xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision 023e41632e065d49bcbe31b3c4b336217f96a271)
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 	/*
72 	 * Largest possible UD header: send with GRH and immediate
73 	 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
74 	 * tag.  (LRH would only use 8 bytes, so Ethernet is the
75 	 * biggest case)
76 	 */
77 	MLX4_IB_UD_HEADER_SIZE		= 82,
78 	MLX4_IB_LSO_HEADER_SPARE	= 128,
79 };
80 
81 struct mlx4_ib_sqp {
82 	struct mlx4_ib_qp	qp;
83 	int			pkey_index;
84 	u32			qkey;
85 	u32			send_psn;
86 	struct ib_ud_header	ud_header;
87 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
88 	struct ib_qp		*roce_v2_gsi;
89 };
90 
91 enum {
92 	MLX4_IB_MIN_SQ_STRIDE	= 6,
93 	MLX4_IB_CACHE_LINE_SIZE	= 64,
94 };
95 
96 enum {
97 	MLX4_RAW_QP_MTU		= 7,
98 	MLX4_RAW_QP_MSGMAX	= 31,
99 };
100 
101 #ifndef ETH_ALEN
102 #define ETH_ALEN        6
103 #endif
104 
105 static const __be32 mlx4_ib_opcode[] = {
106 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
107 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
108 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
109 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
110 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
111 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
112 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
113 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
114 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
115 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
116 	[IB_WR_REG_MR]				= cpu_to_be32(MLX4_OPCODE_FMR),
117 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
118 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
119 };
120 
121 enum mlx4_ib_source_type {
122 	MLX4_IB_QP_SRC	= 0,
123 	MLX4_IB_RWQ_SRC	= 1,
124 };
125 
126 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
127 {
128 	return container_of(mqp, struct mlx4_ib_sqp, qp);
129 }
130 
131 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
132 {
133 	if (!mlx4_is_master(dev->dev))
134 		return 0;
135 
136 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
137 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
138 		8 * MLX4_MFUNC_MAX;
139 }
140 
141 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
142 {
143 	int proxy_sqp = 0;
144 	int real_sqp = 0;
145 	int i;
146 	/* PPF or Native -- real SQP */
147 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
148 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
149 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
150 	if (real_sqp)
151 		return 1;
152 	/* VF or PF -- proxy SQP */
153 	if (mlx4_is_mfunc(dev->dev)) {
154 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
155 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy ||
156 			    qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp1_proxy) {
157 				proxy_sqp = 1;
158 				break;
159 			}
160 		}
161 	}
162 	if (proxy_sqp)
163 		return 1;
164 
165 	return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
166 }
167 
168 /* used for INIT/CLOSE port logic */
169 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
170 {
171 	int proxy_qp0 = 0;
172 	int real_qp0 = 0;
173 	int i;
174 	/* PPF or Native -- real QP0 */
175 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
176 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
177 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
178 	if (real_qp0)
179 		return 1;
180 	/* VF or PF -- proxy QP0 */
181 	if (mlx4_is_mfunc(dev->dev)) {
182 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
183 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy) {
184 				proxy_qp0 = 1;
185 				break;
186 			}
187 		}
188 	}
189 	return proxy_qp0;
190 }
191 
192 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
193 {
194 	return mlx4_buf_offset(&qp->buf, offset);
195 }
196 
197 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
198 {
199 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
200 }
201 
202 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
203 {
204 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
205 }
206 
207 /*
208  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
209  * first four bytes of every 64 byte chunk with 0xffffffff, except for
210  * the very first chunk of the WQE.
211  */
212 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n)
213 {
214 	__be32 *wqe;
215 	int i;
216 	int s;
217 	void *buf;
218 	struct mlx4_wqe_ctrl_seg *ctrl;
219 
220 	buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
221 	ctrl = (struct mlx4_wqe_ctrl_seg *)buf;
222 	s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4;
223 	for (i = 64; i < s; i += 64) {
224 		wqe = buf + i;
225 		*wqe = cpu_to_be32(0xffffffff);
226 	}
227 }
228 
229 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
230 {
231 	struct ib_event event;
232 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
233 
234 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
235 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
236 
237 	if (ibqp->event_handler) {
238 		event.device     = ibqp->device;
239 		event.element.qp = ibqp;
240 		switch (type) {
241 		case MLX4_EVENT_TYPE_PATH_MIG:
242 			event.event = IB_EVENT_PATH_MIG;
243 			break;
244 		case MLX4_EVENT_TYPE_COMM_EST:
245 			event.event = IB_EVENT_COMM_EST;
246 			break;
247 		case MLX4_EVENT_TYPE_SQ_DRAINED:
248 			event.event = IB_EVENT_SQ_DRAINED;
249 			break;
250 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
251 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
252 			break;
253 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
254 			event.event = IB_EVENT_QP_FATAL;
255 			break;
256 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
257 			event.event = IB_EVENT_PATH_MIG_ERR;
258 			break;
259 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
260 			event.event = IB_EVENT_QP_REQ_ERR;
261 			break;
262 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
263 			event.event = IB_EVENT_QP_ACCESS_ERR;
264 			break;
265 		default:
266 			pr_warn("Unexpected event type %d "
267 			       "on QP %06x\n", type, qp->qpn);
268 			return;
269 		}
270 
271 		ibqp->event_handler(&event, ibqp->qp_context);
272 	}
273 }
274 
275 static void mlx4_ib_wq_event(struct mlx4_qp *qp, enum mlx4_event type)
276 {
277 	pr_warn_ratelimited("Unexpected event type %d on WQ 0x%06x. Events are not supported for WQs\n",
278 			    type, qp->qpn);
279 }
280 
281 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
282 {
283 	/*
284 	 * UD WQEs must have a datagram segment.
285 	 * RC and UC WQEs might have a remote address segment.
286 	 * MLX WQEs need two extra inline data segments (for the UD
287 	 * header and space for the ICRC).
288 	 */
289 	switch (type) {
290 	case MLX4_IB_QPT_UD:
291 		return sizeof (struct mlx4_wqe_ctrl_seg) +
292 			sizeof (struct mlx4_wqe_datagram_seg) +
293 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
294 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
295 	case MLX4_IB_QPT_PROXY_SMI:
296 	case MLX4_IB_QPT_PROXY_GSI:
297 		return sizeof (struct mlx4_wqe_ctrl_seg) +
298 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
299 	case MLX4_IB_QPT_TUN_SMI_OWNER:
300 	case MLX4_IB_QPT_TUN_GSI:
301 		return sizeof (struct mlx4_wqe_ctrl_seg) +
302 			sizeof (struct mlx4_wqe_datagram_seg);
303 
304 	case MLX4_IB_QPT_UC:
305 		return sizeof (struct mlx4_wqe_ctrl_seg) +
306 			sizeof (struct mlx4_wqe_raddr_seg);
307 	case MLX4_IB_QPT_RC:
308 		return sizeof (struct mlx4_wqe_ctrl_seg) +
309 			sizeof (struct mlx4_wqe_masked_atomic_seg) +
310 			sizeof (struct mlx4_wqe_raddr_seg);
311 	case MLX4_IB_QPT_SMI:
312 	case MLX4_IB_QPT_GSI:
313 		return sizeof (struct mlx4_wqe_ctrl_seg) +
314 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
315 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
316 					   MLX4_INLINE_ALIGN) *
317 			      sizeof (struct mlx4_wqe_inline_seg),
318 			      sizeof (struct mlx4_wqe_data_seg)) +
319 			ALIGN(4 +
320 			      sizeof (struct mlx4_wqe_inline_seg),
321 			      sizeof (struct mlx4_wqe_data_seg));
322 	default:
323 		return sizeof (struct mlx4_wqe_ctrl_seg);
324 	}
325 }
326 
327 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
328 		       bool is_user, int has_rq, struct mlx4_ib_qp *qp,
329 		       u32 inl_recv_sz)
330 {
331 	/* Sanity check RQ size before proceeding */
332 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
333 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
334 		return -EINVAL;
335 
336 	if (!has_rq) {
337 		if (cap->max_recv_wr || inl_recv_sz)
338 			return -EINVAL;
339 
340 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
341 	} else {
342 		u32 max_inl_recv_sz = dev->dev->caps.max_rq_sg *
343 			sizeof(struct mlx4_wqe_data_seg);
344 		u32 wqe_size;
345 
346 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
347 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge ||
348 				inl_recv_sz > max_inl_recv_sz))
349 			return -EINVAL;
350 
351 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
352 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
353 		wqe_size = qp->rq.max_gs * sizeof(struct mlx4_wqe_data_seg);
354 		qp->rq.wqe_shift = ilog2(max_t(u32, wqe_size, inl_recv_sz));
355 	}
356 
357 	/* leave userspace return values as they were, so as not to break ABI */
358 	if (is_user) {
359 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
360 		cap->max_recv_sge = qp->rq.max_gs;
361 	} else {
362 		cap->max_recv_wr  = qp->rq.max_post =
363 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
364 		cap->max_recv_sge = min(qp->rq.max_gs,
365 					min(dev->dev->caps.max_sq_sg,
366 					    dev->dev->caps.max_rq_sg));
367 	}
368 
369 	return 0;
370 }
371 
372 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
373 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
374 {
375 	int s;
376 
377 	/* Sanity check SQ size before proceeding */
378 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
379 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
380 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
381 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
382 		return -EINVAL;
383 
384 	/*
385 	 * For MLX transport we need 2 extra S/G entries:
386 	 * one for the header and one for the checksum at the end
387 	 */
388 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
389 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
390 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
391 		return -EINVAL;
392 
393 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
394 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
395 		send_wqe_overhead(type, qp->flags);
396 
397 	if (s > dev->dev->caps.max_sq_desc_sz)
398 		return -EINVAL;
399 
400 	qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
401 
402 	/*
403 	 * We need to leave 2 KB + 1 WR of headroom in the SQ to
404 	 * allow HW to prefetch.
405 	 */
406 	qp->sq_spare_wqes = MLX4_IB_SQ_HEADROOM(qp->sq.wqe_shift);
407 	qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr +
408 					    qp->sq_spare_wqes);
409 
410 	qp->sq.max_gs =
411 		(min(dev->dev->caps.max_sq_desc_sz,
412 		     (1 << qp->sq.wqe_shift)) -
413 		 send_wqe_overhead(type, qp->flags)) /
414 		sizeof (struct mlx4_wqe_data_seg);
415 
416 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
417 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
418 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
419 		qp->rq.offset = 0;
420 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
421 	} else {
422 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
423 		qp->sq.offset = 0;
424 	}
425 
426 	cap->max_send_wr  = qp->sq.max_post =
427 		qp->sq.wqe_cnt - qp->sq_spare_wqes;
428 	cap->max_send_sge = min(qp->sq.max_gs,
429 				min(dev->dev->caps.max_sq_sg,
430 				    dev->dev->caps.max_rq_sg));
431 	/* We don't support inline sends for kernel QPs (yet) */
432 	cap->max_inline_data = 0;
433 
434 	return 0;
435 }
436 
437 static int set_user_sq_size(struct mlx4_ib_dev *dev,
438 			    struct mlx4_ib_qp *qp,
439 			    struct mlx4_ib_create_qp *ucmd)
440 {
441 	/* Sanity check SQ size before proceeding */
442 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
443 	    ucmd->log_sq_stride >
444 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
445 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
446 		return -EINVAL;
447 
448 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
449 	qp->sq.wqe_shift = ucmd->log_sq_stride;
450 
451 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
452 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
453 
454 	return 0;
455 }
456 
457 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
458 {
459 	int i;
460 
461 	qp->sqp_proxy_rcv =
462 		kmalloc_array(qp->rq.wqe_cnt, sizeof(struct mlx4_ib_buf),
463 			      GFP_KERNEL);
464 	if (!qp->sqp_proxy_rcv)
465 		return -ENOMEM;
466 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
467 		qp->sqp_proxy_rcv[i].addr =
468 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
469 				GFP_KERNEL);
470 		if (!qp->sqp_proxy_rcv[i].addr)
471 			goto err;
472 		qp->sqp_proxy_rcv[i].map =
473 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
474 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
475 					  DMA_FROM_DEVICE);
476 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
477 			kfree(qp->sqp_proxy_rcv[i].addr);
478 			goto err;
479 		}
480 	}
481 	return 0;
482 
483 err:
484 	while (i > 0) {
485 		--i;
486 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
487 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
488 				    DMA_FROM_DEVICE);
489 		kfree(qp->sqp_proxy_rcv[i].addr);
490 	}
491 	kfree(qp->sqp_proxy_rcv);
492 	qp->sqp_proxy_rcv = NULL;
493 	return -ENOMEM;
494 }
495 
496 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
497 {
498 	int i;
499 
500 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
501 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
502 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
503 				    DMA_FROM_DEVICE);
504 		kfree(qp->sqp_proxy_rcv[i].addr);
505 	}
506 	kfree(qp->sqp_proxy_rcv);
507 }
508 
509 static int qp_has_rq(struct ib_qp_init_attr *attr)
510 {
511 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
512 		return 0;
513 
514 	return !attr->srq;
515 }
516 
517 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
518 {
519 	int i;
520 	for (i = 0; i < dev->caps.num_ports; i++) {
521 		if (qpn == dev->caps.spec_qps[i].qp0_proxy)
522 			return !!dev->caps.spec_qps[i].qp0_qkey;
523 	}
524 	return 0;
525 }
526 
527 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
528 				    struct mlx4_ib_qp *qp)
529 {
530 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
531 	mlx4_counter_free(dev->dev, qp->counter_index->index);
532 	list_del(&qp->counter_index->list);
533 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
534 
535 	kfree(qp->counter_index);
536 	qp->counter_index = NULL;
537 }
538 
539 static int set_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_rss *rss_ctx,
540 		      struct ib_qp_init_attr *init_attr,
541 		      struct mlx4_ib_create_qp_rss *ucmd)
542 {
543 	rss_ctx->base_qpn_tbl_sz = init_attr->rwq_ind_tbl->ind_tbl[0]->wq_num |
544 		(init_attr->rwq_ind_tbl->log_ind_tbl_size << 24);
545 
546 	if ((ucmd->rx_hash_function == MLX4_IB_RX_HASH_FUNC_TOEPLITZ) &&
547 	    (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS_TOP)) {
548 		memcpy(rss_ctx->rss_key, ucmd->rx_hash_key,
549 		       MLX4_EN_RSS_KEY_SIZE);
550 	} else {
551 		pr_debug("RX Hash function is not supported\n");
552 		return (-EOPNOTSUPP);
553 	}
554 
555 	if (ucmd->rx_hash_fields_mask & ~(MLX4_IB_RX_HASH_SRC_IPV4	|
556 					  MLX4_IB_RX_HASH_DST_IPV4	|
557 					  MLX4_IB_RX_HASH_SRC_IPV6	|
558 					  MLX4_IB_RX_HASH_DST_IPV6	|
559 					  MLX4_IB_RX_HASH_SRC_PORT_TCP	|
560 					  MLX4_IB_RX_HASH_DST_PORT_TCP	|
561 					  MLX4_IB_RX_HASH_SRC_PORT_UDP	|
562 					  MLX4_IB_RX_HASH_DST_PORT_UDP  |
563 					  MLX4_IB_RX_HASH_INNER)) {
564 		pr_debug("RX Hash fields_mask has unsupported mask (0x%llx)\n",
565 			 ucmd->rx_hash_fields_mask);
566 		return (-EOPNOTSUPP);
567 	}
568 
569 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) &&
570 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
571 		rss_ctx->flags = MLX4_RSS_IPV4;
572 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) ||
573 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
574 		pr_debug("RX Hash fields_mask is not supported - both IPv4 SRC and DST must be set\n");
575 		return (-EOPNOTSUPP);
576 	}
577 
578 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) &&
579 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
580 		rss_ctx->flags |= MLX4_RSS_IPV6;
581 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) ||
582 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
583 		pr_debug("RX Hash fields_mask is not supported - both IPv6 SRC and DST must be set\n");
584 		return (-EOPNOTSUPP);
585 	}
586 
587 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) &&
588 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
589 		if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS)) {
590 			pr_debug("RX Hash fields_mask for UDP is not supported\n");
591 			return (-EOPNOTSUPP);
592 		}
593 
594 		if (rss_ctx->flags & MLX4_RSS_IPV4)
595 			rss_ctx->flags |= MLX4_RSS_UDP_IPV4;
596 		if (rss_ctx->flags & MLX4_RSS_IPV6)
597 			rss_ctx->flags |= MLX4_RSS_UDP_IPV6;
598 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
599 			pr_debug("RX Hash fields_mask is not supported - UDP must be set with IPv4 or IPv6\n");
600 			return (-EOPNOTSUPP);
601 		}
602 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) ||
603 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
604 		pr_debug("RX Hash fields_mask is not supported - both UDP SRC and DST must be set\n");
605 		return (-EOPNOTSUPP);
606 	}
607 
608 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) &&
609 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
610 		if (rss_ctx->flags & MLX4_RSS_IPV4)
611 			rss_ctx->flags |= MLX4_RSS_TCP_IPV4;
612 		if (rss_ctx->flags & MLX4_RSS_IPV6)
613 			rss_ctx->flags |= MLX4_RSS_TCP_IPV6;
614 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
615 			pr_debug("RX Hash fields_mask is not supported - TCP must be set with IPv4 or IPv6\n");
616 			return (-EOPNOTSUPP);
617 		}
618 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) ||
619 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
620 		pr_debug("RX Hash fields_mask is not supported - both TCP SRC and DST must be set\n");
621 		return (-EOPNOTSUPP);
622 	}
623 
624 	if (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_INNER) {
625 		if (dev->dev->caps.tunnel_offload_mode ==
626 		    MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
627 			/*
628 			 * Hash according to inner headers if exist, otherwise
629 			 * according to outer headers.
