xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision f3a8b664)
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 #include <linux/vmalloc.h>
40 
41 #include <rdma/ib_cache.h>
42 #include <rdma/ib_pack.h>
43 #include <rdma/ib_addr.h>
44 #include <rdma/ib_mad.h>
45 
46 #include <linux/mlx4/driver.h>
47 #include <linux/mlx4/qp.h>
48 
49 #include "mlx4_ib.h"
50 #include <rdma/mlx4-abi.h>
51 
52 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
53 			     struct mlx4_ib_cq *recv_cq);
54 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
55 			       struct mlx4_ib_cq *recv_cq);
56 
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 enum {
80 	MLX4_IB_IBOE_ETHERTYPE		= 0x8915
81 };
82 
83 struct mlx4_ib_sqp {
84 	struct mlx4_ib_qp	qp;
85 	int			pkey_index;
86 	u32			qkey;
87 	u32			send_psn;
88 	struct ib_ud_header	ud_header;
89 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
90 	struct ib_qp		*roce_v2_gsi;
91 };
92 
93 enum {
94 	MLX4_IB_MIN_SQ_STRIDE	= 6,
95 	MLX4_IB_CACHE_LINE_SIZE	= 64,
96 };
97 
98 enum {
99 	MLX4_RAW_QP_MTU		= 7,
100 	MLX4_RAW_QP_MSGMAX	= 31,
101 };
102 
103 #ifndef ETH_ALEN
104 #define ETH_ALEN        6
105 #endif
106 
107 static const __be32 mlx4_ib_opcode[] = {
108 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
109 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
110 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
111 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
112 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
113 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
114 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
115 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
116 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
117 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
118 	[IB_WR_REG_MR]				= cpu_to_be32(MLX4_OPCODE_FMR),
119 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
120 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
121 };
122 
123 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
124 {
125 	return container_of(mqp, struct mlx4_ib_sqp, qp);
126 }
127 
128 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
129 {
130 	if (!mlx4_is_master(dev->dev))
131 		return 0;
132 
133 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
134 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
135 		8 * MLX4_MFUNC_MAX;
136 }
137 
138 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
139 {
140 	int proxy_sqp = 0;
141 	int real_sqp = 0;
142 	int i;
143 	/* PPF or Native -- real SQP */
144 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
145 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
146 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
147 	if (real_sqp)
148 		return 1;
149 	/* VF or PF -- proxy SQP */
150 	if (mlx4_is_mfunc(dev->dev)) {
151 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
152 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] ||
153 			    qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) {
154 				proxy_sqp = 1;
155 				break;
156 			}
157 		}
158 	}
159 	if (proxy_sqp)
160 		return 1;
161 
162 	return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
163 }
164 
165 /* used for INIT/CLOSE port logic */
166 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
167 {
168 	int proxy_qp0 = 0;
169 	int real_qp0 = 0;
170 	int i;
171 	/* PPF or Native -- real QP0 */
172 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
173 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
174 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
175 	if (real_qp0)
176 		return 1;
177 	/* VF or PF -- proxy QP0 */
178 	if (mlx4_is_mfunc(dev->dev)) {
179 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
180 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) {
181 				proxy_qp0 = 1;
182 				break;
183 			}
184 		}
185 	}
186 	return proxy_qp0;
187 }
188 
189 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
190 {
191 	return mlx4_buf_offset(&qp->buf, offset);
192 }
193 
194 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
195 {
196 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
197 }
198 
199 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
200 {
201 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
202 }
203 
204 /*
205  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
206  * first four bytes of every 64 byte chunk with
207  *     0x7FFFFFF | (invalid_ownership_value << 31).
208  *
209  * When the max work request size is less than or equal to the WQE
210  * basic block size, as an optimization, we can stamp all WQEs with
211  * 0xffffffff, and skip the very first chunk of each WQE.
212  */
213 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
214 {
215 	__be32 *wqe;
216 	int i;
217 	int s;
218 	int ind;
219 	void *buf;
220 	__be32 stamp;
221 	struct mlx4_wqe_ctrl_seg *ctrl;
222 
223 	if (qp->sq_max_wqes_per_wr > 1) {
224 		s = roundup(size, 1U << qp->sq.wqe_shift);
225 		for (i = 0; i < s; i += 64) {
226 			ind = (i >> qp->sq.wqe_shift) + n;
227 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
228 						       cpu_to_be32(0xffffffff);
229 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
230 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
231 			*wqe = stamp;
232 		}
233 	} else {
234 		ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
235 		s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4;
236 		for (i = 64; i < s; i += 64) {
237 			wqe = buf + i;
238 			*wqe = cpu_to_be32(0xffffffff);
239 		}
240 	}
241 }
242 
243 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
244 {
245 	struct mlx4_wqe_ctrl_seg *ctrl;
246 	struct mlx4_wqe_inline_seg *inl;
247 	void *wqe;
248 	int s;
249 
250 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
251 	s = sizeof(struct mlx4_wqe_ctrl_seg);
252 
253 	if (qp->ibqp.qp_type == IB_QPT_UD) {
254 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
255 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
256 		memset(dgram, 0, sizeof *dgram);
257 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
258 		s += sizeof(struct mlx4_wqe_datagram_seg);
259 	}
260 
261 	/* Pad the remainder of the WQE with an inline data segment. */
262 	if (size > s) {
263 		inl = wqe + s;
264 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
265 	}
266 	ctrl->srcrb_flags = 0;
267 	ctrl->qpn_vlan.fence_size = size / 16;
268 	/*
269 	 * Make sure descriptor is fully written before setting ownership bit
270 	 * (because HW can start executing as soon as we do).
271 	 */
272 	wmb();
273 
274 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
275 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
276 
277 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
278 }
279 
280 /* Post NOP WQE to prevent wrap-around in the middle of WR */
281 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
282 {
283 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
284 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
285 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
286 		ind += s;
287 	}
288 	return ind;
289 }
290 
291 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
292 {
293 	struct ib_event event;
294 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
295 
296 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
297 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
298 
299 	if (ibqp->event_handler) {
300 		event.device     = ibqp->device;
301 		event.element.qp = ibqp;
302 		switch (type) {
303 		case MLX4_EVENT_TYPE_PATH_MIG:
304 			event.event = IB_EVENT_PATH_MIG;
305 			break;
306 		case MLX4_EVENT_TYPE_COMM_EST:
307 			event.event = IB_EVENT_COMM_EST;
308 			break;
309 		case MLX4_EVENT_TYPE_SQ_DRAINED:
310 			event.event = IB_EVENT_SQ_DRAINED;
311 			break;
312 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
313 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
314 			break;
315 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
316 			event.event = IB_EVENT_QP_FATAL;
317 			break;
318 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
319 			event.event = IB_EVENT_PATH_MIG_ERR;
320 			break;
321 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
322 			event.event = IB_EVENT_QP_REQ_ERR;
323 			break;
324 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
325 			event.event = IB_EVENT_QP_ACCESS_ERR;
326 			break;
327 		default:
328 			pr_warn("Unexpected event type %d "
329 			       "on QP %06x\n", type, qp->qpn);
330 			return;
331 		}
332 
333 		ibqp->event_handler(&event, ibqp->qp_context);
334 	}
335 }
336 
337 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
338 {
339 	/*
340 	 * UD WQEs must have a datagram segment.
341 	 * RC and UC WQEs might have a remote address segment.
342 	 * MLX WQEs need two extra inline data segments (for the UD
343 	 * header and space for the ICRC).
344 	 */
345 	switch (type) {
346 	case MLX4_IB_QPT_UD:
347 		return sizeof (struct mlx4_wqe_ctrl_seg) +
348 			sizeof (struct mlx4_wqe_datagram_seg) +
349 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
350 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
351 	case MLX4_IB_QPT_PROXY_SMI:
352 	case MLX4_IB_QPT_PROXY_GSI:
353 		return sizeof (struct mlx4_wqe_ctrl_seg) +
354 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
355 	case MLX4_IB_QPT_TUN_SMI_OWNER:
356 	case MLX4_IB_QPT_TUN_GSI:
357 		return sizeof (struct mlx4_wqe_ctrl_seg) +
358 			sizeof (struct mlx4_wqe_datagram_seg);
359 
360 	case MLX4_IB_QPT_UC:
361 		return sizeof (struct mlx4_wqe_ctrl_seg) +
362 			sizeof (struct mlx4_wqe_raddr_seg);
363 	case MLX4_IB_QPT_RC:
364 		return sizeof (struct mlx4_wqe_ctrl_seg) +
365 			sizeof (struct mlx4_wqe_masked_atomic_seg) +
366 			sizeof (struct mlx4_wqe_raddr_seg);
367 	case MLX4_IB_QPT_SMI:
368 	case MLX4_IB_QPT_GSI:
369 		return sizeof (struct mlx4_wqe_ctrl_seg) +
370 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
371 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
372 					   MLX4_INLINE_ALIGN) *
373 			      sizeof (struct mlx4_wqe_inline_seg),
374 			      sizeof (struct mlx4_wqe_data_seg)) +
375 			ALIGN(4 +
376 			      sizeof (struct mlx4_wqe_inline_seg),
377 			      sizeof (struct mlx4_wqe_data_seg));
378 	default:
379 		return sizeof (struct mlx4_wqe_ctrl_seg);
380 	}
381 }
382 
383 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
384 		       int is_user, int has_rq, struct mlx4_ib_qp *qp)
385 {
386 	/* Sanity check RQ size before proceeding */
387 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
388 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
389 		return -EINVAL;
390 
391 	if (!has_rq) {
392 		if (cap->max_recv_wr)
393 			return -EINVAL;
394 
395 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
396 	} else {
397 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
398 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
399 			return -EINVAL;
400 
401 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
402 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
403 		qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
404 	}
405 
406 	/* leave userspace return values as they were, so as not to break ABI */
407 	if (is_user) {
408 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
409 		cap->max_recv_sge = qp->rq.max_gs;
410 	} else {
411 		cap->max_recv_wr  = qp->rq.max_post =
412 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
413 		cap->max_recv_sge = min(qp->rq.max_gs,
414 					min(dev->dev->caps.max_sq_sg,
415 					    dev->dev->caps.max_rq_sg));
416 	}
417 
418 	return 0;
419 }
420 
421 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
422 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp,
423 			      bool shrink_wqe)
424 {
425 	int s;
426 
427 	/* Sanity check SQ size before proceeding */
428 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
429 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
430 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
431 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
432 		return -EINVAL;
433 
434 	/*
435 	 * For MLX transport we need 2 extra S/G entries:
436 	 * one for the header and one for the checksum at the end
437 	 */
438 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
439 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
440 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
441 		return -EINVAL;
442 
443 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
444 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
445 		send_wqe_overhead(type, qp->flags);
446 
447 	if (s > dev->dev->caps.max_sq_desc_sz)
448 		return -EINVAL;
449 
450 	/*
451 	 * Hermon supports shrinking WQEs, such that a single work
452 	 * request can include multiple units of 1 << wqe_shift.  This
453 	 * way, work requests can differ in size, and do not have to
454 	 * be a power of 2 in size, saving memory and speeding up send
455 	 * WR posting.  Unfortunately, if we do this then the
456 	 * wqe_index field in CQEs can't be used to look up the WR ID
457 	 * anymore, so we do this only if selective signaling is off.
458 	 *
459 	 * Further, on 32-bit platforms, we can't use vmap() to make
460 	 * the QP buffer virtually contiguous.  Thus we have to use
461 	 * constant-sized WRs to make sure a WR is always fully within
462 	 * a single page-sized chunk.
463 	 *
464 	 * Finally, we use NOP work requests to pad the end of the
465 	 * work queue, to avoid wrap-around in the middle of WR.  We
466 	 * set NEC bit to avoid getting completions with error for
467 	 * these NOP WRs, but since NEC is only supported starting
468 	 * with firmware 2.2.232, we use constant-sized WRs for older
469 	 * firmware.
470 	 *
471 	 * And, since MLX QPs only support SEND, we use constant-sized
472 	 * WRs in this case.
473 	 *
474 	 * We look for the smallest value of wqe_shift such that the
475 	 * resulting number of wqes does not exceed device
476 	 * capabilities.
477 	 *
478 	 * We set WQE size to at least 64 bytes, this way stamping
479 	 * invalidates each WQE.
480 	 */
481 	if (shrink_wqe && dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
482 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
483 	    type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
484 	    !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
485 		      MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
486 		qp->sq.wqe_shift = ilog2(64);
487 	else
488 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
489 
490 	for (;;) {
491 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
492 
493 		/*
494 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
495 		 * allow HW to prefetch.
