xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision 2c6467d2)
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 
45 #include <linux/mlx4/driver.h>
46 #include <linux/mlx4/qp.h>
47 
48 #include "mlx4_ib.h"
49 #include <rdma/mlx4-abi.h>
50 
51 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
52 			     struct mlx4_ib_cq *recv_cq);
53 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
54 			       struct mlx4_ib_cq *recv_cq);
55 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state);
56 
57 enum {
58 	MLX4_IB_ACK_REQ_FREQ	= 8,
59 };
60 
61 enum {
62 	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
63 	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
64 	MLX4_IB_LINK_TYPE_IB		= 0,
65 	MLX4_IB_LINK_TYPE_ETH		= 1
66 };
67 
68 enum {
69 	/*
70 	 * Largest possible UD header: send with GRH and immediate
71 	 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
72 	 * tag.  (LRH would only use 8 bytes, so Ethernet is the
73 	 * biggest case)
74 	 */
75 	MLX4_IB_UD_HEADER_SIZE		= 82,
76 	MLX4_IB_LSO_HEADER_SPARE	= 128,
77 };
78 
79 struct mlx4_ib_sqp {
80 	struct mlx4_ib_qp	qp;
81 	int			pkey_index;
82 	u32			qkey;
83 	u32			send_psn;
84 	struct ib_ud_header	ud_header;
85 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
86 	struct ib_qp		*roce_v2_gsi;
87 };
88 
89 enum {
90 	MLX4_IB_MIN_SQ_STRIDE	= 6,
91 	MLX4_IB_CACHE_LINE_SIZE	= 64,
92 };
93 
94 enum {
95 	MLX4_RAW_QP_MTU		= 7,
96 	MLX4_RAW_QP_MSGMAX	= 31,
97 };
98 
99 #ifndef ETH_ALEN
100 #define ETH_ALEN        6
101 #endif
102 
103 static const __be32 mlx4_ib_opcode[] = {
104 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
105 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
106 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
107 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
108 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
109 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
110 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
111 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
112 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
113 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
114 	[IB_WR_REG_MR]				= cpu_to_be32(MLX4_OPCODE_FMR),
115 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
116 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
117 };
118 
119 enum mlx4_ib_source_type {
120 	MLX4_IB_QP_SRC	= 0,
121 	MLX4_IB_RWQ_SRC	= 1,
122 };
123 
124 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
125 {
126 	return container_of(mqp, struct mlx4_ib_sqp, qp);
127 }
128 
129 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
130 {
131 	if (!mlx4_is_master(dev->dev))
132 		return 0;
133 
134 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
135 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
136 		8 * MLX4_MFUNC_MAX;
137 }
138 
139 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
140 {
141 	int proxy_sqp = 0;
142 	int real_sqp = 0;
143 	int i;
144 	/* PPF or Native -- real SQP */
145 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
146 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
147 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
148 	if (real_sqp)
149 		return 1;
150 	/* VF or PF -- proxy SQP */
151 	if (mlx4_is_mfunc(dev->dev)) {
152 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
153 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy ||
154 			    qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp1_proxy) {
155 				proxy_sqp = 1;
156 				break;
157 			}
158 		}
159 	}
160 	if (proxy_sqp)
161 		return 1;
162 
163 	return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
164 }
165 
166 /* used for INIT/CLOSE port logic */
167 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
168 {
169 	int proxy_qp0 = 0;
170 	int real_qp0 = 0;
171 	int i;
172 	/* PPF or Native -- real QP0 */
173 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
174 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
175 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
176 	if (real_qp0)
177 		return 1;
178 	/* VF or PF -- proxy QP0 */
179 	if (mlx4_is_mfunc(dev->dev)) {
180 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
181 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy) {
182 				proxy_qp0 = 1;
183 				break;
184 			}
185 		}
186 	}
187 	return proxy_qp0;
188 }
189 
190 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
191 {
192 	return mlx4_buf_offset(&qp->buf, offset);
193 }
194 
195 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
196 {
197 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
198 }
199 
200 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
201 {
202 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
203 }
204 
205 /*
206  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
207  * first four bytes of every 64 byte chunk with 0xffffffff, except for
208  * the very first chunk of the WQE.
209  */
210 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n)
211 {
212 	__be32 *wqe;
213 	int i;
214 	int s;
215 	void *buf;
216 	struct mlx4_wqe_ctrl_seg *ctrl;
217 
218 	buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
219 	ctrl = (struct mlx4_wqe_ctrl_seg *)buf;
220 	s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4;
221 	for (i = 64; i < s; i += 64) {
222 		wqe = buf + i;
223 		*wqe = cpu_to_be32(0xffffffff);
224 	}
225 }
226 
227 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
228 {
229 	struct ib_event event;
230 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
231 
232 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
233 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
234 
235 	if (ibqp->event_handler) {
236 		event.device     = ibqp->device;
237 		event.element.qp = ibqp;
238 		switch (type) {
239 		case MLX4_EVENT_TYPE_PATH_MIG:
240 			event.event = IB_EVENT_PATH_MIG;
241 			break;
242 		case MLX4_EVENT_TYPE_COMM_EST:
243 			event.event = IB_EVENT_COMM_EST;
244 			break;
245 		case MLX4_EVENT_TYPE_SQ_DRAINED:
246 			event.event = IB_EVENT_SQ_DRAINED;
247 			break;
248 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
249 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
250 			break;
251 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
252 			event.event = IB_EVENT_QP_FATAL;
253 			break;
254 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
255 			event.event = IB_EVENT_PATH_MIG_ERR;
256 			break;
257 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
258 			event.event = IB_EVENT_QP_REQ_ERR;
259 			break;
260 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
261 			event.event = IB_EVENT_QP_ACCESS_ERR;
262 			break;
263 		default:
264 			pr_warn("Unexpected event type %d "
265 			       "on QP %06x\n", type, qp->qpn);
266 			return;
267 		}
268 
269 		ibqp->event_handler(&event, ibqp->qp_context);
270 	}
271 }
272 
273 static void mlx4_ib_wq_event(struct mlx4_qp *qp, enum mlx4_event type)
274 {
275 	pr_warn_ratelimited("Unexpected event type %d on WQ 0x%06x. Events are not supported for WQs\n",
276 			    type, qp->qpn);
277 }
278 
279 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
280 {
281 	/*
282 	 * UD WQEs must have a datagram segment.
283 	 * RC and UC WQEs might have a remote address segment.
284 	 * MLX WQEs need two extra inline data segments (for the UD
285 	 * header and space for the ICRC).
286 	 */
287 	switch (type) {
288 	case MLX4_IB_QPT_UD:
289 		return sizeof (struct mlx4_wqe_ctrl_seg) +
290 			sizeof (struct mlx4_wqe_datagram_seg) +
291 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
292 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
293 	case MLX4_IB_QPT_PROXY_SMI:
294 	case MLX4_IB_QPT_PROXY_GSI:
295 		return sizeof (struct mlx4_wqe_ctrl_seg) +
296 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
297 	case MLX4_IB_QPT_TUN_SMI_OWNER:
298 	case MLX4_IB_QPT_TUN_GSI:
299 		return sizeof (struct mlx4_wqe_ctrl_seg) +
300 			sizeof (struct mlx4_wqe_datagram_seg);
301 
302 	case MLX4_IB_QPT_UC:
303 		return sizeof (struct mlx4_wqe_ctrl_seg) +
304 			sizeof (struct mlx4_wqe_raddr_seg);
305 	case MLX4_IB_QPT_RC:
306 		return sizeof (struct mlx4_wqe_ctrl_seg) +
307 			sizeof (struct mlx4_wqe_masked_atomic_seg) +
308 			sizeof (struct mlx4_wqe_raddr_seg);
309 	case MLX4_IB_QPT_SMI:
310 	case MLX4_IB_QPT_GSI:
311 		return sizeof (struct mlx4_wqe_ctrl_seg) +
312 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
313 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
314 					   MLX4_INLINE_ALIGN) *
315 			      sizeof (struct mlx4_wqe_inline_seg),
316 			      sizeof (struct mlx4_wqe_data_seg)) +
317 			ALIGN(4 +
318 			      sizeof (struct mlx4_wqe_inline_seg),
319 			      sizeof (struct mlx4_wqe_data_seg));
320 	default:
321 		return sizeof (struct mlx4_wqe_ctrl_seg);
322 	}
323 }
324 
325 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
326 		       bool is_user, int has_rq, struct mlx4_ib_qp *qp,
327 		       u32 inl_recv_sz)
328 {
329 	/* Sanity check RQ size before proceeding */
330 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
331 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
332 		return -EINVAL;
333 
334 	if (!has_rq) {
335 		if (cap->max_recv_wr || inl_recv_sz)
336 			return -EINVAL;
337 
338 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
339 	} else {
340 		u32 max_inl_recv_sz = dev->dev->caps.max_rq_sg *
341 			sizeof(struct mlx4_wqe_data_seg);
342 		u32 wqe_size;
343 
344 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
345 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge ||
346 				inl_recv_sz > max_inl_recv_sz))
347 			return -EINVAL;
348 
349 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
350 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
351 		wqe_size = qp->rq.max_gs * sizeof(struct mlx4_wqe_data_seg);
352 		qp->rq.wqe_shift = ilog2(max_t(u32, wqe_size, inl_recv_sz));
353 	}
354 
355 	/* leave userspace return values as they were, so as not to break ABI */
356 	if (is_user) {
357 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
358 		cap->max_recv_sge = qp->rq.max_gs;
359 	} else {
360 		cap->max_recv_wr  = qp->rq.max_post =
361 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
362 		cap->max_recv_sge = min(qp->rq.max_gs,
363 					min(dev->dev->caps.max_sq_sg,
364 					    dev->dev->caps.max_rq_sg));
365 	}
366 
367 	return 0;
368 }
369 
370 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
371 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
372 {
373 	int s;
374 
375 	/* Sanity check SQ size before proceeding */
376 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
377 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
378 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
379 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
380 		return -EINVAL;
381 
382 	/*
383 	 * For MLX transport we need 2 extra S/G entries:
384 	 * one for the header and one for the checksum at the end
385 	 */
386 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
387 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
388 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
389 		return -EINVAL;
390 
391 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
392 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
393 		send_wqe_overhead(type, qp->flags);
394 
395 	if (s > dev->dev->caps.max_sq_desc_sz)
396 		return -EINVAL;
397 
398 	qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
399 
400 	/*
401 	 * We need to leave 2 KB + 1 WR of headroom in the SQ to
402 	 * allow HW to prefetch.
403 	 */
404 	qp->sq_spare_wqes = MLX4_IB_SQ_HEADROOM(qp->sq.wqe_shift);
405 	qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr +
406 					    qp->sq_spare_wqes);
407 
408 	qp->sq.max_gs =
409 		(min(dev->dev->caps.max_sq_desc_sz,
410 		     (1 << qp->sq.wqe_shift)) -
411 		 send_wqe_overhead(type, qp->flags)) /
412 		sizeof (struct mlx4_wqe_data_seg);
413 
414 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
415 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
416 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
417 		qp->rq.offset = 0;
418 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
419 	} else {
420 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
421 		qp->sq.offset = 0;
422 	}
423 
424 	cap->max_send_wr  = qp->sq.max_post =
425 		qp->sq.wqe_cnt - qp->sq_spare_wqes;
426 	cap->max_send_sge = min(qp->sq.max_gs,
427 				min(dev->dev->caps.max_sq_sg,
428 				    dev->dev->caps.max_rq_sg));
429 	/* We don't support inline sends for kernel QPs (yet) */
430 	cap->max_inline_data = 0;
431 
432 	return 0;
433 }
434 
435 static int set_user_sq_size(struct mlx4_ib_dev *dev,
436 			    struct mlx4_ib_qp *qp,
437 			    struct mlx4_ib_create_qp *ucmd)
438 {
439 	/* Sanity check SQ size before proceeding */
440 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
441 	    ucmd->log_sq_stride >
442 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
443 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
444 		return -EINVAL;
445 
446 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
447 	qp->sq.wqe_shift = ucmd->log_sq_stride;
448 
449 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
450 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
451 
452 	return 0;
453 }
454 
455 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
456 {
457 	int i;
458 
459 	qp->sqp_proxy_rcv =
460 		kmalloc_array(qp->rq.wqe_cnt, sizeof(struct mlx4_ib_buf),
461 			      GFP_KERNEL);
462 	if (!qp->sqp_proxy_rcv)
463 		return -ENOMEM;
464 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
465 		qp->sqp_proxy_rcv[i].addr =
466 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
467 				GFP_KERNEL);
468 		if (!qp->sqp_proxy_rcv[i].addr)
469 			goto err;
470 		qp->sqp_proxy_rcv[i].map =
471 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
472 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
473 					  DMA_FROM_DEVICE);
474 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
475 			kfree(qp->sqp_proxy_rcv[i].addr);
476 			goto err;
477 		}
478 	}
479 	return 0;
480 
481 err:
482 	while (i > 0) {
483 		--i;
484 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
485 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
486 				    DMA_FROM_DEVICE);
487 		kfree(qp->sqp_proxy_rcv[i].addr);
488 	}
489 	kfree(qp->sqp_proxy_rcv);
490 	qp->sqp_proxy_rcv = NULL;
491 	return -ENOMEM;
492 }
493 
494 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
495 {
496 	int i;
497 
498 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
499 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
500 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
501 				    DMA_FROM_DEVICE);
502 		kfree(qp->sqp_proxy_rcv[i].addr);
503 	}
504 	kfree(qp->sqp_proxy_rcv);
505 }
506 
507 static int qp_has_rq(struct ib_qp_init_attr *attr)
508 {
509 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
510 		return 0;
511 
512 	return !attr->srq;
513 }
514 
515 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
516 {
517 	int i;
518 	for (i = 0; i < dev->caps.num_ports; i++) {
519 		if (qpn == dev->caps.spec_qps[i].qp0_proxy)
520 			return !!dev->caps.spec_qps[i].qp0_qkey;
521 	}
522 	return 0;
523 }
524 
525 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
526 				    struct mlx4_ib_qp *qp)
527 {
528 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
529 	mlx4_counter_free(dev->dev, qp->counter_index->index);
530 	list_del(&qp->counter_index->list);
531 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
532 
533 	kfree(qp->counter_index);
534 	qp->counter_index = NULL;
535 }
536 
537 static int set_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_rss *rss_ctx,
538 		      struct ib_qp_init_attr *init_attr,
539 		      struct mlx4_ib_create_qp_rss *ucmd)
540 {
541 	rss_ctx->base_qpn_tbl_sz = init_attr->rwq_ind_tbl->ind_tbl[0]->wq_num |
542 		(init_attr->rwq_ind_tbl->log_ind_tbl_size << 24);
543 
544 	if ((ucmd->rx_hash_function == MLX4_IB_RX_HASH_FUNC_TOEPLITZ) &&
545 	    (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS_TOP)) {
546 		memcpy(rss_ctx->rss_key, ucmd->rx_hash_key,
547 		       MLX4_EN_RSS_KEY_SIZE);
548 	} else {
549 		pr_debug("RX Hash function is not supported\n");
550 		return (-EOPNOTSUPP);
551 	}
552 
553 	if (ucmd->rx_hash_fields_mask & ~(MLX4_IB_RX_HASH_SRC_IPV4	|
554 					  MLX4_IB_RX_HASH_DST_IPV4	|
555 					  MLX4_IB_RX_HASH_SRC_IPV6	|
556 					  MLX4_IB_RX_HASH_DST_IPV6	|
557 					  MLX4_IB_RX_HASH_SRC_PORT_TCP	|
558 					  MLX4_IB_RX_HASH_DST_PORT_TCP	|
559 					  MLX4_IB_RX_HASH_SRC_PORT_UDP	|
560 					  MLX4_IB_RX_HASH_DST_PORT_UDP  |
561 					  MLX4_IB_RX_HASH_INNER)) {
562 		pr_debug("RX Hash fields_mask has unsupported mask (0x%llx)\n",
563 			 ucmd->rx_hash_fields_mask);
564 		return (-EOPNOTSUPP);
565 	}
566 
567 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) &&
568 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
569 		rss_ctx->flags = MLX4_RSS_IPV4;
570 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) ||
571 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
572 		pr_debug("RX Hash fields_mask is not supported - both IPv4 SRC and DST must be set\n");
573 		return (-EOPNOTSUPP);
574 	}
575 
576 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) &&
577 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
578 		rss_ctx->flags |= MLX4_RSS_IPV6;
579 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) ||
580 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
581 		pr_debug("RX Hash fields_mask is not supported - both IPv6 SRC and DST must be set\n");
582 		return (-EOPNOTSUPP);
583 	}
584 
585 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) &&
586 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
587 		if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS)) {
588 			pr_debug("RX Hash fields_mask for UDP is not supported\n");
589 			return (-EOPNOTSUPP);
590 		}
591 
592 		if (rss_ctx->flags & MLX4_RSS_IPV4)
593 			rss_ctx->flags |= MLX4_RSS_UDP_IPV4;
594 		if (rss_ctx->flags & MLX4_RSS_IPV6)
595 			rss_ctx->flags |= MLX4_RSS_UDP_IPV6;
596 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
597 			pr_debug("RX Hash fields_mask is not supported - UDP must be set with IPv4 or IPv6\n");
598 			return (-EOPNOTSUPP);
599 		}
600 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) ||
601 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
602 		pr_debug("RX Hash fields_mask is not supported - both UDP SRC and DST must be set\n");
603 		return (-EOPNOTSUPP);
604 	}
605 
606 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) &&
607 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
608 		if (rss_ctx->flags & MLX4_RSS_IPV4)
609 			rss_ctx->flags |= MLX4_RSS_TCP_IPV4;
610 		if (rss_ctx->flags & MLX4_RSS_IPV6)
611 			rss_ctx->flags |= MLX4_RSS_TCP_IPV6;
612 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
613 			pr_debug("RX Hash fields_mask is not supported - TCP must be set with IPv4 or IPv6\n");
614 			return (-EOPNOTSUPP);
615 		}
616 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) ||
617 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
618 		pr_debug("RX Hash fields_mask is not supported - both TCP SRC and DST must be set\n");
619 		return (-EOPNOTSUPP);
620 	}
621 
622 	if (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_INNER) {
623 		if (dev->dev->caps.tunnel_offload_mode ==
624 		    MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
625 			/*
626 			 * Hash according to inner headers if exist, otherwise
627 			 * according to outer headers.