630 			 */
631 			rss_ctx->flags |= MLX4_RSS_BY_INNER_HEADERS_IPONLY;
632 		} else {
633 			pr_debug("RSS Hash for inner headers isn't supported\n");
634 			return (-EOPNOTSUPP);
635 		}
636 	}
637 
638 	return 0;
639 }
640 
641 static int create_qp_rss(struct mlx4_ib_dev *dev,
642 			 struct ib_qp_init_attr *init_attr,
643 			 struct mlx4_ib_create_qp_rss *ucmd,
644 			 struct mlx4_ib_qp *qp)
645 {
646 	int qpn;
647 	int err;
648 
649 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
650 
651 	err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 0, qp->mqp.usage);
652 	if (err)
653 		return err;
654 
655 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
656 	if (err)
657 		goto err_qpn;
658 
659 	mutex_init(&qp->mutex);
660 
661 	INIT_LIST_HEAD(&qp->gid_list);
662 	INIT_LIST_HEAD(&qp->steering_rules);
663 
664 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
665 	qp->state = IB_QPS_RESET;
666 
667 	/* Set dummy send resources to be compatible with HV and PRM */
668 	qp->sq_no_prefetch = 1;
669 	qp->sq.wqe_cnt = 1;
670 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
671 	qp->buf_size = qp->sq.wqe_cnt << MLX4_IB_MIN_SQ_STRIDE;
672 	qp->mtt = (to_mqp(
673 		   (struct ib_qp *)init_attr->rwq_ind_tbl->ind_tbl[0]))->mtt;
674 
675 	qp->rss_ctx = kzalloc(sizeof(*qp->rss_ctx), GFP_KERNEL);
676 	if (!qp->rss_ctx) {
677 		err = -ENOMEM;
678 		goto err_qp_alloc;
679 	}
680 
681 	err = set_qp_rss(dev, qp->rss_ctx, init_attr, ucmd);
682 	if (err)
683 		goto err;
684 
685 	return 0;
686 
687 err:
688 	kfree(qp->rss_ctx);
689 
690 err_qp_alloc:
691 	mlx4_qp_remove(dev->dev, &qp->mqp);
692 	mlx4_qp_free(dev->dev, &qp->mqp);
693 
694 err_qpn:
695 	mlx4_qp_release_range(dev->dev, qpn, 1);
696 	return err;
697 }
698 
699 static struct ib_qp *_mlx4_ib_create_qp_rss(struct ib_pd *pd,
700 					    struct ib_qp_init_attr *init_attr,
701 					    struct ib_udata *udata)
702 {
703 	struct mlx4_ib_qp *qp;
704 	struct mlx4_ib_create_qp_rss ucmd = {};
705 	size_t required_cmd_sz;
706 	int err;
707 
708 	if (!udata) {
709 		pr_debug("RSS QP with NULL udata\n");
710 		return ERR_PTR(-EINVAL);
711 	}
712 
713 	if (udata->outlen)
714 		return ERR_PTR(-EOPNOTSUPP);
715 
716 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) +
717 					sizeof(ucmd.reserved1);
718 	if (udata->inlen < required_cmd_sz) {
719 		pr_debug("invalid inlen\n");
720 		return ERR_PTR(-EINVAL);
721 	}
722 
723 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
724 		pr_debug("copy failed\n");
725 		return ERR_PTR(-EFAULT);
726 	}
727 
728 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)))
729 		return ERR_PTR(-EOPNOTSUPP);
730 
731 	if (ucmd.comp_mask || ucmd.reserved1)
732 		return ERR_PTR(-EOPNOTSUPP);
733 
734 	if (udata->inlen > sizeof(ucmd) &&
735 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
736 				 udata->inlen - sizeof(ucmd))) {
737 		pr_debug("inlen is not supported\n");
738 		return ERR_PTR(-EOPNOTSUPP);
739 	}
740 
741 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
742 		pr_debug("RSS QP with unsupported QP type %d\n",
743 			 init_attr->qp_type);
744 		return ERR_PTR(-EOPNOTSUPP);
745 	}
746 
747 	if (init_attr->create_flags) {
748 		pr_debug("RSS QP doesn't support create flags\n");
749 		return ERR_PTR(-EOPNOTSUPP);
750 	}
751 
752 	if (init_attr->send_cq || init_attr->cap.max_send_wr) {
753 		pr_debug("RSS QP with unsupported send attributes\n");
754 		return ERR_PTR(-EOPNOTSUPP);
755 	}
756 
757 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
758 	if (!qp)
759 		return ERR_PTR(-ENOMEM);
760 
761 	qp->pri.vid = 0xFFFF;
762 	qp->alt.vid = 0xFFFF;
763 
764 	err = create_qp_rss(to_mdev(pd->device), init_attr, &ucmd, qp);
765 	if (err) {
766 		kfree(qp);
767 		return ERR_PTR(err);
768 	}
769 
770 	qp->ibqp.qp_num = qp->mqp.qpn;
771 
772 	return &qp->ibqp;
773 }
774 
775 /*
776  * This function allocates a WQN from a range which is consecutive and aligned
777  * to its size. In case the range is full, then it creates a new range and
778  * allocates WQN from it. The new range will be used for following allocations.
779  */
780 static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context,
781 			     struct mlx4_ib_qp *qp, int range_size, int *wqn)
782 {
783 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
784 	struct mlx4_wqn_range *range;
785 	int err = 0;
786 
787 	mutex_lock(&context->wqn_ranges_mutex);
788 
789 	range = list_first_entry_or_null(&context->wqn_ranges_list,
790 					 struct mlx4_wqn_range, list);
791 
792 	if (!range || (range->refcount == range->size) || range->dirty) {
793 		range = kzalloc(sizeof(*range), GFP_KERNEL);
794 		if (!range) {
795 			err = -ENOMEM;
796 			goto out;
797 		}
798 
799 		err = mlx4_qp_reserve_range(dev->dev, range_size,
800 					    range_size, &range->base_wqn, 0,
801 					    qp->mqp.usage);
802 		if (err) {
803 			kfree(range);
804 			goto out;
805 		}
806 
807 		range->size = range_size;
808 		list_add(&range->list, &context->wqn_ranges_list);
809 	} else if (range_size != 1) {
810 		/*
811 		 * Requesting a new range (>1) when last range is still open, is
812 		 * not valid.
813 		 */
814 		err = -EINVAL;
815 		goto out;
816 	}
817 
818 	qp->wqn_range = range;
819 
820 	*wqn = range->base_wqn + range->refcount;
821 
822 	range->refcount++;
823 
824 out:
825 	mutex_unlock(&context->wqn_ranges_mutex);
826 
827 	return err;
828 }
829 
830 static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context,
831 				struct mlx4_ib_qp *qp, bool dirty_release)
832 {
833 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
834 	struct mlx4_wqn_range *range;
835 
836 	mutex_lock(&context->wqn_ranges_mutex);
837 
838 	range = qp->wqn_range;
839 
840 	range->refcount--;
841 	if (!range->refcount) {
842 		mlx4_qp_release_range(dev->dev, range->base_wqn,
843 				      range->size);
844 		list_del(&range->list);
845 		kfree(range);
846 	} else if (dirty_release) {
847 	/*
848 	 * A range which one of its WQNs is destroyed, won't be able to be
849 	 * reused for further WQN allocations.
850 	 * The next created WQ will allocate a new range.
851 	 */
852 		range->dirty = 1;
853 	}
854 
855 	mutex_unlock(&context->wqn_ranges_mutex);
856 }
857 
858 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
859 			    enum mlx4_ib_source_type src,
860 			    struct ib_qp_init_attr *init_attr,
861 			    struct ib_udata *udata, int sqpn,
862 			    struct mlx4_ib_qp **caller_qp)
863 {
864 	int qpn;
865 	int err;
866 	struct mlx4_ib_sqp *sqp = NULL;
867 	struct mlx4_ib_qp *qp;
868 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
869 		udata, struct mlx4_ib_ucontext, ibucontext);
870 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
871 	struct mlx4_ib_cq *mcq;
872 	unsigned long flags;
873 	int range_size = 0;
874 
875 	/* When tunneling special qps, we use a plain UD qp */
876 	if (sqpn) {
877 		if (mlx4_is_mfunc(dev->dev) &&
878 		    (!mlx4_is_master(dev->dev) ||
879 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
880 			if (init_attr->qp_type == IB_QPT_GSI)
881 				qp_type = MLX4_IB_QPT_PROXY_GSI;
882 			else {
883 				if (mlx4_is_master(dev->dev) ||
884 				    qp0_enabled_vf(dev->dev, sqpn))
885 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
886 				else
887 					qp_type = MLX4_IB_QPT_PROXY_SMI;
888 			}
889 		}
890 		qpn = sqpn;
891 		/* add extra sg entry for tunneling */
892 		init_attr->cap.max_recv_sge++;
893 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
894 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
895 			container_of(init_attr,
896 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
897 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
898 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
899 		    !mlx4_is_master(dev->dev))
900 			return -EINVAL;
901 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
902 			qp_type = MLX4_IB_QPT_TUN_GSI;
903 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
904 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
905 					     tnl_init->port))
906 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
907 		else
908 			qp_type = MLX4_IB_QPT_TUN_SMI;
909 		/* we are definitely in the PPF here, since we are creating
910 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
911 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
912 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
913 		sqpn = qpn;
914 	}
915 
916 	if (!*caller_qp) {
917 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
918 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
919 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
920 			sqp = kzalloc(sizeof(struct mlx4_ib_sqp), GFP_KERNEL);
921 			if (!sqp)
922 				return -ENOMEM;
923 			qp = &sqp->qp;
924 			qp->pri.vid = 0xFFFF;
925 			qp->alt.vid = 0xFFFF;
926 		} else {
927 			qp = kzalloc(sizeof(struct mlx4_ib_qp), GFP_KERNEL);
928 			if (!qp)
929 				return -ENOMEM;
930 			qp->pri.vid = 0xFFFF;
931 			qp->alt.vid = 0xFFFF;
932 		}
933 	} else
934 		qp = *caller_qp;
935 
936 	qp->mlx4_ib_qp_type = qp_type;
937 
938 	mutex_init(&qp->mutex);
939 	spin_lock_init(&qp->sq.lock);
940 	spin_lock_init(&qp->rq.lock);
941 	INIT_LIST_HEAD(&qp->gid_list);
942 	INIT_LIST_HEAD(&qp->steering_rules);
943 
944 	qp->state	 = IB_QPS_RESET;
945 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
946 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
947 
948 
949 	if (udata) {
950 		union {
951 			struct mlx4_ib_create_qp qp;
952 			struct mlx4_ib_create_wq wq;
953 		} ucmd;
954 		size_t copy_len;
955 		int shift;
956 		int n;
957 
958 		copy_len = (src == MLX4_IB_QP_SRC) ?
959 			   sizeof(struct mlx4_ib_create_qp) :
960 			   min(sizeof(struct mlx4_ib_create_wq), udata->inlen);
961 
962 		if (ib_copy_from_udata(&ucmd, udata, copy_len)) {
963 			err = -EFAULT;
964 			goto err;
965 		}
966 
967 		if (src == MLX4_IB_RWQ_SRC) {
968 			if (ucmd.wq.comp_mask || ucmd.wq.reserved[0] ||
969 			    ucmd.wq.reserved[1] || ucmd.wq.reserved[2]) {
970 				pr_debug("user command isn't supported\n");
971 				err = -EOPNOTSUPP;
972 				goto err;
973 			}
974 
975 			if (ucmd.wq.log_range_size >
976 			    ilog2(dev->dev->caps.max_rss_tbl_sz)) {
977 				pr_debug("WQN range size must be equal or smaller than %d\n",
978 					 dev->dev->caps.max_rss_tbl_sz);
979 				err = -EOPNOTSUPP;
980 				goto err;
981 			}
982 			range_size = 1 << ucmd.wq.log_range_size;
983 		} else {
984 			qp->inl_recv_sz = ucmd.qp.inl_recv_sz;
985 		}
986 
987 		if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
988 			if (!(dev->dev->caps.flags &
989 			      MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
990 				pr_debug("scatter FCS is unsupported\n");
991 				err = -EOPNOTSUPP;
992 				goto err;
993 			}
994 
995 			qp->flags |= MLX4_IB_QP_SCATTER_FCS;
996 		}
997 
998 		err = set_rq_size(dev, &init_attr->cap, udata,
999 				  qp_has_rq(init_attr), qp, qp->inl_recv_sz);
1000 		if (err)
1001 			goto err;
1002 
1003 		if (src == MLX4_IB_QP_SRC) {
1004 			qp->sq_no_prefetch = ucmd.qp.sq_no_prefetch;
1005 
1006 			err = set_user_sq_size(dev, qp,
1007 					       (struct mlx4_ib_create_qp *)
1008 					       &ucmd);
1009 			if (err)
1010 				goto err;
1011 		} else {
1012 			qp->sq_no_prefetch = 1;
1013 			qp->sq.wqe_cnt = 1;
1014 			qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
1015 			/* Allocated buffer expects to have at least that SQ
1016 			 * size.
1017 			 */
1018 			qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
1019 				(qp->sq.wqe_cnt << qp->sq.wqe_shift);
1020 		}
1021 
1022 		qp->umem =
1023 			ib_umem_get(udata,
1024 				    (src == MLX4_IB_QP_SRC) ? ucmd.qp.buf_addr :
1025 							      ucmd.wq.buf_addr,
1026 				    qp->buf_size, 0, 0);
1027 		if (IS_ERR(qp->umem)) {
1028 			err = PTR_ERR(qp->umem);
1029 			goto err;
1030 		}
1031 
1032 		n = ib_umem_page_count(qp->umem);
1033 		shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
1034 		err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
1035 
1036 		if (err)
1037 			goto err_buf;
1038 
1039 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
1040 		if (err)
1041 			goto err_mtt;
1042 
1043 		if (qp_has_rq(init_attr)) {
1044 			err = mlx4_ib_db_map_user(
1045 				context, udata,
1046 				(src == MLX4_IB_QP_SRC) ? ucmd.qp.db_addr :
1047 							  ucmd.wq.db_addr,
1048 				&qp->db);
1049 			if (err)
1050 				goto err_mtt;
1051 		}
1052 		qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
1053 	} else {
1054 		err = set_rq_size(dev, &init_attr->cap, udata,
1055 				  qp_has_rq(init_attr), qp, 0);
1056 		if (err)
1057 			goto err;
1058 
1059 		qp->sq_no_prefetch = 0;
1060 
1061 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
1062 			qp->flags |= MLX4_IB_QP_LSO;
1063 
1064 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1065 			if (dev->steering_support ==
1066 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
1067 				qp->flags |= MLX4_IB_QP_NETIF;
1068 			else
1069 				goto err;
1070 		}
1071 
1072 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
1073 		if (err)
1074 			goto err;
1075 
1076 		if (qp_has_rq(init_attr)) {
1077 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
1078 			if (err)
1079 				goto err;
1080 
1081 			*qp->db.db = 0;
1082 		}
1083 
1084 		if (mlx4_buf_alloc(dev->dev, qp->buf_size,  PAGE_SIZE * 2,
1085 				   &qp->buf)) {
1086 			err = -ENOMEM;
1087 			goto err_db;
1088 		}
1089 
1090 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
1091 				    &qp->mtt);
1092 		if (err)
1093 			goto err_buf;
1094 
1095 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
1096 		if (err)
1097 			goto err_mtt;
1098 
1099 		qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1100 					     sizeof(u64), GFP_KERNEL);
1101 		qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1102 					     sizeof(u64), GFP_KERNEL);
1103 		if (!qp->sq.wrid || !qp->rq.wrid) {
1104 			err = -ENOMEM;
1105 			goto err_wrid;
1106 		}
1107 		qp->mqp.usage = MLX4_RES_USAGE_DRIVER;
1108 	}
1109 
1110 	if (sqpn) {
1111 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1112 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
1113 			if (alloc_proxy_bufs(pd->device, qp)) {
1114 				err = -ENOMEM;
1115 				goto err_wrid;
1116 			}
1117 		}
1118 	} else if (src == MLX4_IB_RWQ_SRC) {
1119 		err = mlx4_ib_alloc_wqn(context, qp, range_size, &qpn);
1120 		if (err)
1121 			goto err_wrid;
1122 	} else {
1123 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
1124 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
1125 		 * VLAN insertion. */
1126 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
1127 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
1128 						    (init_attr->cap.max_send_wr ?