496 		 */
497 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
498 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
499 						    qp->sq_max_wqes_per_wr +
500 						    qp->sq_spare_wqes);
501 
502 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
503 			break;
504 
505 		if (qp->sq_max_wqes_per_wr <= 1)
506 			return -EINVAL;
507 
508 		++qp->sq.wqe_shift;
509 	}
510 
511 	qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
512 			     (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
513 			 send_wqe_overhead(type, qp->flags)) /
514 		sizeof (struct mlx4_wqe_data_seg);
515 
516 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
517 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
518 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
519 		qp->rq.offset = 0;
520 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
521 	} else {
522 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
523 		qp->sq.offset = 0;
524 	}
525 
526 	cap->max_send_wr  = qp->sq.max_post =
527 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
528 	cap->max_send_sge = min(qp->sq.max_gs,
529 				min(dev->dev->caps.max_sq_sg,
530 				    dev->dev->caps.max_rq_sg));
531 	/* We don't support inline sends for kernel QPs (yet) */
532 	cap->max_inline_data = 0;
533 
534 	return 0;
535 }
536 
537 static int set_user_sq_size(struct mlx4_ib_dev *dev,
538 			    struct mlx4_ib_qp *qp,
539 			    struct mlx4_ib_create_qp *ucmd)
540 {
541 	/* Sanity check SQ size before proceeding */
542 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
543 	    ucmd->log_sq_stride >
544 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
545 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
546 		return -EINVAL;
547 
548 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
549 	qp->sq.wqe_shift = ucmd->log_sq_stride;
550 
551 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
552 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
553 
554 	return 0;
555 }
556 
557 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
558 {
559 	int i;
560 
561 	qp->sqp_proxy_rcv =
562 		kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
563 			GFP_KERNEL);
564 	if (!qp->sqp_proxy_rcv)
565 		return -ENOMEM;
566 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
567 		qp->sqp_proxy_rcv[i].addr =
568 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
569 				GFP_KERNEL);
570 		if (!qp->sqp_proxy_rcv[i].addr)
571 			goto err;
572 		qp->sqp_proxy_rcv[i].map =
573 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
574 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
575 					  DMA_FROM_DEVICE);
576 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
577 			kfree(qp->sqp_proxy_rcv[i].addr);
578 			goto err;
579 		}
580 	}
581 	return 0;
582 
583 err:
584 	while (i > 0) {
585 		--i;
586 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
587 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
588 				    DMA_FROM_DEVICE);
589 		kfree(qp->sqp_proxy_rcv[i].addr);
590 	}
591 	kfree(qp->sqp_proxy_rcv);
592 	qp->sqp_proxy_rcv = NULL;
593 	return -ENOMEM;
594 }
595 
596 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
597 {
598 	int i;
599 
600 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
601 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
602 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
603 				    DMA_FROM_DEVICE);
604 		kfree(qp->sqp_proxy_rcv[i].addr);
605 	}
606 	kfree(qp->sqp_proxy_rcv);
607 }
608 
609 static int qp_has_rq(struct ib_qp_init_attr *attr)
610 {
611 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
612 		return 0;
613 
614 	return !attr->srq;
615 }
616 
617 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
618 {
619 	int i;
620 	for (i = 0; i < dev->caps.num_ports; i++) {
621 		if (qpn == dev->caps.qp0_proxy[i])
622 			return !!dev->caps.qp0_qkey[i];
623 	}
624 	return 0;
625 }
626 
627 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
628 				    struct mlx4_ib_qp *qp)
629 {
630 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
631 	mlx4_counter_free(dev->dev, qp->counter_index->index);
632 	list_del(&qp->counter_index->list);
633 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
634 
635 	kfree(qp->counter_index);
636 	qp->counter_index = NULL;
637 }
638 
639 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
640 			    struct ib_qp_init_attr *init_attr,
641 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp,
642 			    gfp_t gfp)
643 {
644 	int qpn;
645 	int err;
646 	struct ib_qp_cap backup_cap;
647 	struct mlx4_ib_sqp *sqp;
648 	struct mlx4_ib_qp *qp;
649 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
650 	struct mlx4_ib_cq *mcq;
651 	unsigned long flags;
652 
653 	/* When tunneling special qps, we use a plain UD qp */
654 	if (sqpn) {
655 		if (mlx4_is_mfunc(dev->dev) &&
656 		    (!mlx4_is_master(dev->dev) ||
657 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
658 			if (init_attr->qp_type == IB_QPT_GSI)
659 				qp_type = MLX4_IB_QPT_PROXY_GSI;
660 			else {
661 				if (mlx4_is_master(dev->dev) ||
662 				    qp0_enabled_vf(dev->dev, sqpn))
663 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
664 				else
665 					qp_type = MLX4_IB_QPT_PROXY_SMI;
666 			}
667 		}
668 		qpn = sqpn;
669 		/* add extra sg entry for tunneling */
670 		init_attr->cap.max_recv_sge++;
671 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
672 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
673 			container_of(init_attr,
674 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
675 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
676 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
677 		    !mlx4_is_master(dev->dev))
678 			return -EINVAL;
679 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
680 			qp_type = MLX4_IB_QPT_TUN_GSI;
681 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
682 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
683 					     tnl_init->port))
684 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
685 		else
686 			qp_type = MLX4_IB_QPT_TUN_SMI;
687 		/* we are definitely in the PPF here, since we are creating
688 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
689 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
690 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
691 		sqpn = qpn;
692 	}
693 
694 	if (!*caller_qp) {
695 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
696 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
697 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
698 			sqp = kzalloc(sizeof (struct mlx4_ib_sqp), gfp);
699 			if (!sqp)
700 				return -ENOMEM;
701 			qp = &sqp->qp;
702 			qp->pri.vid = 0xFFFF;
703 			qp->alt.vid = 0xFFFF;
704 		} else {
705 			qp = kzalloc(sizeof (struct mlx4_ib_qp), gfp);
706 			if (!qp)
707 				return -ENOMEM;
708 			qp->pri.vid = 0xFFFF;
709 			qp->alt.vid = 0xFFFF;
710 		}
711 	} else
712 		qp = *caller_qp;
713 
714 	qp->mlx4_ib_qp_type = qp_type;
715 
716 	mutex_init(&qp->mutex);
717 	spin_lock_init(&qp->sq.lock);
718 	spin_lock_init(&qp->rq.lock);
719 	INIT_LIST_HEAD(&qp->gid_list);
720 	INIT_LIST_HEAD(&qp->steering_rules);
721 
722 	qp->state	 = IB_QPS_RESET;
723 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
724 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
725 
726 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
727 	if (err)
728 		goto err;
729 
730 	if (pd->uobject) {
731 		struct mlx4_ib_create_qp ucmd;
732 
733 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
734 			err = -EFAULT;
735 			goto err;
736 		}
737 
738 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
739 
740 		err = set_user_sq_size(dev, qp, &ucmd);
741 		if (err)
742 			goto err;
743 
744 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
745 				       qp->buf_size, 0, 0);
746 		if (IS_ERR(qp->umem)) {
747 			err = PTR_ERR(qp->umem);
748 			goto err;
749 		}
750 
751 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
752 				    ilog2(qp->umem->page_size), &qp->mtt);
753 		if (err)
754 			goto err_buf;
755 
756 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
757 		if (err)
758 			goto err_mtt;
759 
760 		if (qp_has_rq(init_attr)) {
761 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
762 						  ucmd.db_addr, &qp->db);
763 			if (err)
764 				goto err_mtt;
765 		}
766 	} else {
767 		qp->sq_no_prefetch = 0;
768 
769 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
770 			qp->flags |= MLX4_IB_QP_LSO;
771 
772 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
773 			if (dev->steering_support ==
774 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
775 				qp->flags |= MLX4_IB_QP_NETIF;
776 			else
777 				goto err;
778 		}
779 
780 		memcpy(&backup_cap, &init_attr->cap, sizeof(backup_cap));
781 		err = set_kernel_sq_size(dev, &init_attr->cap,
782 					 qp_type, qp, true);
783 		if (err)
784 			goto err;
785 
786 		if (qp_has_rq(init_attr)) {
787 			err = mlx4_db_alloc(dev->dev, &qp->db, 0, gfp);
788 			if (err)
789 				goto err;
790 
791 			*qp->db.db = 0;
792 		}
793 
794 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, qp->buf_size,
795 				   &qp->buf, gfp)) {
796 			memcpy(&init_attr->cap, &backup_cap,
797 			       sizeof(backup_cap));
798 			err = set_kernel_sq_size(dev, &init_attr->cap, qp_type,
799 						 qp, false);
800 			if (err)
801 				goto err_db;
802 
803 			if (mlx4_buf_alloc(dev->dev, qp->buf_size,
804 					   PAGE_SIZE * 2, &qp->buf, gfp)) {
805 				err = -ENOMEM;
806 				goto err_db;
807 			}
808 		}
809 
810 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
811 				    &qp->mtt);
812 		if (err)
813 			goto err_buf;
814 
815 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf, gfp);
816 		if (err)
817 			goto err_mtt;
818 
819 		qp->sq.wrid = kmalloc_array(qp->sq.wqe_cnt, sizeof(u64),
820 					gfp | __GFP_NOWARN);
821 		if (!qp->sq.wrid)
822 			qp->sq.wrid = __vmalloc(qp->sq.wqe_cnt * sizeof(u64),
823 						gfp, PAGE_KERNEL);
824 		qp->rq.wrid = kmalloc_array(qp->rq.wqe_cnt, sizeof(u64),
825 					gfp | __GFP_NOWARN);
826 		if (!qp->rq.wrid)
827 			qp->rq.wrid = __vmalloc(qp->rq.wqe_cnt * sizeof(u64),
828 						gfp, PAGE_KERNEL);
829 		if (!qp->sq.wrid || !qp->rq.wrid) {
830 			err = -ENOMEM;
831 			goto err_wrid;
832 		}
833 	}
834 
835 	if (sqpn) {
836 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
837 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
838 			if (alloc_proxy_bufs(pd->device, qp)) {
839 				err = -ENOMEM;
840 				goto err_wrid;
841 			}
842 		}
843 	} else {
844 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
845 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
846 		 * VLAN insertion. */
847 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
848 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
849 						    (init_attr->cap.max_send_wr ?
850 						     MLX4_RESERVE_ETH_BF_QP : 0) |
851 						    (init_attr->cap.max_recv_wr ?
852 						     MLX4_RESERVE_A0_QP : 0));
853 		else
854 			if (qp->flags & MLX4_IB_QP_NETIF)
855 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
856 			else
857 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
858 							    &qpn, 0);
859 		if (err)
860 			goto err_proxy;
861 	}
862 
863 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
864 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
865 
866 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp, gfp);
867 	if (err)
868 		goto err_qpn;
869 
870 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
871 		qp->mqp.qpn |= (1 << 23);
872 
873 	/*
874 	 * Hardware wants QPN written in big-endian order (after
875 	 * shifting) for send doorbell.  Precompute this value to save
876 	 * a little bit when posting sends.