628 			 */
629 			rss_ctx->flags |= MLX4_RSS_BY_INNER_HEADERS_IPONLY;
630 		} else {
631 			pr_debug("RSS Hash for inner headers isn't supported\n");
632 			return (-EOPNOTSUPP);
633 		}
634 	}
635 
636 	return 0;
637 }
638 
639 static int create_qp_rss(struct mlx4_ib_dev *dev,
640 			 struct ib_qp_init_attr *init_attr,
641 			 struct mlx4_ib_create_qp_rss *ucmd,
642 			 struct mlx4_ib_qp *qp)
643 {
644 	int qpn;
645 	int err;
646 
647 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
648 
649 	err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 0, qp->mqp.usage);
650 	if (err)
651 		return err;
652 
653 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
654 	if (err)
655 		goto err_qpn;
656 
657 	mutex_init(&qp->mutex);
658 
659 	INIT_LIST_HEAD(&qp->gid_list);
660 	INIT_LIST_HEAD(&qp->steering_rules);
661 
662 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
663 	qp->state = IB_QPS_RESET;
664 
665 	/* Set dummy send resources to be compatible with HV and PRM */
666 	qp->sq_no_prefetch = 1;
667 	qp->sq.wqe_cnt = 1;
668 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
669 	qp->buf_size = qp->sq.wqe_cnt << MLX4_IB_MIN_SQ_STRIDE;
670 	qp->mtt = (to_mqp(
671 		   (struct ib_qp *)init_attr->rwq_ind_tbl->ind_tbl[0]))->mtt;
672 
673 	qp->rss_ctx = kzalloc(sizeof(*qp->rss_ctx), GFP_KERNEL);
674 	if (!qp->rss_ctx) {
675 		err = -ENOMEM;
676 		goto err_qp_alloc;
677 	}
678 
679 	err = set_qp_rss(dev, qp->rss_ctx, init_attr, ucmd);
680 	if (err)
681 		goto err;
682 
683 	return 0;
684 
685 err:
686 	kfree(qp->rss_ctx);
687 
688 err_qp_alloc:
689 	mlx4_qp_remove(dev->dev, &qp->mqp);
690 	mlx4_qp_free(dev->dev, &qp->mqp);
691 
692 err_qpn:
693 	mlx4_qp_release_range(dev->dev, qpn, 1);
694 	return err;
695 }
696 
697 static struct ib_qp *_mlx4_ib_create_qp_rss(struct ib_pd *pd,
698 					    struct ib_qp_init_attr *init_attr,
699 					    struct ib_udata *udata)
700 {
701 	struct mlx4_ib_qp *qp;
702 	struct mlx4_ib_create_qp_rss ucmd = {};
703 	size_t required_cmd_sz;
704 	int err;
705 
706 	if (!udata) {
707 		pr_debug("RSS QP with NULL udata\n");
708 		return ERR_PTR(-EINVAL);
709 	}
710 
711 	if (udata->outlen)
712 		return ERR_PTR(-EOPNOTSUPP);
713 
714 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) +
715 					sizeof(ucmd.reserved1);
716 	if (udata->inlen < required_cmd_sz) {
717 		pr_debug("invalid inlen\n");
718 		return ERR_PTR(-EINVAL);
719 	}
720 
721 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
722 		pr_debug("copy failed\n");
723 		return ERR_PTR(-EFAULT);
724 	}
725 
726 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)))
727 		return ERR_PTR(-EOPNOTSUPP);
728 
729 	if (ucmd.comp_mask || ucmd.reserved1)
730 		return ERR_PTR(-EOPNOTSUPP);
731 
732 	if (udata->inlen > sizeof(ucmd) &&
733 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
734 				 udata->inlen - sizeof(ucmd))) {
735 		pr_debug("inlen is not supported\n");
736 		return ERR_PTR(-EOPNOTSUPP);
737 	}
738 
739 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
740 		pr_debug("RSS QP with unsupported QP type %d\n",
741 			 init_attr->qp_type);
742 		return ERR_PTR(-EOPNOTSUPP);
743 	}
744 
745 	if (init_attr->create_flags) {
746 		pr_debug("RSS QP doesn't support create flags\n");
747 		return ERR_PTR(-EOPNOTSUPP);
748 	}
749 
750 	if (init_attr->send_cq || init_attr->cap.max_send_wr) {
751 		pr_debug("RSS QP with unsupported send attributes\n");
752 		return ERR_PTR(-EOPNOTSUPP);
753 	}
754 
755 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
756 	if (!qp)
757 		return ERR_PTR(-ENOMEM);
758 
759 	qp->pri.vid = 0xFFFF;
760 	qp->alt.vid = 0xFFFF;
761 
762 	err = create_qp_rss(to_mdev(pd->device), init_attr, &ucmd, qp);
763 	if (err) {
764 		kfree(qp);
765 		return ERR_PTR(err);
766 	}
767 
768 	qp->ibqp.qp_num = qp->mqp.qpn;
769 
770 	return &qp->ibqp;
771 }
772 
773 /*
774  * This function allocates a WQN from a range which is consecutive and aligned
775  * to its size. In case the range is full, then it creates a new range and
776  * allocates WQN from it. The new range will be used for following allocations.
777  */
778 static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context,
779 			     struct mlx4_ib_qp *qp, int range_size, int *wqn)
780 {
781 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
782 	struct mlx4_wqn_range *range;
783 	int err = 0;
784 
785 	mutex_lock(&context->wqn_ranges_mutex);
786 
787 	range = list_first_entry_or_null(&context->wqn_ranges_list,
788 					 struct mlx4_wqn_range, list);
789 
790 	if (!range || (range->refcount == range->size) || range->dirty) {
791 		range = kzalloc(sizeof(*range), GFP_KERNEL);
792 		if (!range) {
793 			err = -ENOMEM;
794 			goto out;
795 		}
796 
797 		err = mlx4_qp_reserve_range(dev->dev, range_size,
798 					    range_size, &range->base_wqn, 0,
799 					    qp->mqp.usage);
800 		if (err) {
801 			kfree(range);
802 			goto out;
803 		}
804 
805 		range->size = range_size;
806 		list_add(&range->list, &context->wqn_ranges_list);
807 	} else if (range_size != 1) {
808 		/*
809 		 * Requesting a new range (>1) when last range is still open, is
810 		 * not valid.
811 		 */
812 		err = -EINVAL;
813 		goto out;
814 	}
815 
816 	qp->wqn_range = range;
817 
818 	*wqn = range->base_wqn + range->refcount;
819 
820 	range->refcount++;
821 
822 out:
823 	mutex_unlock(&context->wqn_ranges_mutex);
824 
825 	return err;
826 }
827 
828 static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context,
829 				struct mlx4_ib_qp *qp, bool dirty_release)
830 {
831 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
832 	struct mlx4_wqn_range *range;
833 
834 	mutex_lock(&context->wqn_ranges_mutex);
835 
836 	range = qp->wqn_range;
837 
838 	range->refcount--;
839 	if (!range->refcount) {
840 		mlx4_qp_release_range(dev->dev, range->base_wqn,
841 				      range->size);
842 		list_del(&range->list);
843 		kfree(range);
844 	} else if (dirty_release) {
845 	/*
846 	 * A range which one of its WQNs is destroyed, won't be able to be
847 	 * reused for further WQN allocations.
848 	 * The next created WQ will allocate a new range.
849 	 */
850 		range->dirty = 1;
851 	}
852 
853 	mutex_unlock(&context->wqn_ranges_mutex);
854 }
855 
856 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
857 			    enum mlx4_ib_source_type src,
858 			    struct ib_qp_init_attr *init_attr,
859 			    struct ib_udata *udata, int sqpn,
860 			    struct mlx4_ib_qp **caller_qp)
861 {
862 	int qpn;
863 	int err;
864 	struct mlx4_ib_sqp *sqp = NULL;
865 	struct mlx4_ib_qp *qp;
866 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
867 	struct mlx4_ib_cq *mcq;
868 	unsigned long flags;
869 	int range_size = 0;
870 
871 	/* When tunneling special qps, we use a plain UD qp */
872 	if (sqpn) {
873 		if (mlx4_is_mfunc(dev->dev) &&
874 		    (!mlx4_is_master(dev->dev) ||
875 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
876 			if (init_attr->qp_type == IB_QPT_GSI)
877 				qp_type = MLX4_IB_QPT_PROXY_GSI;
878 			else {
879 				if (mlx4_is_master(dev->dev) ||
880 				    qp0_enabled_vf(dev->dev, sqpn))
881 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
882 				else
883 					qp_type = MLX4_IB_QPT_PROXY_SMI;
884 			}
885 		}
886 		qpn = sqpn;
887 		/* add extra sg entry for tunneling */
888 		init_attr->cap.max_recv_sge++;
889 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
890 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
891 			container_of(init_attr,
892 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
893 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
894 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
895 		    !mlx4_is_master(dev->dev))
896 			return -EINVAL;
897 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
898 			qp_type = MLX4_IB_QPT_TUN_GSI;
899 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
900 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
901 					     tnl_init->port))
902 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
903 		else
904 			qp_type = MLX4_IB_QPT_TUN_SMI;
905 		/* we are definitely in the PPF here, since we are creating
906 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
907 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
908 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
909 		sqpn = qpn;
910 	}
911 
912 	if (!*caller_qp) {
913 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
914 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
915 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
916 			sqp = kzalloc(sizeof(struct mlx4_ib_sqp), GFP_KERNEL);
917 			if (!sqp)
918 				return -ENOMEM;
919 			qp = &sqp->qp;
920 			qp->pri.vid = 0xFFFF;
921 			qp->alt.vid = 0xFFFF;
922 		} else {
923 			qp = kzalloc(sizeof(struct mlx4_ib_qp), GFP_KERNEL);
924 			if (!qp)
925 				return -ENOMEM;
926 			qp->pri.vid = 0xFFFF;
927 			qp->alt.vid = 0xFFFF;
928 		}
929 	} else
930 		qp = *caller_qp;
931 
932 	qp->mlx4_ib_qp_type = qp_type;
933 
934 	mutex_init(&qp->mutex);
935 	spin_lock_init(&qp->sq.lock);
936 	spin_lock_init(&qp->rq.lock);
937 	INIT_LIST_HEAD(&qp->gid_list);
938 	INIT_LIST_HEAD(&qp->steering_rules);
939 
940 	qp->state	 = IB_QPS_RESET;
941 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
942 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
943 
944 
945 	if (udata) {
946 		union {
947 			struct mlx4_ib_create_qp qp;
948 			struct mlx4_ib_create_wq wq;
949 		} ucmd;
950 		size_t copy_len;
951 		int shift;
952 		int n;
953 
954 		copy_len = (src == MLX4_IB_QP_SRC) ?
955 			   sizeof(struct mlx4_ib_create_qp) :
956 			   min(sizeof(struct mlx4_ib_create_wq), udata->inlen);
957 
958 		if (ib_copy_from_udata(&ucmd, udata, copy_len)) {
959 			err = -EFAULT;
960 			goto err;
961 		}
962 
963 		if (src == MLX4_IB_RWQ_SRC) {
964 			if (ucmd.wq.comp_mask || ucmd.wq.reserved[0] ||
965 			    ucmd.wq.reserved[1] || ucmd.wq.reserved[2]) {
966 				pr_debug("user command isn't supported\n");
967 				err = -EOPNOTSUPP;
968 				goto err;
969 			}
970 
971 			if (ucmd.wq.log_range_size >
972 			    ilog2(dev->dev->caps.max_rss_tbl_sz)) {
973 				pr_debug("WQN range size must be equal or smaller than %d\n",
974 					 dev->dev->caps.max_rss_tbl_sz);
975 				err = -EOPNOTSUPP;
976 				goto err;
977 			}
978 			range_size = 1 << ucmd.wq.log_range_size;
979 		} else {
980 			qp->inl_recv_sz = ucmd.qp.inl_recv_sz;
981 		}
982 
983 		if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
984 			if (!(dev->dev->caps.flags &
985 			      MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
986 				pr_debug("scatter FCS is unsupported\n");
987 				err = -EOPNOTSUPP;
988 				goto err;
989 			}
990 
991 			qp->flags |= MLX4_IB_QP_SCATTER_FCS;
992 		}
993 
994 		err = set_rq_size(dev, &init_attr->cap, udata,
995 				  qp_has_rq(init_attr), qp, qp->inl_recv_sz);
996 		if (err)
997 			goto err;
998 
999 		if (src == MLX4_IB_QP_SRC) {
1000 			qp->sq_no_prefetch = ucmd.qp.sq_no_prefetch;
1001 
1002 			err = set_user_sq_size(dev, qp,
1003 					       (struct mlx4_ib_create_qp *)
1004 					       &ucmd);
1005 			if (err)
1006 				goto err;
1007 		} else {
1008 			qp->sq_no_prefetch = 1;
1009 			qp->sq.wqe_cnt = 1;
1010 			qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
1011 			/* Allocated buffer expects to have at least that SQ
1012 			 * size.
1013 			 */
1014 			qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
1015 				(qp->sq.wqe_cnt << qp->sq.wqe_shift);
1016 		}
1017 
1018 		qp->umem = ib_umem_get(pd->uobject->context,
1019 				(src == MLX4_IB_QP_SRC) ? ucmd.qp.buf_addr :
1020 				ucmd.wq.buf_addr, qp->buf_size, 0, 0);
1021 		if (IS_ERR(qp->umem)) {
1022 			err = PTR_ERR(qp->umem);
1023 			goto err;
1024 		}
1025 
1026 		n = ib_umem_page_count(qp->umem);
1027 		shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
1028 		err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
1029 
1030 		if (err)
1031 			goto err_buf;
1032 
1033 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
1034 		if (err)
1035 			goto err_mtt;
1036 
1037 		if (qp_has_rq(init_attr)) {
1038 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
1039 				(src == MLX4_IB_QP_SRC) ? ucmd.qp.db_addr :
1040 				ucmd.wq.db_addr, &qp->db);
1041 			if (err)
1042 				goto err_mtt;
1043 		}
1044 		qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
1045 	} else {
1046 		err = set_rq_size(dev, &init_attr->cap, udata,
1047 				  qp_has_rq(init_attr), qp, 0);
1048 		if (err)
1049 			goto err;
1050 
1051 		qp->sq_no_prefetch = 0;
1052 
1053 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
1054 			qp->flags |= MLX4_IB_QP_LSO;
1055 
1056 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1057 			if (dev->steering_support ==
1058 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
1059 				qp->flags |= MLX4_IB_QP_NETIF;
1060 			else
1061 				goto err;
1062 		}
1063 
1064 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
1065 		if (err)
1066 			goto err;
1067 
1068 		if (qp_has_rq(init_attr)) {
1069 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
1070 			if (err)
1071 				goto err;
1072 
1073 			*qp->db.db = 0;
1074 		}
1075 
1076 		if (mlx4_buf_alloc(dev->dev, qp->buf_size,  PAGE_SIZE * 2,
1077 				   &qp->buf)) {
1078 			err = -ENOMEM;
1079 			goto err_db;
1080 		}
1081 
1082 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
1083 				    &qp->mtt);
1084 		if (err)
1085 			goto err_buf;
1086 
1087 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
1088 		if (err)
1089 			goto err_mtt;
1090 
1091 		qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1092 					     sizeof(u64), GFP_KERNEL);
1093 		qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1094 					     sizeof(u64), GFP_KERNEL);
1095 		if (!qp->sq.wrid || !qp->rq.wrid) {
1096 			err = -ENOMEM;
1097 			goto err_wrid;
1098 		}
1099 		qp->mqp.usage = MLX4_RES_USAGE_DRIVER;
1100 	}
1101 
1102 	if (sqpn) {
1103 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1104 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
1105 			if (alloc_proxy_bufs(pd->device, qp)) {
1106 				err = -ENOMEM;
1107 				goto err_wrid;
1108 			}
1109 		}
1110 	} else if (src == MLX4_IB_RWQ_SRC) {
1111 		err = mlx4_ib_alloc_wqn(to_mucontext(pd->uobject->context), qp,
1112 					range_size, &qpn);
1113 		if (err)
1114 			goto err_wrid;
1115 	} else {
1116 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
1117 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
1118 		 * VLAN insertion. */
1119 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
1120 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
1121 						    (init_attr->cap.max_send_wr ?