1129 						     MLX4_RESERVE_ETH_BF_QP : 0) |
1130 						    (init_attr->cap.max_recv_wr ?
1131 						     MLX4_RESERVE_A0_QP : 0),
1132 						    qp->mqp.usage);
1133 		else
1134 			if (qp->flags & MLX4_IB_QP_NETIF)
1135 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
1136 			else
1137 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
1138 							    &qpn, 0, qp->mqp.usage);
1139 		if (err)
1140 			goto err_proxy;
1141 	}
1142 
1143 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1144 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1145 
1146 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
1147 	if (err)
1148 		goto err_qpn;
1149 
1150 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1151 		qp->mqp.qpn |= (1 << 23);
1152 
1153 	/*
1154 	 * Hardware wants QPN written in big-endian order (after
1155 	 * shifting) for send doorbell.  Precompute this value to save
1156 	 * a little bit when posting sends.
1157 	 */
1158 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1159 
1160 	qp->mqp.event = (src == MLX4_IB_QP_SRC) ? mlx4_ib_qp_event :
1161 						  mlx4_ib_wq_event;
1162 
1163 	if (!*caller_qp)
1164 		*caller_qp = qp;
1165 
1166 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1167 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
1168 			 to_mcq(init_attr->recv_cq));
1169 	/* Maintain device to QPs access, needed for further handling
1170 	 * via reset flow
1171 	 */
1172 	list_add_tail(&qp->qps_list, &dev->qp_list);
1173 	/* Maintain CQ to QPs access, needed for further handling
1174 	 * via reset flow
1175 	 */
1176 	mcq = to_mcq(init_attr->send_cq);
1177 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
1178 	mcq = to_mcq(init_attr->recv_cq);
1179 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
1180 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
1181 			   to_mcq(init_attr->recv_cq));
1182 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1183 	return 0;
1184 
1185 err_qpn:
1186 	if (!sqpn) {
1187 		if (qp->flags & MLX4_IB_QP_NETIF)
1188 			mlx4_ib_steer_qp_free(dev, qpn, 1);
1189 		else if (src == MLX4_IB_RWQ_SRC)
1190 			mlx4_ib_release_wqn(context, qp, 0);
1191 		else
1192 			mlx4_qp_release_range(dev->dev, qpn, 1);
1193 	}
1194 err_proxy:
1195 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1196 		free_proxy_bufs(pd->device, qp);
1197 err_wrid:
1198 	if (udata) {
1199 		if (qp_has_rq(init_attr))
1200 			mlx4_ib_db_unmap_user(context, &qp->db);
1201 	} else {
1202 		kvfree(qp->sq.wrid);
1203 		kvfree(qp->rq.wrid);
1204 	}
1205 
1206 err_mtt:
1207 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1208 
1209 err_buf:
1210 	if (qp->umem)
1211 		ib_umem_release(qp->umem);
1212 	else
1213 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1214 
1215 err_db:
1216 	if (!udata && qp_has_rq(init_attr))
1217 		mlx4_db_free(dev->dev, &qp->db);
1218 
1219 err:
1220 	if (!sqp && !*caller_qp)
1221 		kfree(qp);
1222 	kfree(sqp);
1223 
1224 	return err;
1225 }
1226 
1227 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
1228 {
1229 	switch (state) {
1230 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
1231 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
1232 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
1233 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
1234 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
1235 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
1236 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
1237 	default:		return -1;
1238 	}
1239 }
1240 
1241 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1242 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1243 {
1244 	if (send_cq == recv_cq) {
1245 		spin_lock(&send_cq->lock);
1246 		__acquire(&recv_cq->lock);
1247 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1248 		spin_lock(&send_cq->lock);
1249 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1250 	} else {
1251 		spin_lock(&recv_cq->lock);
1252 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1253 	}
1254 }
1255 
1256 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1257 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1258 {
1259 	if (send_cq == recv_cq) {
1260 		__release(&recv_cq->lock);
1261 		spin_unlock(&send_cq->lock);
1262 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1263 		spin_unlock(&recv_cq->lock);
1264 		spin_unlock(&send_cq->lock);
1265 	} else {
1266 		spin_unlock(&send_cq->lock);
1267 		spin_unlock(&recv_cq->lock);
1268 	}
1269 }
1270 
1271 static void del_gid_entries(struct mlx4_ib_qp *qp)
1272 {
1273 	struct mlx4_ib_gid_entry *ge, *tmp;
1274 
1275 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1276 		list_del(&ge->list);
1277 		kfree(ge);
1278 	}
1279 }
1280 
1281 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
1282 {
1283 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
1284 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
1285 	else
1286 		return to_mpd(qp->ibqp.pd);
1287 }
1288 
1289 static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src,
1290 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1291 {
1292 	switch (qp->ibqp.qp_type) {
1293 	case IB_QPT_XRC_TGT:
1294 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1295 		*recv_cq = *send_cq;
1296 		break;
1297 	case IB_QPT_XRC_INI:
1298 		*send_cq = to_mcq(qp->ibqp.send_cq);
1299 		*recv_cq = *send_cq;
1300 		break;
1301 	default:
1302 		*recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) :
1303 						     to_mcq(qp->ibwq.cq);
1304 		*send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) :
1305 						     *recv_cq;
1306 		break;
1307 	}
1308 }
1309 
1310 static void destroy_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1311 {
1312 	if (qp->state != IB_QPS_RESET) {
1313 		int i;
1314 
1315 		for (i = 0; i < (1 << qp->ibqp.rwq_ind_tbl->log_ind_tbl_size);
1316 		     i++) {
1317 			struct ib_wq *ibwq = qp->ibqp.rwq_ind_tbl->ind_tbl[i];
1318 			struct mlx4_ib_qp *wq =	to_mqp((struct ib_qp *)ibwq);
1319 
1320 			mutex_lock(&wq->mutex);
1321 
1322 			wq->rss_usecnt--;
1323 
1324 			mutex_unlock(&wq->mutex);
1325 		}
1326 
1327 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1328 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1329 			pr_warn("modify QP %06x to RESET failed.\n",
1330 				qp->mqp.qpn);
1331 	}
1332 
1333 	mlx4_qp_remove(dev->dev, &qp->mqp);
1334 	mlx4_qp_free(dev->dev, &qp->mqp);
1335 	mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1336 	del_gid_entries(qp);
1337 	kfree(qp->rss_ctx);
1338 }
1339 
1340 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1341 			      enum mlx4_ib_source_type src, bool is_user)
1342 {
1343 	struct mlx4_ib_cq *send_cq, *recv_cq;
1344 	unsigned long flags;
1345 
1346 	if (qp->state != IB_QPS_RESET) {
1347 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1348 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1349 			pr_warn("modify QP %06x to RESET failed.\n",
1350 			       qp->mqp.qpn);
1351 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1352 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1353 			qp->pri.smac = 0;
1354 			qp->pri.smac_port = 0;
1355 		}
1356 		if (qp->alt.smac) {
1357 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1358 			qp->alt.smac = 0;
1359 		}
1360 		if (qp->pri.vid < 0x1000) {
1361 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1362 			qp->pri.vid = 0xFFFF;
1363 			qp->pri.candidate_vid = 0xFFFF;
1364 			qp->pri.update_vid = 0;
1365 		}
1366 		if (qp->alt.vid < 0x1000) {
1367 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1368 			qp->alt.vid = 0xFFFF;
1369 			qp->alt.candidate_vid = 0xFFFF;
1370 			qp->alt.update_vid = 0;
1371 		}
1372 	}
1373 
1374 	get_cqs(qp, src, &send_cq, &recv_cq);
1375 
1376 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1377 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1378 
1379 	/* del from lists under both locks above to protect reset flow paths */
1380 	list_del(&qp->qps_list);
1381 	list_del(&qp->cq_send_list);
1382 	list_del(&qp->cq_recv_list);
1383 	if (!is_user) {
1384 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1385 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1386 		if (send_cq != recv_cq)
1387 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1388 	}
1389 
1390 	mlx4_qp_remove(dev->dev, &qp->mqp);
1391 
1392 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1393 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1394 
1395 	mlx4_qp_free(dev->dev, &qp->mqp);
1396 
1397 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1398 		if (qp->flags & MLX4_IB_QP_NETIF)
1399 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1400 		else if (src == MLX4_IB_RWQ_SRC)
1401 			mlx4_ib_release_wqn(to_mucontext(
1402 					    qp->ibwq.uobject->context), qp, 1);
1403 		else
1404 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1405 	}
1406 
1407 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1408 
1409 	if (is_user) {
1410 		if (qp->rq.wqe_cnt) {
1411 			struct mlx4_ib_ucontext *mcontext = !src ?
1412 				to_mucontext(qp->ibqp.uobject->context) :
1413 				to_mucontext(qp->ibwq.uobject->context);
1414 			mlx4_ib_db_unmap_user(mcontext, &qp->db);
1415 		}
1416 		ib_umem_release(qp->umem);
1417 	} else {
1418 		kvfree(qp->sq.wrid);
1419 		kvfree(qp->rq.wrid);
1420 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1421 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1422 			free_proxy_bufs(&dev->ib_dev, qp);
1423 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1424 		if (qp->rq.wqe_cnt)
1425 			mlx4_db_free(dev->dev, &qp->db);
1426 	}
1427 
1428 	del_gid_entries(qp);
1429 }
1430 
1431 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1432 {
1433 	/* Native or PPF */
1434 	if (!mlx4_is_mfunc(dev->dev) ||
1435 	    (mlx4_is_master(dev->dev) &&
1436 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1437 		return  dev->dev->phys_caps.base_sqpn +
1438 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1439 			attr->port_num - 1;
1440 	}
1441 	/* PF or VF -- creating proxies */
1442 	if (attr->qp_type == IB_QPT_SMI)
1443 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp0_proxy;
1444 	else
1445 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp1_proxy;
1446 }
1447 
1448 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1449 					struct ib_qp_init_attr *init_attr,
1450 					struct ib_udata *udata)
1451 {
1452 	struct mlx4_ib_qp *qp = NULL;
1453 	int err;
1454 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1455 	u16 xrcdn = 0;
1456 
1457 	if (init_attr->rwq_ind_tbl)
1458 		return _mlx4_ib_create_qp_rss(pd, init_attr, udata);
1459 
1460 	/*
1461 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1462 	 * and only for kernel UD QPs.
1463 	 */
1464 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1465 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1466 					MLX4_IB_SRIOV_TUNNEL_QP |
1467 					MLX4_IB_SRIOV_SQP |
1468 					MLX4_IB_QP_NETIF |
1469 					MLX4_IB_QP_CREATE_ROCE_V2_GSI))
1470 		return ERR_PTR(-EINVAL);
1471 
1472 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1473 		if (init_attr->qp_type != IB_QPT_UD)
1474 			return ERR_PTR(-EINVAL);
1475 	}
1476 
1477 	if (init_attr->create_flags) {
1478 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1479 			return ERR_PTR(-EINVAL);
1480 
1481 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1482 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1483 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1484 		     init_attr->qp_type != IB_QPT_UD) ||
1485 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1486 		     init_attr->qp_type > IB_QPT_GSI) ||
1487 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1488 		     init_attr->qp_type != IB_QPT_GSI))
1489 			return ERR_PTR(-EINVAL);
1490 	}
1491 
1492 	switch (init_attr->qp_type) {
1493 	case IB_QPT_XRC_TGT:
1494 		pd = to_mxrcd(init_attr->xrcd)->pd;
1495 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1496 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1497 		/* fall through */
1498 	case IB_QPT_XRC_INI:
1499 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1500 			return ERR_PTR(-ENOSYS);
1501 		init_attr->recv_cq = init_attr->send_cq;
1502 		/* fall through */
1503 	case IB_QPT_RC:
1504 	case IB_QPT_UC:
1505 	case IB_QPT_RAW_PACKET:
1506 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1507 		if (!qp)
1508 			return ERR_PTR(-ENOMEM);
1509 		qp->pri.vid = 0xFFFF;
1510 		qp->alt.vid = 0xFFFF;
1511 		/* fall through */
1512 	case IB_QPT_UD:
1513 	{
1514 		err = create_qp_common(to_mdev(pd->device), pd,	MLX4_IB_QP_SRC,
1515 				       init_attr, udata, 0, &qp);
1516 		if (err) {
1517 			kfree(qp);
1518 			return ERR_PTR(err);
1519 		}
1520 
1521 		qp->ibqp.qp_num = qp->mqp.qpn;
1522 		qp->xrcdn = xrcdn;
1523 
1524 		break;
1525 	}
1526 	case IB_QPT_SMI:
1527 	case IB_QPT_GSI:
1528 	{
1529 		int sqpn;
1530 
1531 		/* Userspace is not allowed to create special QPs: */
1532 		if (udata)
1533 			return ERR_PTR(-EINVAL);
1534 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1535 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev,
1536 							1, 1, &sqpn, 0,
1537 							MLX4_RES_USAGE_DRIVER);
1538 
1539 			if (res)
1540 				return ERR_PTR(res);
1541 		} else {
1542 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1543 		}
1544 
1545 		err = create_qp_common(to_mdev(pd->device), pd, MLX4_IB_QP_SRC,
1546 				       init_attr, udata, sqpn, &qp);
1547 		if (err)
1548 			return ERR_PTR(err);
1549 
1550 		qp->port	= init_attr->port_num;
1551 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1552 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1553 		break;
1554 	}
1555 	default:
1556 		/* Don't support raw QPs */
1557 		return ERR_PTR(-EINVAL);
1558 	}
1559 
1560 	return &qp->ibqp;
1561 }
1562 
1563 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1564 				struct ib_qp_init_attr *init_attr,
1565 				struct ib_udata *udata) {
1566 	struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1567 	struct ib_qp *ibqp;
1568 	struct mlx4_ib_dev *dev = to_mdev(device);
1569 
1570 	ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1571 
1572 	if (!IS_ERR(ibqp) &&
1573 	    (init_attr->qp_type == IB_QPT_GSI) &&
1574 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1575 		struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1576 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1577 
1578 		if (is_eth &&
1579 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1580 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1581 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1582 
1583 			if (IS_ERR(sqp->roce_v2_gsi)) {
1584 				pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1585 				sqp->roce_v2_gsi = NULL;
1586 			} else {
1587 				sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1588 				sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1589 			}
1590 
1591 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1592 		}
1593 	}
1594 	return ibqp;
1595 }
1596 
1597 static int _mlx4_ib_destroy_qp(struct ib_qp *qp)
1598 {
1599 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1600 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1601 
1602 	if (is_qp0(dev, mqp))
1603 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1604 
1605 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1606 	    dev->qp1_proxy[mqp->port - 1] == mqp) {
1607 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1608 		dev->qp1_proxy[mqp->port - 1] = NULL;
1609 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1610 	}
1611 
1612 	if (mqp->counter_index)
1613 		mlx4_ib_free_qp_counter(dev, mqp);
1614 
1615 	if (qp->rwq_ind_tbl) {
1616 		destroy_qp_rss(dev, mqp);
1617 	} else {
1618 		destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, qp->uobject);
1619 	}
1620 
1621 	if (is_sqp(dev, mqp))
1622 		kfree(to_msqp(mqp));
1623 	else
1624 		kfree(mqp);
1625 
1626 	return 0;
1627 }
1628 
1629 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1630 {
1631 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1632 
1633 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1634 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1635 
1636 		if (sqp->roce_v2_gsi)
1637 			ib_destroy_qp(sqp->roce_v2_gsi);
1638 	}
1639 
1640 	return _mlx4_ib_destroy_qp(qp);
1641 }
1642 
1643 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1644 {
1645 	switch (type) {
1646 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1647 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1648 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1649 	case MLX4_IB_QPT_XRC_INI:
1650 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1651 	case MLX4_IB_QPT_SMI:
1652 	case MLX4_IB_QPT_GSI:
1653 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1654 
1655 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1656 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1657 						MLX4_QP_ST_MLX : -1);
1658 	case MLX4_IB_QPT_PROXY_SMI:
1659 	case MLX4_IB_QPT_TUN_SMI:
1660 	case MLX4_IB_QPT_PROXY_GSI:
1661 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1662 						MLX4_QP_ST_UD : -1);
1663 	default:			return -1;
1664 	}
1665 }
1666 
1667 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1668 				   int attr_mask)
1669 {
1670 	u8 dest_rd_atomic;
1671 	u32 access_flags;
1672 	u32 hw_access_flags = 0;
1673 
1674 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1675 		dest_rd_atomic = attr->max_dest_rd_atomic;
1676 	else
1677 		dest_rd_atomic = qp->resp_depth;
1678 
1679 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1680 		access_flags = attr->qp_access_flags;
1681 	else
1682 		access_flags = qp->atomic_rd_en;
1683 
1684 	if (!dest_rd_atomic)
1685 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1686 
1687 	if (access_flags & IB_ACCESS_REMOTE_READ)
1688 		hw_access_flags |= MLX4_QP_BIT_RRE;
1689 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1690 		hw_access_flags |= MLX4_QP_BIT_RAE;
1691 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1692 		hw_access_flags |= MLX4_QP_BIT_RWE;
1693 
1694 	return cpu_to_be32(hw_access_flags);
1695 }
1696 
1697 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1698 			    int attr_mask)
1699 {
1700 	if (attr_mask & IB_QP_PKEY_INDEX)
1701 		sqp->pkey_index = attr->pkey_index;
1702 	if (attr_mask & IB_QP_QKEY)
1703 		sqp->qkey = attr->qkey;
1704 	if (attr_mask & IB_QP_SQ_PSN)
1705 		sqp->send_psn = attr->sq_psn;
1706 }
1707 
1708 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1709 {
1710 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1711 }
1712 
1713 static int _mlx4_set_path(struct mlx4_ib_dev *dev,
1714 			  const struct rdma_ah_attr *ah,
1715 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1716 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1717 {
1718 	int vidx;
1719 	int smac_index;
1720 	int err;
1721 
1722 	path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
1723 	path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
1724 	if (rdma_ah_get_static_rate(ah)) {
1725 		path->static_rate = rdma_ah_get_static_rate(ah) +
1726 				    MLX4_STAT_RATE_OFFSET;
1727 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1728 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1729 			--path->static_rate;
1730 	} else
1731 		path->static_rate = 0;
1732 
1733 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
1734 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
1735 		int real_sgid_index =
1736 			mlx4_ib_gid_index_to_real_index(dev, grh->sgid_attr);
1737 
1738 		if (real_sgid_index < 0)
1739 			return real_sgid_index;
1740 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1741 			pr_err("sgid_index (%u) too large. max is %d\n",
1742 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1743 			return -1;
1744 		}
1745 
1746 		path->grh_mylmc |= 1 << 7;
1747 		path->mgid_index = real_sgid_index;
1748 		path->hop_limit  = grh->hop_limit;
1749 		path->tclass_flowlabel =
1750 			cpu_to_be32((grh->traffic_class << 20) |
1751 				    (grh->flow_label));
1752 		memcpy(path->rgid, grh->dgid.raw, 16);
1753 	}
1754 
1755 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
1756 		if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
1757 			return -1;
1758 
1759 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1760 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
1761 
1762 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1763 		if (vlan_tag < 0x1000) {
1764 			if (smac_info->vid < 0x1000) {
1765 				/* both valid vlan ids */
1766 				if (smac_info->vid != vlan_tag) {
1767 					/* different VIDs.  unreg old and reg new */
1768 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1769 					if (err)
1770 						return err;
1771 					smac_info->candidate_vid = vlan_tag;
1772 					smac_info->candidate_vlan_index = vidx;
1773 					smac_info->candidate_vlan_port = port;
1774 					smac_info->update_vid = 1;
1775 					path->vlan_index = vidx;
1776 				} else {
1777 					path->vlan_index = smac_info->vlan_index;
1778 				}
1779 			} else {
1780 				/* no current vlan tag in qp */
1781 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1782 				if (err)
1783 					return err;
1784 				smac_info->candidate_vid = vlan_tag;
1785 				smac_info->candidate_vlan_index = vidx;
1786 				smac_info->candidate_vlan_port = port;
1787 				smac_info->update_vid = 1;
1788 				path->vlan_index = vidx;
1789 			}
1790 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1791 			path->fl = 1 << 6;
1792 		} else {
1793 			/* have current vlan tag. unregister it at modify-qp success */
1794 			if (smac_info->vid < 0x1000) {
1795 				smac_info->candidate_vid = 0xFFFF;
1796 				smac_info->update_vid = 1;
1797 			}
1798 		}
1799 
1800 		/* get smac_index for RoCE use.
