xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision 4da722ca)
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 struct mlx4_ib_sqp {
80 	struct mlx4_ib_qp	qp;
81 	int			pkey_index;
82 	u32			qkey;
83 	u32			send_psn;
84 	struct ib_ud_header	ud_header;
85 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
86 	struct ib_qp		*roce_v2_gsi;
87 };
88 
89 enum {
90 	MLX4_IB_MIN_SQ_STRIDE	= 6,
91 	MLX4_IB_CACHE_LINE_SIZE	= 64,
92 };
93 
94 enum {
95 	MLX4_RAW_QP_MTU		= 7,
96 	MLX4_RAW_QP_MSGMAX	= 31,
97 };
98 
99 #ifndef ETH_ALEN
100 #define ETH_ALEN        6
101 #endif
102 
103 static const __be32 mlx4_ib_opcode[] = {
104 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
105 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
106 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
107 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
108 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
109 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
110 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
111 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
112 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
113 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
114 	[IB_WR_REG_MR]				= cpu_to_be32(MLX4_OPCODE_FMR),
115 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
116 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
117 };
118 
119 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
120 {
121 	return container_of(mqp, struct mlx4_ib_sqp, qp);
122 }
123 
124 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
125 {
126 	if (!mlx4_is_master(dev->dev))
127 		return 0;
128 
129 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
130 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
131 		8 * MLX4_MFUNC_MAX;
132 }
133 
134 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
135 {
136 	int proxy_sqp = 0;
137 	int real_sqp = 0;
138 	int i;
139 	/* PPF or Native -- real SQP */
140 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
141 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
142 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
143 	if (real_sqp)
144 		return 1;
145 	/* VF or PF -- proxy SQP */
146 	if (mlx4_is_mfunc(dev->dev)) {
147 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
148 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] ||
149 			    qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) {
150 				proxy_sqp = 1;
151 				break;
152 			}
153 		}
154 	}
155 	if (proxy_sqp)
156 		return 1;
157 
158 	return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
159 }
160 
161 /* used for INIT/CLOSE port logic */
162 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
163 {
164 	int proxy_qp0 = 0;
165 	int real_qp0 = 0;
166 	int i;
167 	/* PPF or Native -- real QP0 */
168 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
169 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
170 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
171 	if (real_qp0)
172 		return 1;
173 	/* VF or PF -- proxy QP0 */
174 	if (mlx4_is_mfunc(dev->dev)) {
175 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
176 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) {
177 				proxy_qp0 = 1;
178 				break;
179 			}
180 		}
181 	}
182 	return proxy_qp0;
183 }
184 
185 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
186 {
187 	return mlx4_buf_offset(&qp->buf, offset);
188 }
189 
190 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
191 {
192 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
193 }
194 
195 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
196 {
197 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
198 }
199 
200 /*
201  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
202  * first four bytes of every 64 byte chunk with
203  *     0x7FFFFFF | (invalid_ownership_value << 31).
204  *
205  * When the max work request size is less than or equal to the WQE
206  * basic block size, as an optimization, we can stamp all WQEs with
207  * 0xffffffff, and skip the very first chunk of each WQE.
208  */
209 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
210 {
211 	__be32 *wqe;
212 	int i;
213 	int s;
214 	int ind;
215 	void *buf;
216 	__be32 stamp;
217 	struct mlx4_wqe_ctrl_seg *ctrl;
218 
219 	if (qp->sq_max_wqes_per_wr > 1) {
220 		s = roundup(size, 1U << qp->sq.wqe_shift);
221 		for (i = 0; i < s; i += 64) {
222 			ind = (i >> qp->sq.wqe_shift) + n;
223 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
224 						       cpu_to_be32(0xffffffff);
225 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
226 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
227 			*wqe = stamp;
228 		}
229 	} else {
230 		ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
231 		s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4;
232 		for (i = 64; i < s; i += 64) {
233 			wqe = buf + i;
234 			*wqe = cpu_to_be32(0xffffffff);
235 		}
236 	}
237 }
238 
239 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
240 {
241 	struct mlx4_wqe_ctrl_seg *ctrl;
242 	struct mlx4_wqe_inline_seg *inl;
243 	void *wqe;
244 	int s;
245 
246 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
247 	s = sizeof(struct mlx4_wqe_ctrl_seg);
248 
249 	if (qp->ibqp.qp_type == IB_QPT_UD) {
250 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
251 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
252 		memset(dgram, 0, sizeof *dgram);
253 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
254 		s += sizeof(struct mlx4_wqe_datagram_seg);
255 	}
256 
257 	/* Pad the remainder of the WQE with an inline data segment. */
258 	if (size > s) {
259 		inl = wqe + s;
260 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
261 	}
262 	ctrl->srcrb_flags = 0;
263 	ctrl->qpn_vlan.fence_size = size / 16;
264 	/*
265 	 * Make sure descriptor is fully written before setting ownership bit
266 	 * (because HW can start executing as soon as we do).
267 	 */
268 	wmb();
269 
270 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
271 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
272 
273 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
274 }
275 
276 /* Post NOP WQE to prevent wrap-around in the middle of WR */
277 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
278 {
279 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
280 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
281 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
282 		ind += s;
283 	}
284 	return ind;
285 }
286 
287 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
288 {
289 	struct ib_event event;
290 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
291 
292 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
293 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
294 
295 	if (ibqp->event_handler) {
296 		event.device     = ibqp->device;
297 		event.element.qp = ibqp;
298 		switch (type) {
299 		case MLX4_EVENT_TYPE_PATH_MIG:
300 			event.event = IB_EVENT_PATH_MIG;
301 			break;
302 		case MLX4_EVENT_TYPE_COMM_EST:
303 			event.event = IB_EVENT_COMM_EST;
304 			break;
305 		case MLX4_EVENT_TYPE_SQ_DRAINED:
306 			event.event = IB_EVENT_SQ_DRAINED;
307 			break;
308 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
309 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
310 			break;
311 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
312 			event.event = IB_EVENT_QP_FATAL;
313 			break;
314 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
315 			event.event = IB_EVENT_PATH_MIG_ERR;
316 			break;
317 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
318 			event.event = IB_EVENT_QP_REQ_ERR;
319 			break;
320 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
321 			event.event = IB_EVENT_QP_ACCESS_ERR;
322 			break;
323 		default:
324 			pr_warn("Unexpected event type %d "
325 			       "on QP %06x\n", type, qp->qpn);
326 			return;
327 		}
328 
329 		ibqp->event_handler(&event, ibqp->qp_context);
330 	}
331 }
332 
333 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
334 {
335 	/*
336 	 * UD WQEs must have a datagram segment.