1122 						     MLX4_RESERVE_ETH_BF_QP : 0) |
1123 						    (init_attr->cap.max_recv_wr ?
1124 						     MLX4_RESERVE_A0_QP : 0),
1125 						    qp->mqp.usage);
1126 		else
1127 			if (qp->flags & MLX4_IB_QP_NETIF)
1128 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
1129 			else
1130 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
1131 							    &qpn, 0, qp->mqp.usage);
1132 		if (err)
1133 			goto err_proxy;
1134 	}
1135 
1136 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1137 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1138 
1139 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
1140 	if (err)
1141 		goto err_qpn;
1142 
1143 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1144 		qp->mqp.qpn |= (1 << 23);
1145 
1146 	/*
1147 	 * Hardware wants QPN written in big-endian order (after
1148 	 * shifting) for send doorbell.  Precompute this value to save
1149 	 * a little bit when posting sends.
1150 	 */
1151 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1152 
1153 	qp->mqp.event = (src == MLX4_IB_QP_SRC) ? mlx4_ib_qp_event :
1154 						  mlx4_ib_wq_event;
1155 
1156 	if (!*caller_qp)
1157 		*caller_qp = qp;
1158 
1159 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1160 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
1161 			 to_mcq(init_attr->recv_cq));
1162 	/* Maintain device to QPs access, needed for further handling
1163 	 * via reset flow
1164 	 */
1165 	list_add_tail(&qp->qps_list, &dev->qp_list);
1166 	/* Maintain CQ to QPs access, needed for further handling
1167 	 * via reset flow
1168 	 */
1169 	mcq = to_mcq(init_attr->send_cq);
1170 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
1171 	mcq = to_mcq(init_attr->recv_cq);
1172 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
1173 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
1174 			   to_mcq(init_attr->recv_cq));
1175 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1176 	return 0;
1177 
1178 err_qpn:
1179 	if (!sqpn) {
1180 		if (qp->flags & MLX4_IB_QP_NETIF)
1181 			mlx4_ib_steer_qp_free(dev, qpn, 1);
1182 		else if (src == MLX4_IB_RWQ_SRC)
1183 			mlx4_ib_release_wqn(to_mucontext(pd->uobject->context),
1184 					    qp, 0);
1185 		else
1186 			mlx4_qp_release_range(dev->dev, qpn, 1);
1187 	}
1188 err_proxy:
1189 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1190 		free_proxy_bufs(pd->device, qp);
1191 err_wrid:
1192 	if (udata) {
1193 		if (qp_has_rq(init_attr))
1194 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
1195 	} else {
1196 		kvfree(qp->sq.wrid);
1197 		kvfree(qp->rq.wrid);
1198 	}
1199 
1200 err_mtt:
1201 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1202 
1203 err_buf:
1204 	if (qp->umem)
1205 		ib_umem_release(qp->umem);
1206 	else
1207 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1208 
1209 err_db:
1210 	if (!udata && qp_has_rq(init_attr))
1211 		mlx4_db_free(dev->dev, &qp->db);
1212 
1213 err:
1214 	if (!sqp && !*caller_qp)
1215 		kfree(qp);
1216 	kfree(sqp);
1217 
1218 	return err;
1219 }
1220 
1221 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
1222 {
1223 	switch (state) {
1224 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
1225 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
1226 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
1227 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
1228 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
1229 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
1230 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
1231 	default:		return -1;
1232 	}
1233 }
1234 
1235 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1236 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1237 {
1238 	if (send_cq == recv_cq) {
1239 		spin_lock(&send_cq->lock);
1240 		__acquire(&recv_cq->lock);
1241 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1242 		spin_lock(&send_cq->lock);
1243 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1244 	} else {
1245 		spin_lock(&recv_cq->lock);
1246 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1247 	}
1248 }
1249 
1250 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1251 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1252 {
1253 	if (send_cq == recv_cq) {
1254 		__release(&recv_cq->lock);
1255 		spin_unlock(&send_cq->lock);
1256 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1257 		spin_unlock(&recv_cq->lock);
1258 		spin_unlock(&send_cq->lock);
1259 	} else {
1260 		spin_unlock(&send_cq->lock);
1261 		spin_unlock(&recv_cq->lock);
1262 	}
1263 }
1264 
1265 static void del_gid_entries(struct mlx4_ib_qp *qp)
1266 {
1267 	struct mlx4_ib_gid_entry *ge, *tmp;
1268 
1269 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1270 		list_del(&ge->list);
1271 		kfree(ge);
1272 	}
1273 }
1274 
1275 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
1276 {
1277 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
1278 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
1279 	else
1280 		return to_mpd(qp->ibqp.pd);
1281 }
1282 
1283 static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src,
1284 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1285 {
1286 	switch (qp->ibqp.qp_type) {
1287 	case IB_QPT_XRC_TGT:
1288 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1289 		*recv_cq = *send_cq;
1290 		break;
1291 	case IB_QPT_XRC_INI:
1292 		*send_cq = to_mcq(qp->ibqp.send_cq);
1293 		*recv_cq = *send_cq;
1294 		break;
1295 	default:
1296 		*recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) :
1297 						     to_mcq(qp->ibwq.cq);
1298 		*send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) :
1299 						     *recv_cq;
1300 		break;
1301 	}
1302 }
1303 
1304 static void destroy_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1305 {
1306 	if (qp->state != IB_QPS_RESET) {
1307 		int i;
1308 
1309 		for (i = 0; i < (1 << qp->ibqp.rwq_ind_tbl->log_ind_tbl_size);
1310 		     i++) {
1311 			struct ib_wq *ibwq = qp->ibqp.rwq_ind_tbl->ind_tbl[i];
1312 			struct mlx4_ib_qp *wq =	to_mqp((struct ib_qp *)ibwq);
1313 
1314 			mutex_lock(&wq->mutex);
1315 
1316 			wq->rss_usecnt--;
1317 
1318 			mutex_unlock(&wq->mutex);
1319 		}
1320 
1321 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1322 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1323 			pr_warn("modify QP %06x to RESET failed.\n",
1324 				qp->mqp.qpn);
1325 	}
1326 
1327 	mlx4_qp_remove(dev->dev, &qp->mqp);
1328 	mlx4_qp_free(dev->dev, &qp->mqp);
1329 	mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1330 	del_gid_entries(qp);
1331 	kfree(qp->rss_ctx);
1332 }
1333 
1334 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1335 			      enum mlx4_ib_source_type src, bool is_user)
1336 {
1337 	struct mlx4_ib_cq *send_cq, *recv_cq;
1338 	unsigned long flags;
1339 
1340 	if (qp->state != IB_QPS_RESET) {
1341 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1342 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1343 			pr_warn("modify QP %06x to RESET failed.\n",
1344 			       qp->mqp.qpn);
1345 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1346 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1347 			qp->pri.smac = 0;
1348 			qp->pri.smac_port = 0;
1349 		}
1350 		if (qp->alt.smac) {
1351 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1352 			qp->alt.smac = 0;
1353 		}
1354 		if (qp->pri.vid < 0x1000) {
1355 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1356 			qp->pri.vid = 0xFFFF;
1357 			qp->pri.candidate_vid = 0xFFFF;
1358 			qp->pri.update_vid = 0;
1359 		}
1360 		if (qp->alt.vid < 0x1000) {
1361 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1362 			qp->alt.vid = 0xFFFF;
1363 			qp->alt.candidate_vid = 0xFFFF;
1364 			qp->alt.update_vid = 0;
1365 		}
1366 	}
1367 
1368 	get_cqs(qp, src, &send_cq, &recv_cq);
1369 
1370 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1371 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1372 
1373 	/* del from lists under both locks above to protect reset flow paths */
1374 	list_del(&qp->qps_list);
1375 	list_del(&qp->cq_send_list);
1376 	list_del(&qp->cq_recv_list);
1377 	if (!is_user) {
1378 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1379 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1380 		if (send_cq != recv_cq)
1381 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1382 	}
1383 
1384 	mlx4_qp_remove(dev->dev, &qp->mqp);
1385 
1386 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1387 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1388 
1389 	mlx4_qp_free(dev->dev, &qp->mqp);
1390 
1391 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1392 		if (qp->flags & MLX4_IB_QP_NETIF)
1393 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1394 		else if (src == MLX4_IB_RWQ_SRC)
1395 			mlx4_ib_release_wqn(to_mucontext(
1396 					    qp->ibwq.uobject->context), qp, 1);
1397 		else
1398 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1399 	}
1400 
1401 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1402 
1403 	if (is_user) {
1404 		if (qp->rq.wqe_cnt) {
1405 			struct mlx4_ib_ucontext *mcontext = !src ?
1406 				to_mucontext(qp->ibqp.uobject->context) :
1407 				to_mucontext(qp->ibwq.uobject->context);
1408 			mlx4_ib_db_unmap_user(mcontext, &qp->db);
1409 		}
1410 		ib_umem_release(qp->umem);
1411 	} else {
1412 		kvfree(qp->sq.wrid);
1413 		kvfree(qp->rq.wrid);
1414 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1415 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1416 			free_proxy_bufs(&dev->ib_dev, qp);
1417 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1418 		if (qp->rq.wqe_cnt)
1419 			mlx4_db_free(dev->dev, &qp->db);
1420 	}
1421 
1422 	del_gid_entries(qp);
1423 }
1424 
1425 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1426 {
1427 	/* Native or PPF */
1428 	if (!mlx4_is_mfunc(dev->dev) ||
1429 	    (mlx4_is_master(dev->dev) &&
1430 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1431 		return  dev->dev->phys_caps.base_sqpn +
1432 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1433 			attr->port_num - 1;
1434 	}
1435 	/* PF or VF -- creating proxies */
1436 	if (attr->qp_type == IB_QPT_SMI)
1437 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp0_proxy;
1438 	else
1439 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp1_proxy;
1440 }
1441 
1442 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1443 					struct ib_qp_init_attr *init_attr,
1444 					struct ib_udata *udata)
1445 {
1446 	struct mlx4_ib_qp *qp = NULL;
1447 	int err;
1448 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1449 	u16 xrcdn = 0;
1450 
1451 	if (init_attr->rwq_ind_tbl)
1452 		return _mlx4_ib_create_qp_rss(pd, init_attr, udata);
1453 
1454 	/*
1455 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1456 	 * and only for kernel UD QPs.
1457 	 */
1458 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1459 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1460 					MLX4_IB_SRIOV_TUNNEL_QP |
1461 					MLX4_IB_SRIOV_SQP |
1462 					MLX4_IB_QP_NETIF |
1463 					MLX4_IB_QP_CREATE_ROCE_V2_GSI))
1464 		return ERR_PTR(-EINVAL);
1465 
1466 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1467 		if (init_attr->qp_type != IB_QPT_UD)
1468 			return ERR_PTR(-EINVAL);
1469 	}
1470 
1471 	if (init_attr->create_flags) {
1472 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1473 			return ERR_PTR(-EINVAL);
1474 
1475 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1476 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1477 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1478 		     init_attr->qp_type != IB_QPT_UD) ||
1479 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1480 		     init_attr->qp_type > IB_QPT_GSI) ||
1481 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1482 		     init_attr->qp_type != IB_QPT_GSI))
1483 			return ERR_PTR(-EINVAL);
1484 	}
1485 
1486 	switch (init_attr->qp_type) {
1487 	case IB_QPT_XRC_TGT:
1488 		pd = to_mxrcd(init_attr->xrcd)->pd;
1489 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1490 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1491 		/* fall through */
1492 	case IB_QPT_XRC_INI:
1493 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1494 			return ERR_PTR(-ENOSYS);
1495 		init_attr->recv_cq = init_attr->send_cq;
1496 		/* fall through */
1497 	case IB_QPT_RC:
1498 	case IB_QPT_UC:
1499 	case IB_QPT_RAW_PACKET:
1500 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1501 		if (!qp)
1502 			return ERR_PTR(-ENOMEM);
1503 		qp->pri.vid = 0xFFFF;
1504 		qp->alt.vid = 0xFFFF;
1505 		/* fall through */
1506 	case IB_QPT_UD:
1507 	{
1508 		err = create_qp_common(to_mdev(pd->device), pd,	MLX4_IB_QP_SRC,
1509 				       init_attr, udata, 0, &qp);
1510 		if (err) {
1511 			kfree(qp);
1512 			return ERR_PTR(err);
1513 		}
1514 
1515 		qp->ibqp.qp_num = qp->mqp.qpn;
1516 		qp->xrcdn = xrcdn;
1517 
1518 		break;
1519 	}
1520 	case IB_QPT_SMI:
1521 	case IB_QPT_GSI:
1522 	{
1523 		int sqpn;
1524 
1525 		/* Userspace is not allowed to create special QPs: */
1526 		if (udata)
1527 			return ERR_PTR(-EINVAL);
1528 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1529 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev,
1530 							1, 1, &sqpn, 0,
1531 							MLX4_RES_USAGE_DRIVER);
1532 
1533 			if (res)
1534 				return ERR_PTR(res);
1535 		} else {
1536 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1537 		}
1538 
1539 		err = create_qp_common(to_mdev(pd->device), pd, MLX4_IB_QP_SRC,
1540 				       init_attr, udata, sqpn, &qp);
1541 		if (err)
1542 			return ERR_PTR(err);
1543 
1544 		qp->port	= init_attr->port_num;
1545 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1546 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1547 		break;
1548 	}
1549 	default:
1550 		/* Don't support raw QPs */
1551 		return ERR_PTR(-EINVAL);
1552 	}
1553 
1554 	return &qp->ibqp;
1555 }
1556 
1557 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1558 				struct ib_qp_init_attr *init_attr,
1559 				struct ib_udata *udata) {
1560 	struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1561 	struct ib_qp *ibqp;
1562 	struct mlx4_ib_dev *dev = to_mdev(device);
1563 
1564 	ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1565 
1566 	if (!IS_ERR(ibqp) &&
1567 	    (init_attr->qp_type == IB_QPT_GSI) &&
1568 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1569 		struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1570 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1571 
1572 		if (is_eth &&
1573 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1574 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1575 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1576 
1577 			if (IS_ERR(sqp->roce_v2_gsi)) {
1578 				pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1579 				sqp->roce_v2_gsi = NULL;
1580 			} else {
1581 				sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1582 				sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1583 			}
1584 
1585 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1586 		}
1587 	}
1588 	return ibqp;
1589 }
1590 
1591 static int _mlx4_ib_destroy_qp(struct ib_qp *qp)
1592 {
1593 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1594 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1595 
1596 	if (is_qp0(dev, mqp))
1597 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1598 
1599 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1600 	    dev->qp1_proxy[mqp->port - 1] == mqp) {
1601 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1602 		dev->qp1_proxy[mqp->port - 1] = NULL;
1603 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1604 	}
1605 
1606 	if (mqp->counter_index)
1607 		mlx4_ib_free_qp_counter(dev, mqp);
1608 
1609 	if (qp->rwq_ind_tbl) {
1610 		destroy_qp_rss(dev, mqp);
1611 	} else {
1612 		destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, qp->uobject);
1613 	}
1614 
1615 	if (is_sqp(dev, mqp))
1616 		kfree(to_msqp(mqp));
1617 	else
1618 		kfree(mqp);
1619 
1620 	return 0;
1621 }
1622 
1623 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1624 {
1625 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1626 
1627 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1628 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1629 
1630 		if (sqp->roce_v2_gsi)
1631 			ib_destroy_qp(sqp->roce_v2_gsi);
1632 	}
1633 
1634 	return _mlx4_ib_destroy_qp(qp);
1635 }
1636 
1637 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1638 {
1639 	switch (type) {
1640 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1641 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1642 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1643 	case MLX4_IB_QPT_XRC_INI:
1644 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1645 	case MLX4_IB_QPT_SMI:
1646 	case MLX4_IB_QPT_GSI:
1647 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1648 
1649 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1650 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1651 						MLX4_QP_ST_MLX : -1);
1652 	case MLX4_IB_QPT_PROXY_SMI:
1653 	case MLX4_IB_QPT_TUN_SMI:
1654 	case MLX4_IB_QPT_PROXY_GSI:
1655 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1656 						MLX4_QP_ST_UD : -1);
1657 	default:			return -1;
1658 	}
1659 }
1660 
1661 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1662 				   int attr_mask)
1663 {
1664 	u8 dest_rd_atomic;
1665 	u32 access_flags;
1666 	u32 hw_access_flags = 0;
1667 
1668 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1669 		dest_rd_atomic = attr->max_dest_rd_atomic;
1670 	else
1671 		dest_rd_atomic = qp->resp_depth;
1672 
1673 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1674 		access_flags = attr->qp_access_flags;
1675 	else
1676 		access_flags = qp->atomic_rd_en;
1677 
1678 	if (!dest_rd_atomic)
1679 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1680 
1681 	if (access_flags & IB_ACCESS_REMOTE_READ)
1682 		hw_access_flags |= MLX4_QP_BIT_RRE;
1683 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1684 		hw_access_flags |= MLX4_QP_BIT_RAE;
1685 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1686 		hw_access_flags |= MLX4_QP_BIT_RWE;
1687 
1688 	return cpu_to_be32(hw_access_flags);
1689 }
1690 
1691 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1692 			    int attr_mask)
1693 {
1694 	if (attr_mask & IB_QP_PKEY_INDEX)
1695 		sqp->pkey_index = attr->pkey_index;
1696 	if (attr_mask & IB_QP_QKEY)
1697 		sqp->qkey = attr->qkey;
1698 	if (attr_mask & IB_QP_SQ_PSN)
1699 		sqp->send_psn = attr->sq_psn;
1700 }
1701 
1702 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1703 {
1704 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1705 }
1706 
1707 static int _mlx4_set_path(struct mlx4_ib_dev *dev,
1708 			  const struct rdma_ah_attr *ah,
1709 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1710 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1711 {
1712 	int vidx;
1713 	int smac_index;
1714 	int err;
1715 
1716 	path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
1717 	path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
1718 	if (rdma_ah_get_static_rate(ah)) {
1719 		path->static_rate = rdma_ah_get_static_rate(ah) +
1720 				    MLX4_STAT_RATE_OFFSET;
1721 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1722 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1723 			--path->static_rate;
1724 	} else
1725 		path->static_rate = 0;
1726 
1727 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
1728 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
1729 		int real_sgid_index =
1730 			mlx4_ib_gid_index_to_real_index(dev, grh->sgid_attr);
1731 
1732 		if (real_sgid_index < 0)
1733 			return real_sgid_index;
1734 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1735 			pr_err("sgid_index (%u) too large. max is %d\n",
1736 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1737 			return -1;
1738 		}
1739 
1740 		path->grh_mylmc |= 1 << 7;
1741 		path->mgid_index = real_sgid_index;
1742 		path->hop_limit  = grh->hop_limit;
1743 		path->tclass_flowlabel =
1744 			cpu_to_be32((grh->traffic_class << 20) |
1745 				    (grh->flow_label));
1746 		memcpy(path->rgid, grh->dgid.raw, 16);
1747 	}
1748 
1749 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
1750 		if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
1751 			return -1;
1752 
1753 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1754 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
1755 
1756 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1757 		if (vlan_tag < 0x1000) {
1758 			if (smac_info->vid < 0x1000) {
1759 				/* both valid vlan ids */
1760 				if (smac_info->vid != vlan_tag) {
1761 					/* different VIDs.  unreg old and reg new */
1762 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1763 					if (err)
1764 						return err;
1765 					smac_info->candidate_vid = vlan_tag;
1766 					smac_info->candidate_vlan_index = vidx;
1767 					smac_info->candidate_vlan_port = port;
1768 					smac_info->update_vid = 1;
1769 					path->vlan_index = vidx;
1770 				} else {
1771 					path->vlan_index = smac_info->vlan_index;
1772 				}
1773 			} else {
1774 				/* no current vlan tag in qp */
1775 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1776 				if (err)
1777 					return err;
1778 				smac_info->candidate_vid = vlan_tag;
1779 				smac_info->candidate_vlan_index = vidx;
1780 				smac_info->candidate_vlan_port = port;
1781 				smac_info->update_vid = 1;
1782 				path->vlan_index = vidx;
1783 			}
1784 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1785 			path->fl = 1 << 6;
1786 		} else {
1787 			/* have current vlan tag. unregister it at modify-qp success */
1788 			if (smac_info->vid < 0x1000) {
1789 				smac_info->candidate_vid = 0xFFFF;
1790 				smac_info->update_vid = 1;
1791 			}
1792 		}
1793 
1794 		/* get smac_index for RoCE use.