877 	 */
878 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
879 
880 	qp->mqp.event = mlx4_ib_qp_event;
881 	if (!*caller_qp)
882 		*caller_qp = qp;
883 
884 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
885 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
886 			 to_mcq(init_attr->recv_cq));
887 	/* Maintain device to QPs access, needed for further handling
888 	 * via reset flow
889 	 */
890 	list_add_tail(&qp->qps_list, &dev->qp_list);
891 	/* Maintain CQ to QPs access, needed for further handling
892 	 * via reset flow
893 	 */
894 	mcq = to_mcq(init_attr->send_cq);
895 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
896 	mcq = to_mcq(init_attr->recv_cq);
897 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
898 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
899 			   to_mcq(init_attr->recv_cq));
900 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
901 	return 0;
902 
903 err_qpn:
904 	if (!sqpn) {
905 		if (qp->flags & MLX4_IB_QP_NETIF)
906 			mlx4_ib_steer_qp_free(dev, qpn, 1);
907 		else
908 			mlx4_qp_release_range(dev->dev, qpn, 1);
909 	}
910 err_proxy:
911 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
912 		free_proxy_bufs(pd->device, qp);
913 err_wrid:
914 	if (pd->uobject) {
915 		if (qp_has_rq(init_attr))
916 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
917 	} else {
918 		kvfree(qp->sq.wrid);
919 		kvfree(qp->rq.wrid);
920 	}
921 
922 err_mtt:
923 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
924 
925 err_buf:
926 	if (pd->uobject)
927 		ib_umem_release(qp->umem);
928 	else
929 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
930 
931 err_db:
932 	if (!pd->uobject && qp_has_rq(init_attr))
933 		mlx4_db_free(dev->dev, &qp->db);
934 
935 err:
936 	if (!*caller_qp)
937 		kfree(qp);
938 	return err;
939 }
940 
941 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
942 {
943 	switch (state) {
944 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
945 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
946 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
947 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
948 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
949 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
950 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
951 	default:		return -1;
952 	}
953 }
954 
955 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
956 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
957 {
958 	if (send_cq == recv_cq) {
959 		spin_lock(&send_cq->lock);
960 		__acquire(&recv_cq->lock);
961 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
962 		spin_lock(&send_cq->lock);
963 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
964 	} else {
965 		spin_lock(&recv_cq->lock);
966 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
967 	}
968 }
969 
970 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
971 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
972 {
973 	if (send_cq == recv_cq) {
974 		__release(&recv_cq->lock);
975 		spin_unlock(&send_cq->lock);
976 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
977 		spin_unlock(&recv_cq->lock);
978 		spin_unlock(&send_cq->lock);
979 	} else {
980 		spin_unlock(&send_cq->lock);
981 		spin_unlock(&recv_cq->lock);
982 	}
983 }
984 
985 static void del_gid_entries(struct mlx4_ib_qp *qp)
986 {
987 	struct mlx4_ib_gid_entry *ge, *tmp;
988 
989 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
990 		list_del(&ge->list);
991 		kfree(ge);
992 	}
993 }
994 
995 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
996 {
997 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
998 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
999 	else
1000 		return to_mpd(qp->ibqp.pd);
1001 }
1002 
1003 static void get_cqs(struct mlx4_ib_qp *qp,
1004 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1005 {
1006 	switch (qp->ibqp.qp_type) {
1007 	case IB_QPT_XRC_TGT:
1008 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1009 		*recv_cq = *send_cq;
1010 		break;
1011 	case IB_QPT_XRC_INI:
1012 		*send_cq = to_mcq(qp->ibqp.send_cq);
1013 		*recv_cq = *send_cq;
1014 		break;
1015 	default:
1016 		*send_cq = to_mcq(qp->ibqp.send_cq);
1017 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
1018 		break;
1019 	}
1020 }
1021 
1022 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1023 			      int is_user)
1024 {
1025 	struct mlx4_ib_cq *send_cq, *recv_cq;
1026 	unsigned long flags;
1027 
1028 	if (qp->state != IB_QPS_RESET) {
1029 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1030 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1031 			pr_warn("modify QP %06x to RESET failed.\n",
1032 			       qp->mqp.qpn);
1033 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1034 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1035 			qp->pri.smac = 0;
1036 			qp->pri.smac_port = 0;
1037 		}
1038 		if (qp->alt.smac) {
1039 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1040 			qp->alt.smac = 0;
1041 		}
1042 		if (qp->pri.vid < 0x1000) {
1043 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1044 			qp->pri.vid = 0xFFFF;
1045 			qp->pri.candidate_vid = 0xFFFF;
1046 			qp->pri.update_vid = 0;
1047 		}
1048 		if (qp->alt.vid < 0x1000) {
1049 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1050 			qp->alt.vid = 0xFFFF;
1051 			qp->alt.candidate_vid = 0xFFFF;
1052 			qp->alt.update_vid = 0;
1053 		}
1054 	}
1055 
1056 	get_cqs(qp, &send_cq, &recv_cq);
1057 
1058 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1059 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1060 
1061 	/* del from lists under both locks above to protect reset flow paths */
1062 	list_del(&qp->qps_list);
1063 	list_del(&qp->cq_send_list);
1064 	list_del(&qp->cq_recv_list);
1065 	if (!is_user) {
1066 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1067 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1068 		if (send_cq != recv_cq)
1069 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1070 	}
1071 
1072 	mlx4_qp_remove(dev->dev, &qp->mqp);
1073 
1074 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1075 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1076 
1077 	mlx4_qp_free(dev->dev, &qp->mqp);
1078 
1079 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1080 		if (qp->flags & MLX4_IB_QP_NETIF)
1081 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1082 		else
1083 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1084 	}
1085 
1086 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1087 
1088 	if (is_user) {
1089 		if (qp->rq.wqe_cnt)
1090 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
1091 					      &qp->db);
1092 		ib_umem_release(qp->umem);
1093 	} else {
1094 		kvfree(qp->sq.wrid);
1095 		kvfree(qp->rq.wrid);
1096 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1097 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1098 			free_proxy_bufs(&dev->ib_dev, qp);
1099 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1100 		if (qp->rq.wqe_cnt)
1101 			mlx4_db_free(dev->dev, &qp->db);
1102 	}
1103 
1104 	del_gid_entries(qp);
1105 }
1106 
1107 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1108 {
1109 	/* Native or PPF */
1110 	if (!mlx4_is_mfunc(dev->dev) ||
1111 	    (mlx4_is_master(dev->dev) &&
1112 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1113 		return  dev->dev->phys_caps.base_sqpn +
1114 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1115 			attr->port_num - 1;
1116 	}
1117 	/* PF or VF -- creating proxies */
1118 	if (attr->qp_type == IB_QPT_SMI)
1119 		return dev->dev->caps.qp0_proxy[attr->port_num - 1];
1120 	else
1121 		return dev->dev->caps.qp1_proxy[attr->port_num - 1];
1122 }
1123 
1124 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1125 					struct ib_qp_init_attr *init_attr,
1126 					struct ib_udata *udata)
1127 {
1128 	struct mlx4_ib_qp *qp = NULL;
1129 	int err;
1130 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1131 	u16 xrcdn = 0;
1132 	gfp_t gfp;
1133 
1134 	gfp = (init_attr->create_flags & MLX4_IB_QP_CREATE_USE_GFP_NOIO) ?
1135 		GFP_NOIO : GFP_KERNEL;
1136 	/*
1137 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1138 	 * and only for kernel UD QPs.
1139 	 */
1140 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1141 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1142 					MLX4_IB_SRIOV_TUNNEL_QP |
1143 					MLX4_IB_SRIOV_SQP |
1144 					MLX4_IB_QP_NETIF |
1145 					MLX4_IB_QP_CREATE_ROCE_V2_GSI |
1146 					MLX4_IB_QP_CREATE_USE_GFP_NOIO))
1147 		return ERR_PTR(-EINVAL);
1148 
1149 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1150 		if (init_attr->qp_type != IB_QPT_UD)
1151 			return ERR_PTR(-EINVAL);
1152 	}
1153 
1154 	if (init_attr->create_flags) {
1155 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1156 			return ERR_PTR(-EINVAL);
1157 
1158 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1159 						 MLX4_IB_QP_CREATE_USE_GFP_NOIO |
1160 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1161 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1162 		     init_attr->qp_type != IB_QPT_UD) ||
1163 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1164 		     init_attr->qp_type > IB_QPT_GSI) ||
1165 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1166 		     init_attr->qp_type != IB_QPT_GSI))
1167 			return ERR_PTR(-EINVAL);
1168 	}
1169 
1170 	switch (init_attr->qp_type) {
1171 	case IB_QPT_XRC_TGT:
1172 		pd = to_mxrcd(init_attr->xrcd)->pd;
1173 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1174 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1175 		/* fall through */
1176 	case IB_QPT_XRC_INI:
1177 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1178 			return ERR_PTR(-ENOSYS);
1179 		init_attr->recv_cq = init_attr->send_cq;
1180 		/* fall through */
1181 	case IB_QPT_RC:
1182 	case IB_QPT_UC:
1183 	case IB_QPT_RAW_PACKET:
1184 		qp = kzalloc(sizeof *qp, gfp);
1185 		if (!qp)
1186 			return ERR_PTR(-ENOMEM);
1187 		qp->pri.vid = 0xFFFF;
1188 		qp->alt.vid = 0xFFFF;
1189 		/* fall through */
1190 	case IB_QPT_UD:
1191 	{
1192 		err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1193 				       udata, 0, &qp, gfp);
1194 		if (err) {
1195 			kfree(qp);
1196 			return ERR_PTR(err);
1197 		}
1198 
1199 		qp->ibqp.qp_num = qp->mqp.qpn;
1200 		qp->xrcdn = xrcdn;
1201 
1202 		break;
1203 	}
1204 	case IB_QPT_SMI:
1205 	case IB_QPT_GSI:
1206 	{
1207 		int sqpn;
1208 
1209 		/* Userspace is not allowed to create special QPs: */
1210 		if (udata)
1211 			return ERR_PTR(-EINVAL);
1212 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1213 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev, 1, 1, &sqpn, 0);
1214 
1215 			if (res)
1216 				return ERR_PTR(res);
1217 		} else {
1218 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1219 		}
1220 
1221 		err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1222 				       sqpn,
1223 				       &qp, gfp);
1224 		if (err)
1225 			return ERR_PTR(err);
1226 
1227 		qp->port	= init_attr->port_num;
1228 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1229 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1230 		break;
1231 	}
1232 	default:
1233 		/* Don't support raw QPs */
1234 		return ERR_PTR(-EINVAL);
1235 	}
1236 
1237 	return &qp->ibqp;
1238 }
1239 
1240 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1241 				struct ib_qp_init_attr *init_attr,
1242 				struct ib_udata *udata) {
1243 	struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1244 	struct ib_qp *ibqp;
1245 	struct mlx4_ib_dev *dev = to_mdev(device);
1246 
1247 	ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1248 
1249 	if (!IS_ERR(ibqp) &&
1250 	    (init_attr->qp_type == IB_QPT_GSI) &&
1251 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1252 		struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1253 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1254 
1255 		if (is_eth &&
1256 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1257 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1258 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1259 
1260 			if (IS_ERR(sqp->roce_v2_gsi)) {
1261 				pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1262 				sqp->roce_v2_gsi = NULL;
1263 			} else {
1264 				sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1265 				sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1266 			}
1267 
1268 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1269 		}
1270 	}
1271 	return ibqp;
1272 }
1273 
1274 static int _mlx4_ib_destroy_qp(struct ib_qp *qp)
1275 {
1276 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1277 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1278 	struct mlx4_ib_pd *pd;
1279 
1280 	if (is_qp0(dev, mqp))
1281 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1282 
1283 	if (dev->qp1_proxy[mqp->port - 1] == mqp) {
1284 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1285 		dev->qp1_proxy[mqp->port - 1] = NULL;
1286 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1287 	}
1288 
1289 	if (mqp->counter_index)
1290 		mlx4_ib_free_qp_counter(dev, mqp);
1291 
1292 	pd = get_pd(mqp);
1293 	destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1294 
1295 	if (is_sqp(dev, mqp))
1296 		kfree(to_msqp(mqp));
1297 	else
1298 		kfree(mqp);
1299 
1300 	return 0;
1301 }
1302 
1303 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1304 {
1305 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1306 
1307 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1308 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1309 
1310 		if (sqp->roce_v2_gsi)
1311 			ib_destroy_qp(sqp->roce_v2_gsi);
1312 	}
1313 
1314 	return _mlx4_ib_destroy_qp(qp);
1315 }
1316 
1317 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1318 {
1319 	switch (type) {
1320 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1321 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1322 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1323 	case MLX4_IB_QPT_XRC_INI:
1324 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1325 	case MLX4_IB_QPT_SMI:
1326 	case MLX4_IB_QPT_GSI:
1327 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1328 
1329 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1330 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1331 						MLX4_QP_ST_MLX : -1);
1332 	case MLX4_IB_QPT_PROXY_SMI:
1333 	case MLX4_IB_QPT_TUN_SMI:
1334 	case MLX4_IB_QPT_PROXY_GSI:
1335 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1336 						MLX4_QP_ST_UD : -1);
1337 	default:			return -1;
1338 	}
1339 }
1340 
1341 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1342 				   int attr_mask)
1343 {
1344 	u8 dest_rd_atomic;
1345 	u32 access_flags;
1346 	u32 hw_access_flags = 0;
1347 
1348 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1349 		dest_rd_atomic = attr->max_dest_rd_atomic;
1350 	else
1351 		dest_rd_atomic = qp->resp_depth;
1352 
1353 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1354 		access_flags = attr->qp_access_flags;
1355 	else
1356 		access_flags = qp->atomic_rd_en;
1357 
1358 	if (!dest_rd_atomic)
1359 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1360 
1361 	if (access_flags & IB_ACCESS_REMOTE_READ)
1362 		hw_access_flags |= MLX4_QP_BIT_RRE;
1363 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1364 		hw_access_flags |= MLX4_QP_BIT_RAE;
1365 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1366 		hw_access_flags |= MLX4_QP_BIT_RWE;
1367 
1368 	return cpu_to_be32(hw_access_flags);
1369 }
1370 
1371 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1372 			    int attr_mask)
1373 {
1374 	if (attr_mask & IB_QP_PKEY_INDEX)
1375 		sqp->pkey_index = attr->pkey_index;
1376 	if (attr_mask & IB_QP_QKEY)
1377 		sqp->qkey = attr->qkey;
1378 	if (attr_mask & IB_QP_SQ_PSN)
1379 		sqp->send_psn = attr->sq_psn;
1380 }
1381 
1382 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1383 {
1384 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1385 }
1386 
1387 static int _mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
1388 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1389 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1390 {
1391 	int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
1392 		IB_LINK_LAYER_ETHERNET;
1393 	int vidx;
1394 	int smac_index;
1395 	int err;
1396 
1397 
1398 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
1399 	path->rlid	    = cpu_to_be16(ah->dlid);
1400 	if (ah->static_rate) {
1401 		path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
1402 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1403 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1404 			--path->static_rate;
1405 	} else
1406 		path->static_rate = 0;
1407 
1408 	if (ah->ah_flags & IB_AH_GRH) {
1409 		int real_sgid_index = mlx4_ib_gid_index_to_real_index(dev,
1410 								      port,
1411 								      ah->grh.sgid_index);
1412 
1413 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1414 			pr_err("sgid_index (%u) too large. max is %d\n",
1415 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1416 			return -1;
1417 		}
1418 
1419 		path->grh_mylmc |= 1 << 7;
1420 		path->mgid_index = real_sgid_index;
1421 		path->hop_limit  = ah->grh.hop_limit;
1422 		path->tclass_flowlabel =
1423 			cpu_to_be32((ah->grh.traffic_class << 20) |
1424 				    (ah->grh.flow_label));
1425 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1426 	}
1427 
1428 	if (is_eth) {
1429 		if (!(ah->ah_flags & IB_AH_GRH))
1430 			return -1;
1431 
1432 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1433 			((port - 1) << 6) | ((ah->sl & 7) << 3);
1434 
1435 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1436 		if (vlan_tag < 0x1000) {
1437 			if (smac_info->vid < 0x1000) {
1438 				/* both valid vlan ids */
1439 				if (smac_info->vid != vlan_tag) {
1440 					/* different VIDs.  unreg old and reg new */
1441 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1442 					if (err)
1443 						return err;
1444 					smac_info->candidate_vid = vlan_tag;
1445 					smac_info->candidate_vlan_index = vidx;
1446 					smac_info->candidate_vlan_port = port;
1447 					smac_info->update_vid = 1;
1448 					path->vlan_index = vidx;
1449 				} else {
1450 					path->vlan_index = smac_info->vlan_index;
1451 				}
1452 			} else {
1453 				/* no current vlan tag in qp */
1454 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1455 				if (err)
1456 					return err;
1457 				smac_info->candidate_vid = vlan_tag;
1458 				smac_info->candidate_vlan_index = vidx;
1459 				smac_info->candidate_vlan_port = port;
1460 				smac_info->update_vid = 1;
1461 				path->vlan_index = vidx;
1462 			}
1463 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1464 			path->fl = 1 << 6;
1465 		} else {
1466 			/* have current vlan tag. unregister it at modify-qp success */
1467 			if (smac_info->vid < 0x1000) {
1468 				smac_info->candidate_vid = 0xFFFF;
1469 				smac_info->update_vid = 1;
1470 			}
1471 		}
1472 
1473 		/* get smac_index for RoCE use.