337 	 * RC and UC WQEs might have a remote address segment.
338 	 * MLX WQEs need two extra inline data segments (for the UD
339 	 * header and space for the ICRC).
340 	 */
341 	switch (type) {
342 	case MLX4_IB_QPT_UD:
343 		return sizeof (struct mlx4_wqe_ctrl_seg) +
344 			sizeof (struct mlx4_wqe_datagram_seg) +
345 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
346 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
347 	case MLX4_IB_QPT_PROXY_SMI:
348 	case MLX4_IB_QPT_PROXY_GSI:
349 		return sizeof (struct mlx4_wqe_ctrl_seg) +
350 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
351 	case MLX4_IB_QPT_TUN_SMI_OWNER:
352 	case MLX4_IB_QPT_TUN_GSI:
353 		return sizeof (struct mlx4_wqe_ctrl_seg) +
354 			sizeof (struct mlx4_wqe_datagram_seg);
355 
356 	case MLX4_IB_QPT_UC:
357 		return sizeof (struct mlx4_wqe_ctrl_seg) +
358 			sizeof (struct mlx4_wqe_raddr_seg);
359 	case MLX4_IB_QPT_RC:
360 		return sizeof (struct mlx4_wqe_ctrl_seg) +
361 			sizeof (struct mlx4_wqe_masked_atomic_seg) +
362 			sizeof (struct mlx4_wqe_raddr_seg);
363 	case MLX4_IB_QPT_SMI:
364 	case MLX4_IB_QPT_GSI:
365 		return sizeof (struct mlx4_wqe_ctrl_seg) +
366 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
367 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
368 					   MLX4_INLINE_ALIGN) *
369 			      sizeof (struct mlx4_wqe_inline_seg),
370 			      sizeof (struct mlx4_wqe_data_seg)) +
371 			ALIGN(4 +
372 			      sizeof (struct mlx4_wqe_inline_seg),
373 			      sizeof (struct mlx4_wqe_data_seg));
374 	default:
375 		return sizeof (struct mlx4_wqe_ctrl_seg);
376 	}
377 }
378 
379 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
380 		       int is_user, int has_rq, struct mlx4_ib_qp *qp)
381 {
382 	/* Sanity check RQ size before proceeding */
383 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
384 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
385 		return -EINVAL;
386 
387 	if (!has_rq) {
388 		if (cap->max_recv_wr)
389 			return -EINVAL;
390 
391 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
392 	} else {
393 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
394 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
395 			return -EINVAL;
396 
397 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
398 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
399 		qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
400 	}
401 
402 	/* leave userspace return values as they were, so as not to break ABI */
403 	if (is_user) {
404 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
405 		cap->max_recv_sge = qp->rq.max_gs;
406 	} else {
407 		cap->max_recv_wr  = qp->rq.max_post =
408 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
409 		cap->max_recv_sge = min(qp->rq.max_gs,
410 					min(dev->dev->caps.max_sq_sg,
411 					    dev->dev->caps.max_rq_sg));
412 	}
413 
414 	return 0;
415 }
416 
417 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
418 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp,
419 			      bool shrink_wqe)
420 {
421 	int s;
422 
423 	/* Sanity check SQ size before proceeding */
424 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
425 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
426 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
427 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
428 		return -EINVAL;
429 
430 	/*
431 	 * For MLX transport we need 2 extra S/G entries:
432 	 * one for the header and one for the checksum at the end
433 	 */
434 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
435 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
436 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
437 		return -EINVAL;
438 
439 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
440 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
441 		send_wqe_overhead(type, qp->flags);
442 
443 	if (s > dev->dev->caps.max_sq_desc_sz)
444 		return -EINVAL;
445 
446 	/*
447 	 * Hermon supports shrinking WQEs, such that a single work
448 	 * request can include multiple units of 1 << wqe_shift.  This
449 	 * way, work requests can differ in size, and do not have to
450 	 * be a power of 2 in size, saving memory and speeding up send
451 	 * WR posting.  Unfortunately, if we do this then the
452 	 * wqe_index field in CQEs can't be used to look up the WR ID
453 	 * anymore, so we do this only if selective signaling is off.
454 	 *
455 	 * Further, on 32-bit platforms, we can't use vmap() to make
456 	 * the QP buffer virtually contiguous.  Thus we have to use
457 	 * constant-sized WRs to make sure a WR is always fully within
458 	 * a single page-sized chunk.
459 	 *
460 	 * Finally, we use NOP work requests to pad the end of the
461 	 * work queue, to avoid wrap-around in the middle of WR.  We
462 	 * set NEC bit to avoid getting completions with error for
463 	 * these NOP WRs, but since NEC is only supported starting
464 	 * with firmware 2.2.232, we use constant-sized WRs for older
465 	 * firmware.
466 	 *
467 	 * And, since MLX QPs only support SEND, we use constant-sized
468 	 * WRs in this case.
469 	 *
470 	 * We look for the smallest value of wqe_shift such that the
471 	 * resulting number of wqes does not exceed device
472 	 * capabilities.
473 	 *
474 	 * We set WQE size to at least 64 bytes, this way stamping
475 	 * invalidates each WQE.