1801 		 * If no smac was yet assigned, register one.
1802 		 * If one was already assigned, but the new mac differs,
1803 		 * unregister the old one and register the new one.
1804 		*/
1805 		if ((!smac_info->smac && !smac_info->smac_port) ||
1806 		    smac_info->smac != smac) {
1807 			/* register candidate now, unreg if needed, after success */
1808 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1809 			if (smac_index >= 0) {
1810 				smac_info->candidate_smac_index = smac_index;
1811 				smac_info->candidate_smac = smac;
1812 				smac_info->candidate_smac_port = port;
1813 			} else {
1814 				return -EINVAL;
1815 			}
1816 		} else {
1817 			smac_index = smac_info->smac_index;
1818 		}
1819 		memcpy(path->dmac, ah->roce.dmac, 6);
1820 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1821 		/* put MAC table smac index for IBoE */
1822 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1823 	} else {
1824 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1825 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
1826 	}
1827 
1828 	return 0;
1829 }
1830 
1831 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1832 			 enum ib_qp_attr_mask qp_attr_mask,
1833 			 struct mlx4_ib_qp *mqp,
1834 			 struct mlx4_qp_path *path, u8 port,
1835 			 u16 vlan_id, u8 *smac)
1836 {
1837 	return _mlx4_set_path(dev, &qp->ah_attr,
1838 			      mlx4_mac_to_u64(smac),
1839 			      vlan_id,
1840 			      path, &mqp->pri, port);
1841 }
1842 
1843 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1844 			     const struct ib_qp_attr *qp,
1845 			     enum ib_qp_attr_mask qp_attr_mask,
1846 			     struct mlx4_ib_qp *mqp,
1847 			     struct mlx4_qp_path *path, u8 port)
1848 {
1849 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1850 			      0,
1851 			      0xffff,
1852 			      path, &mqp->alt, port);
1853 }
1854 
1855 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1856 {
1857 	struct mlx4_ib_gid_entry *ge, *tmp;
1858 
1859 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1860 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1861 			ge->added = 1;
1862 			ge->port = qp->port;
1863 		}
1864 	}
1865 }
1866 
1867 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1868 				    struct mlx4_ib_qp *qp,
1869 				    struct mlx4_qp_context *context)
1870 {
1871 	u64 u64_mac;
1872 	int smac_index;
1873 
1874 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1875 
1876 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1877 	if (!qp->pri.smac && !qp->pri.smac_port) {
1878 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1879 		if (smac_index >= 0) {
1880 			qp->pri.candidate_smac_index = smac_index;
1881 			qp->pri.candidate_smac = u64_mac;
1882 			qp->pri.candidate_smac_port = qp->port;
1883 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1884 		} else {
1885 			return -ENOENT;
1886 		}
1887 	}
1888 	return 0;
1889 }
1890 
1891 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1892 {
1893 	struct counter_index *new_counter_index;
1894 	int err;
1895 	u32 tmp_idx;
1896 
1897 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1898 	    IB_LINK_LAYER_ETHERNET ||
1899 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1900 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1901 		return 0;
1902 
1903 	err = mlx4_counter_alloc(dev->dev, &tmp_idx, MLX4_RES_USAGE_DRIVER);
1904 	if (err)
1905 		return err;
1906 
1907 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1908 	if (!new_counter_index) {
1909 		mlx4_counter_free(dev->dev, tmp_idx);
1910 		return -ENOMEM;
1911 	}
1912 
1913 	new_counter_index->index = tmp_idx;
1914 	new_counter_index->allocated = 1;
1915 	qp->counter_index = new_counter_index;
1916 
1917 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
1918 	list_add_tail(&new_counter_index->list,
1919 		      &dev->counters_table[qp->port - 1].counters_list);
1920 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
1921 
1922 	return 0;
1923 }
1924 
1925 enum {
1926 	MLX4_QPC_ROCE_MODE_1 = 0,
1927 	MLX4_QPC_ROCE_MODE_2 = 2,
1928 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
1929 };
1930 
1931 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
1932 {
1933 	switch (gid_type) {
1934 	case IB_GID_TYPE_ROCE:
1935 		return MLX4_QPC_ROCE_MODE_1;
1936 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
1937 		return MLX4_QPC_ROCE_MODE_2;
1938 	default:
1939 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
1940 	}
1941 }
1942 
1943 /*
1944  * Go over all RSS QP's childes (WQs) and apply their HW state according to
1945  * their logic state if the RSS QP is the first RSS QP associated for the WQ.
1946  */
1947 static int bringup_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, u8 port_num,
1948 			    struct ib_udata *udata)
1949 {
1950 	int err = 0;
1951 	int i;
1952 
1953 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
1954 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
1955 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
1956 
1957 		mutex_lock(&wq->mutex);
1958 
1959 		/* Mlx4_ib restrictions:
1960 		 * WQ's is associated to a port according to the RSS QP it is
1961 		 * associates to.
1962 		 * In case the WQ is associated to a different port by another
1963 		 * RSS QP, return a failure.
1964 		 */
1965 		if ((wq->rss_usecnt > 0) && (wq->port != port_num)) {
1966 			err = -EINVAL;
1967 			mutex_unlock(&wq->mutex);
1968 			break;
1969 		}
1970 		wq->port = port_num;
1971 		if ((wq->rss_usecnt == 0) && (ibwq->state == IB_WQS_RDY)) {
1972 			err = _mlx4_ib_modify_wq(ibwq, IB_WQS_RDY, udata);
1973 			if (err) {
1974 				mutex_unlock(&wq->mutex);
1975 				break;
1976 			}
1977 		}
1978 		wq->rss_usecnt++;
1979 
1980 		mutex_unlock(&wq->mutex);
1981 	}
1982 
1983 	if (i && err) {
1984 		int j;
1985 
1986 		for (j = (i - 1); j >= 0; j--) {
1987 			struct ib_wq *ibwq = ind_tbl->ind_tbl[j];
1988 			struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
1989 
1990 			mutex_lock(&wq->mutex);
1991 
1992 			if ((wq->rss_usecnt == 1) &&
1993 			    (ibwq->state == IB_WQS_RDY))
1994 				if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET,
1995 						       udata))
1996 					pr_warn("failed to reverse WQN=0x%06x\n",
1997 						ibwq->wq_num);
1998 			wq->rss_usecnt--;
1999 
2000 			mutex_unlock(&wq->mutex);
2001 		}
2002 	}
2003 
2004 	return err;
2005 }
2006 
2007 static void bring_down_rss_rwqs(struct ib_rwq_ind_table *ind_tbl,
2008 				struct ib_udata *udata)
2009 {
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 		if ((wq->rss_usecnt == 1) && (ibwq->state == IB_WQS_RDY))
2019 			if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET, udata))
2020 				pr_warn("failed to reverse WQN=%x\n",
2021 					ibwq->wq_num);
2022 		wq->rss_usecnt--;
2023 
2024 		mutex_unlock(&wq->mutex);
2025 	}
2026 }
2027 
2028 static void fill_qp_rss_context(struct mlx4_qp_context *context,
2029 				struct mlx4_ib_qp *qp)
2030 {
2031 	struct mlx4_rss_context *rss_context;
2032 
2033 	rss_context = (void *)context + offsetof(struct mlx4_qp_context,
2034 			pri_path) + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
2035 
2036 	rss_context->base_qpn = cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz);
2037 	rss_context->default_qpn =
2038 		cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz & 0xffffff);
2039 	if (qp->rss_ctx->flags & (MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6))
2040 		rss_context->base_qpn_udp = rss_context->default_qpn;
2041 	rss_context->flags = qp->rss_ctx->flags;
2042 	/* Currently support just toeplitz */
2043 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
2044 
2045 	memcpy(rss_context->rss_key, qp->rss_ctx->rss_key,
2046 	       MLX4_EN_RSS_KEY_SIZE);
2047 }
2048 
2049 static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type,
2050 			       const struct ib_qp_attr *attr, int attr_mask,
2051 			       enum ib_qp_state cur_state,
2052 			       enum ib_qp_state new_state,
2053 			       struct ib_udata *udata)
2054 {
2055 	struct ib_srq  *ibsrq;
2056 	const struct ib_gid_attr *gid_attr = NULL;
2057 	struct ib_rwq_ind_table *rwq_ind_tbl;
2058 	enum ib_qp_type qp_type;
2059 	struct mlx4_ib_dev *dev;
2060 	struct mlx4_ib_qp *qp;
2061 	struct mlx4_ib_pd *pd;
2062 	struct mlx4_ib_cq *send_cq, *recv_cq;
2063 	struct mlx4_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2064 		udata, struct mlx4_ib_ucontext, ibucontext);
2065 	struct mlx4_qp_context *context;
2066 	enum mlx4_qp_optpar optpar = 0;
2067 	int sqd_event;
2068 	int steer_qp = 0;
2069 	int err = -EINVAL;
2070 	int counter_index;
2071 
2072 	if (src_type == MLX4_IB_RWQ_SRC) {
2073 		struct ib_wq *ibwq;
2074 
2075 		ibwq	    = (struct ib_wq *)src;
2076 		ibsrq	    = NULL;
2077 		rwq_ind_tbl = NULL;
2078 		qp_type     = IB_QPT_RAW_PACKET;
2079 		qp	    = to_mqp((struct ib_qp *)ibwq);
2080 		dev	    = to_mdev(ibwq->device);
2081 		pd	    = to_mpd(ibwq->pd);
2082 	} else {
2083 		struct ib_qp *ibqp;
2084 
2085 		ibqp	    = (struct ib_qp *)src;
2086 		ibsrq	    = ibqp->srq;
2087 		rwq_ind_tbl = ibqp->rwq_ind_tbl;
2088 		qp_type     = ibqp->qp_type;
2089 		qp	    = to_mqp(ibqp);
2090 		dev	    = to_mdev(ibqp->device);
2091 		pd	    = get_pd(qp);
2092 	}
2093 
2094 	/* APM is not supported under RoCE */
2095 	if (attr_mask & IB_QP_ALT_PATH &&
2096 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2097 	    IB_LINK_LAYER_ETHERNET)
2098 		return -ENOTSUPP;
2099 
2100 	context = kzalloc(sizeof *context, GFP_KERNEL);
2101 	if (!context)
2102 		return -ENOMEM;
2103 
2104 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
2105 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
2106 
2107 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
2108 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2109 	else {
2110 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
2111 		switch (attr->path_mig_state) {
2112 		case IB_MIG_MIGRATED:
2113 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2114 			break;
2115 		case IB_MIG_REARM:
2116 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
2117 			break;
2118 		case IB_MIG_ARMED:
2119 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
2120 			break;
2121 		}
2122 	}
2123 
2124 	if (qp->inl_recv_sz)
2125 		context->param3 |= cpu_to_be32(1 << 25);
2126 
2127 	if (qp->flags & MLX4_IB_QP_SCATTER_FCS)
2128 		context->param3 |= cpu_to_be32(1 << 29);
2129 
2130 	if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI)
2131 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
2132 	else if (qp_type == IB_QPT_RAW_PACKET)
2133 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
2134 	else if (qp_type == IB_QPT_UD) {
2135 		if (qp->flags & MLX4_IB_QP_LSO)
2136 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
2137 					      ilog2(dev->dev->caps.max_gso_sz);
2138 		else
2139 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 13;
2140 	} else if (attr_mask & IB_QP_PATH_MTU) {
2141 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
2142 			pr_err("path MTU (%u) is invalid\n",
2143 			       attr->path_mtu);
2144 			goto out;
2145 		}
2146 		context->mtu_msgmax = (attr->path_mtu << 5) |
2147 			ilog2(dev->dev->caps.max_msg_sz);
2148 	}
2149 
2150 	if (!rwq_ind_tbl) { /* PRM RSS receive side should be left zeros */
2151 		if (qp->rq.wqe_cnt)
2152 			context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
2153 		context->rq_size_stride |= qp->rq.wqe_shift - 4;
2154 	}
2155 
2156 	if (qp->sq.wqe_cnt)
2157 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
2158 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
2159 
2160 	if (new_state == IB_QPS_RESET && qp->counter_index)
2161 		mlx4_ib_free_qp_counter(dev, qp);
2162 
2163 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2164 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
2165 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
2166 		if (qp_type == IB_QPT_RAW_PACKET)
2167 			context->param3 |= cpu_to_be32(1 << 30);
2168 	}
2169 
2170 	if (ucontext)
2171 		context->usr_page = cpu_to_be32(
2172 			mlx4_to_hw_uar_index(dev->dev, ucontext->uar.index));
2173 	else
2174 		context->usr_page = cpu_to_be32(
2175 			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
2176 
2177 	if (attr_mask & IB_QP_DEST_QPN)
2178 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
2179 
2180 	if (attr_mask & IB_QP_PORT) {
2181 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
2182 		    !(attr_mask & IB_QP_AV)) {
2183 			mlx4_set_sched(&context->pri_path, attr->port_num);
2184 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
2185 		}
2186 	}
2187 
2188 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
2189 		err = create_qp_lb_counter(dev, qp);
2190 		if (err)
2191 			goto out;
2192 
2193 		counter_index =
2194 			dev->counters_table[qp->port - 1].default_counter;
2195 		if (qp->counter_index)
2196 			counter_index = qp->counter_index->index;
2197 
2198 		if (counter_index != -1) {
2199 			context->pri_path.counter_index = counter_index;
2200 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
2201 			if (qp->counter_index) {
2202 				context->pri_path.fl |=
2203 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
2204 				context->pri_path.vlan_control |=
2205 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
2206 			}
2207 		} else
2208 			context->pri_path.counter_index =
2209 				MLX4_SINK_COUNTER_INDEX(dev->dev);
2210 
2211 		if (qp->flags & MLX4_IB_QP_NETIF) {
2212 			mlx4_ib_steer_qp_reg(dev, qp, 1);
2213 			steer_qp = 1;
2214 		}
2215 
2216 		if (qp_type == IB_QPT_GSI) {
2217 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
2218 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
2219 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
2220 
2221 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2222 		}
2223 	}
2224 
2225 	if (attr_mask & IB_QP_PKEY_INDEX) {
2226 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2227 			context->pri_path.disable_pkey_check = 0x40;
2228 		context->pri_path.pkey_index = attr->pkey_index;
2229 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
2230 	}
2231 
2232 	if (attr_mask & IB_QP_AV) {
2233 		u8 port_num = mlx4_is_bonded(dev->dev) ? 1 :
2234 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2235 		u16 vlan = 0xffff;
2236 		u8 smac[ETH_ALEN];
2237 		int is_eth =
2238 			rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
2239 			rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
2240 
2241 		if (is_eth) {
2242 			gid_attr = attr->ah_attr.grh.sgid_attr;
2243 			vlan = rdma_vlan_dev_vlan_id(gid_attr->ndev);
2244 			memcpy(smac, gid_attr->ndev->dev_addr, ETH_ALEN);
2245 		}
2246 
2247 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
2248 				  port_num, vlan, smac))
2249 			goto out;
2250 
2251 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
2252 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
2253 
2254 		if (is_eth &&
2255 		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
2256 			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr->gid_type);
2257 
2258 			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
2259 				err = -EINVAL;
2260 				goto out;
2261 			}
2262 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2263 		}
2264 
2265 	}
2266 
2267 	if (attr_mask & IB_QP_TIMEOUT) {
2268 		context->pri_path.ackto |= attr->timeout << 3;
2269 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
2270 	}
2271 
2272 	if (attr_mask & IB_QP_ALT_PATH) {
2273 		if (attr->alt_port_num == 0 ||
2274 		    attr->alt_port_num > dev->dev->caps.num_ports)
2275 			goto out;
2276 
2277 		if (attr->alt_pkey_index >=
2278 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
2279 			goto out;
2280 
2281 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
2282 				      &context->alt_path,
2283 				      attr->alt_port_num))
2284 			goto out;
2285 
2286 		context->alt_path.pkey_index = attr->alt_pkey_index;
2287 		context->alt_path.ackto = attr->alt_timeout << 3;
2288 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
2289 	}
2290 
2291 	context->pd = cpu_to_be32(pd->pdn);
2292 
2293 	if (!rwq_ind_tbl) {
2294 		context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
2295 		get_cqs(qp, src_type, &send_cq, &recv_cq);
2296 	} else { /* Set dummy CQs to be compatible with HV and PRM */
2297 		send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq);
2298 		recv_cq = send_cq;
2299 	}
2300 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
2301 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
2302 
2303 	/* Set "fast registration enabled" for all kernel QPs */
2304 	if (!ucontext)
2305 		context->params1 |= cpu_to_be32(1 << 11);
2306 
2307 	if (attr_mask & IB_QP_RNR_RETRY) {
2308 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2309 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
2310 	}
2311 
2312 	if (attr_mask & IB_QP_RETRY_CNT) {
2313 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2314 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
2315 	}
2316 
2317 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2318 		if (attr->max_rd_atomic)
2319 			context->params1 |=
2320 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2321 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
2322 	}
2323 
2324 	if (attr_mask & IB_QP_SQ_PSN)
2325 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2326 
2327 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2328 		if (attr->max_dest_rd_atomic)
2329 			context->params2 |=
2330 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2331 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
2332 	}
2333 
2334 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
2335 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
2336 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
2337 	}
2338 
2339 	if (ibsrq)
2340 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
2341 
2342 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
2343 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2344 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
2345 	}
2346 	if (attr_mask & IB_QP_RQ_PSN)