1474 		 * If no smac was yet assigned, register one.
1475 		 * If one was already assigned, but the new mac differs,
1476 		 * unregister the old one and register the new one.
1477 		*/
1478 		if ((!smac_info->smac && !smac_info->smac_port) ||
1479 		    smac_info->smac != smac) {
1480 			/* register candidate now, unreg if needed, after success */
1481 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1482 			if (smac_index >= 0) {
1483 				smac_info->candidate_smac_index = smac_index;
1484 				smac_info->candidate_smac = smac;
1485 				smac_info->candidate_smac_port = port;
1486 			} else {
1487 				return -EINVAL;
1488 			}
1489 		} else {
1490 			smac_index = smac_info->smac_index;
1491 		}
1492 
1493 		memcpy(path->dmac, ah->dmac, 6);
1494 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1495 		/* put MAC table smac index for IBoE */
1496 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1497 	} else {
1498 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1499 			((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1500 	}
1501 
1502 	return 0;
1503 }
1504 
1505 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1506 			 enum ib_qp_attr_mask qp_attr_mask,
1507 			 struct mlx4_ib_qp *mqp,
1508 			 struct mlx4_qp_path *path, u8 port,
1509 			 u16 vlan_id, u8 *smac)
1510 {
1511 	return _mlx4_set_path(dev, &qp->ah_attr,
1512 			      mlx4_mac_to_u64(smac),
1513 			      vlan_id,
1514 			      path, &mqp->pri, port);
1515 }
1516 
1517 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1518 			     const struct ib_qp_attr *qp,
1519 			     enum ib_qp_attr_mask qp_attr_mask,
1520 			     struct mlx4_ib_qp *mqp,
1521 			     struct mlx4_qp_path *path, u8 port)
1522 {
1523 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1524 			      0,
1525 			      0xffff,
1526 			      path, &mqp->alt, port);
1527 }
1528 
1529 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1530 {
1531 	struct mlx4_ib_gid_entry *ge, *tmp;
1532 
1533 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1534 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1535 			ge->added = 1;
1536 			ge->port = qp->port;
1537 		}
1538 	}
1539 }
1540 
1541 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1542 				    struct mlx4_ib_qp *qp,
1543 				    struct mlx4_qp_context *context)
1544 {
1545 	u64 u64_mac;
1546 	int smac_index;
1547 
1548 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1549 
1550 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1551 	if (!qp->pri.smac && !qp->pri.smac_port) {
1552 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1553 		if (smac_index >= 0) {
1554 			qp->pri.candidate_smac_index = smac_index;
1555 			qp->pri.candidate_smac = u64_mac;
1556 			qp->pri.candidate_smac_port = qp->port;
1557 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1558 		} else {
1559 			return -ENOENT;
1560 		}
1561 	}
1562 	return 0;
1563 }
1564 
1565 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1566 {
1567 	struct counter_index *new_counter_index;
1568 	int err;
1569 	u32 tmp_idx;
1570 
1571 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1572 	    IB_LINK_LAYER_ETHERNET ||
1573 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1574 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1575 		return 0;
1576 
1577 	err = mlx4_counter_alloc(dev->dev, &tmp_idx);
1578 	if (err)
1579 		return err;
1580 
1581 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1582 	if (!new_counter_index) {
1583 		mlx4_counter_free(dev->dev, tmp_idx);
1584 		return -ENOMEM;
1585 	}
1586 
1587 	new_counter_index->index = tmp_idx;
1588 	new_counter_index->allocated = 1;
1589 	qp->counter_index = new_counter_index;
1590 
1591 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
1592 	list_add_tail(&new_counter_index->list,
1593 		      &dev->counters_table[qp->port - 1].counters_list);
1594 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
1595 
1596 	return 0;
1597 }
1598 
1599 enum {
1600 	MLX4_QPC_ROCE_MODE_1 = 0,
1601 	MLX4_QPC_ROCE_MODE_2 = 2,
1602 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
1603 };
1604 
1605 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
1606 {
1607 	switch (gid_type) {
1608 	case IB_GID_TYPE_ROCE:
1609 		return MLX4_QPC_ROCE_MODE_1;
1610 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
1611 		return MLX4_QPC_ROCE_MODE_2;
1612 	default:
1613 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
1614 	}
1615 }
1616 
1617 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1618 			       const struct ib_qp_attr *attr, int attr_mask,
1619 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1620 {
1621 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1622 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1623 	struct mlx4_ib_pd *pd;
1624 	struct mlx4_ib_cq *send_cq, *recv_cq;
1625 	struct mlx4_qp_context *context;
1626 	enum mlx4_qp_optpar optpar = 0;
1627 	int sqd_event;
1628 	int steer_qp = 0;
1629 	int err = -EINVAL;
1630 	int counter_index;
1631 
1632 	/* APM is not supported under RoCE */
1633 	if (attr_mask & IB_QP_ALT_PATH &&
1634 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1635 	    IB_LINK_LAYER_ETHERNET)
1636 		return -ENOTSUPP;
1637 
1638 	context = kzalloc(sizeof *context, GFP_KERNEL);
1639 	if (!context)
1640 		return -ENOMEM;
1641 
1642 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1643 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1644 
1645 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1646 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1647 	else {
1648 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
1649 		switch (attr->path_mig_state) {
1650 		case IB_MIG_MIGRATED:
1651 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1652 			break;
1653 		case IB_MIG_REARM:
1654 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1655 			break;
1656 		case IB_MIG_ARMED:
1657 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1658 			break;
1659 		}
1660 	}
1661 
1662 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1663 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1664 	else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1665 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1666 	else if (ibqp->qp_type == IB_QPT_UD) {
1667 		if (qp->flags & MLX4_IB_QP_LSO)
1668 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
1669 					      ilog2(dev->dev->caps.max_gso_sz);
1670 		else
1671 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1672 	} else if (attr_mask & IB_QP_PATH_MTU) {
1673 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1674 			pr_err("path MTU (%u) is invalid\n",
1675 			       attr->path_mtu);
1676 			goto out;
1677 		}
1678 		context->mtu_msgmax = (attr->path_mtu << 5) |
1679 			ilog2(dev->dev->caps.max_msg_sz);
1680 	}
1681 
1682 	if (qp->rq.wqe_cnt)
1683 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1684 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
1685 
1686 	if (qp->sq.wqe_cnt)
1687 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1688 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
1689 
1690 	if (new_state == IB_QPS_RESET && qp->counter_index)
1691 		mlx4_ib_free_qp_counter(dev, qp);
1692 
1693 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1694 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1695 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1696 		if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1697 			context->param3 |= cpu_to_be32(1 << 30);
1698 	}
1699 
1700 	if (qp->ibqp.uobject)
1701 		context->usr_page = cpu_to_be32(
1702 			mlx4_to_hw_uar_index(dev->dev,
1703 					     to_mucontext(ibqp->uobject->context)->uar.index));
1704 	else
1705 		context->usr_page = cpu_to_be32(
1706 			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
1707 
1708 	if (attr_mask & IB_QP_DEST_QPN)
1709 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1710 
1711 	if (attr_mask & IB_QP_PORT) {
1712 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1713 		    !(attr_mask & IB_QP_AV)) {
1714 			mlx4_set_sched(&context->pri_path, attr->port_num);
1715 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1716 		}
1717 	}
1718 
1719 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1720 		err = create_qp_lb_counter(dev, qp);
1721 		if (err)
1722 			goto out;
1723 
1724 		counter_index =
1725 			dev->counters_table[qp->port - 1].default_counter;
1726 		if (qp->counter_index)
1727 			counter_index = qp->counter_index->index;
1728 
1729 		if (counter_index != -1) {
1730 			context->pri_path.counter_index = counter_index;
1731 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1732 			if (qp->counter_index) {
1733 				context->pri_path.fl |=
1734 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
1735 				context->pri_path.vlan_control |=
1736 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
1737 			}
1738 		} else
1739 			context->pri_path.counter_index =
1740 				MLX4_SINK_COUNTER_INDEX(dev->dev);
1741 
1742 		if (qp->flags & MLX4_IB_QP_NETIF) {
1743 			mlx4_ib_steer_qp_reg(dev, qp, 1);
1744 			steer_qp = 1;
1745 		}
1746 
1747 		if (ibqp->qp_type == IB_QPT_GSI) {
1748 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
1749 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
1750 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
1751 
1752 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
1753 		}
1754 	}
1755 
1756 	if (attr_mask & IB_QP_PKEY_INDEX) {
1757 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1758 			context->pri_path.disable_pkey_check = 0x40;
1759 		context->pri_path.pkey_index = attr->pkey_index;
1760 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1761 	}
1762 
1763 	if (attr_mask & IB_QP_AV) {
1764 		u8 port_num = mlx4_is_bonded(to_mdev(ibqp->device)->dev) ? 1 :
1765 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1766 		union ib_gid gid;
1767 		struct ib_gid_attr gid_attr;
1768 		u16 vlan = 0xffff;
1769 		u8 smac[ETH_ALEN];
1770 		int status = 0;
1771 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
1772 			attr->ah_attr.ah_flags & IB_AH_GRH;
1773 
1774 		if (is_eth) {
1775 			int index = attr->ah_attr.grh.sgid_index;
1776 
1777 			status = ib_get_cached_gid(ibqp->device, port_num,
1778 						   index, &gid, &gid_attr);
1779 			if (!status && !memcmp(&gid, &zgid, sizeof(gid)))
1780 				status = -ENOENT;
1781 			if (!status && gid_attr.ndev) {
1782 				vlan = rdma_vlan_dev_vlan_id(gid_attr.ndev);
1783 				memcpy(smac, gid_attr.ndev->dev_addr, ETH_ALEN);
1784 				dev_put(gid_attr.ndev);
1785 			}
1786 		}
1787 		if (status)
1788 			goto out;
1789 
1790 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
1791 				  port_num, vlan, smac))
1792 			goto out;
1793 
1794 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1795 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
1796 
1797 		if (is_eth &&
1798 		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
1799 			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr.gid_type);
1800 
1801 			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
1802 				err = -EINVAL;
1803 				goto out;
1804 			}
1805 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
1806 		}
1807 
1808 	}
1809 
1810 	if (attr_mask & IB_QP_TIMEOUT) {
1811 		context->pri_path.ackto |= attr->timeout << 3;
1812 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1813 	}
1814 
1815 	if (attr_mask & IB_QP_ALT_PATH) {
1816 		if (attr->alt_port_num == 0 ||
1817 		    attr->alt_port_num > dev->dev->caps.num_ports)
1818 			goto out;
1819 
1820 		if (attr->alt_pkey_index >=
1821 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
1822 			goto out;
1823 
1824 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
1825 				      &context->alt_path,
1826 				      attr->alt_port_num))
1827 			goto out;
1828 
1829 		context->alt_path.pkey_index = attr->alt_pkey_index;
1830 		context->alt_path.ackto = attr->alt_timeout << 3;
1831 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1832 	}
1833 
1834 	pd = get_pd(qp);
1835 	get_cqs(qp, &send_cq, &recv_cq);
1836 	context->pd       = cpu_to_be32(pd->pdn);
1837 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1838 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1839 	context->params1  = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1840 
1841 	/* Set "fast registration enabled" for all kernel QPs */
1842 	if (!qp->ibqp.uobject)
1843 		context->params1 |= cpu_to_be32(1 << 11);
1844 
1845 	if (attr_mask & IB_QP_RNR_RETRY) {
1846 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1847 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1848 	}
1849 
1850 	if (attr_mask & IB_QP_RETRY_CNT) {
1851 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1852 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1853 	}
1854 
1855 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1856 		if (attr->max_rd_atomic)
1857 			context->params1 |=
1858 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1859 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1860 	}
1861 
1862 	if (attr_mask & IB_QP_SQ_PSN)
1863 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1864 
1865 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1866 		if (attr->max_dest_rd_atomic)
1867 			context->params2 |=
1868 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1869 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1870 	}
1871 
1872 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1873 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1874 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1875 	}
