLX sends */
3766 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3767 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3768 			     qp->mlx4_ib_qp_type &
3769 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3770 			set_mlx_icrc_seg(dseg + 1);
3771 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3772 		}
3773 
3774 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3775 			set_data_seg(dseg, wr->sg_list + i);
3776 
3777 		/*
3778 		 * Possibly overwrite stamping in cacheline with LSO
3779 		 * segment only after making sure all data segments
3780 		 * are written.
3781 		 */
3782 		wmb();
3783 		*lso_wqe = lso_hdr_sz;
3784 
3785 		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3786 					     MLX4_WQE_CTRL_FENCE : 0) | size;
3787 
3788 		/*
3789 		 * Make sure descriptor is fully written before
3790 		 * setting ownership bit (because HW can start
3791 		 * executing as soon as we do).
3792 		 */
3793 		wmb();
3794 
3795 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3796 			*bad_wr = wr;
3797 			err = -EINVAL;
3798 			goto out;
3799 		}
3800 
3801 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3802 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3803 
3804 		/*
3805 		 * We can improve latency by not stamping the last
3806 		 * send queue WQE until after ringing the doorbell, so
3807 		 * only stamp here if there are still more WQEs to post.
3808 		 */
3809 		if (wr->next)
3810 			stamp_send_wqe(qp, ind + qp->sq_spare_wqes);
3811 		ind++;
3812 	}
3813 
3814 out:
3815 	if (likely(nreq)) {
3816 		qp->sq.head += nreq;
3817 
3818 		/*
3819 		 * Make sure that descriptors are written before
3820 		 * doorbell record.
3821 		 */
3822 		wmb();
3823 
3824 		writel_relaxed(qp->doorbell_qpn,
3825 			to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3826 
3827 		stamp_send_wqe(qp, ind + qp->sq_spare_wqes - 1);
3828 
3829 		qp->sq_next_wqe = ind;
3830 	}
3831 
3832 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3833 
3834 	return err;
3835 }
3836 
3837 int mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3838 		      const struct ib_send_wr **bad_wr)
3839 {
3840 	return _mlx4_ib_post_send(ibqp, wr, bad_wr, false);
3841 }
3842 
3843 static int _mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3844 			      const struct ib_recv_wr **bad_wr, bool drain)
3845 {
3846 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3847 	struct mlx4_wqe_data_seg *scat;
3848 	unsigned long flags;
3849 	int err = 0;
3850 	int nreq;
3851 	int ind;
3852 	int max_gs;
3853 	int i;
3854 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3855 
3856 	max_gs = qp->rq.max_gs;
3857 	spin_lock_irqsave(&qp->rq.lock, flags);
3858 
3859 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3860 	    !drain) {
3861 		err = -EIO;
3862 		*bad_wr = wr;
3863 		nreq = 0;
3864 		goto out;
3865 	}
3866 
3867 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3868 
3869 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3870 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3871 			err = -ENOMEM;
3872 			*bad_wr = wr;
3873 			goto out;
3874 		}
3875 
3876 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3877 			err = -EINVAL;
3878 			*bad_wr = wr;
3879 			goto out;
3880 		}
3881 
3882 		scat = get_recv_wqe(qp, ind);
3883 
3884 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3885 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3886 			ib_dma_sync_single_for_device(ibqp->device,
3887 						      qp->sqp_proxy_rcv[ind].map,
3888 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3889 						      DMA_FROM_DEVICE);
3890 			scat->byte_count =
3891 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3892 			/* use dma lkey from upper layer entry */
3893 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3894 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3895 			scat++;
3896 			max_gs--;
3897 		}
3898 
3899 		for (i = 0; i < wr->num_sge; ++i)
3900 			__set_data_seg(scat + i, wr->sg_list + i);
3901 
3902 		if (i < max_gs) {
3903 			scat[i].byte_count = 0;
3904 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3905 			scat[i].addr       = 0;
3906 		}
3907 
3908 		qp->rq.wrid[ind] = wr->wr_id;
3909 
3910 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3911 	}
3912 
3913 out:
3914 	if (likely(nreq)) {
3915 		qp->rq.head += nreq;
3916 
3917 		/*
3918 		 * Make sure that descriptors are written before
3919 		 * doorbell record.