2347 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2348 
2349 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
2350 	if (attr_mask & IB_QP_QKEY) {
2351 		if (qp->mlx4_ib_qp_type &
2352 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
2353 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2354 		else {
2355 			if (mlx4_is_mfunc(dev->dev) &&
2356 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
2357 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
2358 			    MLX4_RESERVED_QKEY_BASE) {
2359 				pr_err("Cannot use reserved QKEY"
2360 				       " 0x%x (range 0xffff0000..0xffffffff"
2361 				       " is reserved)\n", attr->qkey);
2362 				err = -EINVAL;
2363 				goto out;
2364 			}
2365 			context->qkey = cpu_to_be32(attr->qkey);
2366 		}
2367 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
2368 	}
2369 
2370 	if (ibsrq)
2371 		context->srqn = cpu_to_be32(1 << 24 |
2372 					    to_msrq(ibsrq)->msrq.srqn);
2373 
2374 	if (qp->rq.wqe_cnt &&
2375 	    cur_state == IB_QPS_RESET &&
2376 	    new_state == IB_QPS_INIT)
2377 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2378 
2379 	if (cur_state == IB_QPS_INIT &&
2380 	    new_state == IB_QPS_RTR  &&
2381 	    (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI ||
2382 	     qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) {
2383 		context->pri_path.sched_queue = (qp->port - 1) << 6;
2384 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2385 		    qp->mlx4_ib_qp_type &
2386 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
2387 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
2388 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
2389 				context->pri_path.fl = 0x80;
2390 		} else {
2391 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2392 				context->pri_path.fl = 0x80;
2393 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
2394 		}
2395 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2396 		    IB_LINK_LAYER_ETHERNET) {
2397 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
2398 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
2399 				context->pri_path.feup = 1 << 7; /* don't fsm */
2400 			/* handle smac_index */
2401 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
2402 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
2403 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
2404 				err = handle_eth_ud_smac_index(dev, qp, context);
2405 				if (err) {
2406 					err = -EINVAL;
2407 					goto out;
2408 				}
2409 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
2410 					dev->qp1_proxy[qp->port - 1] = qp;
2411 			}
2412 		}
2413 	}
2414 
2415 	if (qp_type == IB_QPT_RAW_PACKET) {
2416 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
2417 					MLX4_IB_LINK_TYPE_ETH;
2418 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
2419 			/* set QP to receive both tunneled & non-tunneled packets */
2420 			if (!rwq_ind_tbl)
2421 				context->srqn = cpu_to_be32(7 << 28);
2422 		}
2423 	}
2424 
2425 	if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
2426 		int is_eth = rdma_port_get_link_layer(
2427 				&dev->ib_dev, qp->port) ==
2428 				IB_LINK_LAYER_ETHERNET;
2429 		if (is_eth) {
2430 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
2431 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
2432 		}
2433 	}
2434 
2435 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2436 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2437 		sqd_event = 1;
2438 	else
2439 		sqd_event = 0;
2440 
2441 	if (!ucontext &&
2442 	    cur_state == IB_QPS_RESET &&
2443 	    new_state == IB_QPS_INIT)
2444 		context->rlkey_roce_mode |= (1 << 4);
2445 
2446 	/*
2447 	 * Before passing a kernel QP to the HW, make sure that the
2448 	 * ownership bits of the send queue are set and the SQ
2449 	 * headroom is stamped so that the hardware doesn't start
2450 	 * processing stale work requests.
2451 	 */
2452 	if (!ucontext &&
2453 	    cur_state == IB_QPS_RESET &&
2454 	    new_state == IB_QPS_INIT) {
2455 		struct mlx4_wqe_ctrl_seg *ctrl;
2456 		int i;
2457 
2458 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
2459 			ctrl = get_send_wqe(qp, i);
2460 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
2461 			ctrl->qpn_vlan.fence_size =
2462 				1 << (qp->sq.wqe_shift - 4);
2463 			stamp_send_wqe(qp, i);
2464 		}
2465 	}
2466 
2467 	if (rwq_ind_tbl	&&
2468 	    cur_state == IB_QPS_RESET &&
2469 	    new_state == IB_QPS_INIT) {
2470 		fill_qp_rss_context(context, qp);
2471 		context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET);
2472 	}
2473 
2474 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2475 			     to_mlx4_state(new_state), context, optpar,
2476 			     sqd_event, &qp->mqp);
2477 	if (err)
2478 		goto out;
2479 
2480 	qp->state = new_state;
2481 
2482 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2483 		qp->atomic_rd_en = attr->qp_access_flags;
2484 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2485 		qp->resp_depth = attr->max_dest_rd_atomic;
2486 	if (attr_mask & IB_QP_PORT) {
2487 		qp->port = attr->port_num;
2488 		update_mcg_macs(dev, qp);
2489 	}
2490 	if (attr_mask & IB_QP_ALT_PATH)
2491 		qp->alt_port = attr->alt_port_num;
2492 
2493 	if (is_sqp(dev, qp))
2494 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2495 
2496 	/*
2497 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
2498 	 * QP0 to RESET or ERROR, bring the link back down.
2499 	 */
2500 	if (is_qp0(dev, qp)) {
2501 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2502 			if (mlx4_INIT_PORT(dev->dev, qp->port))
2503 				pr_warn("INIT_PORT failed for port %d\n",
2504 				       qp->port);
2505 
2506 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2507 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2508 			mlx4_CLOSE_PORT(dev->dev, qp->port);
2509 	}
2510 
2511 	/*
2512 	 * If we moved a kernel QP to RESET, clean up all old CQ
2513 	 * entries and reinitialize the QP.
2514 	 */
2515 	if (new_state == IB_QPS_RESET) {
2516 		if (!ucontext) {
2517 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2518 					 ibsrq ? to_msrq(ibsrq) : NULL);
2519 			if (send_cq != recv_cq)
2520 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2521 
2522 			qp->rq.head = 0;
2523 			qp->rq.tail = 0;
2524 			qp->sq.head = 0;
2525 			qp->sq.tail = 0;
2526 			qp->sq_next_wqe = 0;
2527 			if (qp->rq.wqe_cnt)
2528 				*qp->db.db  = 0;
2529 
2530 			if (qp->flags & MLX4_IB_QP_NETIF)
2531 				mlx4_ib_steer_qp_reg(dev, qp, 0);
2532 		}
2533 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2534 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2535 			qp->pri.smac = 0;
2536 			qp->pri.smac_port = 0;
2537 		}
2538 		if (qp->alt.smac) {
2539 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2540 			qp->alt.smac = 0;
2541 		}
2542 		if (qp->pri.vid < 0x1000) {
2543 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2544 			qp->pri.vid = 0xFFFF;
2545 			qp->pri.candidate_vid = 0xFFFF;
2546 			qp->pri.update_vid = 0;
2547 		}
2548 
2549 		if (qp->alt.vid < 0x1000) {
2550 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2551 			qp->alt.vid = 0xFFFF;
2552 			qp->alt.candidate_vid = 0xFFFF;
2553 			qp->alt.update_vid = 0;
2554 		}
2555 	}
2556 out:
2557 	if (err && qp->counter_index)
2558 		mlx4_ib_free_qp_counter(dev, qp);
2559 	if (err && steer_qp)
2560 		mlx4_ib_steer_qp_reg(dev, qp, 0);
2561 	kfree(context);
2562 	if (qp->pri.candidate_smac ||
2563 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2564 		if (err) {
2565 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2566 		} else {
2567 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2568 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2569 			qp->pri.smac = qp->pri.candidate_smac;
2570 			qp->pri.smac_index = qp->pri.candidate_smac_index;
2571 			qp->pri.smac_port = qp->pri.candidate_smac_port;
2572 		}
2573 		qp->pri.candidate_smac = 0;
2574 		qp->pri.candidate_smac_index = 0;
2575 		qp->pri.candidate_smac_port = 0;
2576 	}
2577 	if (qp->alt.candidate_smac) {
2578 		if (err) {
2579 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2580 		} else {
2581 			if (qp->alt.smac)
2582 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2583 			qp->alt.smac = qp->alt.candidate_smac;
2584 			qp->alt.smac_index = qp->alt.candidate_smac_index;
2585 			qp->alt.smac_port = qp->alt.candidate_smac_port;
2586 		}
2587 		qp->alt.candidate_smac = 0;
2588 		qp->alt.candidate_smac_index = 0;
2589 		qp->alt.candidate_smac_port = 0;
2590 	}
2591 
2592 	if (qp->pri.update_vid) {
2593 		if (err) {
2594 			if (qp->pri.candidate_vid < 0x1000)
2595 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2596 						     qp->pri.candidate_vid);
2597 		} else {
2598 			if (qp->pri.vid < 0x1000)
2599 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2600 						     qp->pri.vid);
2601 			qp->pri.vid = qp->pri.candidate_vid;
2602 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2603 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
2604 		}
2605 		qp->pri.candidate_vid = 0xFFFF;
2606 		qp->pri.update_vid = 0;
2607 	}
2608 
2609 	if (qp->alt.update_vid) {
2610 		if (err) {
2611 			if (qp->alt.candidate_vid < 0x1000)
2612 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2613 						     qp->alt.candidate_vid);
2614 		} else {
2615 			if (qp->alt.vid < 0x1000)
2616 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2617 						     qp->alt.vid);
2618 			qp->alt.vid = qp->alt.candidate_vid;
2619 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2620 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2621 		}
2622 		qp->alt.candidate_vid = 0xFFFF;
2623 		qp->alt.update_vid = 0;
2624 	}
2625 
2626 	return err;
2627 }
2628 
2629 enum {
2630 	MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK = (IB_QP_STATE	|
2631 					      IB_QP_PORT),
2632 };
2633 
2634 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2635 			      int attr_mask, struct ib_udata *udata)
2636 {
2637 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2638 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2639 	enum ib_qp_state cur_state, new_state;
2640 	int err = -EINVAL;
2641 	mutex_lock(&qp->mutex);
2642 
2643 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2644 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2645 
2646 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2647 				attr_mask)) {
2648 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2649 			 "for transition %d to %d. qp_type %d,"
2650 			 " attr_mask 0x%x\n",
2651 			 ibqp->qp_num, cur_state, new_state,
2652 			 ibqp->qp_type, attr_mask);
2653 		goto out;
2654 	}
2655 
2656 	if (ibqp->rwq_ind_tbl) {
2657 		if (!(((cur_state == IB_QPS_RESET) &&
2658 		       (new_state == IB_QPS_INIT)) ||
2659 		      ((cur_state == IB_QPS_INIT)  &&
2660 		       (new_state == IB_QPS_RTR)))) {
2661 			pr_debug("qpn 0x%x: RSS QP unsupported transition %d to %d\n",
2662 				 ibqp->qp_num, cur_state, new_state);
2663 
2664 			err = -EOPNOTSUPP;
2665 			goto out;
2666 		}
2667 
2668 		if (attr_mask & ~MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK) {
2669 			pr_debug("qpn 0x%x: RSS QP unsupported attribute mask 0x%x for transition %d to %d\n",
2670 				 ibqp->qp_num, attr_mask, cur_state, new_state);
2671 
2672 			err = -EOPNOTSUPP;
2673 			goto out;
2674 		}
2675 	}
2676 
2677 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2678 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2679 			if ((ibqp->qp_type == IB_QPT_RC) ||
2680 			    (ibqp->qp_type == IB_QPT_UD) ||
2681 			    (ibqp->qp_type == IB_QPT_UC) ||
2682 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2683 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2684 				attr->port_num = mlx4_ib_bond_next_port(dev);
2685 			}
2686 		} else {
2687 			/* no sense in changing port_num
2688 			 * when ports are bonded */
2689 			attr_mask &= ~IB_QP_PORT;
2690 		}
2691 	}
2692 
2693 	if ((attr_mask & IB_QP_PORT) &&
2694 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2695 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2696 			 "for transition %d to %d. qp_type %d\n",
2697 			 ibqp->qp_num, attr->port_num, cur_state,
2698 			 new_state, ibqp->qp_type);
2699 		goto out;
2700 	}
2701 
2702 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2703 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2704 	     IB_LINK_LAYER_ETHERNET))
2705 		goto out;
2706 
2707 	if (attr_mask & IB_QP_PKEY_INDEX) {
2708 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2709 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2710 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2711 				 "for transition %d to %d. qp_type %d\n",
2712 				 ibqp->qp_num, attr->pkey_index, cur_state,
2713 				 new_state, ibqp->qp_type);
2714 			goto out;
2715 		}
2716 	}
2717 
2718 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2719 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2720 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2721 			 "Transition %d to %d. qp_type %d\n",
2722 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2723 			 new_state, ibqp->qp_type);
2724 		goto out;
2725 	}
2726 
2727 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2728 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2729 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2730 			 "Transition %d to %d. qp_type %d\n",
2731 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2732 			 new_state, ibqp->qp_type);
2733 		goto out;
2734 	}
2735 
2736 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2737 		err = 0;
2738 		goto out;
2739 	}
2740 
2741 	if (ibqp->rwq_ind_tbl && (new_state == IB_QPS_INIT)) {
2742 		err = bringup_rss_rwqs(ibqp->rwq_ind_tbl, attr->port_num,
2743 				       udata);
2744 		if (err)
2745 			goto out;
2746 	}
2747 
2748 	err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask,
2749 				  cur_state, new_state, udata);
2750 
2751 	if (ibqp->rwq_ind_tbl && err)
2752 		bring_down_rss_rwqs(ibqp->rwq_ind_tbl, udata);
2753 
2754 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2755 		attr->port_num = 1;
2756 
2757 out:
2758 	mutex_unlock(&qp->mutex);
2759 	return err;
2760 }
2761 
2762 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2763 		      int attr_mask, struct ib_udata *udata)
2764 {
2765 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2766 	int ret;
2767 
2768 	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2769 
2770 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2771 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
2772 		int err = 0;
2773 
2774 		if (sqp->roce_v2_gsi)
2775 			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2776 		if (err)
2777 			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2778 			       err);
2779 	}
2780 	return ret;
2781 }
2782 
2783 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2784 {
2785 	int i;
2786 	for (i = 0; i < dev->caps.num_ports; i++) {
2787 		if (qpn == dev->caps.spec_qps[i].qp0_proxy ||
2788 		    qpn == dev->caps.spec_qps[i].qp0_tunnel) {
2789 			*qkey = dev->caps.spec_qps[i].qp0_qkey;
2790 			return 0;
2791 		}
2792 	}
2793 	return -EINVAL;
2794 }
2795 
2796 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2797 				  const struct ib_ud_wr *wr,
2798 				  void *wqe, unsigned *mlx_seg_len)
2799 {
2800 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2801 	struct ib_device *ib_dev = &mdev->ib_dev;
2802 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2803 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2804 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2805 	u16 pkey;
2806 	u32 qkey;
2807 	int send_size;
2808 	int header_size;
2809 	int spc;
2810 	int i;
2811 
2812 	if (wr->wr.opcode != IB_WR_SEND)
2813 		return -EINVAL;
2814 
2815 	send_size = 0;
2816 
2817 	for (i = 0; i < wr->wr.num_sge; ++i)
2818 		send_size += wr->wr.sg_list[i].length;
2819 
2820 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2821 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2822 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2823 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2824 
2825 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2826 
2827 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2828 		sqp->ud_header.lrh.service_level =
2829 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2830 		sqp->ud_header.lrh.destination_lid =
2831 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2832 		sqp->ud_header.lrh.source_lid =
2833 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2834 	}
2835 
2836 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2837 
2838 	/* force loopback */
2839 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2840 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2841 
2842 	sqp->ud_header.lrh.virtual_lane    = 0;
2843 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2844 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2845 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2846 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2847 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2848 	else
2849 		sqp->ud_header.bth.destination_qpn =
2850 			cpu_to_be32(mdev->dev->caps.spec_qps[sqp->qp.port - 1].qp0_tunnel);
2851 
2852 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2853 	if (mlx4_is_master(mdev->dev)) {
2854 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2855 			return -EINVAL;
2856 	} else {
2857 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2858 			return -EINVAL;
2859 	}
2860 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2861 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2862 
2863 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2864 	sqp->ud_header.immediate_present = 0;
2865 
2866 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2867 
2868 	/*
2869 	 * Inline data segments may not cross a 64 byte boundary.  If
2870 	 * our UD header is bigger than the space available up to the
2871 	 * next 64 byte boundary in the WQE, use two inline data
2872 	 * segments to hold the UD header.