1795 		 * If no smac was yet assigned, register one.
1796 		 * If one was already assigned, but the new mac differs,
1797 		 * unregister the old one and register the new one.
1798 		*/
1799 		if ((!smac_info->smac && !smac_info->smac_port) ||
1800 		    smac_info->smac != smac) {
1801 			/* register candidate now, unreg if needed, after success */
1802 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1803 			if (smac_index >= 0) {
1804 				smac_info->candidate_smac_index = smac_index;
1805 				smac_info->candidate_smac = smac;
1806 				smac_info->candidate_smac_port = port;
1807 			} else {
1808 				return -EINVAL;
1809 			}
1810 		} else {
1811 			smac_index = smac_info->smac_index;
1812 		}
1813 		memcpy(path->dmac, ah->roce.dmac, 6);
1814 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1815 		/* put MAC table smac index for IBoE */
1816 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1817 	} else {
1818 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1819 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
1820 	}
1821 
1822 	return 0;
1823 }
1824 
1825 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1826 			 enum ib_qp_attr_mask qp_attr_mask,
1827 			 struct mlx4_ib_qp *mqp,
1828 			 struct mlx4_qp_path *path, u8 port,
1829 			 u16 vlan_id, u8 *smac)
1830 {
1831 	return _mlx4_set_path(dev, &qp->ah_attr,
1832 			      mlx4_mac_to_u64(smac),
1833 			      vlan_id,
1834 			      path, &mqp->pri, port);
1835 }
1836 
1837 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1838 			     const struct ib_qp_attr *qp,
1839 			     enum ib_qp_attr_mask qp_attr_mask,
1840 			     struct mlx4_ib_qp *mqp,
1841 			     struct mlx4_qp_path *path, u8 port)
1842 {
1843 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1844 			      0,
1845 			      0xffff,
1846 			      path, &mqp->alt, port);
1847 }
1848 
1849 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1850 {
1851 	struct mlx4_ib_gid_entry *ge, *tmp;
1852 
1853 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1854 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1855 			ge->added = 1;
1856 			ge->port = qp->port;
1857 		}
1858 	}
1859 }
1860 
1861 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1862 				    struct mlx4_ib_qp *qp,
1863 				    struct mlx4_qp_context *context)
1864 {
1865 	u64 u64_mac;
1866 	int smac_index;
1867 
1868 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1869 
1870 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1871 	if (!qp->pri.smac && !qp->pri.smac_port) {
1872 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1873 		if (smac_index >= 0) {
1874 			qp->pri.candidate_smac_index = smac_index;
1875 			qp->pri.candidate_smac = u64_mac;
1876 			qp->pri.candidate_smac_port = qp->port;
1877 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1878 		} else {
1879 			return -ENOENT;
1880 		}
1881 	}
1882 	return 0;
1883 }
1884 
1885 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1886 {
1887 	struct counter_index *new_counter_index;
1888 	int err;
1889 	u32 tmp_idx;
1890 
1891 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1892 	    IB_LINK_LAYER_ETHERNET ||
1893 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1894 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1895 		return 0;
1896 
1897 	err = mlx4_counter_alloc(dev->dev, &tmp_idx, MLX4_RES_USAGE_DRIVER);
1898 	if (err)
1899 		return err;
1900 
1901 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1902 	if (!new_counter_index) {
1903 		mlx4_counter_free(dev->dev, tmp_idx);
1904 		return -ENOMEM;
1905 	}
1906 
1907 	new_counter_index->index = tmp_idx;
1908 	new_counter_index->allocated = 1;
1909 	qp->counter_index = new_counter_index;
1910 
1911 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
1912 	list_add_tail(&new_counter_index->list,
1913 		      &dev->counters_table[qp->port - 1].counters_list);
1914 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
1915 
1916 	return 0;
1917 }
1918 
1919 enum {
1920 	MLX4_QPC_ROCE_MODE_1 = 0,
1921 	MLX4_QPC_ROCE_MODE_2 = 2,
1922 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
1923 };
1924 
1925 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
1926 {
1927 	switch (gid_type) {
1928 	case IB_GID_TYPE_ROCE:
1929 		return MLX4_QPC_ROCE_MODE_1;
1930 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
1931 		return MLX4_QPC_ROCE_MODE_2;
1932 	default:
1933 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
1934 	}
1935 }
1936 
1937 /*
1938  * Go over all RSS QP's childes (WQs) and apply their HW state according to
1939  * their logic state if the RSS QP is the first RSS QP associated for the WQ.
1940  */
1941 static int bringup_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, u8 port_num)
1942 {
1943 	int err = 0;
1944 	int i;
1945 
1946 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
1947 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
1948 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
1949 
1950 		mutex_lock(&wq->mutex);
1951 
1952 		/* Mlx4_ib restrictions:
1953 		 * WQ's is associated to a port according to the RSS QP it is
1954 		 * associates to.
1955 		 * In case the WQ is associated to a different port by another
1956 		 * RSS QP, return a failure.
1957 		 */
1958 		if ((wq->rss_usecnt > 0) && (wq->port != port_num)) {
1959 			err = -EINVAL;
1960 			mutex_unlock(&wq->mutex);
1961 			break;
1962 		}
1963 		wq->port = port_num;
1964 		if ((wq->rss_usecnt == 0) && (ibwq->state == IB_WQS_RDY)) {
1965 			err = _mlx4_ib_modify_wq(ibwq, IB_WQS_RDY);
1966 			if (err) {
1967 				mutex_unlock(&wq->mutex);
1968 				break;
1969 			}
1970 		}
1971 		wq->rss_usecnt++;
1972 
1973 		mutex_unlock(&wq->mutex);
1974 	}
1975 
1976 	if (i && err) {
1977 		int j;
1978 
1979 		for (j = (i - 1); j >= 0; j--) {
1980 			struct ib_wq *ibwq = ind_tbl->ind_tbl[j];
1981 			struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
1982 
1983 			mutex_lock(&wq->mutex);
1984 
1985 			if ((wq->rss_usecnt == 1) &&
1986 			    (ibwq->state == IB_WQS_RDY))
1987 				if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET))
1988 					pr_warn("failed to reverse WQN=0x%06x\n",
1989 						ibwq->wq_num);
1990 			wq->rss_usecnt--;
1991 
1992 			mutex_unlock(&wq->mutex);
1993 		}
1994 	}
1995 
1996 	return err;
1997 }
1998 
1999 static void bring_down_rss_rwqs(struct ib_rwq_ind_table *ind_tbl)
2000 {
2001 	int i;
2002 
2003 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2004 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2005 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2006 
2007 		mutex_lock(&wq->mutex);
2008 
2009 		if ((wq->rss_usecnt == 1) && (ibwq->state == IB_WQS_RDY))
2010 			if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET))
2011 				pr_warn("failed to reverse WQN=%x\n",
2012 					ibwq->wq_num);
2013 		wq->rss_usecnt--;
2014 
2015 		mutex_unlock(&wq->mutex);
2016 	}
2017 }
2018 
2019 static void fill_qp_rss_context(struct mlx4_qp_context *context,
2020 				struct mlx4_ib_qp *qp)
2021 {
2022 	struct mlx4_rss_context *rss_context;
2023 
2024 	rss_context = (void *)context + offsetof(struct mlx4_qp_context,
2025 			pri_path) + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
2026 
2027 	rss_context->base_qpn = cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz);
2028 	rss_context->default_qpn =
2029 		cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz & 0xffffff);
2030 	if (qp->rss_ctx->flags & (MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6))
2031 		rss_context->base_qpn_udp = rss_context->default_qpn;
2032 	rss_context->flags = qp->rss_ctx->flags;
2033 	/* Currently support just toeplitz */
2034 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
2035 
2036 	memcpy(rss_context->rss_key, qp->rss_ctx->rss_key,
2037 	       MLX4_EN_RSS_KEY_SIZE);
2038 }
2039 
2040 static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type,
2041 			       const struct ib_qp_attr *attr, int attr_mask,
2042 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
2043 {
2044 	struct ib_uobject *ibuobject;
2045 	struct ib_srq  *ibsrq;
2046 	const struct ib_gid_attr *gid_attr = NULL;
2047 	struct ib_rwq_ind_table *rwq_ind_tbl;
2048 	enum ib_qp_type qp_type;
2049 	struct mlx4_ib_dev *dev;
2050 	struct mlx4_ib_qp *qp;
2051 	struct mlx4_ib_pd *pd;
2052 	struct mlx4_ib_cq *send_cq, *recv_cq;
2053 	struct mlx4_qp_context *context;
2054 	enum mlx4_qp_optpar optpar = 0;
2055 	int sqd_event;
2056 	int steer_qp = 0;
2057 	int err = -EINVAL;
2058 	int counter_index;
2059 
2060 	if (src_type == MLX4_IB_RWQ_SRC) {
2061 		struct ib_wq *ibwq;
2062 
2063 		ibwq	    = (struct ib_wq *)src;
2064 		ibuobject   = ibwq->uobject;
2065 		ibsrq	    = NULL;
2066 		rwq_ind_tbl = NULL;
2067 		qp_type     = IB_QPT_RAW_PACKET;
2068 		qp	    = to_mqp((struct ib_qp *)ibwq);
2069 		dev	    = to_mdev(ibwq->device);
2070 		pd	    = to_mpd(ibwq->pd);
2071 	} else {
2072 		struct ib_qp *ibqp;
2073 
2074 		ibqp	    = (struct ib_qp *)src;
2075 		ibuobject   = ibqp->uobject;
2076 		ibsrq	    = ibqp->srq;
2077 		rwq_ind_tbl = ibqp->rwq_ind_tbl;
2078 		qp_type     = ibqp->qp_type;
2079 		qp	    = to_mqp(ibqp);
2080 		dev	    = to_mdev(ibqp->device);
2081 		pd	    = get_pd(qp);
2082 	}
2083 
2084 	/* APM is not supported under RoCE */
2085 	if (attr_mask & IB_QP_ALT_PATH &&
2086 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2087 	    IB_LINK_LAYER_ETHERNET)
2088 		return -ENOTSUPP;
2089 
2090 	context = kzalloc(sizeof *context, GFP_KERNEL);
2091 	if (!context)
2092 		return -ENOMEM;
2093 
2094 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
2095 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
2096 
2097 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
2098 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2099 	else {
2100 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
2101 		switch (attr->path_mig_state) {
2102 		case IB_MIG_MIGRATED:
2103 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2104 			break;
2105 		case IB_MIG_REARM:
2106 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
2107 			break;
2108 		case IB_MIG_ARMED:
2109 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
2110 			break;
2111 		}
2112 	}
2113 
2114 	if (qp->inl_recv_sz)
2115 		context->param3 |= cpu_to_be32(1 << 25);
2116 
2117 	if (qp->flags & MLX4_IB_QP_SCATTER_FCS)
2118 		context->param3 |= cpu_to_be32(1 << 29);
2119 
2120 	if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI)
2121 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
2122 	else if (qp_type == IB_QPT_RAW_PACKET)
2123 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
2124 	else if (qp_type == IB_QPT_UD) {
2125 		if (qp->flags & MLX4_IB_QP_LSO)
2126 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
2127 					      ilog2(dev->dev->caps.max_gso_sz);
2128 		else
2129 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 13;
2130 	} else if (attr_mask & IB_QP_PATH_MTU) {
2131 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
2132 			pr_err("path MTU (%u) is invalid\n",
2133 			       attr->path_mtu);
2134 			goto out;
2135 		}
2136 		context->mtu_msgmax = (attr->path_mtu << 5) |
2137 			ilog2(dev->dev->caps.max_msg_sz);
2138 	}
2139 
2140 	if (!rwq_ind_tbl) { /* PRM RSS receive side should be left zeros */
2141 		if (qp->rq.wqe_cnt)
2142 			context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
2143 		context->rq_size_stride |= qp->rq.wqe_shift - 4;
2144 	}
2145 
2146 	if (qp->sq.wqe_cnt)
2147 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
2148 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
2149 
2150 	if (new_state == IB_QPS_RESET && qp->counter_index)
2151 		mlx4_ib_free_qp_counter(dev, qp);
2152 
2153 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2154 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
2155 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
2156 		if (qp_type == IB_QPT_RAW_PACKET)
2157 			context->param3 |= cpu_to_be32(1 << 30);
2158 	}
2159 
2160 	if (ibuobject)
2161 		context->usr_page = cpu_to_be32(
2162 			mlx4_to_hw_uar_index(dev->dev,
2163 					     to_mucontext(ibuobject->context)
2164 					     ->uar.index));
2165 	else
2166 		context->usr_page = cpu_to_be32(
2167 			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
2168 
2169 	if (attr_mask & IB_QP_DEST_QPN)
2170 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
2171 
2172 	if (attr_mask & IB_QP_PORT) {
2173 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
2174 		    !(attr_mask & IB_QP_AV)) {
2175 			mlx4_set_sched(&context->pri_path, attr->port_num);
2176 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
2177 		}
2178 	}
2179 
2180 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
2181 		err = create_qp_lb_counter(dev, qp);
2182 		if (err)
2183 			goto out;
2184 
2185 		counter_index =
2186 			dev->counters_table[qp->port - 1].default_counter;
2187 		if (qp->counter_index)
2188 			counter_index = qp->counter_index->index;
2189 
2190 		if (counter_index != -1) {
2191 			context->pri_path.counter_index = counter_index;
2192 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
2193 			if (qp->counter_index) {
2194 				context->pri_path.fl |=
2195 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
2196 				context->pri_path.vlan_control |=
2197 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
2198 			}
2199 		} else
2200 			context->pri_path.counter_index =
2201 				MLX4_SINK_COUNTER_INDEX(dev->dev);
2202 
2203 		if (qp->flags & MLX4_IB_QP_NETIF) {
2204 			mlx4_ib_steer_qp_reg(dev, qp, 1);
2205 			steer_qp = 1;
2206 		}
2207 
2208 		if (qp_type == IB_QPT_GSI) {
2209 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
2210 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
2211 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
2212 
2213 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2214 		}
2215 	}
2216 
2217 	if (attr_mask & IB_QP_PKEY_INDEX) {
2218 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2219 			context->pri_path.disable_pkey_check = 0x40;
2220 		context->pri_path.pkey_index = attr->pkey_index;
2221 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
2222 	}
2223 
2224 	if (attr_mask & IB_QP_AV) {
2225 		u8 port_num = mlx4_is_bonded(dev->dev) ? 1 :
2226 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2227 		u16 vlan = 0xffff;
2228 		u8 smac[ETH_ALEN];
2229 		int is_eth =
2230 			rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
2231 			rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
2232 
2233 		if (is_eth) {
2234 			gid_attr = attr->ah_attr.grh.sgid_attr;
2235 			vlan = rdma_vlan_dev_vlan_id(gid_attr->ndev);
2236 			memcpy(smac, gid_attr->ndev->dev_addr, ETH_ALEN);
2237 		}
2238 
2239 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
2240 				  port_num, vlan, smac))
2241 			goto out;
2242 
2243 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
2244 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
2245 
2246 		if (is_eth &&
2247 		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
2248 			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr->gid_type);
2249 
2250 			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
2251 				err = -EINVAL;
2252 				goto out;
2253 			}
2254 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2255 		}
2256 
2257 	}
2258 
2259 	if (attr_mask & IB_QP_TIMEOUT) {
2260 		context->pri_path.ackto |= attr->timeout << 3;
2261 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
2262 	}
2263 
2264 	if (attr_mask & IB_QP_ALT_PATH) {
2265 		if (attr->alt_port_num == 0 ||
2266 		    attr->alt_port_num > dev->dev->caps.num_ports)
2267 			goto out;
2268 
2269 		if (attr->alt_pkey_index >=
2270 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
2271 			goto out;
2272 
2273 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
2274 				      &context->alt_path,
2275 				      attr->alt_port_num))
2276 			goto out;
2277 
2278 		context->alt_path.pkey_index = attr->alt_pkey_index;
2279 		context->alt_path.ackto = attr->alt_timeout << 3;
2280 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
2281 	}
2282 
2283 	context->pd = cpu_to_be32(pd->pdn);
2284 
2285 	if (!rwq_ind_tbl) {
2286 		context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
2287 		get_cqs(qp, src_type, &send_cq, &recv_cq);
2288 	} else { /* Set dummy CQs to be compatible with HV and PRM */
2289 		send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq);
2290 		recv_cq = send_cq;
2291 	}
2292 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
2293 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
2294 
2295 	/* Set "fast registration enabled" for all kernel QPs */
2296 	if (!ibuobject)
2297 		context->params1 |= cpu_to_be32(1 << 11);
2298 
2299 	if (attr_mask & IB_QP_RNR_RETRY) {
2300 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2301 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
2302 	}
2303 
2304 	if (attr_mask & IB_QP_RETRY_CNT) {
2305 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2306 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
2307 	}
2308 
2309 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2310 		if (attr->max_rd_atomic)
2311 			context->params1 |=
2312 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2313 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
2314 	}
2315 
2316 	if (attr_mask & IB_QP_SQ_PSN)
2317 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2318 
2319 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2320 		if (attr->max_dest_rd_atomic)
2321 			context->params2 |=
2322 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2323 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
2324 	}
2325 
2326 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
2327 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
2328 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
2329 	}
2330 
2331 	if (ibsrq)
2332 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
2333 
2334 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
2335 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2336 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