3920 		 */
3921 		wmb();
3922 
3923 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3924 	}
3925 
3926 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3927 
3928 	return err;
3929 }
3930 
3931 int mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3932 		      const struct ib_recv_wr **bad_wr)
3933 {
3934 	return _mlx4_ib_post_recv(ibqp, wr, bad_wr, false);
3935 }
3936 
3937 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3938 {
3939 	switch (mlx4_state) {
3940 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3941 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3942 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3943 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3944 	case MLX4_QP_STATE_SQ_DRAINING:
3945 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3946 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3947 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3948 	default:		     return -1;
3949 	}
3950 }
3951 
3952 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3953 {
3954 	switch (mlx4_mig_state) {
3955 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3956 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3957 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3958 	default: return -1;
3959 	}
3960 }
3961 
3962 static int to_ib_qp_access_flags(int mlx4_flags)
3963 {
3964 	int ib_flags = 0;
3965 
3966 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3967 		ib_flags |= IB_ACCESS_REMOTE_READ;
3968 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3969 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3970 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3971 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3972 
3973 	return ib_flags;
3974 }
3975 
3976 static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
3977 			    struct rdma_ah_attr *ah_attr,
3978 			    struct mlx4_qp_path *path)
3979 {
3980 	struct mlx4_dev *dev = ibdev->dev;
3981 	u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
3982 
3983 	memset(ah_attr, 0, sizeof(*ah_attr));
3984 	if (port_num == 0 || port_num > dev->caps.num_ports)
3985 		return;
3986 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
3987 
3988 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
3989 		rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
3990 			       ((path->sched_queue & 4) << 1));
3991 	else
3992 		rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
3993 	rdma_ah_set_port_num(ah_attr, port_num);
3994 
3995 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
3996 	rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
3997 	rdma_ah_set_static_rate(ah_attr,
3998 				path->static_rate ? path->static_rate - 5 : 0);
3999 	if (path->grh_mylmc & (1 << 7)) {
4000 		rdma_ah_set_grh(ah_attr, NULL,
4001 				be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
4002 				path->mgid_index,
4003 				path->hop_limit,
4004 				(be32_to_cpu(path->tclass_flowlabel)
4005 				 >> 20) & 0xff);
4006 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
4007 	}
4008 }
4009 
4010 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
4011 		     struct ib_qp_init_attr *qp_init_attr)
4012 {
4013 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
4014 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
4015 	struct mlx4_qp_context context;
4016 	int mlx4_state;
4017 	int err = 0;
4018 
4019 	if (ibqp->rwq_ind_tbl)
4020 		return -EOPNOTSUPP;
4021 
4022 	mutex_lock(&qp->mutex);
4023 
4024 	if (qp->state == IB_QPS_RESET) {
4025 		qp_attr->qp_state = IB_QPS_RESET;
4026 		goto done;
4027 	}
4028 
4029 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
4030 	if (err) {
4031 		err = -EINVAL;
4032 		goto out;
4033 	}
4034 
4035 	mlx4_state = be32_to_cpu(context.flags) >> 28;
4036 
4037 	qp->state		     = to_ib_qp_state(mlx4_state);
4038 	qp_attr->qp_state	     = qp->state;
4039 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
4040 	qp_attr->path_mig_state	     =
4041 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
4042 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
4043 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
4044 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
4045 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
4046 	qp_attr->qp_access_flags     =
4047 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
4048 
4049 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC ||
4050 	    qp->ibqp.qp_type == IB_QPT_XRC_INI ||
4051 	    qp->ibqp.qp_type == IB_QPT_XRC_TGT) {
4052 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
4053 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
4054 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
4055 		qp_attr->alt_port_num	=
4056 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
4057 	}
4058 
4059 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
4060 	if (qp_attr->qp_state == IB_QPS_INIT)
4061 		qp_attr->port_num = qp->port;
4062 	else
4063 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
4064 
4065 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4066 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
4067 
4068 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
4069 
4070 	qp_attr->max_dest_rd_atomic =
4071 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
4072 	qp_attr->min_rnr_timer	    =
4073 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
4074 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
4075 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
4076 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
4077 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
4078 
4079 done:
4080 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4081 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4082 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4083 
4084 	if (!ibqp->uobject) {
4085 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
4086 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4087 	} else {
4088 		qp_attr->cap.max_send_wr  = 0;
4089 		qp_attr->cap.max_send_sge = 0;
4090 	}
4091 
4092 	/*
4093 	 * We don't support inline sends for kernel QPs (yet), and we
4094 	 * don't know what userspace's value should be.