2873 	 */
2874 	spc = MLX4_INLINE_ALIGN -
2875 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2876 	if (header_size <= spc) {
2877 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2878 		memcpy(inl + 1, sqp->header_buf, header_size);
2879 		i = 1;
2880 	} else {
2881 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2882 		memcpy(inl + 1, sqp->header_buf, spc);
2883 
2884 		inl = (void *) (inl + 1) + spc;
2885 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2886 		/*
2887 		 * Need a barrier here to make sure all the data is
2888 		 * visible before the byte_count field is set.
2889 		 * Otherwise the HCA prefetcher could grab the 64-byte
2890 		 * chunk with this inline segment and get a valid (!=
2891 		 * 0xffffffff) byte count but stale data, and end up
2892 		 * generating a packet with bad headers.
2893 		 *
2894 		 * The first inline segment's byte_count field doesn't
2895 		 * need a barrier, because it comes after a
2896 		 * control/MLX segment and therefore is at an offset
2897 		 * of 16 mod 64.
2898 		 */
2899 		wmb();
2900 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2901 		i = 2;
2902 	}
2903 
2904 	*mlx_seg_len =
2905 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2906 	return 0;
2907 }
2908 
2909 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
2910 {
2911 	union sl2vl_tbl_to_u64 tmp_vltab;
2912 	u8 vl;
2913 
2914 	if (sl > 15)
2915 		return 0xf;
2916 	tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
2917 	vl = tmp_vltab.sl8[sl >> 1];
2918 	if (sl & 1)
2919 		vl &= 0x0f;
2920 	else
2921 		vl >>= 4;
2922 	return vl;
2923 }
2924 
2925 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
2926 				int index, union ib_gid *gid,
2927 				enum ib_gid_type *gid_type)
2928 {
2929 	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
2930 	struct mlx4_port_gid_table *port_gid_table;
2931 	unsigned long flags;
2932 
2933 	port_gid_table = &iboe->gids[port_num - 1];
2934 	spin_lock_irqsave(&iboe->lock, flags);
2935 	memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
2936 	*gid_type = port_gid_table->gids[index].gid_type;
2937 	spin_unlock_irqrestore(&iboe->lock, flags);
2938 	if (rdma_is_zero_gid(gid))
2939 		return -ENOENT;
2940 
2941 	return 0;
2942 }
2943 
2944 #define MLX4_ROCEV2_QP1_SPORT 0xC000
2945 static int build_mlx_header(struct mlx4_ib_sqp *sqp, const struct ib_ud_wr *wr,
2946 			    void *wqe, unsigned *mlx_seg_len)
2947 {
2948 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
2949 	struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
2950 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2951 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2952 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2953 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2954 	union ib_gid sgid;
2955 	u16 pkey;
2956 	int send_size;
2957 	int header_size;
2958 	int spc;
2959 	int i;
2960 	int err = 0;
2961 	u16 vlan = 0xffff;
2962 	bool is_eth;
2963 	bool is_vlan = false;
2964 	bool is_grh;
2965 	bool is_udp = false;
2966 	int ip_version = 0;
2967 
2968 	send_size = 0;
2969 	for (i = 0; i < wr->wr.num_sge; ++i)
2970 		send_size += wr->wr.sg_list[i].length;
2971 
2972 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2973 	is_grh = mlx4_ib_ah_grh_present(ah);
2974 	if (is_eth) {
2975 		enum ib_gid_type gid_type;
2976 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2977 			/* When multi-function is enabled, the ib_core gid
2978 			 * indexes don't necessarily match the hw ones, so
2979 			 * we must use our own cache */
2980 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2981 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
2982 							   ah->av.ib.gid_index, &sgid.raw[0]);
2983 			if (err)
2984 				return err;
2985 		} else  {
2986 			err = fill_gid_by_hw_index(ibdev, sqp->qp.port,
2987 					    ah->av.ib.gid_index,
2988 					    &sgid, &gid_type);
2989 			if (!err) {
2990 				is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
2991 				if (is_udp) {
2992 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
2993 						ip_version = 4;
2994 					else
2995 						ip_version = 6;
2996 					is_grh = false;
2997 				}
2998 			} else {
2999 				return err;
3000 			}
3001 		}
3002 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
3003 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
3004 			is_vlan = 1;
3005 		}
3006 	}
3007 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
3008 			  ip_version, is_udp, 0, &sqp->ud_header);
3009 	if (err)
3010 		return err;
3011 
3012 	if (!is_eth) {
3013 		sqp->ud_header.lrh.service_level =
3014 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
3015 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
3016 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
3017 	}
3018 
3019 	if (is_grh || (ip_version == 6)) {
3020 		sqp->ud_header.grh.traffic_class =
3021 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3022 		sqp->ud_header.grh.flow_label    =
3023 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
3024 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
3025 		if (is_eth) {
3026 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
3027 		} else {
3028 			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3029 				/* When multi-function is enabled, the ib_core gid
3030 				 * indexes don't necessarily match the hw ones, so
3031 				 * we must use our own cache
3032 				 */
3033 				sqp->ud_header.grh.source_gid.global.subnet_prefix =
3034 					cpu_to_be64(atomic64_read(&(to_mdev(ib_dev)->sriov.
3035 								    demux[sqp->qp.port - 1].
3036 								    subnet_prefix)));
3037 				sqp->ud_header.grh.source_gid.global.interface_id =
3038 					to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
3039 						       guid_cache[ah->av.ib.gid_index];
3040 			} else {
3041 				sqp->ud_header.grh.source_gid =
3042 					ah->ibah.sgid_attr->gid;
3043 			}
3044 		}
3045 		memcpy(sqp->ud_header.grh.destination_gid.raw,
3046 		       ah->av.ib.dgid, 16);
3047 	}
3048 
3049 	if (ip_version == 4) {
3050 		sqp->ud_header.ip4.tos =
3051 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3052 		sqp->ud_header.ip4.id = 0;
3053 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
3054 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
3055 
3056 		memcpy(&sqp->ud_header.ip4.saddr,
3057 		       sgid.raw + 12, 4);
3058 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
3059 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
3060 	}
3061 
3062 	if (is_udp) {
3063 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
3064 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
3065 		sqp->ud_header.udp.csum = 0;
3066 	}
3067 
3068 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
3069 
3070 	if (!is_eth) {
3071 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
3072 					  (sqp->ud_header.lrh.destination_lid ==
3073 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
3074 					  (sqp->ud_header.lrh.service_level << 8));
3075 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
3076 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
3077 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
3078 	}
3079 
3080 	switch (wr->wr.opcode) {
3081 	case IB_WR_SEND:
3082 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
3083 		sqp->ud_header.immediate_present = 0;
3084 		break;
3085 	case IB_WR_SEND_WITH_IMM:
3086 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
3087 		sqp->ud_header.immediate_present = 1;
3088 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
3089 		break;
3090 	default:
3091 		return -EINVAL;
3092 	}
3093 
3094 	if (is_eth) {
3095 		struct in6_addr in6;
3096 		u16 ether_type;
3097 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
3098 
3099 		ether_type = (!is_udp) ? ETH_P_IBOE:
3100 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
3101 
3102 		mlx->sched_prio = cpu_to_be16(pcp);
3103 
3104 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
3105 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
3106 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
3107 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
3108 		memcpy(&in6, sgid.raw, sizeof(in6));
3109 
3110 
3111 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
3112 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
3113 		if (!is_vlan) {
3114 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
3115 		} else {
3116 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
3117 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
3118 		}
3119 	} else {
3120 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 :
3121 							sl_to_vl(to_mdev(ib_dev),
3122 								 sqp->ud_header.lrh.service_level,
3123 								 sqp->qp.port);
3124 		if (sqp->qp.ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
3125 			return -EINVAL;
3126 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
3127 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
3128 	}
3129 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
3130 	if (!sqp->qp.ibqp.qp_num)
3131 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
3132 	else
3133 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
3134 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
3135 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
3136 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
3137 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
3138 					       sqp->qkey : wr->remote_qkey);
3139 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
3140 
3141 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3142 
3143 	if (0) {
3144 		pr_err("built UD header of size %d:\n", header_size);
3145 		for (i = 0; i < header_size / 4; ++i) {
3146 			if (i % 8 == 0)
3147 				pr_err("  [%02x] ", i * 4);
3148 			pr_cont(" %08x",
3149 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
3150 			if ((i + 1) % 8 == 0)
3151 				pr_cont("\n");
3152 		}
3153 		pr_err("\n");
3154 	}
3155 
3156 	/*
3157 	 * Inline data segments may not cross a 64 byte boundary.  If
3158 	 * our UD header is bigger than the space available up to the
3159 	 * next 64 byte boundary in the WQE, use two inline data
3160 	 * segments to hold the UD header.
3161 	 */
3162 	spc = MLX4_INLINE_ALIGN -
3163 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3164 	if (header_size <= spc) {
3165 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3166 		memcpy(inl + 1, sqp->header_buf, header_size);
3167 		i = 1;
3168 	} else {
3169 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3170 		memcpy(inl + 1, sqp->header_buf, spc);
3171 
3172 		inl = (void *) (inl + 1) + spc;
3173 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3174 		/*
3175 		 * Need a barrier here to make sure all the data is
3176 		 * visible before the byte_count field is set.
3177 		 * Otherwise the HCA prefetcher could grab the 64-byte
3178 		 * chunk with this inline segment and get a valid (!=
3179 		 * 0xffffffff) byte count but stale data, and end up
3180 		 * generating a packet with bad headers.
3181 		 *
3182 		 * The first inline segment's byte_count field doesn't
3183 		 * need a barrier, because it comes after a
3184 		 * control/MLX segment and therefore is at an offset
3185 		 * of 16 mod 64.
3186 		 */
3187 		wmb();
3188 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3189 		i = 2;
3190 	}
3191 
3192 	*mlx_seg_len =
3193 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3194 	return 0;
3195 }
3196 
3197 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3198 {
3199 	unsigned cur;
3200 	struct mlx4_ib_cq *cq;
3201 
3202 	cur = wq->head - wq->tail;
3203 	if (likely(cur + nreq < wq->max_post))
3204 		return 0;
3205 
3206 	cq = to_mcq(ib_cq);
3207 	spin_lock(&cq->lock);
3208 	cur = wq->head - wq->tail;
3209 	spin_unlock(&cq->lock);
3210 
3211 	return cur + nreq >= wq->max_post;
3212 }
3213 
3214 static __be32 convert_access(int acc)
3215 {
3216 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
3217 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
3218 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
3219 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
3220 	       (acc & IB_ACCESS_REMOTE_READ   ?
3221 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
3222 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
3223 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
3224 }
3225 
3226 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
3227 			const struct ib_reg_wr *wr)
3228 {
3229 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
3230 
3231 	fseg->flags		= convert_access(wr->access);
3232 	fseg->mem_key		= cpu_to_be32(wr->key);
3233 	fseg->buf_list		= cpu_to_be64(mr->page_map);
3234 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
3235 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
3236 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
3237 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
3238 	fseg->reserved[0]	= 0;
3239 	fseg->reserved[1]	= 0;
3240 }
3241 
3242 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
3243 {
3244 	memset(iseg, 0, sizeof(*iseg));
3245 	iseg->mem_key = cpu_to_be32(rkey);
3246 }
3247 
3248 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
3249 					  u64 remote_addr, u32 rkey)
3250 {
3251 	rseg->raddr    = cpu_to_be64(remote_addr);
3252 	rseg->rkey     = cpu_to_be32(rkey);
3253 	rseg->reserved = 0;
3254 }
3255 
3256 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
3257 			   const struct ib_atomic_wr *wr)
3258 {
3259 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
3260 		aseg->swap_add = cpu_to_be64(wr->swap);
3261 		aseg->compare  = cpu_to_be64(wr->compare_add);
3262 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
3263 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3264 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
3265 	} else {
3266 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3267 		aseg->compare  = 0;
3268 	}
3269 
3270 }
3271 
3272 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
3273 				  const struct ib_atomic_wr *wr)
3274 {
3275 	aseg->swap_add		= cpu_to_be64(wr->swap);
3276 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
3277 	aseg->compare		= cpu_to_be64(wr->compare_add);
3278 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
3279 }
3280 
3281 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
3282 			     const struct ib_ud_wr *wr)
3283 {
3284 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
3285 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
3286 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
3287 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
3288 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
3289 }
3290 
3291 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
3292 				    struct mlx4_wqe_datagram_seg *dseg,
3293 				    const struct ib_ud_wr *wr,
3294 				    enum mlx4_ib_qp_type qpt)
3295 {
3296 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
3297 	struct mlx4_av sqp_av = {0};
3298 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
3299 
3300 	/* force loopback */
3301 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
3302 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
3303 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
3304 			cpu_to_be32(0xf0000000);
3305 
3306 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
3307 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
3308 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp1_tunnel);
3309 	else
3310 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp0_tunnel);
3311 	/* Use QKEY from the QP context, which is set by master */
3312 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
3313 }
3314 
3315 static void build_tunnel_header(const struct ib_ud_wr *wr, void *wqe,
3316 				unsigned *mlx_seg_len)
3317 {
3318 	struct mlx4_wqe_inline_seg *inl = wqe;
3319 	struct mlx4_ib_tunnel_header hdr;
3320 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3321 	int spc;
3322 	int i;
3323 
3324 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
3325 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
3326 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
3327 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
3328 	memcpy(hdr.mac, ah->av.eth.mac, 6);
3329 	hdr.vlan = ah->av.eth.vlan;
3330 
3331 	spc = MLX4_INLINE_ALIGN -
3332 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3333 	if (sizeof (hdr) <= spc) {
3334 		memcpy(inl + 1, &hdr, sizeof (hdr));
3335 		wmb();
3336 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
3337 		i = 1;
3338 	} else {
3339 		memcpy(inl + 1, &hdr, spc);
3340 		wmb();
3341 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3342 
3343 		inl = (void *) (inl + 1) + spc;
3344 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
3345 		wmb();
3346 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
3347 		i = 2;
3348 	}
3349 
3350 	*mlx_seg_len =
3351 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
3352 }
3353 
3354 static void set_mlx_icrc_seg(void *dseg)
3355 {
3356 	u32 *t = dseg;
3357 	struct mlx4_wqe_inline_seg *iseg = dseg;
3358 
3359 	t[1] = 0;
3360 
3361 	/*
3362 	 * Need a barrier here before writing the byte_count field to
3363 	 * make sure that all the data is visible before the
3364 	 * byte_count field is set.  Otherwise, if the segment begins
3365 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3366 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3367 	 * stale data, and end up sending the wrong data.