476 	 */
477 	if (shrink_wqe && dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
478 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
479 	    type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
480 	    !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
481 		      MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
482 		qp->sq.wqe_shift = ilog2(64);
483 	else
484 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
485 
486 	for (;;) {
487 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
488 
489 		/*
490 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
491 		 * allow HW to prefetch.
492 		 */
493 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
494 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
495 						    qp->sq_max_wqes_per_wr +
496 						    qp->sq_spare_wqes);
497 
498 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
499 			break;
500 
501 		if (qp->sq_max_wqes_per_wr <= 1)
502 			return -EINVAL;
503 
504 		++qp->sq.wqe_shift;
505 	}
506 
507 	qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
508 			     (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
509 			 send_wqe_overhead(type, qp->flags)) /
510 		sizeof (struct mlx4_wqe_data_seg);
511 
512 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
513 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
514 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
515 		qp->rq.offset = 0;
516 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
517 	} else {
518 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
519 		qp->sq.offset = 0;
520 	}
521 
522 	cap->max_send_wr  = qp->sq.max_post =
523 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
524 	cap->max_send_sge = min(qp->sq.max_gs,
525 				min(dev->dev->caps.max_sq_sg,
526 				    dev->dev->caps.max_rq_sg));
527 	/* We don't support inline sends for kernel QPs (yet) */
528 	cap->max_inline_data = 0;
529 
530 	return 0;
531 }
532 
533 static int set_user_sq_size(struct mlx4_ib_dev *dev,
534 			    struct mlx4_ib_qp *qp,
535 			    struct mlx4_ib_create_qp *ucmd)
536 {
537 	/* Sanity check SQ size before proceeding */
538 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
539 	    ucmd->log_sq_stride >
540 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
541 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
542 		return -EINVAL;
543 
544 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
545 	qp->sq.wqe_shift = ucmd->log_sq_stride;
546 
547 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
548 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
549 
550 	return 0;
551 }
552 
553 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
554 {
555 	int i;
556 
557 	qp->sqp_proxy_rcv =
558 		kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
559 			GFP_KERNEL);
560 	if (!qp->sqp_proxy_rcv)
561 		return -ENOMEM;
562 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
563 		qp->sqp_proxy_rcv[i].addr =
564 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
565 				GFP_KERNEL);
566 		if (!qp->sqp_proxy_rcv[i].addr)
567 			goto err;
568 		qp->sqp_proxy_rcv[i].map =
569 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
570 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
571 					  DMA_FROM_DEVICE);
572 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
573 			kfree(qp->sqp_proxy_rcv[i].addr);
574 			goto err;
575 		}
576 	}
577 	return 0;
578 
579 err:
580 	while (i > 0) {
581 		--i;
582 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
583 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
584 				    DMA_FROM_DEVICE);
585 		kfree(qp->sqp_proxy_rcv[i].addr);
586 	}
587 	kfree(qp->sqp_proxy_rcv);
588 	qp->sqp_proxy_rcv = NULL;
589 	return -ENOMEM;
590 }
591 
592 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
593 {
594 	int i;
595 
596 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
597 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
598 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
599 				    DMA_FROM_DEVICE);
600 		kfree(qp->sqp_proxy_rcv[i].addr);
601 	}
602 	kfree(qp->sqp_proxy_rcv);
603 }
604 
605 static int qp_has_rq(struct ib_qp_init_attr *attr)
606 {
607 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
608 		return 0;
609 
610 	return !attr->srq;
611 }
612 
613 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
614 {
615 	int i;
616 	for (i = 0; i < dev->caps.num_ports; i++) {
617 		if (qpn == dev->caps.qp0_proxy[i])
618 			return !!dev->caps.qp0_qkey[i];
619 	}
620 	return 0;
621 }
622 
623 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
624 				    struct mlx4_ib_qp *qp)
625 {
626 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
627 	mlx4_counter_free(dev->dev, qp->counter_index->index);
628 	list_del(&qp->counter_index->list);
629 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
630 
631 	kfree(qp->counter_index);
632 	qp->counter_index = NULL;
633 }
634 
635 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
636 			    struct ib_qp_init_attr *init_attr,
637 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp,
638 			    gfp_t gfp)
639 {
640 	int qpn;
641 	int err;
642 	struct ib_qp_cap backup_cap;
643 	struct mlx4_ib_sqp *sqp = NULL;
644 	struct mlx4_ib_qp *qp;
645 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
646 	struct mlx4_ib_cq *mcq;
647 	unsigned long flags;
648 
649 	/* When tunneling special qps, we use a plain UD qp */
650 	if (sqpn) {
651 		if (mlx4_is_mfunc(dev->dev) &&
652 		    (!mlx4_is_master(dev->dev) ||
653 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
654 			if (init_attr->qp_type == IB_QPT_GSI)
655 				qp_type = MLX4_IB_QPT_PROXY_GSI;
656 			else {
657 				if (mlx4_is_master(dev->dev) ||
658 				    qp0_enabled_vf(dev->dev, sqpn))
659 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
660 				else
661 					qp_type = MLX4_IB_QPT_PROXY_SMI;
662 			}
663 		}
664 		qpn = sqpn;
665 		/* add extra sg entry for tunneling */
666 		init_attr->cap.max_recv_sge++;
667 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
668 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
669 			container_of(init_attr,
670 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