4095 	 */
4096 	qp_attr->cap.max_inline_data = 0;
4097 
4098 	qp_init_attr->cap	     = qp_attr->cap;
4099 
4100 	qp_init_attr->create_flags = 0;
4101 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4102 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4103 
4104 	if (qp->flags & MLX4_IB_QP_LSO)
4105 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
4106 
4107 	if (qp->flags & MLX4_IB_QP_NETIF)
4108 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
4109 
4110 	qp_init_attr->sq_sig_type =
4111 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
4112 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4113 
4114 out:
4115 	mutex_unlock(&qp->mutex);
4116 	return err;
4117 }
4118 
4119 struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
4120 				struct ib_wq_init_attr *init_attr,
4121 				struct ib_udata *udata)
4122 {
4123 	struct mlx4_dev *dev = to_mdev(pd->device)->dev;
4124 	struct ib_qp_init_attr ib_qp_init_attr = {};
4125 	struct mlx4_ib_qp *qp;
4126 	struct mlx4_ib_create_wq ucmd;
4127 	int err, required_cmd_sz;
4128 
4129 	if (!udata)
4130 		return ERR_PTR(-EINVAL);
4131 
4132 	required_cmd_sz = offsetof(typeof(ucmd), comp_mask) +
4133 			  sizeof(ucmd.comp_mask);
4134 	if (udata->inlen < required_cmd_sz) {
4135 		pr_debug("invalid inlen\n");
4136 		return ERR_PTR(-EINVAL);
4137 	}
4138 
4139 	if (udata->inlen > sizeof(ucmd) &&
4140 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4141 				 udata->inlen - sizeof(ucmd))) {
4142 		pr_debug("inlen is not supported\n");
4143 		return ERR_PTR(-EOPNOTSUPP);
4144 	}
4145 
4146 	if (udata->outlen)
4147 		return ERR_PTR(-EOPNOTSUPP);
4148 
4149 	if (init_attr->wq_type != IB_WQT_RQ) {
4150 		pr_debug("unsupported wq type %d\n", init_attr->wq_type);
4151 		return ERR_PTR(-EOPNOTSUPP);
4152 	}
4153 
4154 	if (init_attr->create_flags & ~IB_WQ_FLAGS_SCATTER_FCS ||
4155 	    !(dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
4156 		pr_debug("unsupported create_flags %u\n",
4157 			 init_attr->create_flags);
4158 		return ERR_PTR(-EOPNOTSUPP);
4159 	}
4160 
4161 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
4162 	if (!qp)
4163 		return ERR_PTR(-ENOMEM);
4164 
4165 	mutex_init(&qp->mutex);
4166 	qp->pri.vid = 0xFFFF;
4167 	qp->alt.vid = 0xFFFF;
4168 
4169 	ib_qp_init_attr.qp_context = init_attr->wq_context;
4170 	ib_qp_init_attr.qp_type = IB_QPT_RAW_PACKET;
4171 	ib_qp_init_attr.cap.max_recv_wr = init_attr->max_wr;
4172 	ib_qp_init_attr.cap.max_recv_sge = init_attr->max_sge;
4173 	ib_qp_init_attr.recv_cq = init_attr->cq;
4174 	ib_qp_init_attr.send_cq = ib_qp_init_attr.recv_cq; /* Dummy CQ */
4175 
4176 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS)
4177 		ib_qp_init_attr.create_flags |= IB_QP_CREATE_SCATTER_FCS;
4178 
4179 	err = create_rq(pd, &ib_qp_init_attr, udata, qp);
4180 	if (err) {
4181 		kfree(qp);
4182 		return ERR_PTR(err);
4183 	}
4184 
4185 	qp->ibwq.event_handler = init_attr->event_handler;
4186 	qp->ibwq.wq_num = qp->mqp.qpn;
4187 	qp->ibwq.state = IB_WQS_RESET;
4188 
4189 	return &qp->ibwq;
4190 }
4191 
4192 static int ib_wq2qp_state(enum ib_wq_state state)
4193 {
4194 	switch (state) {
4195 	case IB_WQS_RESET:
4196 		return IB_QPS_RESET;
4197 	case IB_WQS_RDY:
4198 		return IB_QPS_RTR;
4199 	default:
4200 		return IB_QPS_ERR;
4201 	}
4202 }
4203 
4204 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state,
4205 			      struct ib_udata *udata)
4206 {
4207 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4208 	enum ib_qp_state qp_cur_state;
4209 	enum ib_qp_state qp_new_state;
4210 	int attr_mask;
4211 	int err;
4212 
4213 	/* ib_qp.state represents the WQ HW state while ib_wq.state represents
4214 	 * the WQ logic state.