1876 
1877 	if (ibqp->srq)
1878 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1879 
1880 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1881 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1882 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1883 	}
1884 	if (attr_mask & IB_QP_RQ_PSN)
1885 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1886 
1887 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1888 	if (attr_mask & IB_QP_QKEY) {
1889 		if (qp->mlx4_ib_qp_type &
1890 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1891 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1892 		else {
1893 			if (mlx4_is_mfunc(dev->dev) &&
1894 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1895 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1896 			    MLX4_RESERVED_QKEY_BASE) {
1897 				pr_err("Cannot use reserved QKEY"
1898 				       " 0x%x (range 0xffff0000..0xffffffff"
1899 				       " is reserved)\n", attr->qkey);
1900 				err = -EINVAL;
1901 				goto out;
1902 			}
1903 			context->qkey = cpu_to_be32(attr->qkey);
1904 		}
1905 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
1906 	}
1907 
1908 	if (ibqp->srq)
1909 		context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1910 
1911 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1912 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1913 
1914 	if (cur_state == IB_QPS_INIT &&
1915 	    new_state == IB_QPS_RTR  &&
1916 	    (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1917 	     ibqp->qp_type == IB_QPT_UD ||
1918 	     ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1919 		context->pri_path.sched_queue = (qp->port - 1) << 6;
1920 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1921 		    qp->mlx4_ib_qp_type &
1922 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1923 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1924 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1925 				context->pri_path.fl = 0x80;
1926 		} else {
1927 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1928 				context->pri_path.fl = 0x80;
1929 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1930 		}
1931 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1932 		    IB_LINK_LAYER_ETHERNET) {
1933 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
1934 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
1935 				context->pri_path.feup = 1 << 7; /* don't fsm */
1936 			/* handle smac_index */
1937 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
1938 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
1939 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
1940 				err = handle_eth_ud_smac_index(dev, qp, context);
1941 				if (err) {
1942 					err = -EINVAL;
1943 					goto out;
1944 				}
1945 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1946 					dev->qp1_proxy[qp->port - 1] = qp;
1947 			}
1948 		}
1949 	}
1950 
1951 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1952 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
1953 					MLX4_IB_LINK_TYPE_ETH;
1954 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1955 			/* set QP to receive both tunneled & non-tunneled packets */
1956 			if (!(context->flags & cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET)))
1957 				context->srqn = cpu_to_be32(7 << 28);
1958 		}
1959 	}
1960 
1961 	if (ibqp->qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
1962 		int is_eth = rdma_port_get_link_layer(
1963 				&dev->ib_dev, qp->port) ==
1964 				IB_LINK_LAYER_ETHERNET;
1965 		if (is_eth) {
1966 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
1967 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
1968 		}
1969 	}
1970 
1971 
1972 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1973 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1974 		sqd_event = 1;
1975 	else
1976 		sqd_event = 0;
1977 
1978 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1979 		context->rlkey_roce_mode |= (1 << 4);
1980 
1981 	/*
1982 	 * Before passing a kernel QP to the HW, make sure that the
1983 	 * ownership bits of the send queue are set and the SQ
1984 	 * headroom is stamped so that the hardware doesn't start
1985 	 * processing stale work requests.
1986 	 */
1987 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1988 		struct mlx4_wqe_ctrl_seg *ctrl;
1989 		int i;
1990 
1991 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1992 			ctrl = get_send_wqe(qp, i);
1993 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
1994 			if (qp->sq_max_wqes_per_wr == 1)
1995 				ctrl->qpn_vlan.fence_size =
1996 						1 << (qp->sq.wqe_shift - 4);
1997 
1998 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1999 		}
2000 	}
2001 
2002 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2003 			     to_mlx4_state(new_state), context, optpar,
2004 			     sqd_event, &qp->mqp);
2005 	if (err)
2006 		goto out;
2007 
2008 	qp->state = new_state;
2009 
2010 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2011 		qp->atomic_rd_en = attr->qp_access_flags;
2012 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2013 		qp->resp_depth = attr->max_dest_rd_atomic;
2014 	if (attr_mask & IB_QP_PORT) {
2015 		qp->port = attr->port_num;
2016 		update_mcg_macs(dev, qp);
2017 	}
2018 	if (attr_mask & IB_QP_ALT_PATH)
2019 		qp->alt_port = attr->alt_port_num;
2020 
2021 	if (is_sqp(dev, qp))
2022 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2023 
2024 	/*
2025 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
2026 	 * QP0 to RESET or ERROR, bring the link back down.
2027 	 */
2028 	if (is_qp0(dev, qp)) {
2029 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2030 			if (mlx4_INIT_PORT(dev->dev, qp->port))
2031 				pr_warn("INIT_PORT failed for port %d\n",
2032 				       qp->port);
2033 
2034 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2035 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2036 			mlx4_CLOSE_PORT(dev->dev, qp->port);
2037 	}
2038 
2039 	/*
2040 	 * If we moved a kernel QP to RESET, clean up all old CQ
2041 	 * entries and reinitialize the QP.
2042 	 */
2043 	if (new_state == IB_QPS_RESET) {
2044 		if (!ibqp->uobject) {
2045 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2046 					 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
2047 			if (send_cq != recv_cq)
2048 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2049 
2050 			qp->rq.head = 0;
2051 			qp->rq.tail = 0;
2052 			qp->sq.head = 0;
2053 			qp->sq.tail = 0;
2054 			qp->sq_next_wqe = 0;
2055 			if (qp->rq.wqe_cnt)
2056 				*qp->db.db  = 0;
2057 
2058 			if (qp->flags & MLX4_IB_QP_NETIF)
2059 				mlx4_ib_steer_qp_reg(dev, qp, 0);
2060 		}
2061 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2062 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2063 			qp->pri.smac = 0;
2064 			qp->pri.smac_port = 0;
2065 		}
2066 		if (qp->alt.smac) {
2067 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2068 			qp->alt.smac = 0;
2069 		}
2070 		if (qp->pri.vid < 0x1000) {
2071 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2072 			qp->pri.vid = 0xFFFF;
2073 			qp->pri.candidate_vid = 0xFFFF;
2074 			qp->pri.update_vid = 0;
2075 		}
2076 
2077 		if (qp->alt.vid < 0x1000) {
2078 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2079 			qp->alt.vid = 0xFFFF;
2080 			qp->alt.candidate_vid = 0xFFFF;
2081 			qp->alt.update_vid = 0;
2082 		}
2083 	}
2084 out:
2085 	if (err && qp->counter_index)
2086 		mlx4_ib_free_qp_counter(dev, qp);
2087 	if (err && steer_qp)
2088 		mlx4_ib_steer_qp_reg(dev, qp, 0);
2089 	kfree(context);
2090 	if (qp->pri.candidate_smac ||
2091 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2092 		if (err) {
2093 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2094 		} else {
2095 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2096 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2097 			qp->pri.smac = qp->pri.candidate_smac;
2098 			qp->pri.smac_index = qp->pri.candidate_smac_index;
2099 			qp->pri.smac_port = qp->pri.candidate_smac_port;
2100 		}
2101 		qp->pri.candidate_smac = 0;
2102 		qp->pri.candidate_smac_index = 0;
2103 		qp->pri.candidate_smac_port = 0;
2104 	}
2105 	if (qp->alt.candidate_smac) {
2106 		if (err) {
2107 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2108 		} else {
2109 			if (qp->alt.smac)
2110 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2111 			qp->alt.smac = qp->alt.candidate_smac;
2112 			qp->alt.smac_index = qp->alt.candidate_smac_index;
2113 			qp->alt.smac_port = qp->alt.candidate_smac_port;
2114 		}
2115 		qp->alt.candidate_smac = 0;
2116 		qp->alt.candidate_smac_index = 0;
2117 		qp->alt.candidate_smac_port = 0;
2118 	}
2119 
2120 	if (qp->pri.update_vid) {
2121 		if (err) {
2122 			if (qp->pri.candidate_vid < 0x1000)
2123 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2124 						     qp->pri.candidate_vid);
2125 		} else {
2126 			if (qp->pri.vid < 0x1000)
2127 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2128 						     qp->pri.vid);
2129 			qp->pri.vid = qp->pri.candidate_vid;
2130 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2131 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
2132 		}
2133 		qp->pri.candidate_vid = 0xFFFF;
2134 		qp->pri.update_vid = 0;
2135 	}
2136 
2137 	if (qp->alt.update_vid) {
2138 		if (err) {
2139 			if (qp->alt.candidate_vid < 0x1000)
2140 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2141 						     qp->alt.candidate_vid);
2142 		} else {
2143 			if (qp->alt.vid < 0x1000)
2144 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2145 						     qp->alt.vid);
2146 			qp->alt.vid = qp->alt.candidate_vid;
2147 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2148 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2149 		}
2150 		qp->alt.candidate_vid = 0xFFFF;
2151 		qp->alt.update_vid = 0;
2152 	}
2153 
2154 	return err;
2155 }
2156 
2157 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2158 			      int attr_mask, struct ib_udata *udata)
2159 {
2160 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2161 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2162 	enum ib_qp_state cur_state, new_state;
2163 	int err = -EINVAL;
2164 	int ll;
2165 	mutex_lock(&qp->mutex);
2166 
2167 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2168 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2169 
2170 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2171 		ll = IB_LINK_LAYER_UNSPECIFIED;
2172 	} else {
2173 		int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2174 		ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2175 	}
2176 
2177 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2178 				attr_mask, ll)) {
2179 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2180 			 "for transition %d to %d. qp_type %d,"
2181 			 " attr_mask 0x%x\n",
2182 			 ibqp->qp_num, cur_state, new_state,
2183 			 ibqp->qp_type, attr_mask);
2184 		goto out;
2185 	}
2186 
2187 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2188 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2189 			if ((ibqp->qp_type == IB_QPT_RC) ||
2190 			    (ibqp->qp_type == IB_QPT_UD) ||
2191 			    (ibqp->qp_type == IB_QPT_UC) ||
2192 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2193 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2194 				attr->port_num = mlx4_ib_bond_next_port(dev);
2195 			}
2196 		} else {
2197 			/* no sense in changing port_num
2198 			 * when ports are bonded */
2199 			attr_mask &= ~IB_QP_PORT;
2200 		}
2201 	}
2202 
2203 	if ((attr_mask & IB_QP_PORT) &&
2204 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2205 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2206 			 "for transition %d to %d. qp_type %d\n",
2207 			 ibqp->qp_num, attr->port_num, cur_state,
2208 			 new_state, ibqp->qp_type);
2209 		goto out;
2210 	}
2211 
2212 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2213 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2214 	     IB_LINK_LAYER_ETHERNET))
2215 		goto out;
2216 
2217 	if (attr_mask & IB_QP_PKEY_INDEX) {
2218 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2219 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2220 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2221 				 "for transition %d to %d. qp_type %d\n",
2222 				 ibqp->qp_num, attr->pkey_index, cur_state,
2223 				 new_state, ibqp->qp_type);
2224 			goto out;
2225 		}
2226 	}
2227 
2228 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2229 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2230 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2231 			 "Transition %d to %d. qp_type %d\n",
2232 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2233 			 new_state, ibqp->qp_type);
2234 		goto out;
2235 	}
2236 
2237 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2238 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2239 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2240 			 "Transition %d to %d. qp_type %d\n",
2241 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2242 			 new_state, ibqp->qp_type);
2243 		goto out;
2244 	}
2245 
2246 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2247 		err = 0;
2248 		goto out;
2249 	}
2250 
2251 	err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2252 
2253 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2254 		attr->port_num = 1;
2255 
2256 out:
2257 	mutex_unlock(&qp->mutex);
2258 	return err;
2259 }
2260 
2261 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2262 		      int attr_mask, struct ib_udata *udata)