671 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
672 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
673 		    !mlx4_is_master(dev->dev))
674 			return -EINVAL;
675 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
676 			qp_type = MLX4_IB_QPT_TUN_GSI;
677 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
678 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
679 					     tnl_init->port))
680 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
681 		else
682 			qp_type = MLX4_IB_QPT_TUN_SMI;
683 		/* we are definitely in the PPF here, since we are creating
684 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
685 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
686 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
687 		sqpn = qpn;
688 	}
689 
690 	if (!*caller_qp) {
691 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
692 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
693 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
694 			sqp = kzalloc(sizeof (struct mlx4_ib_sqp), gfp);
695 			if (!sqp)
696 				return -ENOMEM;
697 			qp = &sqp->qp;
698 			qp->pri.vid = 0xFFFF;
699 			qp->alt.vid = 0xFFFF;
700 		} else {
701 			qp = kzalloc(sizeof (struct mlx4_ib_qp), gfp);
702 			if (!qp)
703 				return -ENOMEM;
704 			qp->pri.vid = 0xFFFF;
705 			qp->alt.vid = 0xFFFF;
706 		}
707 	} else
708 		qp = *caller_qp;
709 
710 	qp->mlx4_ib_qp_type = qp_type;
711 
712 	mutex_init(&qp->mutex);
713 	spin_lock_init(&qp->sq.lock);
714 	spin_lock_init(&qp->rq.lock);
715 	INIT_LIST_HEAD(&qp->gid_list);
716 	INIT_LIST_HEAD(&qp->steering_rules);
717 
718 	qp->state	 = IB_QPS_RESET;
719 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
720 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
721 
722 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
723 	if (err)
724 		goto err;
725 
726 	if (pd->uobject) {
727 		struct mlx4_ib_create_qp ucmd;
728 
729 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
730 			err = -EFAULT;
731 			goto err;
732 		}
733 
734 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
735 
736 		err = set_user_sq_size(dev, qp, &ucmd);
737 		if (err)
738 			goto err;
739 
740 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
741 				       qp->buf_size, 0, 0);
742 		if (IS_ERR(qp->umem)) {
743 			err = PTR_ERR(qp->umem);
744 			goto err;
745 		}
746 
747 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
748 				    qp->umem->page_shift, &qp->mtt);
749 		if (err)
750 			goto err_buf;
751 
752 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
753 		if (err)
754 			goto err_mtt;
755 
756 		if (qp_has_rq(init_attr)) {
757 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
758 						  ucmd.db_addr, &qp->db);
759 			if (err)
760 				goto err_mtt;
761 		}
762 	} else {
763 		qp->sq_no_prefetch = 0;
764 
765 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
766 			qp->flags |= MLX4_IB_QP_LSO;
767 
768 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
769 			if (dev->steering_support ==
770 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
771 				qp->flags |= MLX4_IB_QP_NETIF;
772 			else
773 				goto err;
774 		}
775 
776 		memcpy(&backup_cap, &init_attr->cap, sizeof(backup_cap));
777 		err = set_kernel_sq_size(dev, &init_attr->cap,
778 					 qp_type, qp, true);
779 		if (err)
780 			goto err;
781 
782 		if (qp_has_rq(init_attr)) {
783 			err = mlx4_db_alloc(dev->dev, &qp->db, 0, gfp);
784 			if (err)
785 				goto err;
786 
787 			*qp->db.db = 0;
788 		}
789 
790 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, qp->buf_size,
791 				   &qp->buf, gfp)) {
792 			memcpy(&init_attr->cap, &backup_cap,
793 			       sizeof(backup_cap));
794 			err = set_kernel_sq_size(dev, &init_attr->cap, qp_type,
795 						 qp, false);
796 			if (err)
797 				goto err_db;
798 
799 			if (mlx4_buf_alloc(dev->dev, qp->buf_size,
800 					   PAGE_SIZE * 2, &qp->buf, gfp)) {
801 				err = -ENOMEM;
802 				goto err_db;
803 			}
804 		}
805 
806 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
807 				    &qp->mtt);
808 		if (err)
809 			goto err_buf;
810 
811 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf, gfp);
812 		if (err)
813 			goto err_mtt;
814 
815 		qp->sq.wrid = kmalloc_array(qp->sq.wqe_cnt, sizeof(u64),
816 					gfp | __GFP_NOWARN);
817 		if (!qp->sq.wrid)
818 			qp->sq.wrid = __vmalloc(qp->sq.wqe_cnt * sizeof(u64),
819 						gfp, PAGE_KERNEL);
820 		qp->rq.wrid = kmalloc_array(qp->rq.wqe_cnt, sizeof(u64),
821 					gfp | __GFP_NOWARN);
822 		if (!qp->rq.wrid)
823 			qp->rq.wrid = __vmalloc(qp->rq.wqe_cnt * sizeof(u64),
824 						gfp, PAGE_KERNEL);
825 		if (!qp->sq.wrid || !qp->rq.wrid) {
826 			err = -ENOMEM;
827 			goto err_wrid;
828 		}
829 	}
830 
831 	if (sqpn) {
832 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
833 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
834 			if (alloc_proxy_bufs(pd->device, qp)) {
835 				err = -ENOMEM;
836 				goto err_wrid;
837 			}
838 		}
839 	} else {
840 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
841 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
842 		 * VLAN insertion. */
843 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
844 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
845 						    (init_attr->cap.max_send_wr ?
846 						     MLX4_RESERVE_ETH_BF_QP : 0) |
847 						    (init_attr->cap.max_recv_wr ?
848 						     MLX4_RESERVE_A0_QP : 0));
849 		else
850 			if (qp->flags & MLX4_IB_QP_NETIF)
851 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
852 			else
853 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
854 							    &qpn, 0);
855 		if (err)
856 			goto err_proxy;
857 	}
858 
859 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
860 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
861 
862 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp, gfp);
863 	if (err)
864 		goto err_qpn;
865 
866 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
867 		qp->mqp.qpn |= (1 << 23);
868 
869 	/*
870 	 * Hardware wants QPN written in big-endian order (after
871 	 * shifting) for send doorbell.  Precompute this value to save
872 	 * a little bit when posting sends.