2337 	}
2338 	if (attr_mask & IB_QP_RQ_PSN)
2339 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2340 
2341 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
2342 	if (attr_mask & IB_QP_QKEY) {
2343 		if (qp->mlx4_ib_qp_type &
2344 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
2345 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2346 		else {
2347 			if (mlx4_is_mfunc(dev->dev) &&
2348 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
2349 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
2350 			    MLX4_RESERVED_QKEY_BASE) {
2351 				pr_err("Cannot use reserved QKEY"
2352 				       " 0x%x (range 0xffff0000..0xffffffff"
2353 				       " is reserved)\n", attr->qkey);
2354 				err = -EINVAL;
2355 				goto out;
2356 			}
2357 			context->qkey = cpu_to_be32(attr->qkey);
2358 		}
2359 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
2360 	}
2361 
2362 	if (ibsrq)
2363 		context->srqn = cpu_to_be32(1 << 24 |
2364 					    to_msrq(ibsrq)->msrq.srqn);
2365 
2366 	if (qp->rq.wqe_cnt &&
2367 	    cur_state == IB_QPS_RESET &&
2368 	    new_state == IB_QPS_INIT)
2369 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2370 
2371 	if (cur_state == IB_QPS_INIT &&
2372 	    new_state == IB_QPS_RTR  &&
2373 	    (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI ||
2374 	     qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) {
2375 		context->pri_path.sched_queue = (qp->port - 1) << 6;
2376 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2377 		    qp->mlx4_ib_qp_type &
2378 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
2379 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
2380 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
2381 				context->pri_path.fl = 0x80;
2382 		} else {
2383 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2384 				context->pri_path.fl = 0x80;
2385 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
2386 		}
2387 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2388 		    IB_LINK_LAYER_ETHERNET) {
2389 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
2390 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
2391 				context->pri_path.feup = 1 << 7; /* don't fsm */
2392 			/* handle smac_index */
2393 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
2394 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
2395 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
2396 				err = handle_eth_ud_smac_index(dev, qp, context);
2397 				if (err) {
2398 					err = -EINVAL;
2399 					goto out;
2400 				}
2401 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
2402 					dev->qp1_proxy[qp->port - 1] = qp;
2403 			}
2404 		}
2405 	}
2406 
2407 	if (qp_type == IB_QPT_RAW_PACKET) {
2408 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
2409 					MLX4_IB_LINK_TYPE_ETH;
2410 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
2411 			/* set QP to receive both tunneled & non-tunneled packets */
2412 			if (!rwq_ind_tbl)
2413 				context->srqn = cpu_to_be32(7 << 28);
2414 		}
2415 	}
2416 
2417 	if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
2418 		int is_eth = rdma_port_get_link_layer(
2419 				&dev->ib_dev, qp->port) ==
2420 				IB_LINK_LAYER_ETHERNET;
2421 		if (is_eth) {
2422 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
2423 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
2424 		}
2425 	}
2426 
2427 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2428 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2429 		sqd_event = 1;
2430 	else
2431 		sqd_event = 0;
2432 
2433 	if (!ibuobject &&
2434 	    cur_state == IB_QPS_RESET &&
2435 	    new_state == IB_QPS_INIT)
2436 		context->rlkey_roce_mode |= (1 << 4);
2437 
2438 	/*
2439 	 * Before passing a kernel QP to the HW, make sure that the
2440 	 * ownership bits of the send queue are set and the SQ
2441 	 * headroom is stamped so that the hardware doesn't start
2442 	 * processing stale work requests.
2443 	 */
2444 	if (!ibuobject &&
2445 	    cur_state == IB_QPS_RESET &&
2446 	    new_state == IB_QPS_INIT) {
2447 		struct mlx4_wqe_ctrl_seg *ctrl;
2448 		int i;
2449 
2450 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
2451 			ctrl = get_send_wqe(qp, i);
2452 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
2453 			ctrl->qpn_vlan.fence_size =
2454 				1 << (qp->sq.wqe_shift - 4);
2455 			stamp_send_wqe(qp, i);
2456 		}
2457 	}
2458 
2459 	if (rwq_ind_tbl	&&
2460 	    cur_state == IB_QPS_RESET &&
2461 	    new_state == IB_QPS_INIT) {
2462 		fill_qp_rss_context(context, qp);
2463 		context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET);
2464 	}
2465 
2466 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2467 			     to_mlx4_state(new_state), context, optpar,
2468 			     sqd_event, &qp->mqp);
2469 	if (err)
2470 		goto out;
2471 
2472 	qp->state = new_state;
2473 
2474 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2475 		qp->atomic_rd_en = attr->qp_access_flags;
2476 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2477 		qp->resp_depth = attr->max_dest_rd_atomic;
2478 	if (attr_mask & IB_QP_PORT) {
2479 		qp->port = attr->port_num;
2480 		update_mcg_macs(dev, qp);
2481 	}
2482 	if (attr_mask & IB_QP_ALT_PATH)
2483 		qp->alt_port = attr->alt_port_num;
2484 
2485 	if (is_sqp(dev, qp))
2486 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2487 
2488 	/*
2489 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
2490 	 * QP0 to RESET or ERROR, bring the link back down.
2491 	 */
2492 	if (is_qp0(dev, qp)) {
2493 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2494 			if (mlx4_INIT_PORT(dev->dev, qp->port))
2495 				pr_warn("INIT_PORT failed for port %d\n",
2496 				       qp->port);
2497 
2498 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2499 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2500 			mlx4_CLOSE_PORT(dev->dev, qp->port);
2501 	}
2502 
2503 	/*
2504 	 * If we moved a kernel QP to RESET, clean up all old CQ
2505 	 * entries and reinitialize the QP.
2506 	 */
2507 	if (new_state == IB_QPS_RESET) {
2508 		if (!ibuobject) {
2509 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2510 					 ibsrq ? to_msrq(ibsrq) : NULL);
2511 			if (send_cq != recv_cq)
2512 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2513 
2514 			qp->rq.head = 0;
2515 			qp->rq.tail = 0;
2516 			qp->sq.head = 0;
2517 			qp->sq.tail = 0;
2518 			qp->sq_next_wqe = 0;
2519 			if (qp->rq.wqe_cnt)
2520 				*qp->db.db  = 0;
2521 
2522 			if (qp->flags & MLX4_IB_QP_NETIF)
2523 				mlx4_ib_steer_qp_reg(dev, qp, 0);
2524 		}
2525 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2526 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2527 			qp->pri.smac = 0;
2528 			qp->pri.smac_port = 0;
2529 		}
2530 		if (qp->alt.smac) {
2531 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2532 			qp->alt.smac = 0;
2533 		}
2534 		if (qp->pri.vid < 0x1000) {
2535 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2536 			qp->pri.vid = 0xFFFF;
2537 			qp->pri.candidate_vid = 0xFFFF;
2538 			qp->pri.update_vid = 0;
2539 		}
2540 
2541 		if (qp->alt.vid < 0x1000) {
2542 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2543 			qp->alt.vid = 0xFFFF;
2544 			qp->alt.candidate_vid = 0xFFFF;
2545 			qp->alt.update_vid = 0;
2546 		}
2547 	}
2548 out:
2549 	if (err && qp->counter_index)
2550 		mlx4_ib_free_qp_counter(dev, qp);
2551 	if (err && steer_qp)
2552 		mlx4_ib_steer_qp_reg(dev, qp, 0);
2553 	kfree(context);
2554 	if (qp->pri.candidate_smac ||
2555 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2556 		if (err) {
2557 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2558 		} else {
2559 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2560 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2561 			qp->pri.smac = qp->pri.candidate_smac;
2562 			qp->pri.smac_index = qp->pri.candidate_smac_index;
2563 			qp->pri.smac_port = qp->pri.candidate_smac_port;
2564 		}
2565 		qp->pri.candidate_smac = 0;
2566 		qp->pri.candidate_smac_index = 0;
2567 		qp->pri.candidate_smac_port = 0;
2568 	}
2569 	if (qp->alt.candidate_smac) {
2570 		if (err) {
2571 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2572 		} else {
2573 			if (qp->alt.smac)
2574 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2575 			qp->alt.smac = qp->alt.candidate_smac;
2576 			qp->alt.smac_index = qp->alt.candidate_smac_index;
2577 			qp->alt.smac_port = qp->alt.candidate_smac_port;
2578 		}
2579 		qp->alt.candidate_smac = 0;
2580 		qp->alt.candidate_smac_index = 0;
2581 		qp->alt.candidate_smac_port = 0;
2582 	}
2583 
2584 	if (qp->pri.update_vid) {
2585 		if (err) {
2586 			if (qp->pri.candidate_vid < 0x1000)
2587 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2588 						     qp->pri.candidate_vid);
2589 		} else {
2590 			if (qp->pri.vid < 0x1000)
2591 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2592 						     qp->pri.vid);
2593 			qp->pri.vid = qp->pri.candidate_vid;
2594 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2595 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
2596 		}
2597 		qp->pri.candidate_vid = 0xFFFF;
2598 		qp->pri.update_vid = 0;
2599 	}
2600 
2601 	if (qp->alt.update_vid) {
2602 		if (err) {
2603 			if (qp->alt.candidate_vid < 0x1000)
2604 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2605 						     qp->alt.candidate_vid);
2606 		} else {
2607 			if (qp->alt.vid < 0x1000)
2608 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2609 						     qp->alt.vid);
2610 			qp->alt.vid = qp->alt.candidate_vid;
2611 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2612 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2613 		}
2614 		qp->alt.candidate_vid = 0xFFFF;
2615 		qp->alt.update_vid = 0;
2616 	}
2617 
2618 	return err;
2619 }
2620 
2621 enum {
2622 	MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK = (IB_QP_STATE	|
2623 					      IB_QP_PORT),
2624 };
2625 
2626 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2627 			      int attr_mask, struct ib_udata *udata)
2628 {
2629 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2630 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2631 	enum ib_qp_state cur_state, new_state;
2632 	int err = -EINVAL;
2633 	mutex_lock(&qp->mutex);
2634 
2635 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2636 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2637 
2638 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2639 				attr_mask)) {
2640 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2641 			 "for transition %d to %d. qp_type %d,"
2642 			 " attr_mask 0x%x\n",
2643 			 ibqp->qp_num, cur_state, new_state,
2644 			 ibqp->qp_type, attr_mask);
2645 		goto out;
2646 	}
2647 
2648 	if (ibqp->rwq_ind_tbl) {
2649 		if (!(((cur_state == IB_QPS_RESET) &&
2650 		       (new_state == IB_QPS_INIT)) ||
2651 		      ((cur_state == IB_QPS_INIT)  &&
2652 		       (new_state == IB_QPS_RTR)))) {
2653 			pr_debug("qpn 0x%x: RSS QP unsupported transition %d to %d\n",
2654 				 ibqp->qp_num, cur_state, new_state);
2655 
2656 			err = -EOPNOTSUPP;
2657 			goto out;
2658 		}
2659 
2660 		if (attr_mask & ~MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK) {
2661 			pr_debug("qpn 0x%x: RSS QP unsupported attribute mask 0x%x for transition %d to %d\n",
2662 				 ibqp->qp_num, attr_mask, cur_state, new_state);
2663 
2664 			err = -EOPNOTSUPP;
2665 			goto out;
2666 		}
2667 	}
2668 
2669 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2670 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2671 			if ((ibqp->qp_type == IB_QPT_RC) ||
2672 			    (ibqp->qp_type == IB_QPT_UD) ||
2673 			    (ibqp->qp_type == IB_QPT_UC) ||
2674 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2675 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2676 				attr->port_num = mlx4_ib_bond_next_port(dev);
2677 			}
2678 		} else {
2679 			/* no sense in changing port_num
2680 			 * when ports are bonded */
2681 			attr_mask &= ~IB_QP_PORT;
2682 		}
2683 	}
2684 
2685 	if ((attr_mask & IB_QP_PORT) &&
2686 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2687 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2688 			 "for transition %d to %d. qp_type %d\n",
2689 			 ibqp->qp_num, attr->port_num, cur_state,
2690 			 new_state, ibqp->qp_type);
2691 		goto out;
2692 	}
2693 
2694 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2695 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2696 	     IB_LINK_LAYER_ETHERNET))
2697 		goto out;
2698 
2699 	if (attr_mask & IB_QP_PKEY_INDEX) {
2700 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2701 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2702 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2703 				 "for transition %d to %d. qp_type %d\n",
2704 				 ibqp->qp_num, attr->pkey_index, cur_state,
2705 				 new_state, ibqp->qp_type);
2706 			goto out;
2707 		}
2708 	}
2709 
2710 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2711 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2712 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2713 			 "Transition %d to %d. qp_type %d\n",
2714 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2715 			 new_state, ibqp->qp_type);
2716 		goto out;
2717 	}
2718 
2719 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2720 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2721 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2722 			 "Transition %d to %d. qp_type %d\n",
2723 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2724 			 new_state, ibqp->qp_type);
2725 		goto out;
2726 	}
2727 
2728 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2729 		err = 0;
2730 		goto out;
2731 	}
2732 
2733 	if (ibqp->rwq_ind_tbl && (new_state == IB_QPS_INIT)) {
2734 		err = bringup_rss_rwqs(ibqp->rwq_ind_tbl, attr->port_num);
2735 		if (err)
2736 			goto out;
2737 	}
2738 
2739 	err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask,
2740 				  cur_state, new_state);
2741 
2742 	if (ibqp->rwq_ind_tbl && err)
2743 		bring_down_rss_rwqs(ibqp->rwq_ind_tbl);
2744 
2745 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2746 		attr->port_num = 1;
2747 
2748 out:
2749 	mutex_unlock(&qp->mutex);
2750 	return err;
2751 }
2752 
2753 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2754 		      int attr_mask, struct ib_udata *udata)
2755 {
2756 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2757 	int ret;
2758 
2759 	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2760 
2761 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2762 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
2763 		int err = 0;
2764 
2765 		if (sqp->roce_v2_gsi)
2766 			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2767 		if (err)
2768 			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2769 			       err);
2770 	}
2771 	return ret;
2772 }
2773 
2774 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2775 {
2776 	int i;
2777 	for (i = 0; i < dev->caps.num_ports; i++) {
2778 		if (qpn == dev->caps.spec_qps[i].qp0_proxy ||
2779 		    qpn == dev->caps.spec_qps[i].qp0_tunnel) {
2780 			*qkey = dev->caps.spec_qps[i].qp0_qkey;
2781 			return 0;
2782 		}
2783 	}
2784 	return -EINVAL;
2785 }
2786 
2787 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2788 				  const struct ib_ud_wr *wr,
2789 				  void *wqe, unsigned *mlx_seg_len)
2790 {
2791 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2792 	struct ib_device *ib_dev = &mdev->ib_dev;
2793 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2794 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2795 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2796 	u16 pkey;
2797 	u32 qkey;
2798 	int send_size;
2799 	int header_size;
2800 	int spc;
2801 	int i;
2802 
2803 	if (wr->wr.opcode != IB_WR_SEND)
2804 		return -EINVAL;
2805 
2806 	send_size = 0;
2807 
2808 	for (i = 0; i < wr->wr.num_sge; ++i)
2809 		send_size += wr->wr.sg_list[i].length;
2810 
2811 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2812 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2813 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2814 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2815 
2816 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2817 
2818 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2819 		sqp->ud_header.lrh.service_level =
2820 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2821 		sqp->ud_header.lrh.destination_lid =
2822 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2823 		sqp->ud_header.lrh.source_lid =
2824 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2825 	}
2826 
2827 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2828 
2829 	/* force loopback */
2830 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2831 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2832 
2833 	sqp->ud_header.lrh.virtual_lane    = 0;
2834 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2835 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2836 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2837 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2838 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2839 	else
2840 		sqp->ud_header.bth.destination_qpn =
2841 			cpu_to_be32(mdev->dev->caps.spec_qps[sqp->qp.port - 1].qp0_tunnel);
2842 
2843 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2844 	if (mlx4_is_master(mdev->dev)) {
2845 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2846 			return -EINVAL;
2847 	} else {
2848 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2849 			return -EINVAL;
2850 	}
2851 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2852 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2853 
2854 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2855 	sqp->ud_header.immediate_present = 0;
2856 
2857 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2858 
2859 	/*
2860 	 * Inline data segments may not cross a 64 byte boundary.  If
2861 	 * our UD header is bigger than the space available up to the
2862 	 * next 64 byte boundary in the WQE, use two inline data
2863 	 * segments to hold the UD header.