3368 	 */
3369 	wmb();
3370 
3371 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
3372 }
3373 
3374 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3375 {
3376 	dseg->lkey       = cpu_to_be32(sg->lkey);
3377 	dseg->addr       = cpu_to_be64(sg->addr);
3378 
3379 	/*
3380 	 * Need a barrier here before writing the byte_count field to
3381 	 * make sure that all the data is visible before the
3382 	 * byte_count field is set.  Otherwise, if the segment begins
3383 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3384 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3385 	 * stale data, and end up sending the wrong data.
3386 	 */
3387 	wmb();
3388 
3389 	dseg->byte_count = cpu_to_be32(sg->length);
3390 }
3391 
3392 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3393 {
3394 	dseg->byte_count = cpu_to_be32(sg->length);
3395 	dseg->lkey       = cpu_to_be32(sg->lkey);
3396 	dseg->addr       = cpu_to_be64(sg->addr);
3397 }
3398 
3399 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe,
3400 			 const struct ib_ud_wr *wr, struct mlx4_ib_qp *qp,
3401 			 unsigned *lso_seg_len, __be32 *lso_hdr_sz, __be32 *blh)
3402 {
3403 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
3404 
3405 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
3406 		*blh = cpu_to_be32(1 << 6);
3407 
3408 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
3409 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
3410 		return -EINVAL;
3411 
3412 	memcpy(wqe->header, wr->header, wr->hlen);
3413 
3414 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
3415 	*lso_seg_len = halign;
3416 	return 0;
3417 }
3418 
3419 static __be32 send_ieth(const struct ib_send_wr *wr)
3420 {
3421 	switch (wr->opcode) {
3422 	case IB_WR_SEND_WITH_IMM:
3423 	case IB_WR_RDMA_WRITE_WITH_IMM:
3424 		return wr->ex.imm_data;
3425 
3426 	case IB_WR_SEND_WITH_INV:
3427 		return cpu_to_be32(wr->ex.invalidate_rkey);
3428 
3429 	default:
3430 		return 0;
3431 	}
3432 }
3433 
3434 static void add_zero_len_inline(void *wqe)
3435 {
3436 	struct mlx4_wqe_inline_seg *inl = wqe;
3437 	memset(wqe, 0, 16);
3438 	inl->byte_count = cpu_to_be32(1 << 31);
3439 }
3440 
3441 static int _mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3442 			      const struct ib_send_wr **bad_wr, bool drain)
3443 {
3444 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3445 	void *wqe;
3446 	struct mlx4_wqe_ctrl_seg *ctrl;
3447 	struct mlx4_wqe_data_seg *dseg;
3448 	unsigned long flags;
3449 	int nreq;
3450 	int err = 0;
3451 	unsigned ind;
3452 	int uninitialized_var(size);
3453 	unsigned uninitialized_var(seglen);
3454 	__be32 dummy;
3455 	__be32 *lso_wqe;
3456 	__be32 uninitialized_var(lso_hdr_sz);
3457 	__be32 blh;
3458 	int i;
3459 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3460 
3461 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
3462 		struct mlx4_ib_sqp *sqp = to_msqp(qp);
3463 
3464 		if (sqp->roce_v2_gsi) {
3465 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
3466 			enum ib_gid_type gid_type;
3467 			union ib_gid gid;
3468 
3469 			if (!fill_gid_by_hw_index(mdev, sqp->qp.port,
3470 					   ah->av.ib.gid_index,
3471 					   &gid, &gid_type))
3472 				qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
3473 						to_mqp(sqp->roce_v2_gsi) : qp;
3474 			else
3475 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
3476 				       ah->av.ib.gid_index);
3477 		}
3478 	}
3479 
3480 	spin_lock_irqsave(&qp->sq.lock, flags);
3481 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3482 	    !drain) {
3483 		err = -EIO;
3484 		*bad_wr = wr;
3485 		nreq = 0;
3486 		goto out;
3487 	}
3488 
3489 	ind = qp->sq_next_wqe;
3490 
3491 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3492 		lso_wqe = &dummy;
3493 		blh = 0;
3494 
3495 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
3496 			err = -ENOMEM;
3497 			*bad_wr = wr;
3498 			goto out;
3499 		}
3500 
3501 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3502 			err = -EINVAL;
3503 			*bad_wr = wr;
3504 			goto out;
3505 		}
3506 
3507 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3508 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3509 
3510 		ctrl->srcrb_flags =
3511 			(wr->send_flags & IB_SEND_SIGNALED ?
3512 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3513 			(wr->send_flags & IB_SEND_SOLICITED ?
3514 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3515 			((wr->send_flags & IB_SEND_IP_CSUM) ?
3516 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3517 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3518 			qp->sq_signal_bits;
3519 
3520 		ctrl->imm = send_ieth(wr);
3521 
3522 		wqe += sizeof *ctrl;
3523 		size = sizeof *ctrl / 16;
3524 
3525 		switch (qp->mlx4_ib_qp_type) {
3526 		case MLX4_IB_QPT_RC:
3527 		case MLX4_IB_QPT_UC:
3528 			switch (wr->opcode) {
3529 			case IB_WR_ATOMIC_CMP_AND_SWP:
3530 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3531 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3532 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3533 					      atomic_wr(wr)->rkey);
3534 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3535 
3536 				set_atomic_seg(wqe, atomic_wr(wr));
3537 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
3538 
3539 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3540 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3541 
3542 				break;
3543 
3544 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3545 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3546 					      atomic_wr(wr)->rkey);
3547 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3548 
3549 				set_masked_atomic_seg(wqe, atomic_wr(wr));
3550 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
3551 
3552 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3553 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3554 
3555 				break;
3556 
3557 			case IB_WR_RDMA_READ:
3558 			case IB_WR_RDMA_WRITE:
3559 			case IB_WR_RDMA_WRITE_WITH_IMM:
3560 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3561 					      rdma_wr(wr)->rkey);
3562 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3563 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3564 				break;
3565 
3566 			case IB_WR_LOCAL_INV:
3567 				ctrl->srcrb_flags |=
3568 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3569 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3570 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3571 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3572 				break;
3573 
3574 			case IB_WR_REG_MR:
3575 				ctrl->srcrb_flags |=
3576 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3577 				set_reg_seg(wqe, reg_wr(wr));
3578 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3579 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3580 				break;
3581 
3582 			default:
3583 				/* No extra segments required for sends */
3584 				break;
3585 			}
3586 			break;
3587 
3588 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3589 			err =  build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3590 					ctrl, &seglen);
3591 			if (unlikely(err)) {
3592 				*bad_wr = wr;
3593 				goto out;
3594 			}
3595 			wqe  += seglen;
3596 			size += seglen / 16;
3597 			break;
3598 		case MLX4_IB_QPT_TUN_SMI:
3599 		case MLX4_IB_QPT_TUN_GSI:
3600 			/* this is a UD qp used in MAD responses to slaves. */
3601 			set_datagram_seg(wqe, ud_wr(wr));
3602 			/* set the forced-loopback bit in the data seg av */
3603 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3604 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3605 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3606 			break;
3607 		case MLX4_IB_QPT_UD:
3608 			set_datagram_seg(wqe, ud_wr(wr));
3609 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3610 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3611 
3612 			if (wr->opcode == IB_WR_LSO) {
3613 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3614 						&lso_hdr_sz, &blh);
3615 				if (unlikely(err)) {
3616 					*bad_wr = wr;
3617 					goto out;
3618 				}
3619 				lso_wqe = (__be32 *) wqe;
3620 				wqe  += seglen;
3621 				size += seglen / 16;
3622 			}
3623 			break;
3624 
3625 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3626 			err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3627 					ctrl, &seglen);
3628 			if (unlikely(err)) {
3629 				*bad_wr = wr;
3630 				goto out;
3631 			}
3632 			wqe  += seglen;
3633 			size += seglen / 16;
3634 			/* to start tunnel header on a cache-line boundary */
3635 			add_zero_len_inline(wqe);
3636 			wqe += 16;
3637 			size++;
3638 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3639 			wqe  += seglen;
3640 			size += seglen / 16;
3641 			break;
3642 		case MLX4_IB_QPT_PROXY_SMI:
3643 		case MLX4_IB_QPT_PROXY_GSI:
3644 			/* If we are tunneling special qps, this is a UD qp.
3645 			 * In this case we first add a UD segment targeting
3646 			 * the tunnel qp, and then add a header with address
3647 			 * information */
3648 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3649 						ud_wr(wr),
3650 						qp->mlx4_ib_qp_type);
3651 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3652 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3653 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3654 			wqe  += seglen;
3655 			size += seglen / 16;
3656 			break;
3657 
3658 		case MLX4_IB_QPT_SMI:
3659 		case MLX4_IB_QPT_GSI:
3660 			err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
3661 					&seglen);
3662 			if (unlikely(err)) {
3663 				*bad_wr = wr;
3664 				goto out;
3665 			}
3666 			wqe  += seglen;
3667 			size += seglen / 16;
3668 			break;
3669 
3670 		default:
3671 			break;
3672 		}
3673 
3674 		/*
3675 		 * Write data segments in reverse order, so as to
3676 		 * overwrite cacheline stamp last within each
3677 		 * cacheline.  This avoids issues with WQE
3678 		 * prefetching.
3679 		 */
3680 
3681 		dseg = wqe;
3682 		dseg += wr->num_sge - 1;
3683 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3684 
3685 		/* Add one more inline data segment for ICRC for MLX sends */
3686 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3687 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3688 			     qp->mlx4_ib_qp_type &
3689 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3690 			set_mlx_icrc_seg(dseg + 1);
3691 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3692 		}
3693 
3694 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3695 			set_data_seg(dseg, wr->sg_list + i);
3696 
3697 		/*
3698 		 * Possibly overwrite stamping in cacheline with LSO
3699 		 * segment only after making sure all data segments
3700 		 * are written.
3701 		 */
3702 		wmb();
3703 		*lso_wqe = lso_hdr_sz;
3704 
3705 		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3706 					     MLX4_WQE_CTRL_FENCE : 0) | size;
3707 
3708 		/*
3709 		 * Make sure descriptor is fully written before
3710 		 * setting ownership bit (because HW can start
3711 		 * executing as soon as we do).
3712 		 */
3713 		wmb();
3714 
3715 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3716 			*bad_wr = wr;
3717 			err = -EINVAL;
3718 			goto out;
3719 		}
3720 
3721 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3722 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3723 
3724 		/*
3725 		 * We can improve latency by not stamping the last
3726 		 * send queue WQE until after ringing the doorbell, so
3727 		 * only stamp here if there are still more WQEs to post.
3728 		 */
3729 		if (wr->next)
3730 			stamp_send_wqe(qp, ind + qp->sq_spare_wqes);
3731 		ind++;
3732 	}
3733 
3734 out:
3735 	if (likely(nreq)) {
3736 		qp->sq.head += nreq;
3737 
3738 		/*
3739 		 * Make sure that descriptors are written before
3740 		 * doorbell record.
3741 		 */
3742 		wmb();
3743 
3744 		writel_relaxed(qp->doorbell_qpn,
3745 			to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3746 
3747 		/*
3748 		 * Make sure doorbells don't leak out of SQ spinlock
3749 		 * and reach the HCA out of order.
3750 		 */
3751 		mmiowb();
3752 
3753 		stamp_send_wqe(qp, ind + qp->sq_spare_wqes - 1);
3754 
3755 		qp->sq_next_wqe = ind;
3756 	}
3757 
3758 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3759 
3760 	return err;
3761 }
3762 
3763 int mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3764 		      const struct ib_send_wr **bad_wr)
3765 {
3766 	return _mlx4_ib_post_send(ibqp, wr, bad_wr, false);
3767 }
3768 
3769 static int _mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3770 			      const struct ib_recv_wr **bad_wr, bool drain)
3771 {
3772 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3773 	struct mlx4_wqe_data_seg *scat;
3774 	unsigned long flags;
3775 	int err = 0;
3776 	int nreq;
3777 	int ind;
3778 	int max_gs;
3779 	int i;
3780 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3781 
3782 	max_gs = qp->rq.max_gs;
3783 	spin_lock_irqsave(&qp->rq.lock, flags);
3784 
3785 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3786 	    !drain) {
3787 		err = -EIO;
3788 		*bad_wr = wr;
3789 		nreq = 0;
3790 		goto out;
3791 	}
3792 
3793 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3794 
3795 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3796 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3797 			err = -ENOMEM;
3798 			*bad_wr = wr;
3799 			goto out;
3800 		}
3801 
3802 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3803 			err = -EINVAL;
3804 			*bad_wr = wr;
3805 			goto out;
3806 		}
3807 
3808 		scat = get_recv_wqe(qp, ind);
3809 
3810 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3811 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3812 			ib_dma_sync_single_for_device(ibqp->device,
3813 						      qp->sqp_proxy_rcv[ind].map,
3814 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3815 						      DMA_FROM_DEVICE);
3816 			scat->byte_count =
3817 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3818 			/* use dma lkey from upper layer entry */
3819 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3820 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3821 			scat++;
3822 			max_gs--;
3823 		}
3824 
3825 		for (i = 0; i < wr->num_sge; ++i)
3826 			__set_data_seg(scat + i, wr->sg_list + i);
3827 
3828 		if (i < max_gs) {
3829 			scat[i].byte_count = 0;
3830 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3831 			scat[i].addr       = 0;
3832 		}
3833 
3834 		qp->rq.wrid[ind] = wr->wr_id;
3835 
3836 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3837 	}
3838 
3839 out:
3840 	if (likely(nreq)) {
3841 		qp->rq.head += nreq;
3842 
3843 		/*
3844 		 * Make sure that descriptors are written before
3845 		 * doorbell record.
3846 		 */
3847 		wmb();
3848 
3849 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3850 	}
3851 
3852 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3853 
3854 	return err;
3855 }
3856 
3857 int mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3858 		      const struct ib_recv_wr **bad_wr)
3859 {
3860 	return _mlx4_ib_post_recv(ibqp, wr, bad_wr, false);
3861 }
3862 
3863 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3864 {
3865 	switch (mlx4_state) {
3866 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3867 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3868 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3869 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3870 	case MLX4_QP_STATE_SQ_DRAINING:
3871 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3872 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3873 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3874 	default:		     return -1;
3875 	}
3876 }
3877 
3878 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3879 {
3880 	switch (mlx4_mig_state) {
3881 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3882 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3883 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3884 	default: return -1;
3885 	}
3886 }
3887 
3888 static int to_ib_qp_access_flags(int mlx4_flags)
3889 {
3890 	int ib_flags = 0;
3891 
3892 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3893 		ib_flags |= IB_ACCESS_REMOTE_READ;
3894 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3895 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3896 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3897 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3898 
3899 	return ib_flags;
3900 }
3901 
3902 static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
3903 			    struct rdma_ah_attr *ah_attr,
3904 			    struct mlx4_qp_path *path)
3905 {
3906 	struct mlx4_dev *dev = ibdev->dev;
3907 	u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
3908 
3909 	memset(ah_attr, 0, sizeof(*ah_attr));
3910 	if (port_num == 0 || port_num > dev->caps.num_ports)
3911 		return;
3912 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
3913 
3914 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
3915 		rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
3916 			       ((path->sched_queue & 4) << 1));
3917 	else
3918 		rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
3919 	rdma_ah_set_port_num(ah_attr, port_num);
3920 
3921 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
3922 	rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
3923 	rdma_ah_set_static_rate(ah_attr,
3924 				path->static_rate ? path->static_rate - 5 : 0);
3925 	if (path->grh_mylmc & (1 << 7)) {
3926 		rdma_ah_set_grh(ah_attr, NULL,
3927 				be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
3928 				path->mgid_index,
3929 				path->hop_limit,
3930 				(be32_to_cpu(path->tclass_flowlabel)
3931 				 >> 20) & 0xff);
3932 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
3933 	}
3934 }
3935 
3936 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3937 		     struct ib_qp_init_attr *qp_init_attr)
3938 {
3939 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3940 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3941 	struct mlx4_qp_context context;
3942 	int mlx4_state;
3943 	int err = 0;
3944 
3945 	if (ibqp->rwq_ind_tbl)
3946 		return -EOPNOTSUPP;
3947 
3948 	mutex_lock(&qp->mutex);
3949 
3950 	if (qp->state == IB_QPS_RESET) {
3951 		qp_attr->qp_state = IB_QPS_RESET;
3952 		goto done;
3953 	}
3954 
3955 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3956 	if (err) {
3957 		err = -EINVAL;
3958 		goto out;
3959 	}
3960 
3961 	mlx4_state = be32_to_cpu(context.flags) >> 28;
3962 
3963 	qp->state		     = to_ib_qp_state(mlx4_state);
3964 	qp_attr->qp_state	     = qp->state;
3965 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
3966 	qp_attr->path_mig_state	     =
3967 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3968 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
3969 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3970 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
3971 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
3972 	qp_attr->qp_access_flags     =
3973 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
3974 
3975 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3976 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3977 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3978 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3979 		qp_attr->alt_port_num	=
3980 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
3981 	}
3982 
3983 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3984 	if (qp_attr->qp_state == IB_QPS_INIT)
3985 		qp_attr->port_num = qp->port;
3986 	else
3987 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3988 
3989 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3990 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3991 
3992 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3993 
3994 	qp_attr->max_dest_rd_atomic =
3995 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3996 	qp_attr->min_rnr_timer	    =
3997 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3998 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
3999 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
4000 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
4001 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
4002 
4003 done:
4004 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4005 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4006 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4007 
4008 	if (!ibqp->uobject) {
4009 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
4010 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4011 	} else {
4012 		qp_attr->cap.max_send_wr  = 0;
4013 		qp_attr->cap.max_send_sge = 0;
4014 	}
4015 
4016 	/*
4017 	 * We don't support inline sends for kernel QPs (yet), and we
4018 	 * don't know what userspace's value should be.