873 	 */
874 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
875 
876 	qp->mqp.event = mlx4_ib_qp_event;
877 	if (!*caller_qp)
878 		*caller_qp = qp;
879 
880 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
881 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
882 			 to_mcq(init_attr->recv_cq));
883 	/* Maintain device to QPs access, needed for further handling
884 	 * via reset flow
885 	 */
886 	list_add_tail(&qp->qps_list, &dev->qp_list);
887 	/* Maintain CQ to QPs access, needed for further handling
888 	 * via reset flow
889 	 */
890 	mcq = to_mcq(init_attr->send_cq);
891 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
892 	mcq = to_mcq(init_attr->recv_cq);
893 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
894 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
895 			   to_mcq(init_attr->recv_cq));
896 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
897 	return 0;
898 
899 err_qpn:
900 	if (!sqpn) {
901 		if (qp->flags & MLX4_IB_QP_NETIF)
902 			mlx4_ib_steer_qp_free(dev, qpn, 1);
903 		else
904 			mlx4_qp_release_range(dev->dev, qpn, 1);
905 	}
906 err_proxy:
907 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
908 		free_proxy_bufs(pd->device, qp);
909 err_wrid:
910 	if (pd->uobject) {
911 		if (qp_has_rq(init_attr))
912 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
913 	} else {
914 		kvfree(qp->sq.wrid);
915 		kvfree(qp->rq.wrid);
916 	}
917 
918 err_mtt:
919 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
920 
921 err_buf:
922 	if (pd->uobject)
923 		ib_umem_release(qp->umem);
924 	else
925 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
926 
927 err_db:
928 	if (!pd->uobject && qp_has_rq(init_attr))
929 		mlx4_db_free(dev->dev, &qp->db);
930 
931 err:
932 	if (sqp)
933 		kfree(sqp);
934 	else if (!*caller_qp)
935 		kfree(qp);
936 	return err;
937 }
938 
939 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
940 {
941 	switch (state) {
942 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
943 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
944 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
945 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
946 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
947 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
948 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
949 	default:		return -1;
950 	}
951 }
952 
953 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
954 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
955 {
956 	if (send_cq == recv_cq) {
957 		spin_lock(&send_cq->lock);
958 		__acquire(&recv_cq->lock);
959 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
960 		spin_lock(&send_cq->lock);
961 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
962 	} else {
963 		spin_lock(&recv_cq->lock);
964 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
965 	}
966 }
967 
968 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
969 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
970 {
971 	if (send_cq == recv_cq) {
972 		__release(&recv_cq->lock);
973 		spin_unlock(&send_cq->lock);
974 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
975 		spin_unlock(&recv_cq->lock);
976 		spin_unlock(&send_cq->lock);
977 	} else {
978 		spin_unlock(&send_cq->lock);
979 		spin_unlock(&recv_cq->lock);
980 	}
981 }
982 
983 static void del_gid_entries(struct mlx4_ib_qp *qp)
984 {
985 	struct mlx4_ib_gid_entry *ge, *tmp;
986 
987 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
988 		list_del(&ge->list);
989 		kfree(ge);
990 	}
991 }
992 
993 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
994 {
995 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
996 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
997 	else
998 		return to_mpd(qp->ibqp.pd);
999 }
1000 
1001 static void get_cqs(struct mlx4_ib_qp *qp,
1002 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1003 {
1004 	switch (qp->ibqp.qp_type) {
1005 	case IB_QPT_XRC_TGT:
1006 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1007 		*recv_cq = *send_cq;
1008 		break;
1009 	case IB_QPT_XRC_INI:
1010 		*send_cq = to_mcq(qp->ibqp.send_cq);
1011 		*recv_cq = *send_cq;
1012 		break;
1013 	default:
1014 		*send_cq = to_mcq(qp->ibqp.send_cq);
1015 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
1016 		break;
1017 	}
1018 }
1019 
1020 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1021 			      int is_user)
1022 {
1023 	struct mlx4_ib_cq *send_cq, *recv_cq;
1024 	unsigned long flags;
1025 
1026 	if (qp->state != IB_QPS_RESET) {
1027 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1028 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1029 			pr_warn("modify QP %06x to RESET failed.\n",
1030 			       qp->mqp.qpn);
1031 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1032 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1033 			qp->pri.smac = 0;
1034 			qp->pri.smac_port = 0;
1035 		}
1036 		if (qp->alt.smac) {
1037 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1038 			qp->alt.smac = 0;
1039 		}
1040 		if (qp->pri.vid < 0x1000) {
1041 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1042 			qp->pri.vid = 0xFFFF;
1043 			qp->pri.candidate_vid = 0xFFFF;
1044 			qp->pri.update_vid = 0;
1045 		}
1046 		if (qp->alt.vid < 0x1000) {
1047 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1048 			qp->alt.vid = 0xFFFF;
1049 			qp->alt.candidate_vid = 0xFFFF;
1050 			qp->alt.update_vid = 0;
1051 		}
1052 	}
1053 
1054 	get_cqs(qp, &send_cq, &recv_cq);
1055 
1056 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1057 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1058 
1059 	/* del from lists under both locks above to protect reset flow paths */
1060 	list_del(&qp->qps_list);
1061 	list_del(&qp->cq_send_list);
1062 	list_del(&qp->cq_recv_list);
1063 	if (!is_user) {
1064 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1065 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1066 		if (send_cq != recv_cq)
1067 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1068 	}
1069 
1070 	mlx4_qp_remove(dev->dev, &qp->mqp);
1071 
1072 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1073 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1074 
1075 	mlx4_qp_free(dev->dev, &qp->mqp);
1076 
1077 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1078 		if (qp->flags & MLX4_IB_QP_NETIF)
1079 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1080 		else
1081 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1082 	}
1083 
1084 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1085 
1086 	if (is_user) {
1087 		if (qp->rq.wqe_cnt)
1088 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
1089 					      &qp->db);
1090 		ib_umem_release(qp->umem);
1091 	} else {
1092 		kvfree(qp->sq.wrid);
1093 		kvfree(qp->rq.wrid);
1094 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1095 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1096 			free_proxy_bufs(&dev->ib_dev, qp);
1097 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1098 		if (qp->rq.wqe_cnt)
1099 			mlx4_db_free(dev->dev, &qp->db);
1100 	}
1101 
1102 	del_gid_entries(qp);
1103 }
1104 
1105 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1106 {
1107 	/* Native or PPF */
1108 	if (!mlx4_is_mfunc(dev->dev) ||
1109 	    (mlx4_is_master(dev->dev) &&
1110 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1111 		return  dev->dev->phys_caps.base_sqpn +
1112 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1113 			attr->port_num - 1;
1114 	}
1115 	/* PF or VF -- creating proxies */
1116 	if (attr->qp_type == IB_QPT_SMI)
1117 		return dev->dev->caps.qp0_proxy[attr->port_num - 1];
1118 	else
1119 		return dev->dev->caps.qp1_proxy[attr->port_num - 1];
1120 }
1121 
1122 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1123 					struct ib_qp_init_attr *init_attr,
1124 					struct ib_udata *udata)
1125 {
1126 	struct mlx4_ib_qp *qp = NULL;
1127 	int err;
1128 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1129 	u16 xrcdn = 0;
1130 	gfp_t gfp;
1131 
1132 	gfp = (init_attr->create_flags & MLX4_IB_QP_CREATE_USE_GFP_NOIO) ?