2864 	 */
2865 	spc = MLX4_INLINE_ALIGN -
2866 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2867 	if (header_size <= spc) {
2868 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2869 		memcpy(inl + 1, sqp->header_buf, header_size);
2870 		i = 1;
2871 	} else {
2872 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2873 		memcpy(inl + 1, sqp->header_buf, spc);
2874 
2875 		inl = (void *) (inl + 1) + spc;
2876 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2877 		/*
2878 		 * Need a barrier here to make sure all the data is
2879 		 * visible before the byte_count field is set.
2880 		 * Otherwise the HCA prefetcher could grab the 64-byte
2881 		 * chunk with this inline segment and get a valid (!=
2882 		 * 0xffffffff) byte count but stale data, and end up
2883 		 * generating a packet with bad headers.
2884 		 *
2885 		 * The first inline segment's byte_count field doesn't
2886 		 * need a barrier, because it comes after a
2887 		 * control/MLX segment and therefore is at an offset
2888 		 * of 16 mod 64.
2889 		 */
2890 		wmb();
2891 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2892 		i = 2;
2893 	}
2894 
2895 	*mlx_seg_len =
2896 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2897 	return 0;
2898 }
2899 
2900 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
2901 {
2902 	union sl2vl_tbl_to_u64 tmp_vltab;
2903 	u8 vl;
2904 
2905 	if (sl > 15)
2906 		return 0xf;
2907 	tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
2908 	vl = tmp_vltab.sl8[sl >> 1];
2909 	if (sl & 1)
2910 		vl &= 0x0f;
2911 	else
2912 		vl >>= 4;
2913 	return vl;
2914 }
2915 
2916 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
2917 				int index, union ib_gid *gid,
2918 				enum ib_gid_type *gid_type)
2919 {
2920 	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
2921 	struct mlx4_port_gid_table *port_gid_table;
2922 	unsigned long flags;
2923 
2924 	port_gid_table = &iboe->gids[port_num - 1];
2925 	spin_lock_irqsave(&iboe->lock, flags);
2926 	memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
2927 	*gid_type = port_gid_table->gids[index].gid_type;
2928 	spin_unlock_irqrestore(&iboe->lock, flags);
2929 	if (rdma_is_zero_gid(gid))
2930 		return -ENOENT;
2931 
2932 	return 0;
2933 }
2934 
2935 #define MLX4_ROCEV2_QP1_SPORT 0xC000
2936 static int build_mlx_header(struct mlx4_ib_sqp *sqp, const struct ib_ud_wr *wr,
2937 			    void *wqe, unsigned *mlx_seg_len)
2938 {
2939 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
2940 	struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
2941 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2942 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2943 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2944 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2945 	union ib_gid sgid;
2946 	u16 pkey;
2947 	int send_size;
2948 	int header_size;
2949 	int spc;
2950 	int i;
2951 	int err = 0;
2952 	u16 vlan = 0xffff;
2953 	bool is_eth;
2954 	bool is_vlan = false;
2955 	bool is_grh;
2956 	bool is_udp = false;
2957 	int ip_version = 0;
2958 
2959 	send_size = 0;
2960 	for (i = 0; i < wr->wr.num_sge; ++i)
2961 		send_size += wr->wr.sg_list[i].length;
2962 
2963 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2964 	is_grh = mlx4_ib_ah_grh_present(ah);
2965 	if (is_eth) {
2966 		enum ib_gid_type gid_type;
2967 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2968 			/* When multi-function is enabled, the ib_core gid
2969 			 * indexes don't necessarily match the hw ones, so
2970 			 * we must use our own cache */
2971 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2972 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
2973 							   ah->av.ib.gid_index, &sgid.raw[0]);
2974 			if (err)
2975 				return err;
2976 		} else  {
2977 			err = fill_gid_by_hw_index(ibdev, sqp->qp.port,
2978 					    ah->av.ib.gid_index,
2979 					    &sgid, &gid_type);
2980 			if (!err) {
2981 				is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
2982 				if (is_udp) {
2983 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
2984 						ip_version = 4;
2985 					else
2986 						ip_version = 6;
2987 					is_grh = false;
2988 				}
2989 			} else {
2990 				return err;
2991 			}
2992 		}
2993 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
2994 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
2995 			is_vlan = 1;
2996 		}
2997 	}
2998 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
2999 			  ip_version, is_udp, 0, &sqp->ud_header);
3000 	if (err)
3001 		return err;
3002 
3003 	if (!is_eth) {
3004 		sqp->ud_header.lrh.service_level =
3005 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
3006 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
3007 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
3008 	}
3009 
3010 	if (is_grh || (ip_version == 6)) {
3011 		sqp->ud_header.grh.traffic_class =
3012 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3013 		sqp->ud_header.grh.flow_label    =
3014 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
3015 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
3016 		if (is_eth) {
3017 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
3018 		} else {
3019 			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3020 				/* When multi-function is enabled, the ib_core gid
3021 				 * indexes don't necessarily match the hw ones, so
3022 				 * we must use our own cache
3023 				 */
3024 				sqp->ud_header.grh.source_gid.global.subnet_prefix =
3025 					cpu_to_be64(atomic64_read(&(to_mdev(ib_dev)->sriov.
3026 								    demux[sqp->qp.port - 1].
3027 								    subnet_prefix)));
3028 				sqp->ud_header.grh.source_gid.global.interface_id =
3029 					to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
3030 						       guid_cache[ah->av.ib.gid_index];
3031 			} else {
3032 				sqp->ud_header.grh.source_gid =
3033 					ah->ibah.sgid_attr->gid;
3034 			}
3035 		}
3036 		memcpy(sqp->ud_header.grh.destination_gid.raw,
3037 		       ah->av.ib.dgid, 16);
3038 	}
3039 
3040 	if (ip_version == 4) {
3041 		sqp->ud_header.ip4.tos =
3042 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3043 		sqp->ud_header.ip4.id = 0;
3044 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
3045 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
3046 
3047 		memcpy(&sqp->ud_header.ip4.saddr,
3048 		       sgid.raw + 12, 4);
3049 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
3050 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
3051 	}
3052 
3053 	if (is_udp) {
3054 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
3055 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
3056 		sqp->ud_header.udp.csum = 0;
3057 	}
3058 
3059 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
3060 
3061 	if (!is_eth) {
3062 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
3063 					  (sqp->ud_header.lrh.destination_lid ==
3064 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
3065 					  (sqp->ud_header.lrh.service_level << 8));
3066 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
3067 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
3068 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
3069 	}
3070 
3071 	switch (wr->wr.opcode) {
3072 	case IB_WR_SEND:
3073 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
3074 		sqp->ud_header.immediate_present = 0;
3075 		break;
3076 	case IB_WR_SEND_WITH_IMM:
3077 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
3078 		sqp->ud_header.immediate_present = 1;
3079 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
3080 		break;
3081 	default:
3082 		return -EINVAL;
3083 	}
3084 
3085 	if (is_eth) {
3086 		struct in6_addr in6;
3087 		u16 ether_type;
3088 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
3089 
3090 		ether_type = (!is_udp) ? ETH_P_IBOE:
3091 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
3092 
3093 		mlx->sched_prio = cpu_to_be16(pcp);
3094 
3095 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
3096 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
3097 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
3098 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
3099 		memcpy(&in6, sgid.raw, sizeof(in6));
3100 
3101 
3102 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
3103 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
3104 		if (!is_vlan) {
3105 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
3106 		} else {
3107 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
3108 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
3109 		}
3110 	} else {
3111 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 :
3112 							sl_to_vl(to_mdev(ib_dev),
3113 								 sqp->ud_header.lrh.service_level,
3114 								 sqp->qp.port);
3115 		if (sqp->qp.ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
3116 			return -EINVAL;
3117 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
3118 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
3119 	}
3120 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
3121 	if (!sqp->qp.ibqp.qp_num)
3122 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
3123 	else
3124 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
3125 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
3126 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
3127 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
3128 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
3129 					       sqp->qkey : wr->remote_qkey);
3130 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
3131 
3132 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3133 
3134 	if (0) {
3135 		pr_err("built UD header of size %d:\n", header_size);
3136 		for (i = 0; i < header_size / 4; ++i) {
3137 			if (i % 8 == 0)
3138 				pr_err("  [%02x] ", i * 4);
3139 			pr_cont(" %08x",
3140 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
3141 			if ((i + 1) % 8 == 0)
3142 				pr_cont("\n");
3143 		}
3144 		pr_err("\n");
3145 	}
3146 
3147 	/*
3148 	 * Inline data segments may not cross a 64 byte boundary.  If
3149 	 * our UD header is bigger than the space available up to the
3150 	 * next 64 byte boundary in the WQE, use two inline data
3151 	 * segments to hold the UD header.
3152 	 */
3153 	spc = MLX4_INLINE_ALIGN -
3154 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3155 	if (header_size <= spc) {
3156 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3157 		memcpy(inl + 1, sqp->header_buf, header_size);
3158 		i = 1;
3159 	} else {
3160 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3161 		memcpy(inl + 1, sqp->header_buf, spc);
3162 
3163 		inl = (void *) (inl + 1) + spc;
3164 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3165 		/*
3166 		 * Need a barrier here to make sure all the data is
3167 		 * visible before the byte_count field is set.
3168 		 * Otherwise the HCA prefetcher could grab the 64-byte
3169 		 * chunk with this inline segment and get a valid (!=
3170 		 * 0xffffffff) byte count but stale data, and end up
3171 		 * generating a packet with bad headers.
3172 		 *
3173 		 * The first inline segment's byte_count field doesn't
3174 		 * need a barrier, because it comes after a
3175 		 * control/MLX segment and therefore is at an offset
3176 		 * of 16 mod 64.
3177 		 */
3178 		wmb();
3179 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3180 		i = 2;
3181 	}
3182 
3183 	*mlx_seg_len =
3184 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3185 	return 0;
3186 }
3187 
3188 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3189 {
3190 	unsigned cur;
3191 	struct mlx4_ib_cq *cq;
3192 
3193 	cur = wq->head - wq->tail;
3194 	if (likely(cur + nreq < wq->max_post))
3195 		return 0;
3196 
3197 	cq = to_mcq(ib_cq);
3198 	spin_lock(&cq->lock);
3199 	cur = wq->head - wq->tail;
3200 	spin_unlock(&cq->lock);
3201 
3202 	return cur + nreq >= wq->max_post;
3203 }
3204 
3205 static __be32 convert_access(int acc)
3206 {
3207 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
3208 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
3209 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
3210 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
3211 	       (acc & IB_ACCESS_REMOTE_READ   ?
3212 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
3213 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
3214 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
3215 }
3216 
3217 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
3218 			const struct ib_reg_wr *wr)
3219 {
3220 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
3221 
3222 	fseg->flags		= convert_access(wr->access);
3223 	fseg->mem_key		= cpu_to_be32(wr->key);
3224 	fseg->buf_list		= cpu_to_be64(mr->page_map);
3225 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
3226 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
3227 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
3228 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
3229 	fseg->reserved[0]	= 0;
3230 	fseg->reserved[1]	= 0;
3231 }
3232 
3233 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
3234 {
3235 	memset(iseg, 0, sizeof(*iseg));
3236 	iseg->mem_key = cpu_to_be32(rkey);
3237 }
3238 
3239 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
3240 					  u64 remote_addr, u32 rkey)
3241 {
3242 	rseg->raddr    = cpu_to_be64(remote_addr);
3243 	rseg->rkey     = cpu_to_be32(rkey);
3244 	rseg->reserved = 0;
3245 }
3246 
3247 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
3248 			   const struct ib_atomic_wr *wr)
3249 {
3250 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
3251 		aseg->swap_add = cpu_to_be64(wr->swap);
3252 		aseg->compare  = cpu_to_be64(wr->compare_add);
3253 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
3254 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3255 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
3256 	} else {
3257 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3258 		aseg->compare  = 0;
3259 	}
3260 
3261 }
3262 
3263 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
3264 				  const struct ib_atomic_wr *wr)
3265 {
3266 	aseg->swap_add		= cpu_to_be64(wr->swap);
3267 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
3268 	aseg->compare		= cpu_to_be64(wr->compare_add);
3269 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
3270 }
3271 
3272 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
3273 			     const struct ib_ud_wr *wr)
3274 {
3275 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
3276 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
3277 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
3278 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
3279 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
3280 }
3281 
3282 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
3283 				    struct mlx4_wqe_datagram_seg *dseg,
3284 				    const struct ib_ud_wr *wr,
3285 				    enum mlx4_ib_qp_type qpt)
3286 {
3287 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
3288 	struct mlx4_av sqp_av = {0};
3289 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
3290 
3291 	/* force loopback */
3292 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
3293 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
3294 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
3295 			cpu_to_be32(0xf0000000);
3296 
3297 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
3298 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
3299 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp1_tunnel);
3300 	else
3301 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp0_tunnel);
3302 	/* Use QKEY from the QP context, which is set by master */
3303 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
3304 }
3305 
3306 static void build_tunnel_header(const struct ib_ud_wr *wr, void *wqe,
3307 				unsigned *mlx_seg_len)
3308 {
3309 	struct mlx4_wqe_inline_seg *inl = wqe;
3310 	struct mlx4_ib_tunnel_header hdr;
3311 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3312 	int spc;
3313 	int i;
3314 
3315 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
3316 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
3317 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
3318 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
3319 	memcpy(hdr.mac, ah->av.eth.mac, 6);
3320 	hdr.vlan = ah->av.eth.vlan;
3321 
3322 	spc = MLX4_INLINE_ALIGN -
3323 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3324 	if (sizeof (hdr) <= spc) {
3325 		memcpy(inl + 1, &hdr, sizeof (hdr));
3326 		wmb();
3327 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
3328 		i = 1;
3329 	} else {
3330 		memcpy(inl + 1, &hdr, spc);
3331 		wmb();
3332 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3333 
3334 		inl = (void *) (inl + 1) + spc;
3335 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
3336 		wmb();
3337 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
3338 		i = 2;
3339 	}
3340 
3341 	*mlx_seg_len =
3342 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
3343 }
3344 
3345 static void set_mlx_icrc_seg(void *dseg)
3346 {
3347 	u32 *t = dseg;
3348 	struct mlx4_wqe_inline_seg *iseg = dseg;
3349 
3350 	t[1] = 0;
3351 
3352 	/*
3353 	 * Need a barrier here before writing the byte_count field to
3354 	 * make sure that all the data is visible before the
3355 	 * byte_count field is set.  Otherwise, if the segment begins
3356 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3357 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3358 	 * stale data, and end up sending the wrong data.