2263 {
2264 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2265 	int ret;
2266 
2267 	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2268 
2269 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2270 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
2271 		int err = 0;
2272 
2273 		if (sqp->roce_v2_gsi)
2274 			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2275 		if (err)
2276 			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2277 			       err);
2278 	}
2279 	return ret;
2280 }
2281 
2282 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2283 {
2284 	int i;
2285 	for (i = 0; i < dev->caps.num_ports; i++) {
2286 		if (qpn == dev->caps.qp0_proxy[i] ||
2287 		    qpn == dev->caps.qp0_tunnel[i]) {
2288 			*qkey = dev->caps.qp0_qkey[i];
2289 			return 0;
2290 		}
2291 	}
2292 	return -EINVAL;
2293 }
2294 
2295 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2296 				  struct ib_ud_wr *wr,
2297 				  void *wqe, unsigned *mlx_seg_len)
2298 {
2299 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2300 	struct ib_device *ib_dev = &mdev->ib_dev;
2301 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2302 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2303 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2304 	u16 pkey;
2305 	u32 qkey;
2306 	int send_size;
2307 	int header_size;
2308 	int spc;
2309 	int i;
2310 
2311 	if (wr->wr.opcode != IB_WR_SEND)
2312 		return -EINVAL;
2313 
2314 	send_size = 0;
2315 
2316 	for (i = 0; i < wr->wr.num_sge; ++i)
2317 		send_size += wr->wr.sg_list[i].length;
2318 
2319 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2320 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2321 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2322 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2323 
2324 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2325 
2326 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2327 		sqp->ud_header.lrh.service_level =
2328 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2329 		sqp->ud_header.lrh.destination_lid =
2330 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2331 		sqp->ud_header.lrh.source_lid =
2332 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2333 	}
2334 
2335 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2336 
2337 	/* force loopback */
2338 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2339 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2340 
2341 	sqp->ud_header.lrh.virtual_lane    = 0;
2342 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2343 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2344 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2345 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2346 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2347 	else
2348 		sqp->ud_header.bth.destination_qpn =
2349 			cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
2350 
2351 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2352 	if (mlx4_is_master(mdev->dev)) {
2353 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2354 			return -EINVAL;
2355 	} else {
2356 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2357 			return -EINVAL;
2358 	}
2359 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2360 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2361 
2362 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2363 	sqp->ud_header.immediate_present = 0;
2364 
2365 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2366 
2367 	/*
2368 	 * Inline data segments may not cross a 64 byte boundary.  If
2369 	 * our UD header is bigger than the space available up to the
2370 	 * next 64 byte boundary in the WQE, use two inline data
2371 	 * segments to hold the UD header.
2372 	 */
2373 	spc = MLX4_INLINE_ALIGN -
2374 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2375 	if (header_size <= spc) {
2376 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2377 		memcpy(inl + 1, sqp->header_buf, header_size);
2378 		i = 1;
2379 	} else {
2380 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2381 		memcpy(inl + 1, sqp->header_buf, spc);
2382 
2383 		inl = (void *) (inl + 1) + spc;
2384 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2385 		/*
2386 		 * Need a barrier here to make sure all the data is
2387 		 * visible before the byte_count field is set.
2388 		 * Otherwise the HCA prefetcher could grab the 64-byte
2389 		 * chunk with this inline segment and get a valid (!=
2390 		 * 0xffffffff) byte count but stale data, and end up
2391 		 * generating a packet with bad headers.
2392 		 *
2393 		 * The first inline segment's byte_count field doesn't
2394 		 * need a barrier, because it comes after a
2395 		 * control/MLX segment and therefore is at an offset
2396 		 * of 16 mod 64.
2397 		 */
2398 		wmb();
2399 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2400 		i = 2;
2401 	}
2402 
2403 	*mlx_seg_len =
2404 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2405 	return 0;
2406 }
2407 
2408 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
2409 {
2410 	union sl2vl_tbl_to_u64 tmp_vltab;
2411 	u8 vl;
2412 
2413 	if (sl > 15)
2414 		return 0xf;
2415 	tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
2416 	vl = tmp_vltab.sl8[sl >> 1];
2417 	if (sl & 1)
2418 		vl &= 0x0f;
2419 	else
2420 		vl >>= 4;
2421 	return vl;
2422 }
2423 
2424 #define MLX4_ROCEV2_QP1_SPORT 0xC000
2425 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_ud_wr *wr,
2426 			    void *wqe, unsigned *mlx_seg_len)
2427 {
2428 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
2429 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2430 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2431 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2432 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2433 	union ib_gid sgid;
2434 	u16 pkey;
2435 	int send_size;
2436 	int header_size;
2437 	int spc;
2438 	int i;
2439 	int err = 0;
2440 	u16 vlan = 0xffff;
2441 	bool is_eth;
2442 	bool is_vlan = false;
2443 	bool is_grh;
2444 	bool is_udp = false;
2445 	int ip_version = 0;
2446 
2447 	send_size = 0;
2448 	for (i = 0; i < wr->wr.num_sge; ++i)
2449 		send_size += wr->wr.sg_list[i].length;
2450 
2451 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2452 	is_grh = mlx4_ib_ah_grh_present(ah);
2453 	if (is_eth) {
2454 		struct ib_gid_attr gid_attr;
2455 
2456 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2457 			/* When multi-function is enabled, the ib_core gid
2458 			 * indexes don't necessarily match the hw ones, so
2459 			 * we must use our own cache */
2460 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2461 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
2462 							   ah->av.ib.gid_index, &sgid.raw[0]);
2463 			if (err)
2464 				return err;
2465 		} else  {
2466 			err = ib_get_cached_gid(ib_dev,
2467 						be32_to_cpu(ah->av.ib.port_pd) >> 24,
2468 						ah->av.ib.gid_index, &sgid,
2469 						&gid_attr);
2470 			if (!err) {
2471 				if (gid_attr.ndev)
2472 					dev_put(gid_attr.ndev);
2473 				if (!memcmp(&sgid, &zgid, sizeof(sgid)))
2474 					err = -ENOENT;
2475 			}
2476 			if (!err) {
2477 				is_udp = gid_attr.gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
2478 				if (is_udp) {
2479 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
2480 						ip_version = 4;
2481 					else
2482 						ip_version = 6;
2483 					is_grh = false;
2484 				}
2485 			} else {
2486 				return err;
2487 			}
2488 		}
2489 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
2490 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
2491 			is_vlan = 1;
2492 		}
2493 	}
2494 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
2495 			  ip_version, is_udp, 0, &sqp->ud_header);
2496 	if (err)
2497 		return err;
2498 
2499 	if (!is_eth) {
2500 		sqp->ud_header.lrh.service_level =
2501 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2502 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
2503 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2504 	}
2505 
2506 	if (is_grh || (ip_version == 6)) {
2507 		sqp->ud_header.grh.traffic_class =
2508 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2509 		sqp->ud_header.grh.flow_label    =
2510 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
2511 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
2512 		if (is_eth) {
2513 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
2514 		} else {
2515 			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2516 				/* When multi-function is enabled, the ib_core gid
2517 				 * indexes don't necessarily match the hw ones, so
2518 				 * we must use our own cache
2519 				 */
2520 				sqp->ud_header.grh.source_gid.global.subnet_prefix =
2521 					cpu_to_be64(atomic64_read(&(to_mdev(ib_dev)->sriov.
2522 								    demux[sqp->qp.port - 1].
2523 								    subnet_prefix)));
2524 				sqp->ud_header.grh.source_gid.global.interface_id =
2525 					to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2526 						       guid_cache[ah->av.ib.gid_index];
2527 			} else {
2528 				ib_get_cached_gid(ib_dev,
2529 						  be32_to_cpu(ah->av.ib.port_pd) >> 24,
2530 						  ah->av.ib.gid_index,
2531 						  &sqp->ud_header.grh.source_gid, NULL);
2532 			}
2533 		}
2534 		memcpy(sqp->ud_header.grh.destination_gid.raw,
2535 		       ah->av.ib.dgid, 16);
2536 	}
2537 
2538 	if (ip_version == 4) {
2539 		sqp->ud_header.ip4.tos =
2540 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2541 		sqp->ud_header.ip4.id = 0;
2542 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
2543 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
2544 
2545 		memcpy(&sqp->ud_header.ip4.saddr,
2546 		       sgid.raw + 12, 4);
2547 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
2548 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
2549 	}
2550 
2551 	if (is_udp) {
2552 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
2553 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
2554 		sqp->ud_header.udp.csum = 0;
2555 	}
2556 
2557 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2558 
2559 	if (!is_eth) {
2560 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
2561 					  (sqp->ud_header.lrh.destination_lid ==
2562 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
2563 					  (sqp->ud_header.lrh.service_level << 8));
2564 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
2565 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
2566 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
2567 	}
2568 
2569 	switch (wr->wr.opcode) {
2570 	case IB_WR_SEND:
2571 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
2572 		sqp->ud_header.immediate_present = 0;
2573 		break;
2574 	case IB_WR_SEND_WITH_IMM:
2575 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2576 		sqp->ud_header.immediate_present = 1;
2577 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
2578 		break;
2579 	default:
2580 		return -EINVAL;
2581 	}
2582 
2583 	if (is_eth) {
2584 		struct in6_addr in6;
2585 		u16 ether_type;
2586 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
2587 
2588 		ether_type = (!is_udp) ? MLX4_IB_IBOE_ETHERTYPE :
2589 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
2590 
2591 		mlx->sched_prio = cpu_to_be16(pcp);
2592 
2593 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
2594 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
2595 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
2596 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
2597 		memcpy(&in6, sgid.raw, sizeof(in6));
2598 
2599 
2600 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
2601 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
2602 		if (!is_vlan) {
2603 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
2604 		} else {
2605 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
2606 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
2607 		}
2608 	} else {
2609 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 :
2610 							sl_to_vl(to_mdev(ib_dev),
2611 								 sqp->ud_header.lrh.service_level,
2612 								 sqp->qp.port);
2613 		if (sqp->qp.ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
2614 			return -EINVAL;
2615 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
2616 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
2617 	}
2618 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2619 	if (!sqp->qp.ibqp.qp_num)
2620 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
2621 	else
2622 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
2623 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2624 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2625 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2626 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
2627 					       sqp->qkey : wr->remote_qkey);
2628 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
2629 
2630 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2631 
2632 	if (0) {
2633 		pr_err("built UD header of size %d:\n", header_size);
2634 		for (i = 0; i < header_size / 4; ++i) {
2635 			if (i % 8 == 0)
2636 				pr_err("  [%02x] ", i * 4);
2637 			pr_cont(" %08x",
2638 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
2639 			if ((i + 1) % 8 == 0)
2640 				pr_cont("\n");
2641 		}
2642 		pr_err("\n");
2643 	}
2644 
2645 	/*
2646 	 * Inline data segments may not cross a 64 byte boundary.  If
2647 	 * our UD header is bigger than the space available up to the
2648 	 * next 64 byte boundary in the WQE, use two inline data
2649 	 * segments to hold the UD header.