4215 	 */
4216 	qp_cur_state = qp->state;
4217 	qp_new_state = ib_wq2qp_state(new_state);
4218 
4219 	if (ib_wq2qp_state(new_state) == qp_cur_state)
4220 		return 0;
4221 
4222 	if (new_state == IB_WQS_RDY) {
4223 		struct ib_qp_attr attr = {};
4224 
4225 		attr.port_num = qp->port;
4226 		attr_mask = IB_QP_PORT;
4227 
4228 		err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, &attr,
4229 					  attr_mask, IB_QPS_RESET, IB_QPS_INIT,
4230 					  udata);
4231 		if (err) {
4232 			pr_debug("WQN=0x%06x failed to apply RST->INIT on the HW QP\n",
4233 				 ibwq->wq_num);
4234 			return err;
4235 		}
4236 
4237 		qp_cur_state = IB_QPS_INIT;
4238 	}
4239 
4240 	attr_mask = 0;
4241 	err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, attr_mask,
4242 				  qp_cur_state,  qp_new_state, udata);
4243 
4244 	if (err && (qp_cur_state == IB_QPS_INIT)) {
4245 		qp_new_state = IB_QPS_RESET;
4246 		if (__mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL,
4247 					attr_mask, IB_QPS_INIT, IB_QPS_RESET,
4248 					udata)) {
4249 			pr_warn("WQN=0x%06x failed with reverting HW's resources failure\n",
4250 				ibwq->wq_num);
4251 			qp_new_state = IB_QPS_INIT;
4252 		}
4253 	}
4254 
4255 	qp->state = qp_new_state;
4256 
4257 	return err;
4258 }
4259 
4260 int mlx4_ib_modify_wq(struct ib_wq *ibwq, struct ib_wq_attr *wq_attr,
4261 		      u32 wq_attr_mask, struct ib_udata *udata)
4262 {
4263 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4264 	struct mlx4_ib_modify_wq ucmd = {};
4265 	size_t required_cmd_sz;
4266 	enum ib_wq_state cur_state, new_state;
4267 	int err = 0;
4268 
4269 	required_cmd_sz = offsetof(typeof(ucmd), reserved) +
4270 				   sizeof(ucmd.reserved);
4271 	if (udata->inlen < required_cmd_sz)
4272 		return -EINVAL;
4273 
4274 	if (udata->inlen > sizeof(ucmd) &&
4275 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4276 				 udata->inlen - sizeof(ucmd)))
4277 		return -EOPNOTSUPP;
4278 
4279 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4280 		return -EFAULT;
4281 
4282 	if (ucmd.comp_mask || ucmd.reserved)
4283 		return -EOPNOTSUPP;
4284 
4285 	if (wq_attr_mask & IB_WQ_FLAGS)
4286 		return -EOPNOTSUPP;
4287 
4288 	cur_state = wq_attr->curr_wq_state;
4289 	new_state = wq_attr->wq_state;
4290 
4291 	if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR))
4292 		return -EINVAL;
4293 
4294 	if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET))
4295 		return -EINVAL;
4296 
4297 	/* Need to protect against the parent RSS which also may modify WQ
4298 	 * state.