4019 	 */
4020 	qp_attr->cap.max_inline_data = 0;
4021 
4022 	qp_init_attr->cap	     = qp_attr->cap;
4023 
4024 	qp_init_attr->create_flags = 0;
4025 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4026 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4027 
4028 	if (qp->flags & MLX4_IB_QP_LSO)
4029 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
4030 
4031 	if (qp->flags & MLX4_IB_QP_NETIF)
4032 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
4033 
4034 	qp_init_attr->sq_sig_type =
4035 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
4036 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4037 
4038 out:
4039 	mutex_unlock(&qp->mutex);
4040 	return err;
4041 }
4042 
4043 struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
4044 				struct ib_wq_init_attr *init_attr,
4045 				struct ib_udata *udata)
4046 {
4047 	struct mlx4_ib_dev *dev;
4048 	struct ib_qp_init_attr ib_qp_init_attr;
4049 	struct mlx4_ib_qp *qp;
4050 	struct mlx4_ib_create_wq ucmd;
4051 	int err, required_cmd_sz;
4052 
4053 	if (!udata)
4054 		return ERR_PTR(-EINVAL);
4055 
4056 	required_cmd_sz = offsetof(typeof(ucmd), comp_mask) +
4057 			  sizeof(ucmd.comp_mask);
4058 	if (udata->inlen < required_cmd_sz) {
4059 		pr_debug("invalid inlen\n");
4060 		return ERR_PTR(-EINVAL);
4061 	}
4062 
4063 	if (udata->inlen > sizeof(ucmd) &&
4064 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4065 				 udata->inlen - sizeof(ucmd))) {
4066 		pr_debug("inlen is not supported\n");
4067 		return ERR_PTR(-EOPNOTSUPP);
4068 	}
4069 
4070 	if (udata->outlen)
4071 		return ERR_PTR(-EOPNOTSUPP);
4072 
4073 	dev = to_mdev(pd->device);
4074 
4075 	if (init_attr->wq_type != IB_WQT_RQ) {
4076 		pr_debug("unsupported wq type %d\n", init_attr->wq_type);
4077 		return ERR_PTR(-EOPNOTSUPP);
4078 	}
4079 
4080 	if (init_attr->create_flags & ~IB_WQ_FLAGS_SCATTER_FCS) {
4081 		pr_debug("unsupported create_flags %u\n",
4082 			 init_attr->create_flags);
4083 		return ERR_PTR(-EOPNOTSUPP);
4084 	}
4085 
4086 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
4087 	if (!qp)
4088 		return ERR_PTR(-ENOMEM);
4089 
4090 	qp->pri.vid = 0xFFFF;
4091 	qp->alt.vid = 0xFFFF;
4092 
4093 	memset(&ib_qp_init_attr, 0, sizeof(ib_qp_init_attr));
4094 	ib_qp_init_attr.qp_context = init_attr->wq_context;
4095 	ib_qp_init_attr.qp_type = IB_QPT_RAW_PACKET;
4096 	ib_qp_init_attr.cap.max_recv_wr = init_attr->max_wr;
4097 	ib_qp_init_attr.cap.max_recv_sge = init_attr->max_sge;
4098 	ib_qp_init_attr.recv_cq = init_attr->cq;
4099 	ib_qp_init_attr.send_cq = ib_qp_init_attr.recv_cq; /* Dummy CQ */
4100 
4101 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS)
4102 		ib_qp_init_attr.create_flags |= IB_QP_CREATE_SCATTER_FCS;
4103 
4104 	err = create_qp_common(dev, pd, MLX4_IB_RWQ_SRC, &ib_qp_init_attr,
4105 			       udata, 0, &qp);
4106 	if (err) {
4107 		kfree(qp);
4108 		return ERR_PTR(err);
4109 	}
4110 
4111 	qp->ibwq.event_handler = init_attr->event_handler;
4112 	qp->ibwq.wq_num = qp->mqp.qpn;
4113 	qp->ibwq.state = IB_WQS_RESET;
4114 
4115 	return &qp->ibwq;
4116 }
4117 
4118 static int ib_wq2qp_state(enum ib_wq_state state)
4119 {
4120 	switch (state) {
4121 	case IB_WQS_RESET:
4122 		return IB_QPS_RESET;
4123 	case IB_WQS_RDY:
4124 		return IB_QPS_RTR;
4125 	default:
4126 		return IB_QPS_ERR;
4127 	}
4128 }
4129 
4130 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state,
4131 			      struct ib_udata *udata)
4132 {
4133 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4134 	enum ib_qp_state qp_cur_state;
4135 	enum ib_qp_state qp_new_state;
4136 	int attr_mask;
4137 	int err;
4138 
4139 	/* ib_qp.state represents the WQ HW state while ib_wq.state represents
4140 	 * the WQ logic state.
4141 	 */
4142 	qp_cur_state = qp->state;
4143 	qp_new_state = ib_wq2qp_state(new_state);
4144 
4145 	if (ib_wq2qp_state(new_state) == qp_cur_state)
4146 		return 0;
4147 
4148 	if (new_state == IB_WQS_RDY) {
4149 		struct ib_qp_attr attr = {};
4150 
4151 		attr.port_num = qp->port;
4152 		attr_mask = IB_QP_PORT;
4153 
4154 		err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, &attr,
4155 					  attr_mask, IB_QPS_RESET, IB_QPS_INIT,
4156 					  udata);
4157 		if (err) {
4158 			pr_debug("WQN=0x%06x failed to apply RST->INIT on the HW QP\n",
4159 				 ibwq->wq_num);
4160 			return err;
4161 		}
4162 
4163 		qp_cur_state = IB_QPS_INIT;
4164 	}
4165 
4166 	attr_mask = 0;
4167 	err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, attr_mask,
4168 				  qp_cur_state,  qp_new_state, udata);
4169 
4170 	if (err && (qp_cur_state == IB_QPS_INIT)) {
4171 		qp_new_state = IB_QPS_RESET;
4172 		if (__mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL,
4173 					attr_mask, IB_QPS_INIT, IB_QPS_RESET,
4174 					udata)) {
4175 			pr_warn("WQN=0x%06x failed with reverting HW's resources failure\n",
4176 				ibwq->wq_num);
4177 			qp_new_state = IB_QPS_INIT;
4178 		}
4179 	}
4180 
4181 	qp->state = qp_new_state;
4182 
4183 	return err;
4184 }
4185 
4186 int mlx4_ib_modify_wq(struct ib_wq *ibwq, struct ib_wq_attr *wq_attr,
4187 		      u32 wq_attr_mask, struct ib_udata *udata)
4188 {
4189 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4190 	struct mlx4_ib_modify_wq ucmd = {};
4191 	size_t required_cmd_sz;
4192 	enum ib_wq_state cur_state, new_state;
4193 	int err = 0;
4194 
4195 	required_cmd_sz = offsetof(typeof(ucmd), reserved) +
4196 				   sizeof(ucmd.reserved);
4197 	if (udata->inlen < required_cmd_sz)
4198 		return -EINVAL;
4199 
4200 	if (udata->inlen > sizeof(ucmd) &&
4201 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4202 				 udata->inlen - sizeof(ucmd)))
4203 		return -EOPNOTSUPP;
4204 
4205 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4206 		return -EFAULT;
4207 
4208 	if (ucmd.comp_mask || ucmd.reserved)
4209 		return -EOPNOTSUPP;
4210 
4211 	if (wq_attr_mask & IB_WQ_FLAGS)
4212 		return -EOPNOTSUPP;
4213 
4214 	cur_state = wq_attr_mask & IB_WQ_CUR_STATE ? wq_attr->curr_wq_state :
4215 						     ibwq->state;
4216 	new_state = wq_attr_mask & IB_WQ_STATE ? wq_attr->wq_state : cur_state;
4217 
4218 	if (cur_state  < IB_WQS_RESET || cur_state  > IB_WQS_ERR ||
4219 	    new_state < IB_WQS_RESET || new_state > IB_WQS_ERR)
4220 		return -EINVAL;
4221 
4222 	if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR))
4223 		return -EINVAL;
4224 
4225 	if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET))
4226 		return -EINVAL;
4227 
4228 	/* Need to protect against the parent RSS which also may modify WQ
4229 	 * state.
4230 	 */
4231 	mutex_lock(&qp->mutex);
4232 
4233 	/* Can update HW state only if a RSS QP has already associated to this
4234 	 * WQ, so we can apply its port on the WQ.
4235 	 */
4236 	if (qp->rss_usecnt)
4237 		err = _mlx4_ib_modify_wq(ibwq, new_state, udata);
4238 
4239 	if (!err)
4240 		ibwq->state = new_state;
4241 
4242 	mutex_unlock(&qp->mutex);
4243 
4244 	return err;
4245 }
4246 
4247 int mlx4_ib_destroy_wq(struct ib_wq *ibwq)
4248 {
4249 	struct mlx4_ib_dev *dev = to_mdev(ibwq->device);
4250 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4251 
4252 	if (qp->counter_index)
4253 		mlx4_ib_free_qp_counter(dev, qp);
4254 
4255 	destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, 1);
4256 
4257 	kfree(qp);
4258 
4259 	return 0;
4260 }
4261 
4262 struct ib_rwq_ind_table
4263 *mlx4_ib_create_rwq_ind_table(struct ib_device *device,
4264 			      struct ib_rwq_ind_table_init_attr *init_attr,
4265 			      struct ib_udata *udata)
4266 {
4267 	struct ib_rwq_ind_table *rwq_ind_table;
4268 	struct mlx4_ib_create_rwq_ind_tbl_resp resp = {};
4269 	unsigned int ind_tbl_size = 1 << init_attr->log_ind_tbl_size;
4270 	unsigned int base_wqn;
4271 	size_t min_resp_len;
4272 	int i;
4273 	int err;
4274 
4275 	if (udata->inlen > 0 &&
4276 	    !ib_is_udata_cleared(udata, 0,
4277 				 udata->inlen))
4278 		return ERR_PTR(-EOPNOTSUPP);
4279 
4280 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4281 	if (udata->outlen && udata->outlen < min_resp_len)
4282 		return ERR_PTR(-EINVAL);
4283 
4284 	if (ind_tbl_size >
4285 	    device->attrs.rss_caps.max_rwq_indirection_table_size) {
4286 		pr_debug("log_ind_tbl_size = %d is bigger than supported = %d\n",
4287 			 ind_tbl_size,
4288 			 device->attrs.rss_caps.max_rwq_indirection_table_size);
4289 		return ERR_PTR(-EINVAL);
4290 	}
4291 
4292 	base_wqn = init_attr->ind_tbl[0]->wq_num;
4293 
4294 	if (base_wqn % ind_tbl_size) {
4295 		pr_debug("WQN=0x%x isn't aligned with indirection table size\n",
4296 			 base_wqn);
4297 		return ERR_PTR(-EINVAL);
4298 	}
4299 
4300 	for (i = 1; i < ind_tbl_size; i++) {
4301 		if (++base_wqn != init_attr->ind_tbl[i]->wq_num) {
4302 			pr_debug("indirection table's WQNs aren't consecutive\n");
4303 			return ERR_PTR(-EINVAL);
4304 		}
4305 	}
4306 
4307 	rwq_ind_table = kzalloc(sizeof(*rwq_ind_table), GFP_KERNEL);
4308 	if (!rwq_ind_table)
4309 		return ERR_PTR(-ENOMEM);
4310 
4311 	if (udata->outlen) {
4312 		resp.response_length = offsetof(typeof(resp), response_length) +
4313 					sizeof(resp.response_length);
4314 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4315 		if (err)
4316 			goto err;
4317 	}
4318 
4319 	return rwq_ind_table;
4320 
4321 err:
4322 	kfree(rwq_ind_table);
4323 	return ERR_PTR(err);
4324 }
4325 
4326 int mlx4_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4327 {
4328 	kfree(ib_rwq_ind_tbl);
4329 	return 0;
4330 }
4331 
4332 struct mlx4_ib_drain_cqe {
4333 	struct ib_cqe cqe;
4334 	struct completion done;
4335 };
4336 
4337 static void mlx4_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
4338 {
4339 	struct mlx4_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
4340 						     struct mlx4_ib_drain_cqe,
4341 						     cqe);
4342 
4343 	complete(&cqe->done);
4344 }
4345 
4346 /* This function returns only once the drained WR was completed */
4347 static void handle_drain_completion(struct ib_cq *cq,
4348 				    struct mlx4_ib_drain_cqe *sdrain,
4349 				    struct mlx4_ib_dev *dev)
4350 {
4351 	struct mlx4_dev *mdev = dev->dev;
4352 
4353 	if (cq->poll_ctx == IB_POLL_DIRECT) {
4354 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
4355 			ib_process_cq_direct(cq, -1);
4356 		return;
4357 	}
4358 
4359 	if (mdev->persist->state == MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4360 		struct mlx4_ib_cq *mcq = to_mcq(cq);
4361 		bool triggered = false;
4362 		unsigned long flags;
4363 
4364 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
4365 		/* Make sure that the CQ handler won't run if wasn't run yet */
4366 		if (!mcq->mcq.reset_notify_added)
4367 			mcq->mcq.reset_notify_added = 1;
4368 		else
4369 			triggered = true;
4370 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
4371 
4372 		if (triggered) {
4373 			/* Wait for any scheduled/running task to be ended */
4374 			switch (cq->poll_ctx) {
4375 			case IB_POLL_SOFTIRQ:
4376 				irq_poll_disable(&cq->iop);
4377 				irq_poll_enable(&cq->iop);
4378 				break;
4379 			case IB_POLL_WORKQUEUE:
4380 				cancel_work_sync(&cq->work);
4381 				break;
4382 			default:
4383 				WARN_ON_ONCE(1);
4384 			}
4385 		}
4386 
4387 		/* Run the CQ handler - this makes sure that the drain WR will
4388 		 * be processed if wasn't processed yet.
4389 		 */
4390 		mcq->mcq.comp(&mcq->mcq);
4391 	}
4392 
4393 	wait_for_completion(&sdrain->done);
4394 }
4395 
4396 void mlx4_ib_drain_sq(struct ib_qp *qp)
4397 {
4398 	struct ib_cq *cq = qp->send_cq;
4399 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4400 	struct mlx4_ib_drain_cqe sdrain;
4401 	const struct ib_send_wr *bad_swr;
4402 	struct ib_rdma_wr swr = {
4403 		.wr = {
4404 			.next = NULL,
4405 			{ .wr_cqe	= &sdrain.cqe, },
4406 			.opcode	= IB_WR_RDMA_WRITE,
4407 		},
4408 	};
4409 	int ret;
4410 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
4411 	struct mlx4_dev *mdev = dev->dev;
4412 
4413 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4414 	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4415 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4416 		return;
4417 	}
4418 
4419 	sdrain.cqe.done = mlx4_ib_drain_qp_done;
4420 	init_completion(&sdrain.done);
4421 
4422 	ret = _mlx4_ib_post_send(qp, &swr.wr, &bad_swr, true);
4423 	if (ret) {
4424 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4425 		return;
4426 	}
4427 
4428 	handle_drain_completion(cq, &sdrain, dev);
4429 }
4430 
4431 void mlx4_ib_drain_rq(struct ib_qp *qp)
4432 {
4433 	struct ib_cq *cq = qp->recv_cq;
4434 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4435 	struct mlx4_ib_drain_cqe rdrain;
4436 	struct ib_recv_wr rwr = {};
4437 	const struct ib_recv_wr *bad_rwr;
4438 	int ret;
4439 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
4440 	struct mlx4_dev *mdev = dev->dev;
4441 
4442 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4443 	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4444 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4445 		return;
4446 	}
4447 
4448 	rwr.wr_cqe = &rdrain.cqe;
4449 	rdrain.cqe.done = mlx4_ib_drain_qp_done;
4450 	init_completion(&rdrain.done);
4451 
4452 	ret = _mlx4_ib_post_recv(qp, &rwr, &bad_rwr, true);
4453 	if (ret) {
4454 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4455 		return;
4456 	}
4457 
4458 	handle_drain_completion(cq, &rdrain, dev);
4459 }
4460