2650 	 */
2651 	spc = MLX4_INLINE_ALIGN -
2652 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2653 	if (header_size <= spc) {
2654 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2655 		memcpy(inl + 1, sqp->header_buf, header_size);
2656 		i = 1;
2657 	} else {
2658 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2659 		memcpy(inl + 1, sqp->header_buf, spc);
2660 
2661 		inl = (void *) (inl + 1) + spc;
2662 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2663 		/*
2664 		 * Need a barrier here to make sure all the data is
2665 		 * visible before the byte_count field is set.
2666 		 * Otherwise the HCA prefetcher could grab the 64-byte
2667 		 * chunk with this inline segment and get a valid (!=
2668 		 * 0xffffffff) byte count but stale data, and end up
2669 		 * generating a packet with bad headers.
2670 		 *
2671 		 * The first inline segment's byte_count field doesn't
2672 		 * need a barrier, because it comes after a
2673 		 * control/MLX segment and therefore is at an offset
2674 		 * of 16 mod 64.
2675 		 */
2676 		wmb();
2677 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2678 		i = 2;
2679 	}
2680 
2681 	*mlx_seg_len =
2682 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2683 	return 0;
2684 }
2685 
2686 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2687 {
2688 	unsigned cur;
2689 	struct mlx4_ib_cq *cq;
2690 
2691 	cur = wq->head - wq->tail;
2692 	if (likely(cur + nreq < wq->max_post))
2693 		return 0;
2694 
2695 	cq = to_mcq(ib_cq);
2696 	spin_lock(&cq->lock);
2697 	cur = wq->head - wq->tail;
2698 	spin_unlock(&cq->lock);
2699 
2700 	return cur + nreq >= wq->max_post;
2701 }
2702 
2703 static __be32 convert_access(int acc)
2704 {
2705 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
2706 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
2707 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
2708 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
2709 	       (acc & IB_ACCESS_REMOTE_READ   ?
2710 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
2711 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
2712 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
2713 }
2714 
2715 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
2716 			struct ib_reg_wr *wr)
2717 {
2718 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
2719 
2720 	fseg->flags		= convert_access(wr->access);
2721 	fseg->mem_key		= cpu_to_be32(wr->key);
2722 	fseg->buf_list		= cpu_to_be64(mr->page_map);
2723 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
2724 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
2725 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
2726 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
2727 	fseg->reserved[0]	= 0;
2728 	fseg->reserved[1]	= 0;
2729 }
2730 
2731 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2732 {
2733 	memset(iseg, 0, sizeof(*iseg));
2734 	iseg->mem_key = cpu_to_be32(rkey);
2735 }
2736 
2737 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2738 					  u64 remote_addr, u32 rkey)
2739 {
2740 	rseg->raddr    = cpu_to_be64(remote_addr);
2741 	rseg->rkey     = cpu_to_be32(rkey);
2742 	rseg->reserved = 0;
2743 }
2744 
2745 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
2746 		struct ib_atomic_wr *wr)
2747 {
2748 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2749 		aseg->swap_add = cpu_to_be64(wr->swap);
2750 		aseg->compare  = cpu_to_be64(wr->compare_add);
2751 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2752 		aseg->swap_add = cpu_to_be64(wr->compare_add);
2753 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
2754 	} else {
2755 		aseg->swap_add = cpu_to_be64(wr->compare_add);
2756 		aseg->compare  = 0;
2757 	}
2758 
2759 }
2760 
2761 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2762 				  struct ib_atomic_wr *wr)
2763 {
2764 	aseg->swap_add		= cpu_to_be64(wr->swap);
2765 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
2766 	aseg->compare		= cpu_to_be64(wr->compare_add);
2767 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
2768 }
2769 
2770 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2771 			     struct ib_ud_wr *wr)
2772 {
2773 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
2774 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
2775 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
2776 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
2777 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
2778 }
2779 
2780 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2781 				    struct mlx4_wqe_datagram_seg *dseg,
2782 				    struct ib_ud_wr *wr,
2783 				    enum mlx4_ib_qp_type qpt)
2784 {
2785 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
2786 	struct mlx4_av sqp_av = {0};
2787 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
2788 
2789 	/* force loopback */
2790 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2791 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2792 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2793 			cpu_to_be32(0xf0000000);
2794 
2795 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2796 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
2797 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2798 	else
2799 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp0_tunnel[port - 1]);
2800 	/* Use QKEY from the QP context, which is set by master */
2801 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2802 }
2803 
2804 static void build_tunnel_header(struct ib_ud_wr *wr, void *wqe, unsigned *mlx_seg_len)
2805 {
2806 	struct mlx4_wqe_inline_seg *inl = wqe;
2807 	struct mlx4_ib_tunnel_header hdr;
2808 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2809 	int spc;
2810 	int i;
2811 
2812 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2813 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
2814 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
2815 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
2816 	memcpy(hdr.mac, ah->av.eth.mac, 6);
2817 	hdr.vlan = ah->av.eth.vlan;
2818 
2819 	spc = MLX4_INLINE_ALIGN -
2820 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2821 	if (sizeof (hdr) <= spc) {
2822 		memcpy(inl + 1, &hdr, sizeof (hdr));
2823 		wmb();
2824 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
2825 		i = 1;
2826 	} else {
2827 		memcpy(inl + 1, &hdr, spc);
2828 		wmb();
2829 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2830 
2831 		inl = (void *) (inl + 1) + spc;
2832 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2833 		wmb();
2834 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
2835 		i = 2;
2836 	}
2837 
2838 	*mlx_seg_len =
2839 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2840 }
2841 
2842 static void set_mlx_icrc_seg(void *dseg)
2843 {
2844 	u32 *t = dseg;
2845 	struct mlx4_wqe_inline_seg *iseg = dseg;
2846 
2847 	t[1] = 0;
2848 
2849 	/*
2850 	 * Need a barrier here before writing the byte_count field to
2851 	 * make sure that all the data is visible before the
2852 	 * byte_count field is set.  Otherwise, if the segment begins
2853 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2854 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2855 	 * stale data, and end up sending the wrong data.
2856 	 */
2857 	wmb();
2858 
2859 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2860 }
2861 
2862 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2863 {
2864 	dseg->lkey       = cpu_to_be32(sg->lkey);
2865 	dseg->addr       = cpu_to_be64(sg->addr);
2866 
2867 	/*
2868 	 * Need a barrier here before writing the byte_count field to
2869 	 * make sure that all the data is visible before the
2870 	 * byte_count field is set.  Otherwise, if the segment begins
2871 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2872 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2873 	 * stale data, and end up sending the wrong data.
2874 	 */
2875 	wmb();
2876 
2877 	dseg->byte_count = cpu_to_be32(sg->length);
2878 }
2879 
2880 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2881 {
2882 	dseg->byte_count = cpu_to_be32(sg->length);
2883 	dseg->lkey       = cpu_to_be32(sg->lkey);
2884 	dseg->addr       = cpu_to_be64(sg->addr);
2885 }
2886 
2887 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_ud_wr *wr,
2888 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2889 			 __be32 *lso_hdr_sz, __be32 *blh)
2890 {
2891 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
2892 
2893 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2894 		*blh = cpu_to_be32(1 << 6);
2895 
2896 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2897 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
2898 		return -EINVAL;
2899 
2900 	memcpy(wqe->header, wr->header, wr->hlen);
2901 
2902 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
2903 	*lso_seg_len = halign;
2904 	return 0;
2905 }
2906 
2907 static __be32 send_ieth(struct ib_send_wr *wr)
2908 {
2909 	switch (wr->opcode) {
2910 	case IB_WR_SEND_WITH_IMM:
2911 	case IB_WR_RDMA_WRITE_WITH_IMM:
2912 		return wr->ex.imm_data;
2913 
2914 	case IB_WR_SEND_WITH_INV:
2915 		return cpu_to_be32(wr->ex.invalidate_rkey);
2916 
2917 	default:
2918 		return 0;
2919 	}
2920 }
2921 
2922 static void add_zero_len_inline(void *wqe)
2923 {
2924 	struct mlx4_wqe_inline_seg *inl = wqe;
2925 	memset(wqe, 0, 16);
2926 	inl->byte_count = cpu_to_be32(1 << 31);
2927 }
2928 
2929 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2930 		      struct ib_send_wr **bad_wr)
2931 {
2932 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2933 	void *wqe;
2934 	struct mlx4_wqe_ctrl_seg *ctrl;
2935 	struct mlx4_wqe_data_seg *dseg;
2936 	unsigned long flags;
2937 	int nreq;
2938 	int err = 0;
2939 	unsigned ind;
2940 	int uninitialized_var(stamp);
2941 	int uninitialized_var(size);
2942 	unsigned uninitialized_var(seglen);
2943 	__be32 dummy;
2944 	__be32 *lso_wqe;
2945 	__be32 uninitialized_var(lso_hdr_sz);
2946 	__be32 blh;
2947 	int i;
2948 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
2949 
2950 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2951 		struct mlx4_ib_sqp *sqp = to_msqp(qp);
2952 
2953 		if (sqp->roce_v2_gsi) {
2954 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
2955 			struct ib_gid_attr gid_attr;
2956 			union ib_gid gid;
2957 
2958 			if (!ib_get_cached_gid(ibqp->device,
2959 					       be32_to_cpu(ah->av.ib.port_pd) >> 24,
2960 					       ah->av.ib.gid_index, &gid,
2961 					       &gid_attr)) {
2962 				if (gid_attr.ndev)
2963 					dev_put(gid_attr.ndev);
2964 				qp = (gid_attr.gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
2965 					to_mqp(sqp->roce_v2_gsi) : qp;
2966 			} else {
2967 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
2968 				       ah->av.ib.gid_index);
2969 			}
2970 		}
2971 	}
2972 
2973 	spin_lock_irqsave(&qp->sq.lock, flags);
2974 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
2975 		err = -EIO;
2976 		*bad_wr = wr;
2977 		nreq = 0;
2978 		goto out;
2979 	}
2980 
2981 	ind = qp->sq_next_wqe;
2982 
2983 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2984 		lso_wqe = &dummy;
2985 		blh = 0;
2986 
2987 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2988 			err = -ENOMEM;
2989 			*bad_wr = wr;
2990 			goto out;
2991 		}
2992 
2993 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
2994 			err = -EINVAL;
2995 			*bad_wr = wr;
2996 			goto out;
2997 		}
2998 
2999 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3000 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3001 
3002 		ctrl->srcrb_flags =
3003 			(wr->send_flags & IB_SEND_SIGNALED ?
3004 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3005 			(wr->send_flags & IB_SEND_SOLICITED ?
3006 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3007 			((wr->send_flags & IB_SEND_IP_CSUM) ?