4299 	 */
4300 	mutex_lock(&qp->mutex);
4301 
4302 	/* Can update HW state only if a RSS QP has already associated to this
4303 	 * WQ, so we can apply its port on the WQ.
4304 	 */
4305 	if (qp->rss_usecnt)
4306 		err = _mlx4_ib_modify_wq(ibwq, new_state, udata);
4307 
4308 	if (!err)
4309 		ibwq->state = new_state;
4310 
4311 	mutex_unlock(&qp->mutex);
4312 
4313 	return err;
4314 }
4315 
4316 int mlx4_ib_destroy_wq(struct ib_wq *ibwq, struct ib_udata *udata)
4317 {
4318 	struct mlx4_ib_dev *dev = to_mdev(ibwq->device);
4319 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4320 
4321 	if (qp->counter_index)
4322 		mlx4_ib_free_qp_counter(dev, qp);
4323 
4324 	destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, udata);
4325 
4326 	kfree(qp);
4327 	return 0;
4328 }
4329 
4330 int mlx4_ib_create_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table,
4331 				 struct ib_rwq_ind_table_init_attr *init_attr,
4332 				 struct ib_udata *udata)
4333 {
4334 	struct mlx4_ib_create_rwq_ind_tbl_resp resp = {};
4335 	unsigned int ind_tbl_size = 1 << init_attr->log_ind_tbl_size;
4336 	struct ib_device *device = rwq_ind_table->device;
4337 	unsigned int base_wqn;
4338 	size_t min_resp_len;
4339 	int i, err = 0;
4340 
4341 	if (udata->inlen > 0 &&
4342 	    !ib_is_udata_cleared(udata, 0,
4343 				 udata->inlen))
4344 		return -EOPNOTSUPP;
4345 
4346 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4347 	if (udata->outlen && udata->outlen < min_resp_len)
4348 		return -EINVAL;
4349 
4350 	if (ind_tbl_size >
4351 	    device->attrs.rss_caps.max_rwq_indirection_table_size) {
4352 		pr_debug("log_ind_tbl_size = %d is bigger than supported = %d\n",
4353 			 ind_tbl_size,
4354 			 device->attrs.rss_caps.max_rwq_indirection_table_size);
4355 		return -EINVAL;
4356 	}
4357 
4358 	base_wqn = init_attr->ind_tbl[0]->wq_num;
4359 
4360 	if (base_wqn % ind_tbl_size) {
4361 		pr_debug("WQN=0x%x isn't aligned with indirection table size\n",
4362 			 base_wqn);
4363 		return -EINVAL;
4364 	}
4365 
4366 	for (i = 1; i < ind_tbl_size; i++) {
4367 		if (++base_wqn != init_attr->ind_tbl[i]->wq_num) {
4368 			pr_debug("indirection table's WQNs aren't consecutive\n");
4369 			return -EINVAL;
4370 		}
4371 	}
4372 
4373 	if (udata->outlen) {
4374 		resp.response_length = offsetof(typeof(resp), response_length) +
4375 					sizeof(resp.response_length);
4376 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4377 	}
4378 
4379 	return err;
4380 }
4381 
4382 struct mlx4_ib_drain_cqe {
4383 	struct ib_cqe cqe;
4384 	struct completion done;
4385 };
4386 
4387 static void mlx4_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
4388 {
4389 	struct mlx4_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
4390 						     struct mlx4_ib_drain_cqe,
4391 						     cqe);
4392 
4393 	complete(&cqe->done);
4394 }
4395 
4396 /* This function returns only once the drained WR was completed */
4397 static void handle_drain_completion(struct ib_cq *cq,
4398 				    struct mlx4_ib_drain_cqe *sdrain,
4399 				    struct mlx4_ib_dev *dev)
4400 {
4401 	struct mlx4_dev *mdev = dev->dev;
4402 
4403 	if (cq->poll_ctx == IB_POLL_DIRECT) {
4404 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
4405 			ib_process_cq_direct(cq, -1);
4406 		return;
4407 	}
4408 