1133 		GFP_NOIO : GFP_KERNEL;
1134 	/*
1135 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1136 	 * and only for kernel UD QPs.
1137 	 */
1138 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1139 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1140 					MLX4_IB_SRIOV_TUNNEL_QP |
1141 					MLX4_IB_SRIOV_SQP |
1142 					MLX4_IB_QP_NETIF |
1143 					MLX4_IB_QP_CREATE_ROCE_V2_GSI |
1144 					MLX4_IB_QP_CREATE_USE_GFP_NOIO))
1145 		return ERR_PTR(-EINVAL);
1146 
1147 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1148 		if (init_attr->qp_type != IB_QPT_UD)
1149 			return ERR_PTR(-EINVAL);
1150 	}
1151 
1152 	if (init_attr->create_flags) {
1153 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1154 			return ERR_PTR(-EINVAL);
1155 
1156 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1157 						 MLX4_IB_QP_CREATE_USE_GFP_NOIO |
1158 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1159 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1160 		     init_attr->qp_type != IB_QPT_UD) ||
1161 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1162 		     init_attr->qp_type > IB_QPT_GSI) ||
1163 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1164 		     init_attr->qp_type != IB_QPT_GSI))
1165 			return ERR_PTR(-EINVAL);
1166 	}
1167 
1168 	switch (init_attr->qp_type) {
1169 	case IB_QPT_XRC_TGT:
1170 		pd = to_mxrcd(init_attr->xrcd)->pd;
1171 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1172 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1173 		/* fall through */
1174 	case IB_QPT_XRC_INI:
1175 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1176 			return ERR_PTR(-ENOSYS);
1177 		init_attr->recv_cq = init_attr->send_cq;
1178 		/* fall through */
1179 	case IB_QPT_RC:
1180 	case IB_QPT_UC:
1181 	case IB_QPT_RAW_PACKET:
1182 		qp = kzalloc(sizeof *qp, gfp);
1183 		if (!qp)
1184 			return ERR_PTR(-ENOMEM);
1185 		qp->pri.vid = 0xFFFF;
1186 		qp->alt.vid = 0xFFFF;
1187 		/* fall through */
1188 	case IB_QPT_UD:
1189 	{
1190 		err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1191 				       udata, 0, &qp, gfp);
1192 		if (err) {
1193 			kfree(qp);
1194 			return ERR_PTR(err);
1195 		}
1196 
1197 		qp->ibqp.qp_num = qp->mqp.qpn;
1198 		qp->xrcdn = xrcdn;
1199 
1200 		break;
1201 	}
1202 	case IB_QPT_SMI:
1203 	case IB_QPT_GSI:
1204 	{
1205 		int sqpn;
1206 
1207 		/* Userspace is not allowed to create special QPs: */
1208 		if (udata)
1209 			return ERR_PTR(-EINVAL);
1210 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1211 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev, 1, 1, &sqpn, 0);
1212 
1213 			if (res)
1214 				return ERR_PTR(res);
1215 		} else {
1216 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1217 		}
1218 
1219 		err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1220 				       sqpn,
1221 				       &qp, gfp);
1222 		if (err)
1223 			return ERR_PTR(err);
1224 
1225 		qp->port	= init_attr->port_num;
1226 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1227 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1228 		break;
1229 	}
1230 	default:
1231 		/* Don't support raw QPs */
1232 		return ERR_PTR(-EINVAL);
1233 	}
1234 
1235 	return &qp->ibqp;
1236 }
1237 
1238 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1239 				struct ib_qp_init_attr *init_attr,
1240 				struct ib_udata *udata) {
1241 	struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1242 	struct ib_qp *ibqp;
1243 	struct mlx4_ib_dev *dev = to_mdev(device);
1244 
1245 	ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1246 
1247 	if (!IS_ERR(ibqp) &&
1248 	    (init_attr->qp_type == IB_QPT_GSI) &&
1249 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1250 		struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1251 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1252 
1253 		if (is_eth &&
1254 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1255 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1256 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1257 
1258 			if (IS_ERR(sqp->roce_v2_gsi)) {
1259 				pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1260 				sqp->roce_v2_gsi = NULL;
1261 			} else {
1262 				sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1263 				sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1264 			}
1265 
1266 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1267 		}
1268 	}
1269 	return ibqp;
1270 }
1271 
1272 static int _mlx4_ib_destroy_qp(struct ib_qp *qp)
1273 {
1274 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1275 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1276 	struct mlx4_ib_pd *pd;
1277 
1278 	if (is_qp0(dev, mqp))
1279 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1280 
1281 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1282 	    dev->qp1_proxy[mqp->port - 1] == mqp) {
1283 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1284 		dev->qp1_proxy[mqp->port - 1] = NULL;
1285 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1286 	}
1287 
1288 	if (mqp->counter_index)
1289 		mlx4_ib_free_qp_counter(dev, mqp);
1290 
1291 	pd = get_pd(mqp);
1292 	destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1293 
1294 	if (is_sqp(dev, mqp))
1295 		kfree(to_msqp(mqp));
1296 	else
1297 		kfree(mqp);
1298 
1299 	return 0;
1300 }
1301 
1302 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1303 {
1304 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1305 
1306 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1307 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1308 