3359 	 */
3360 	wmb();
3361 
3362 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
3363 }
3364 
3365 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3366 {
3367 	dseg->lkey       = cpu_to_be32(sg->lkey);
3368 	dseg->addr       = cpu_to_be64(sg->addr);
3369 
3370 	/*
3371 	 * Need a barrier here before writing the byte_count field to
3372 	 * make sure that all the data is visible before the
3373 	 * byte_count field is set.  Otherwise, if the segment begins
3374 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3375 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3376 	 * stale data, and end up sending the wrong data.
3377 	 */
3378 	wmb();
3379 
3380 	dseg->byte_count = cpu_to_be32(sg->length);
3381 }
3382 
3383 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3384 {
3385 	dseg->byte_count = cpu_to_be32(sg->length);
3386 	dseg->lkey       = cpu_to_be32(sg->lkey);
3387 	dseg->addr       = cpu_to_be64(sg->addr);
3388 }
3389 
3390 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe,
3391 			 const struct ib_ud_wr *wr, struct mlx4_ib_qp *qp,
3392 			 unsigned *lso_seg_len, __be32 *lso_hdr_sz, __be32 *blh)
3393 {
3394 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
3395 
3396 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
3397 		*blh = cpu_to_be32(1 << 6);
3398 
3399 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
3400 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
3401 		return -EINVAL;
3402 
3403 	memcpy(wqe->header, wr->header, wr->hlen);
3404 
3405 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
3406 	*lso_seg_len = halign;
3407 	return 0;
3408 }
3409 
3410 static __be32 send_ieth(const struct ib_send_wr *wr)
3411 {
3412 	switch (wr->opcode) {
3413 	case IB_WR_SEND_WITH_IMM:
3414 	case IB_WR_RDMA_WRITE_WITH_IMM:
3415 		return wr->ex.imm_data;
3416 
3417 	case IB_WR_SEND_WITH_INV:
3418 		return cpu_to_be32(wr->ex.invalidate_rkey);
3419 
3420 	default:
3421 		return 0;
3422 	}
3423 }
3424 
3425 static void add_zero_len_inline(void *wqe)
3426 {
3427 	struct mlx4_wqe_inline_seg *inl = wqe;
3428 	memset(wqe, 0, 16);
3429 	inl->byte_count = cpu_to_be32(1 << 31);
3430 }
3431 
3432 static int _mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3433 			      const struct ib_send_wr **bad_wr, bool drain)
3434 {
3435 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3436 	void *wqe;
3437 	struct mlx4_wqe_ctrl_seg *ctrl;
3438 	struct mlx4_wqe_data_seg *dseg;
3439 	unsigned long flags;
3440 	int nreq;
3441 	int err = 0;
3442 	unsigned ind;
3443 	int uninitialized_var(size);
3444 	unsigned uninitialized_var(seglen);
3445 	__be32 dummy;
3446 	__be32 *lso_wqe;
3447 	__be32 uninitialized_var(lso_hdr_sz);
3448 	__be32 blh;
3449 	int i;
3450 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3451 
3452 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
3453 		struct mlx4_ib_sqp *sqp = to_msqp(qp);
3454 
3455 		if (sqp->roce_v2_gsi) {
3456 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
3457 			enum ib_gid_type gid_type;
3458 			union ib_gid gid;
3459 
3460 			if (!fill_gid_by_hw_index(mdev, sqp->qp.port,
3461 					   ah->av.ib.gid_index,
3462 					   &gid, &gid_type))
3463 				qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
3464 						to_mqp(sqp->roce_v2_gsi) : qp;
3465 			else
3466 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
3467 				       ah->av.ib.gid_index);
3468 		}
3469 	}
3470 
3471 	spin_lock_irqsave(&qp->sq.lock, flags);
3472 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3473 	    !drain) {
3474 		err = -EIO;
3475 		*bad_wr = wr;
3476 		nreq = 0;
3477 		goto out;
3478 	}
3479 
3480 	ind = qp->sq_next_wqe;
3481 
3482 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3483 		lso_wqe = &dummy;
3484 		blh = 0;
3485 
3486 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
3487 			err = -ENOMEM;
3488 			*bad_wr = wr;
3489 			goto out;
3490 		}
3491 
3492 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3493 			err = -EINVAL;
3494 			*bad_wr = wr;
3495 			goto out;
3496 		}
3497 
3498 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3499 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3500 
3501 		ctrl->srcrb_flags =
3502 			(wr->send_flags & IB_SEND_SIGNALED ?
3503 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3504 			(wr->send_flags & IB_SEND_SOLICITED ?
3505 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3506 			((wr->send_flags & IB_SEND_IP_CSUM) ?
3507 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3508 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3509 			qp->sq_signal_bits;
3510 
3511 		ctrl->imm = send_ieth(wr);
3512 
3513 		wqe += sizeof *ctrl;
3514 		size = sizeof *ctrl / 16;
3515 
3516 		switch (qp->mlx4_ib_qp_type) {
3517 		case MLX4_IB_QPT_RC:
3518 		case MLX4_IB_QPT_UC:
3519 			switch (wr->opcode) {
3520 			case IB_WR_ATOMIC_CMP_AND_SWP:
3521 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3522 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3523 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3524 					      atomic_wr(wr)->rkey);
3525 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3526 
3527 				set_atomic_seg(wqe, atomic_wr(wr));
3528 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
3529 
3530 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3531 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3532 
3533 				break;
3534 
3535 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3536 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3537 					      atomic_wr(wr)->rkey);
3538 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3539 
3540 				set_masked_atomic_seg(wqe, atomic_wr(wr));
3541 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
3542 
3543 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3544 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3545 
3546 				break;
3547 
3548 			case IB_WR_RDMA_READ:
3549 			case IB_WR_RDMA_WRITE:
3550 			case IB_WR_RDMA_WRITE_WITH_IMM:
3551 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3552 					      rdma_wr(wr)->rkey);
3553 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3554 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3555 				break;
3556 
3557 			case IB_WR_LOCAL_INV:
3558 				ctrl->srcrb_flags |=
3559 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3560 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3561 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3562 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3563 				break;
3564 
3565 			case IB_WR_REG_MR:
3566 				ctrl->srcrb_flags |=
3567 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3568 				set_reg_seg(wqe, reg_wr(wr));
3569 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3570 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3571 				break;
3572 
3573 			default:
3574 				/* No extra segments required for sends */
3575 				break;
3576 			}
3577 			break;
3578 
3579 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3580 			err =  build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3581 					ctrl, &seglen);
3582 			if (unlikely(err)) {
3583 				*bad_wr = wr;
3584 				goto out;
3585 			}
3586 			wqe  += seglen;
3587 			size += seglen / 16;
3588 			break;
3589 		case MLX4_IB_QPT_TUN_SMI:
3590 		case MLX4_IB_QPT_TUN_GSI:
3591 			/* this is a UD qp used in MAD responses to slaves. */
3592 			set_datagram_seg(wqe, ud_wr(wr));
3593 			/* set the forced-loopback bit in the data seg av */
3594 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3595 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3596 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3597 			break;
3598 		case MLX4_IB_QPT_UD:
3599 			set_datagram_seg(wqe, ud_wr(wr));
3600 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3601 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3602 
3603 			if (wr->opcode == IB_WR_LSO) {
3604 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3605 						&lso_hdr_sz, &blh);
3606 				if (unlikely(err)) {
3607 					*bad_wr = wr;
3608 					goto out;
3609 				}
3610 				lso_wqe = (__be32 *) wqe;
3611 				wqe  += seglen;
3612 				size += seglen / 16;
3613 			}
3614 			break;
3615 
3616 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3617 			err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3618 					ctrl, &seglen);
3619 			if (unlikely(err)) {
3620 				*bad_wr = wr;
3621 				goto out;
3622 			}
3623 			wqe  += seglen;
3624 			size += seglen / 16;
3625 			/* to start tunnel header on a cache-line boundary */
3626 			add_zero_len_inline(wqe);
3627 			wqe += 16;
3628 			size++;
3629 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3630 			wqe  += seglen;
3631 			size += seglen / 16;
3632 			break;
3633 		case MLX4_IB_QPT_PROXY_SMI:
3634 		case MLX4_IB_QPT_PROXY_GSI:
3635 			/* If we are tunneling special qps, this is a UD qp.
3636 			 * In this case we first add a UD segment targeting
3637 			 * the tunnel qp, and then add a header with address
3638 			 * information */
3639 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3640 						ud_wr(wr),
3641 						qp->mlx4_ib_qp_type);
3642 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3643 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3644 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3645 			wqe  += seglen;
3646 			size += seglen / 16;
3647 			break;
3648 
3649 		case MLX4_IB_QPT_SMI:
3650 		case MLX4_IB_QPT_GSI:
3651 			err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
3652 					&seglen);
3653 			if (unlikely(err)) {
3654 				*bad_wr = wr;
3655 				goto out;
3656 			}
3657 			wqe  += seglen;
3658 			size += seglen / 16;
3659 			break;
3660 
3661 		default:
3662 			break;
3663 		}
3664 
3665 		/*
3666 		 * Write data segments in reverse order, so as to
3667 		 * overwrite cacheline stamp last within each
3668 		 * cacheline.  This avoids issues with WQE
3669 		 * prefetching.
3670 		 */
3671 
3672 		dseg = wqe;
3673 		dseg += wr->num_sge - 1;
3674 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3675 
3676 		/* Add one more inline data segment for ICRC for MLX sends */
3677 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3678 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3679 			     qp->mlx4_ib_qp_type &
3680 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3681 			set_mlx_icrc_seg(dseg + 1);
3682 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3683 		}
3684 
3685 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3686 			set_data_seg(dseg, wr->sg_list + i);
3687 
3688 		/*
3689 		 * Possibly overwrite stamping in cacheline with LSO
3690 		 * segment only after making sure all data segments
3691 		 * are written.
3692 		 */
3693 		wmb();
3694 		*lso_wqe = lso_hdr_sz;
3695 
3696 		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3697 					     MLX4_WQE_CTRL_FENCE : 0) | size;
3698 
3699 		/*
3700 		 * Make sure descriptor is fully written before
3701 		 * setting ownership bit (because HW can start
3702 		 * executing as soon as we do).
3703 		 */
3704 		wmb();
3705 
3706 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3707 			*bad_wr = wr;
3708 			err = -EINVAL;
3709 			goto out;
3710 		}
3711 
3712 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3713 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3714 
3715 		/*
3716 		 * We can improve latency by not stamping the last
3717 		 * send queue WQE until after ringing the doorbell, so
3718 		 * only stamp here if there are still more WQEs to post.
3719 		 */
3720 		if (wr->next)
3721 			stamp_send_wqe(qp, ind + qp->sq_spare_wqes);
3722 		ind++;
3723 	}
3724 
3725 out:
3726 	if (likely(nreq)) {
3727 		qp->sq.head += nreq;
3728 
3729 		/*
3730 		 * Make sure that descriptors are written before
3731 		 * doorbell record.
3732 		 */
3733 		wmb();
3734 
3735 		writel_relaxed(qp->doorbell_qpn,
3736 			to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3737 
3738 		/*
3739 		 * Make sure doorbells don't leak out of SQ spinlock
3740 		 * and reach the HCA out of order.
3741 		 */
3742 		mmiowb();
3743 
3744 		stamp_send_wqe(qp, ind + qp->sq_spare_wqes - 1);
3745 
3746 		qp->sq_next_wqe = ind;
3747 	}
3748 
3749 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3750 
3751 	return err;
3752 }
3753 
3754 int mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3755 		      const struct ib_send_wr **bad_wr)
3756 {
3757 	return _mlx4_ib_post_send(ibqp, wr, bad_wr, false);
3758 }
3759 
3760 static int _mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3761 			      const struct ib_recv_wr **bad_wr, bool drain)
3762 {
3763 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3764 	struct mlx4_wqe_data_seg *scat;
3765 	unsigned long flags;
3766 	int err = 0;
3767 	int nreq;
3768 	int ind;
3769 	int max_gs;
3770 	int i;
3771 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3772 
3773 	max_gs = qp->rq.max_gs;
3774 	spin_lock_irqsave(&qp->rq.lock, flags);
3775 
3776 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3777 	    !drain) {
3778 		err = -EIO;
3779 		*bad_wr = wr;
3780 		nreq = 0;
3781 		goto out;
3782 	}
3783 
3784 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3785 
3786 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3787 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3788 			err = -ENOMEM;
3789 			*bad_wr = wr;
3790 			goto out;
3791 		}
3792 
3793 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3794 			err = -EINVAL;
3795 			*bad_wr = wr;
3796 			goto out;
3797 		}
3798 
3799 		scat = get_recv_wqe(qp, ind);
3800 
3801 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3802 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3803 			ib_dma_sync_single_for_device(ibqp->device,
3804 						      qp->sqp_proxy_rcv[ind].map,
3805 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3806 						      DMA_FROM_DEVICE);
3807 			scat->byte_count =
3808 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3809 			/* use dma lkey from upper layer entry */
3810 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3811 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3812 			scat++;
3813 			max_gs--;
3814 		}
3815 
3816 		for (i = 0; i < wr->num_sge; ++i)
3817 			__set_data_seg(scat + i, wr->sg_list + i);
3818 
3819 		if (i < max_gs) {
3820 			scat[i].byte_count = 0;
3821 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3822 			scat[i].addr       = 0;
3823 		}
3824 
3825 		qp->rq.wrid[ind] = wr->wr_id;
3826 
3827 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3828 	}
3829 
3830 out:
3831 	if (likely(nreq)) {
3832 		qp->rq.head += nreq;
3833 
3834 		/*
3835 		 * Make sure that descriptors are written before
3836 		 * doorbell record.
3837 		 */
3838 		wmb();
3839 
3840 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3841 	}
3842 
3843 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3844 
3845 	return err;
3846 }
3847 
3848 int mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3849 		      const struct ib_recv_wr **bad_wr)
3850 {
3851 	return _mlx4_ib_post_recv(ibqp, wr, bad_wr, false);
3852 }
3853 
3854 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3855 {
3856 	switch (mlx4_state) {
3857 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3858 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3859 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3860 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3861 	case MLX4_QP_STATE_SQ_DRAINING:
3862 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3863 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3864 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3865 	default:		     return -1;
3866 	}
3867 }
3868 
3869 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3870 {
3871 	switch (mlx4_mig_state) {
3872 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3873 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3874 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3875 	default: return -1;
3876 	}
3877 }
3878 
3879 static int to_ib_qp_access_flags(int mlx4_flags)
3880 {
3881 	int ib_flags = 0;
3882 
3883 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3884 		ib_flags |= IB_ACCESS_REMOTE_READ;
3885 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3886 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3887 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3888 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3889 
3890 	return ib_flags;
3891 }
3892 
3893 static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
3894 			    struct rdma_ah_attr *ah_attr,
3895 			    struct mlx4_qp_path *path)
3896 {
3897 	struct mlx4_dev *dev = ibdev->dev;
3898 	u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
3899 
3900 	memset(ah_attr, 0, sizeof(*ah_attr));
3901 	if (port_num == 0 || port_num > dev->caps.num_ports)
3902 		return;
3903 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
3904 
3905 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
3906 		rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
3907 			       ((path->sched_queue & 4) << 1));
3908 	else
3909 		rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
3910 	rdma_ah_set_port_num(ah_attr, port_num);
3911 
3912 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
3913 	rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
3914 	rdma_ah_set_static_rate(ah_attr,
3915 				path->static_rate ? path->static_rate - 5 : 0);
3916 	if (path->grh_mylmc & (1 << 7)) {
3917 		rdma_ah_set_grh(ah_attr, NULL,
3918 				be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
3919 				path->mgid_index,
3920 				path->hop_limit,
3921 				(be32_to_cpu(path->tclass_flowlabel)
3922 				 >> 20) & 0xff);
3923 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
3924 	}
3925 }
3926 
3927 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3928 		     struct ib_qp_init_attr *qp_init_attr)
3929 {
3930 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3931 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3932 	struct mlx4_qp_context context;
3933 	int mlx4_state;
3934 	int err = 0;
3935 
3936 	if (ibqp->rwq_ind_tbl)
3937 		return -EOPNOTSUPP;
3938 
3939 	mutex_lock(&qp->mutex);
3940 
3941 	if (qp->state == IB_QPS_RESET) {
3942 		qp_attr->qp_state = IB_QPS_RESET;
3943 		goto done;
3944 	}
3945 
3946 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3947 	if (err) {
3948 		err = -EINVAL;
3949 		goto out;
3950 	}
3951 
3952 	mlx4_state = be32_to_cpu(context.flags) >> 28;
3953 
3954 	qp->state		     = to_ib_qp_state(mlx4_state);
3955 	qp_attr->qp_state	     = qp->state;
3956 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
3957 	qp_attr->path_mig_state	     =
3958 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3959 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
3960 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3961 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
3962 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
3963 	qp_attr->qp_access_flags     =
3964 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
3965 
3966 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3967 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3968 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3969 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3970 		qp_attr->alt_port_num	=
3971 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
3972 	}
3973 
3974 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3975 	if (qp_attr->qp_state == IB_QPS_INIT)
3976 		qp_attr->port_num = qp->port;
3977 	else
3978 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3979 
3980 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3981 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3982 
3983 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3984 
3985 	qp_attr->max_dest_rd_atomic =
3986 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3987 	qp_attr->min_rnr_timer	    =
3988 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3989 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
3990 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
3991 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
3992 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
3993 
3994 done:
3995 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
3996 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
3997 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
3998 
3999 	if (!ibqp->uobject) {
4000 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
4001 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4002 	} else {
4003 		qp_attr->cap.max_send_wr  = 0;
4004 		qp_attr->cap.max_send_sge = 0;
4005 	}
4006 
4007 	/*
4008 	 * We don't support inline sends for kernel QPs (yet), and we
4009 	 * don't know what userspace's value should be.