4409 	if (mdev->persist->state == MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4410 		struct mlx4_ib_cq *mcq = to_mcq(cq);
4411 		bool triggered = false;
4412 		unsigned long flags;
4413 
4414 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
4415 		/* Make sure that the CQ handler won't run if wasn't run yet */
4416 		if (!mcq->mcq.reset_notify_added)
4417 			mcq->mcq.reset_notify_added = 1;
4418 		else
4419 			triggered = true;
4420 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
4421 
4422 		if (triggered) {
4423 			/* Wait for any scheduled/running task to be ended */
4424 			switch (cq->poll_ctx) {
4425 			case IB_POLL_SOFTIRQ:
4426 				irq_poll_disable(&cq->iop);
4427 				irq_poll_enable(&cq->iop);
4428 				break;
4429 			case IB_POLL_WORKQUEUE:
4430 				cancel_work_sync(&cq->work);
4431 				break;
4432 			default:
4433 				WARN_ON_ONCE(1);
4434 			}
4435 		}
4436 
4437 		/* Run the CQ handler - this makes sure that the drain WR will
4438 		 * be processed if wasn't processed yet.
4439 		 */
4440 		mcq->mcq.comp(&mcq->mcq);
4441 	}
4442 
4443 	wait_for_completion(&sdrain->done);
4444 }
4445 
4446 void mlx4_ib_drain_sq(struct ib_qp *qp)
4447 {
4448 	struct ib_cq *cq = qp->send_cq;
4449 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4450 	struct mlx4_ib_drain_cqe sdrain;
4451 	const struct ib_send_wr *bad_swr;
4452 	struct ib_rdma_wr swr = {
4453 		.wr = {
4454 			.next = NULL,
4455 			{ .wr_cqe	= &sdrain.cqe, },
4456 			.opcode	= IB_WR_RDMA_WRITE,
4457 		},
4458 	};
4459 	int ret;
4460 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
4461 	struct mlx4_dev *mdev = dev->dev;
4462 
4463 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4464 	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4465 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4466 		return;
4467 	}
4468 
4469 	sdrain.cqe.done = mlx4_ib_drain_qp_done;
4470 	init_completion(&sdrain.done);
4471 
4472 	ret = _mlx4_ib_post_send(qp, &swr.wr, &bad_swr, true);
4473 	if (ret) {
4474 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4475 		return;
4476 	}
4477 
4478 	handle_drain_completion(cq, &sdrain, dev);
4479 }
4480 
4481 void mlx4_ib_drain_rq(struct ib_qp *qp)
4482 {
4483 	struct ib_cq *cq = qp->recv_cq;
4484 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4485 	struct mlx4_ib_drain_cqe rdrain;
4486 	struct ib_recv_wr rwr = {};
4487 	const struct ib_recv_wr *bad_rwr;
4488 	int ret;
4489 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
4490 	struct mlx4_dev *mdev = dev->dev;
4491 
4492 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4493 	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4494 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4495 		return;
4496 	}
4497 
4498 	rwr.wr_cqe = &rdrain.cqe;
4499 	rdrain.cqe.done = mlx4_ib_drain_qp_done;
4500 	init_completion(&rdrain.done);
4501 
4502 	ret = _mlx4_ib_post_recv(qp, &rwr, &bad_rwr, true);
4503 	if (ret) {
4504 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4505 		return;
4506 	}
4507 
4508 	handle_drain_completion(cq, &rdrain, dev);
4509 }
4510 
4511 int mlx4_ib_qp_event_init(void)
4512 {
4513 	mlx4_ib_qp_event_wq = alloc_ordered_workqueue("mlx4_ib_qp_event_wq", 0);
4514 	if (!mlx4_ib_qp_event_wq)
4515 		return -ENOMEM;
4516 
4517 	return 0;
4518 }
4519 
4520 void mlx4_ib_qp_event_cleanup(void)
4521 {
4522 	destroy_workqueue(mlx4_ib_qp_event_wq);
4523 }
4524