1309 		if (sqp->roce_v2_gsi)
1310 			ib_destroy_qp(sqp->roce_v2_gsi);
1311 	}
1312 
1313 	return _mlx4_ib_destroy_qp(qp);
1314 }
1315 
1316 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1317 {
1318 	switch (type) {
1319 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1320 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1321 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1322 	case MLX4_IB_QPT_XRC_INI:
1323 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1324 	case MLX4_IB_QPT_SMI:
1325 	case MLX4_IB_QPT_GSI:
1326 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1327 
1328 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1329 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1330 						MLX4_QP_ST_MLX : -1);
1331 	case MLX4_IB_QPT_PROXY_SMI:
1332 	case MLX4_IB_QPT_TUN_SMI:
1333 	case MLX4_IB_QPT_PROXY_GSI:
1334 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1335 						MLX4_QP_ST_UD : -1);
1336 	default:			return -1;
1337 	}
1338 }
1339 
1340 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1341 				   int attr_mask)
1342 {
1343 	u8 dest_rd_atomic;
1344 	u32 access_flags;
1345 	u32 hw_access_flags = 0;
1346 
1347 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1348 		dest_rd_atomic = attr->max_dest_rd_atomic;
1349 	else
1350 		dest_rd_atomic = qp->resp_depth;
1351 
1352 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1353 		access_flags = attr->qp_access_flags;
1354 	else
1355 		access_flags = qp->atomic_rd_en;
1356 
1357 	if (!dest_rd_atomic)
1358 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1359 
1360 	if (access_flags & IB_ACCESS_REMOTE_READ)
1361 		hw_access_flags |= MLX4_QP_BIT_RRE;
1362 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1363 		hw_access_flags |= MLX4_QP_BIT_RAE;
1364 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1365 		hw_access_flags |= MLX4_QP_BIT_RWE;
1366 
1367 	return cpu_to_be32(hw_access_flags);
1368 }
1369 
1370 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1371 			    int attr_mask)
1372 {
1373 	if (attr_mask & IB_QP_PKEY_INDEX)
1374 		sqp->pkey_index = attr->pkey_index;
1375 	if (attr_mask & IB_QP_QKEY)
1376 		sqp->qkey = attr->qkey;
1377 	if (attr_mask & IB_QP_SQ_PSN)
1378 		sqp->send_psn = attr->sq_psn;
1379 }
1380 
1381 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1382 {
1383 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1384 }
1385 
1386 static int _mlx4_set_path(struct mlx4_ib_dev *dev,
1387 			  const struct rdma_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 vidx;
1392 	int smac_index;
1393 	int err;
1394 
1395 	path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
1396 	path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
1397 	if (rdma_ah_get_static_rate(ah)) {
1398 		path->static_rate = rdma_ah_get_static_rate(ah) +
1399 				    MLX4_STAT_RATE_OFFSET;
1400 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1401 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1402 			--path->static_rate;
1403 	} else
1404 		path->static_rate = 0;
1405 
1406 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
1407 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
1408 		int real_sgid_index =
1409 			mlx4_ib_gid_index_to_real_index(dev, port,
1410 							grh->sgid_index);
1411 
1412 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1413 			pr_err("sgid_index (%u) too large. max is %d\n",
1414 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1415 			return -1;
1416 		}
1417 
1418 		path->grh_mylmc |= 1 << 7;
1419 		path->mgid_index = real_sgid_index;
1420 		path->hop_limit  = grh->hop_limit;
1421 		path->tclass_flowlabel =
1422 			cpu_to_be32((grh->traffic_class << 20) |
1423 				    (grh->flow_label));
1424 		memcpy(path->rgid, grh->dgid.raw, 16);
1425 	}
1426 
1427 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
1428 		if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
1429 			return -1;
1430 
1431 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1432 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
1433 
1434 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1435 		if (vlan_tag < 0x1000) {
1436 			if (smac_info->vid < 0x1000) {
1437 				/* both valid vlan ids */
1438 				if (smac_info->vid != vlan_tag) {
1439 					/* different VIDs.  unreg old and reg new */
1440 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1441 					if (err)
1442 						return err;
1443 					smac_info->candidate_vid = vlan_tag;
1444 					smac_info->candidate_vlan_index = vidx;
1445 					smac_info->candidate_vlan_port = port;
1446 					smac_info->update_vid = 1;
1447 					path->vlan_index = vidx;
1448 				} else {
1449 					path->vlan_index = smac_info->vlan_index;
1450 				}
1451 			} else {
1452 				/* no current vlan tag in qp */
1453 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1454 				if (err)
1455 					return err;
1456 				smac_info->candidate_vid = vlan_tag;
1457 				smac_info->candidate_vlan_index = vidx;
1458 				smac_info->candidate_vlan_port = port;
1459 				smac_info->update_vid = 1;
1460 				path->vlan_index = vidx;
1461 			}
1462 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1463 			path->fl = 1 << 6;
1464 		} else {
1465 			/* have current vlan tag. unregister it at modify-qp success */
1466 			if (smac_info->vid < 0x1000) {
1467 				smac_info->candidate_vid = 0xFFFF;
1468 				smac_info->update_vid = 1;
1469 			}
1470 		}
1471 
1472 		/* get smac_index for RoCE use.
1473 		 * If no smac was yet assigned, register one.