3008 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3009 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3010 			qp->sq_signal_bits;
3011 
3012 		ctrl->imm = send_ieth(wr);
3013 
3014 		wqe += sizeof *ctrl;
3015 		size = sizeof *ctrl / 16;
3016 
3017 		switch (qp->mlx4_ib_qp_type) {
3018 		case MLX4_IB_QPT_RC:
3019 		case MLX4_IB_QPT_UC:
3020 			switch (wr->opcode) {
3021 			case IB_WR_ATOMIC_CMP_AND_SWP:
3022 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3023 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3024 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3025 					      atomic_wr(wr)->rkey);
3026 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3027 
3028 				set_atomic_seg(wqe, atomic_wr(wr));
3029 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
3030 
3031 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3032 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3033 
3034 				break;
3035 
3036 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3037 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3038 					      atomic_wr(wr)->rkey);
3039 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3040 
3041 				set_masked_atomic_seg(wqe, atomic_wr(wr));
3042 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
3043 
3044 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3045 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3046 
3047 				break;
3048 
3049 			case IB_WR_RDMA_READ:
3050 			case IB_WR_RDMA_WRITE:
3051 			case IB_WR_RDMA_WRITE_WITH_IMM:
3052 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3053 					      rdma_wr(wr)->rkey);
3054 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3055 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3056 				break;
3057 
3058 			case IB_WR_LOCAL_INV:
3059 				ctrl->srcrb_flags |=
3060 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3061 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3062 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3063 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3064 				break;
3065 
3066 			case IB_WR_REG_MR:
3067 				ctrl->srcrb_flags |=
3068 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3069 				set_reg_seg(wqe, reg_wr(wr));
3070 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3071 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3072 				break;
3073 
3074 			default:
3075 				/* No extra segments required for sends */
3076 				break;
3077 			}
3078 			break;
3079 
3080 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3081 			err =  build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3082 					ctrl, &seglen);
3083 			if (unlikely(err)) {
3084 				*bad_wr = wr;
3085 				goto out;
3086 			}
3087 			wqe  += seglen;
3088 			size += seglen / 16;
3089 			break;
3090 		case MLX4_IB_QPT_TUN_SMI:
3091 		case MLX4_IB_QPT_TUN_GSI:
3092 			/* this is a UD qp used in MAD responses to slaves. */
3093 			set_datagram_seg(wqe, ud_wr(wr));
3094 			/* set the forced-loopback bit in the data seg av */
3095 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3096 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3097 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3098 			break;
3099 		case MLX4_IB_QPT_UD:
3100 			set_datagram_seg(wqe, ud_wr(wr));
3101 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3102 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3103 
3104 			if (wr->opcode == IB_WR_LSO) {
3105 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3106 						&lso_hdr_sz, &blh);
3107 				if (unlikely(err)) {
3108 					*bad_wr = wr;
3109 					goto out;
3110 				}
3111 				lso_wqe = (__be32 *) wqe;
3112 				wqe  += seglen;
3113 				size += seglen / 16;
3114 			}
3115 			break;
3116 
3117 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3118 			err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3119 					ctrl, &seglen);
3120 			if (unlikely(err)) {
3121 				*bad_wr = wr;
3122 				goto out;
3123 			}
3124 			wqe  += seglen;
3125 			size += seglen / 16;
3126 			/* to start tunnel header on a cache-line boundary */
3127 			add_zero_len_inline(wqe);
3128 			wqe += 16;
3129 			size++;
3130 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3131 			wqe  += seglen;
3132 			size += seglen / 16;
3133 			break;
3134 		case MLX4_IB_QPT_PROXY_SMI:
3135 		case MLX4_IB_QPT_PROXY_GSI:
3136 			/* If we are tunneling special qps, this is a UD qp.
3137 			 * In this case we first add a UD segment targeting
3138 			 * the tunnel qp, and then add a header with address
3139 			 * information */
3140 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3141 						ud_wr(wr),
3142 						qp->mlx4_ib_qp_type);
3143 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3144 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3145 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3146 			wqe  += seglen;
3147 			size += seglen / 16;
3148 			break;
3149 
3150 		case MLX4_IB_QPT_SMI:
3151 		case MLX4_IB_QPT_GSI:
3152 			err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
3153 					&seglen);
3154 			if (unlikely(err)) {
3155 				*bad_wr = wr;
3156 				goto out;
3157 			}
3158 			wqe  += seglen;
3159 			size += seglen / 16;
3160 			break;
3161 
3162 		default:
3163 			break;
3164 		}
3165 
3166 		/*
3167 		 * Write data segments in reverse order, so as to
3168 		 * overwrite cacheline stamp last within each
3169 		 * cacheline.  This avoids issues with WQE
3170 		 * prefetching.
3171 		 */
3172 
3173 		dseg = wqe;
3174 		dseg += wr->num_sge - 1;
3175 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3176 
3177 		/* Add one more inline data segment for ICRC for MLX sends */
3178 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3179 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3180 			     qp->mlx4_ib_qp_type &
3181 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3182 			set_mlx_icrc_seg(dseg + 1);
3183 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3184 		}
3185 
3186 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3187 			set_data_seg(dseg, wr->sg_list + i);
3188 
3189 		/*
3190 		 * Possibly overwrite stamping in cacheline with LSO
3191 		 * segment only after making sure all data segments
3192 		 * are written.
3193 		 */
3194 		wmb();
3195 		*lso_wqe = lso_hdr_sz;
3196 
3197 		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3198 					     MLX4_WQE_CTRL_FENCE : 0) | size;
3199 
3200 		/*
3201 		 * Make sure descriptor is fully written before
3202 		 * setting ownership bit (because HW can start
3203 		 * executing as soon as we do).
3204 		 */
3205 		wmb();
3206 
3207 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3208 			*bad_wr = wr;
3209 			err = -EINVAL;
3210 			goto out;
3211 		}
3212 
3213 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3214 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3215 
3216 		stamp = ind + qp->sq_spare_wqes;
3217 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
3218 
3219 		/*
3220 		 * We can improve latency by not stamping the last
3221 		 * send queue WQE until after ringing the doorbell, so
3222 		 * only stamp here if there are still more WQEs to post.
3223 		 *
3224 		 * Same optimization applies to padding with NOP wqe
3225 		 * in case of WQE shrinking (used to prevent wrap-around
3226 		 * in the middle of WR).
3227 		 */
3228 		if (wr->next) {
3229 			stamp_send_wqe(qp, stamp, size * 16);
3230 			ind = pad_wraparound(qp, ind);
3231 		}
3232 	}
3233 
3234 out:
3235 	if (likely(nreq)) {
3236 		qp->sq.head += nreq;
3237 
3238 		/*
3239 		 * Make sure that descriptors are written before
3240 		 * doorbell record.
3241 		 */
3242 		wmb();
3243 
3244 		writel(qp->doorbell_qpn,
3245 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3246 
3247 		/*
3248 		 * Make sure doorbells don't leak out of SQ spinlock
3249 		 * and reach the HCA out of order.
3250 		 */
3251 		mmiowb();
3252 
3253 		stamp_send_wqe(qp, stamp, size * 16);
3254 
3255 		ind = pad_wraparound(qp, ind);
3256 		qp->sq_next_wqe = ind;
3257 	}
3258 
3259 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3260 
3261 	return err;
3262 }
3263 
3264 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3265 		      struct ib_recv_wr **bad_wr)
3266 {
3267 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3268 	struct mlx4_wqe_data_seg *scat;
3269 	unsigned long flags;
3270 	int err = 0;
3271 	int nreq;
3272 	int ind;
3273 	int max_gs;
3274 	int i;
3275 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3276 
3277 	max_gs = qp->rq.max_gs;
3278 	spin_lock_irqsave(&qp->rq.lock, flags);
3279 
3280 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3281 		err = -EIO;
3282 		*bad_wr = wr;
3283 		nreq = 0;
3284 		goto out;
3285 	}
3286 
3287 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3288 
3289 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3290 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3291 			err = -ENOMEM;
3292 			*bad_wr = wr;
3293 			goto out;
3294 		}
3295 
3296 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3297 			err = -EINVAL;
3298 			*bad_wr = wr;
3299 			goto out;
3300 		}
3301 
3302 		scat = get_recv_wqe(qp, ind);
3303 
3304 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3305 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3306 			ib_dma_sync_single_for_device(ibqp->device,
3307 						      qp->sqp_proxy_rcv[ind].map,
3308 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3309 						      DMA_FROM_DEVICE);
3310 			scat->byte_count =
3311 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3312 			/* use dma lkey from upper layer entry */
3313 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3314 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3315 			scat++;
3316 			max_gs--;
3317 		}
3318 
3319 		for (i = 0; i < wr->num_sge; ++i)
3320 			__set_data_seg(scat + i, wr->sg_list + i);
3321 
3322 		if (i < max_gs) {
3323 			scat[i].byte_count = 0;
3324 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3325 			scat[i].addr       = 0;
3326 		}
3327 
3328 		qp->rq.wrid[ind] = wr->wr_id;
3329 
3330 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3331 	}
3332 
3333 out:
3334 	if (likely(nreq)) {
3335 		qp->rq.head += nreq;
3336 
3337 		/*
3338 		 * Make sure that descriptors are written before
3339 		 * doorbell record.
3340 		 */
3341 		wmb();
3342 
3343 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3344 	}
3345 
3346 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3347 
3348 	return err;
3349 }
3350 
3351 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3352 {
3353 	switch (mlx4_state) {
3354 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3355 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3356 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3357 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3358 	case MLX4_QP_STATE_SQ_DRAINING:
3359 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3360 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3361 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3362 	default:		     return -1;
3363 	}
3364 }
3365 
3366 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3367 {
3368 	switch (mlx4_mig_state) {
3369 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3370 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3371 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3372 	default: return -1;
3373 	}
3374 }
3375 
3376 static int to_ib_qp_access_flags(int mlx4_flags)
3377 {
3378 	int ib_flags = 0;
3379 
3380 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3381 		ib_flags |= IB_ACCESS_REMOTE_READ;
3382 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3383 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3384 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3385 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3386 
3387 	return ib_flags;
3388 }
3389 
3390 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
3391 				struct mlx4_qp_path *path)
3392 {
3393 	struct mlx4_dev *dev = ibdev->dev;
3394 	int is_eth;
3395 
3396 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
3397 	ib_ah_attr->port_num	  = path->sched_queue & 0x40 ? 2 : 1;
3398 
3399 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
3400 		return;
3401 
3402 	is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
3403 		IB_LINK_LAYER_ETHERNET;
3404 	if (is_eth)
3405 		ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
3406 		((path->sched_queue & 4) << 1);
3407 	else
3408 		ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
3409 
3410 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
3411 	ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
3412 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
3413 	ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
3414 	if (ib_ah_attr->ah_flags) {
3415 		ib_ah_attr->grh.sgid_index = path->mgid_index;
3416 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
3417 		ib_ah_attr->grh.traffic_class =
3418 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3419 		ib_ah_attr->grh.flow_label =
3420 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3421 		memcpy(ib_ah_attr->grh.dgid.raw,
3422 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
3423 	}
3424 }
3425 
3426 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3427 		     struct ib_qp_init_attr *qp_init_attr)
3428 {
3429 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3430 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3431 	struct mlx4_qp_context context;
3432 	int mlx4_state;
3433 	int err = 0;
3434 
3435 	mutex_lock(&qp->mutex);
3436 
3437 	if (qp->state == IB_QPS_RESET) {
3438 		qp_attr->qp_state = IB_QPS_RESET;
3439 		goto done;
3440 	}
3441 
3442 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3443 	if (err) {
3444 		err = -EINVAL;
3445 		goto out;
3446 	}
3447 
3448 	mlx4_state = be32_to_cpu(context.flags) >> 28;
3449 
3450 	qp->state		     = to_ib_qp_state(mlx4_state);
3451 	qp_attr->qp_state	     = qp->state;
3452 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
3453 	qp_attr->path_mig_state	     =
3454 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3455 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
3456 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3457 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
3458 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
3459 	qp_attr->qp_access_flags     =
3460 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
3461 
3462 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3463 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3464 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3465 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3466 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
3467 	}
3468 
3469 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3470 	if (qp_attr->qp_state == IB_QPS_INIT)
3471 		qp_attr->port_num = qp->port;
3472 	else
3473 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3474 
3475 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3476 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3477 
3478 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3479 
3480 	qp_attr->max_dest_rd_atomic =
3481 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3482 	qp_attr->min_rnr_timer	    =
3483 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3484 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
3485 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
3486 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
3487 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
3488 
3489 done:
3490 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
3491 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
3492 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
3493 
3494 	if (!ibqp->uobject) {
3495 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
3496 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
3497 	} else {
3498 		qp_attr->cap.max_send_wr  = 0;
3499 		qp_attr->cap.max_send_sge = 0;
3500 	}
3501 
3502 	/*
3503 	 * We don't support inline sends for kernel QPs (yet), and we
3504 	 * don't know what userspace's value should be.
3505 	 */
3506 	qp_attr->cap.max_inline_data = 0;
3507 
3508 	qp_init_attr->cap	     = qp_attr->cap;
3509 
3510 	qp_init_attr->create_flags = 0;
3511 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3512 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3513 
3514 	if (qp->flags & MLX4_IB_QP_LSO)
3515 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
3516 
3517 	if (qp->flags & MLX4_IB_QP_NETIF)
3518 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
3519 
3520 	qp_init_attr->sq_sig_type =
3521 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
3522 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3523 
3524 out:
3525 	mutex_unlock(&qp->mutex);
3526 	return err;
3527 }
3528 
3529