4010 	 */
4011 	qp_attr->cap.max_inline_data = 0;
4012 
4013 	qp_init_attr->cap	     = qp_attr->cap;
4014 
4015 	qp_init_attr->create_flags = 0;
4016 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4017 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4018 
4019 	if (qp->flags & MLX4_IB_QP_LSO)
4020 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
4021 
4022 	if (qp->flags & MLX4_IB_QP_NETIF)
4023 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
4024 
4025 	qp_init_attr->sq_sig_type =
4026 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
4027 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4028 
4029 out:
4030 	mutex_unlock(&qp->mutex);
4031 	return err;
4032 }
4033 
4034 struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
4035 				struct ib_wq_init_attr *init_attr,
4036 				struct ib_udata *udata)
4037 {
4038 	struct mlx4_ib_dev *dev;
4039 	struct ib_qp_init_attr ib_qp_init_attr;
4040 	struct mlx4_ib_qp *qp;
4041 	struct mlx4_ib_create_wq ucmd;
4042 	int err, required_cmd_sz;
4043 
4044 	if (!udata)
4045 		return ERR_PTR(-EINVAL);
4046 
4047 	required_cmd_sz = offsetof(typeof(ucmd), comp_mask) +
4048 			  sizeof(ucmd.comp_mask);
4049 	if (udata->inlen < required_cmd_sz) {
4050 		pr_debug("invalid inlen\n");
4051 		return ERR_PTR(-EINVAL);
4052 	}
4053 
4054 	if (udata->inlen > sizeof(ucmd) &&
4055 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4056 				 udata->inlen - sizeof(ucmd))) {
4057 		pr_debug("inlen is not supported\n");
4058 		return ERR_PTR(-EOPNOTSUPP);
4059 	}
4060 
4061 	if (udata->outlen)
4062 		return ERR_PTR(-EOPNOTSUPP);
4063 
4064 	dev = to_mdev(pd->device);
4065 
4066 	if (init_attr->wq_type != IB_WQT_RQ) {
4067 		pr_debug("unsupported wq type %d\n", init_attr->wq_type);
4068 		return ERR_PTR(-EOPNOTSUPP);
4069 	}
4070 
4071 	if (init_attr->create_flags & ~IB_WQ_FLAGS_SCATTER_FCS) {
4072 		pr_debug("unsupported create_flags %u\n",
4073 			 init_attr->create_flags);
4074 		return ERR_PTR(-EOPNOTSUPP);
4075 	}
4076 
4077 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
4078 	if (!qp)
4079 		return ERR_PTR(-ENOMEM);
4080 
4081 	qp->pri.vid = 0xFFFF;
4082 	qp->alt.vid = 0xFFFF;
4083 
4084 	memset(&ib_qp_init_attr, 0, sizeof(ib_qp_init_attr));
4085 	ib_qp_init_attr.qp_context = init_attr->wq_context;
4086 	ib_qp_init_attr.qp_type = IB_QPT_RAW_PACKET;
4087 	ib_qp_init_attr.cap.max_recv_wr = init_attr->max_wr;
4088 	ib_qp_init_attr.cap.max_recv_sge = init_attr->max_sge;
4089 	ib_qp_init_attr.recv_cq = init_attr->cq;
4090 	ib_qp_init_attr.send_cq = ib_qp_init_attr.recv_cq; /* Dummy CQ */
4091 
4092 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS)
4093 		ib_qp_init_attr.create_flags |= IB_QP_CREATE_SCATTER_FCS;
4094 
4095 	err = create_qp_common(dev, pd, MLX4_IB_RWQ_SRC, &ib_qp_init_attr,
4096 			       udata, 0, &qp);
4097 	if (err) {
4098 		kfree(qp);
4099 		return ERR_PTR(err);
4100 	}
4101 
4102 	qp->ibwq.event_handler = init_attr->event_handler;
4103 	qp->ibwq.wq_num = qp->mqp.qpn;
4104 	qp->ibwq.state = IB_WQS_RESET;
4105 
4106 	return &qp->ibwq;
4107 }
4108 
4109 static int ib_wq2qp_state(enum ib_wq_state state)
4110 {
4111 	switch (state) {
4112 	case IB_WQS_RESET:
4113 		return IB_QPS_RESET;
4114 	case IB_WQS_RDY:
4115 		return IB_QPS_RTR;
4116 	default:
4117 		return IB_QPS_ERR;
4118 	}
4119 }
4120 
4121 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state)
4122 {
4123 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4124 	enum ib_qp_state qp_cur_state;
4125 	enum ib_qp_state qp_new_state;
4126 	int attr_mask;
4127 	int err;
4128 
4129 	/* ib_qp.state represents the WQ HW state while ib_wq.state represents
4130 	 * the WQ logic state.
4131 	 */
4132 	qp_cur_state = qp->state;
4133 	qp_new_state = ib_wq2qp_state(new_state);
4134 
4135 	if (ib_wq2qp_state(new_state) == qp_cur_state)
4136 		return 0;
4137 
4138 	if (new_state == IB_WQS_RDY) {
4139 		struct ib_qp_attr attr = {};
4140 
4141 		attr.port_num = qp->port;
4142 		attr_mask = IB_QP_PORT;
4143 
4144 		err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, &attr,
4145 					  attr_mask, IB_QPS_RESET, IB_QPS_INIT);
4146 		if (err) {
4147 			pr_debug("WQN=0x%06x failed to apply RST->INIT on the HW QP\n",
4148 				 ibwq->wq_num);
4149 			return err;
4150 		}
4151 
4152 		qp_cur_state = IB_QPS_INIT;
4153 	}
4154 
4155 	attr_mask = 0;
4156 	err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, attr_mask,
4157 				  qp_cur_state,  qp_new_state);
4158 
4159 	if (err && (qp_cur_state == IB_QPS_INIT)) {
4160 		qp_new_state = IB_QPS_RESET;
4161 		if (__mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL,
4162 					attr_mask, IB_QPS_INIT, IB_QPS_RESET)) {
4163 			pr_warn("WQN=0x%06x failed with reverting HW's resources failure\n",
4164 				ibwq->wq_num);
4165 			qp_new_state = IB_QPS_INIT;
4166 		}
4167 	}
4168 
4169 	qp->state = qp_new_state;
4170 
4171 	return err;
4172 }
4173 
4174 int mlx4_ib_modify_wq(struct ib_wq *ibwq, struct ib_wq_attr *wq_attr,
4175 		      u32 wq_attr_mask, struct ib_udata *udata)
4176 {
4177 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4178 	struct mlx4_ib_modify_wq ucmd = {};
4179 	size_t required_cmd_sz;
4180 	enum ib_wq_state cur_state, new_state;
4181 	int err = 0;
4182 
4183 	required_cmd_sz = offsetof(typeof(ucmd), reserved) +
4184 				   sizeof(ucmd.reserved);
4185 	if (udata->inlen < required_cmd_sz)
4186 		return -EINVAL;
4187 
4188 	if (udata->inlen > sizeof(ucmd) &&
4189 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4190 				 udata->inlen - sizeof(ucmd)))
4191 		return -EOPNOTSUPP;
4192 
4193 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4194 		return -EFAULT;
4195 
4196 	if (ucmd.comp_mask || ucmd.reserved)
4197 		return -EOPNOTSUPP;
4198 
4199 	if (wq_attr_mask & IB_WQ_FLAGS)
4200 		return -EOPNOTSUPP;
4201 
4202 	cur_state = wq_attr_mask & IB_WQ_CUR_STATE ? wq_attr->curr_wq_state :
4203 						     ibwq->state;
4204 	new_state = wq_attr_mask & IB_WQ_STATE ? wq_attr->wq_state : cur_state;
4205 
4206 	if (cur_state  < IB_WQS_RESET || cur_state  > IB_WQS_ERR ||
4207 	    new_state < IB_WQS_RESET || new_state > IB_WQS_ERR)
4208 		return -EINVAL;
4209 
4210 	if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR))
4211 		return -EINVAL;
4212 
4213 	if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET))
4214 		return -EINVAL;
4215 
4216 	/* Need to protect against the parent RSS which also may modify WQ
4217 	 * state.
4218 	 */
4219 	mutex_lock(&qp->mutex);
4220 
4221 	/* Can update HW state only if a RSS QP has already associated to this
4222 	 * WQ, so we can apply its port on the WQ.
4223 	 */
4224 	if (qp->rss_usecnt)
4225 		err = _mlx4_ib_modify_wq(ibwq, new_state);
4226 
4227 	if (!err)
4228 		ibwq->state = new_state;
4229 
4230 	mutex_unlock(&qp->mutex);
4231 
4232 	return err;
4233 }
4234 
4235 int mlx4_ib_destroy_wq(struct ib_wq *ibwq)
4236 {
4237 	struct mlx4_ib_dev *dev = to_mdev(ibwq->device);
4238 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4239 
4240 	if (qp->counter_index)
4241 		mlx4_ib_free_qp_counter(dev, qp);
4242 
4243 	destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, 1);
4244 
4245 	kfree(qp);
4246 
4247 	return 0;
4248 }
4249 
4250 struct ib_rwq_ind_table
4251 *mlx4_ib_create_rwq_ind_table(struct ib_device *device,
4252 			      struct ib_rwq_ind_table_init_attr *init_attr,
4253 			      struct ib_udata *udata)
4254 {
4255 	struct ib_rwq_ind_table *rwq_ind_table;
4256 	struct mlx4_ib_create_rwq_ind_tbl_resp resp = {};
4257 	unsigned int ind_tbl_size = 1 << init_attr->log_ind_tbl_size;
4258 	unsigned int base_wqn;
4259 	size_t min_resp_len;
4260 	int i;
4261 	int err;
4262 
4263 	if (udata->inlen > 0 &&
4264 	    !ib_is_udata_cleared(udata, 0,
4265 				 udata->inlen))
4266 		return ERR_PTR(-EOPNOTSUPP);
4267 
4268 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4269 	if (udata->outlen && udata->outlen < min_resp_len)
4270 		return ERR_PTR(-EINVAL);
4271 
4272 	if (ind_tbl_size >
4273 	    device->attrs.rss_caps.max_rwq_indirection_table_size) {
4274 		pr_debug("log_ind_tbl_size = %d is bigger than supported = %d\n",
4275 			 ind_tbl_size,
4276 			 device->attrs.rss_caps.max_rwq_indirection_table_size);
4277 		return ERR_PTR(-EINVAL);
4278 	}
4279 
4280 	base_wqn = init_attr->ind_tbl[0]->wq_num;
4281 
4282 	if (base_wqn % ind_tbl_size) {
4283 		pr_debug("WQN=0x%x isn't aligned with indirection table size\n",
4284 			 base_wqn);
4285 		return ERR_PTR(-EINVAL);
4286 	}
4287 
4288 	for (i = 1; i < ind_tbl_size; i++) {
4289 		if (++base_wqn != init_attr->ind_tbl[i]->wq_num) {
4290 			pr_debug("indirection table's WQNs aren't consecutive\n");
4291 			return ERR_PTR(-EINVAL);
4292 		}
4293 	}
4294 
4295 	rwq_ind_table = kzalloc(sizeof(*rwq_ind_table), GFP_KERNEL);
4296 	if (!rwq_ind_table)
4297 		return ERR_PTR(-ENOMEM);
4298 
4299 	if (udata->outlen) {
4300 		resp.response_length = offsetof(typeof(resp), response_length) +
4301 					sizeof(resp.response_length);
4302 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4303 		if (err)
4304 			goto err;
4305 	}
4306 
4307 	return rwq_ind_table;
4308 
4309 err:
4310 	kfree(rwq_ind_table);
4311 	return ERR_PTR(err);
4312 }
4313 
4314 int mlx4_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4315 {
4316 	kfree(ib_rwq_ind_tbl);
4317 	return 0;
4318 }
4319 
4320 struct mlx4_ib_drain_cqe {
4321 	struct ib_cqe cqe;
4322 	struct completion done;
4323 };
4324 
4325 static void mlx4_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
4326 {
4327 	struct mlx4_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
4328 						     struct mlx4_ib_drain_cqe,
4329 						     cqe);
4330 
4331 	complete(&cqe->done);
4332 }
4333 
4334 /* This function returns only once the drained WR was completed */
4335 static void handle_drain_completion(struct ib_cq *cq,
4336 				    struct mlx4_ib_drain_cqe *sdrain,
4337 				    struct mlx4_ib_dev *dev)
4338 {
4339 	struct mlx4_dev *mdev = dev->dev;
4340 
4341 	if (cq->poll_ctx == IB_POLL_DIRECT) {
4342 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
4343 			ib_process_cq_direct(cq, -1);
4344 		return;
4345 	}
4346 
4347 	if (mdev->persist->state == MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4348 		struct mlx4_ib_cq *mcq = to_mcq(cq);
4349 		bool triggered = false;
4350 		unsigned long flags;
4351 
4352 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
4353 		/* Make sure that the CQ handler won't run if wasn't run yet */
4354 		if (!mcq->mcq.reset_notify_added)
4355 			mcq->mcq.reset_notify_added = 1;
4356 		else
4357 			triggered = true;
4358 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
4359 
4360 		if (triggered) {
4361 			/* Wait for any scheduled/running task to be ended */
4362 			switch (cq->poll_ctx) {
4363 			case IB_POLL_SOFTIRQ:
4364 				irq_poll_disable(&cq->iop);
4365 				irq_poll_enable(&cq->iop);
4366 				break;
4367 			case IB_POLL_WORKQUEUE:
4368 				cancel_work_sync(&cq->work);
4369 				break;
4370 			default:
4371 				WARN_ON_ONCE(1);
4372 			}
4373 		}
4374 
4375 		/* Run the CQ handler - this makes sure that the drain WR will
4376 		 * be processed if wasn't processed yet.
4377 		 */
4378 		mcq->mcq.comp(&mcq->mcq);
4379 	}
4380 
4381 	wait_for_completion(&sdrain->done);
4382 }
4383 
4384 void mlx4_ib_drain_sq(struct ib_qp *qp)
4385 {
4386 	struct ib_cq *cq = qp->send_cq;
4387 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4388 	struct mlx4_ib_drain_cqe sdrain;
4389 	const struct ib_send_wr *bad_swr;
4390 	struct ib_rdma_wr swr = {
4391 		.wr = {
4392 			.next = NULL,
4393 			{ .wr_cqe	= &sdrain.cqe, },
4394 			.opcode	= IB_WR_RDMA_WRITE,
4395 		},
4396 	};
4397 	int ret;
4398 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
4399 	struct mlx4_dev *mdev = dev->dev;
4400 
4401 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4402 	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4403 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4404 		return;
4405 	}
4406 
4407 	sdrain.cqe.done = mlx4_ib_drain_qp_done;
4408 	init_completion(&sdrain.done);
4409 
4410 	ret = _mlx4_ib_post_send(qp, &swr.wr, &bad_swr, true);
4411 	if (ret) {
4412 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4413 		return;
4414 	}
4415 
4416 	handle_drain_completion(cq, &sdrain, dev);
4417 }
4418 
4419 void mlx4_ib_drain_rq(struct ib_qp *qp)
4420 {
4421 	struct ib_cq *cq = qp->recv_cq;
4422 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4423 	struct mlx4_ib_drain_cqe rdrain;
4424 	struct ib_recv_wr rwr = {};
4425 	const struct ib_recv_wr *bad_rwr;
4426 	int ret;
4427 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
4428 	struct mlx4_dev *mdev = dev->dev;
4429 
4430 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4431 	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4432 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4433 		return;
4434 	}
4435 
4436 	rwr.wr_cqe = &rdrain.cqe;
4437 	rdrain.cqe.done = mlx4_ib_drain_qp_done;
4438 	init_completion(&rdrain.done);
4439 
4440 	ret = _mlx4_ib_post_recv(qp, &rwr, &bad_rwr, true);
4441 	if (ret) {
4442 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4443 		return;
4444 	}
4445 
4446 	handle_drain_completion(cq, &rdrain, dev);
4447 }
4448