1474 		 * If one was already assigned, but the new mac differs,
1475 		 * unregister the old one and register the new one.
1476 		*/
1477 		if ((!smac_info->smac && !smac_info->smac_port) ||
1478 		    smac_info->smac != smac) {
1479 			/* register candidate now, unreg if needed, after success */
1480 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1481 			if (smac_index >= 0) {
1482 				smac_info->candidate_smac_index = smac_index;
1483 				smac_info->candidate_smac = smac;
1484 				smac_info->candidate_smac_port = port;
1485 			} else {
1486 				return -EINVAL;
1487 			}
1488 		} else {
1489 			smac_index = smac_info->smac_index;
1490 		}
1491 		memcpy(path->dmac, ah->roce.dmac, 6);
1492 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1493 		/* put MAC table smac index for IBoE */
1494 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1495 	} else {
1496 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1497 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
1498 	}
1499 
1500 	return 0;
1501 }
1502 
1503 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1504 			 enum ib_qp_attr_mask qp_attr_mask,
1505 			 struct mlx4_ib_qp *mqp,
1506 			 struct mlx4_qp_path *path, u8 port,
1507 			 u16 vlan_id, u8 *smac)
1508 {
1509 	return _mlx4_set_path(dev, &qp->ah_attr,
1510 			      mlx4_mac_to_u64(smac),
1511 			      vlan_id,
1512 			      path, &mqp->pri, port);
1513 }
1514 
1515 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1516 			     const struct ib_qp_attr *qp,
1517 			     enum ib_qp_attr_mask qp_attr_mask,
1518 			     struct mlx4_ib_qp *mqp,
1519 			     struct mlx4_qp_path *path, u8 port)
1520 {
1521 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1522 			      0,
1523 			      0xffff,
1524 			      path, &mqp->alt, port);
1525 }
1526 
1527 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1528 {
1529 	struct mlx4_ib_gid_entry *ge, *tmp;
1530 
1531 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1532 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1533 			ge->added = 1;
1534 			ge->port = qp->port;
1535 		}
1536 	}
1537 }
1538 
1539 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1540 				    struct mlx4_ib_qp *qp,
1541 				    struct mlx4_qp_context *context)
1542 {
1543 	u64 u64_mac;
1544 	int smac_index;
1545 
1546 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1547 
1548 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1549 	if (!qp->pri.smac && !qp->pri.smac_port) {
1550 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1551 		if (smac_index >= 0) {
1552 			qp->pri.candidate_smac_index = smac_index;
1553 			qp->pri.candidate_smac = u64_mac;
1554 			qp->pri.candidate_smac_port = qp->port;
1555 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1556 		} else {
1557 			return -ENOENT;
1558 		}
1559 	}
1560 	return 0;
1561 }
1562 
1563 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1564 {
1565 	struct counter_index *new_counter_index;
1566 	int err;
1567 	u32 tmp_idx;
1568 
1569 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1570 	    IB_LINK_LAYER_ETHERNET ||
1571 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1572 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1573 		return 0;
1574 
1575 	err = mlx4_counter_alloc(dev->dev, &tmp_idx);
1576 	if (err)
1577 		return err;
1578 
1579 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1580 	if (!new_counter_index) {
1581 		mlx4_counter_free(dev->dev, tmp_idx);
1582 		return -ENOMEM;
1583 	}
1584 
1585 	new_counter_index->index = tmp_idx;
1586 	new_counter_index->allocated = 1;
1587 	qp->counter_index = new_counter_index;
1588 
1589 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
1590 	list_add_tail(&new_counter_index->list,
1591 		      &dev->counters_table[qp->port - 1].counters_list);
1592 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
1593 
1594 	return 0;
1595 }
1596 
1597 enum {
1598 	MLX4_QPC_ROCE_MODE_1 = 0,
1599 	MLX4_QPC_ROCE_MODE_2 = 2,
1600 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
1601 };
1602 
1603 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
1604 {
1605 	switch (gid_type) {
1606 	case IB_GID_TYPE_ROCE:
1607 		return MLX4_QPC_ROCE_MODE_1;
1608 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
1609 		return MLX4_QPC_ROCE_MODE_2;
1610 	default:
1611 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
1612 	}
1613 }
1614 
1615 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1616 			       const struct ib_qp_attr *attr, int attr_mask,
1617 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1618 {
1619 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1620 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1621 	struct mlx4_ib_pd *pd;
1622 	struct mlx4_ib_cq *send_cq, *recv_cq;
1623 	struct mlx4_qp_context *context;
1624 	enum mlx4_qp_optpar optpar = 0;
1625 	int sqd_event;
1626 	int steer_qp = 0;
1627 	int err = -EINVAL;
1628 	int counter_index;
1629 
1630 	/* APM is not supported under RoCE */
1631 	if (attr_mask & IB_QP_ALT_PATH &&
1632 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1633 	    IB_LINK_LAYER_ETHERNET)
1634 		return -ENOTSUPP;
1635 
1636 	context = kzalloc(sizeof *context, GFP_KERNEL);
1637 	if (!context)
1638 		return -ENOMEM;
1639 
1640 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1641 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1642 
1643 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1644 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1645 	else {
1646 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
1647 		switch (attr->path_mig_state) {
1648 		case IB_MIG_MIGRATED:
1649 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1650 			break;
1651 		case IB_MIG_REARM:
1652 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1653 			break;
1654 		case IB_MIG_ARMED:
1655 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1656 			break;
1657 		}
1658 	}
1659 
1660 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1661 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1662 	else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1663 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1664 	else if (ibqp->qp_type == IB_QPT_UD) {
1665 		if (qp->flags & MLX4_IB_QP_LSO)
1666 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
1667 					      ilog2(dev->dev->caps.max_gso_sz);
1668 		else
1669 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1670 	} else if (attr_mask & IB_QP_PATH_MTU) {
1671 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1672 			pr_err("path MTU (%u) is invalid\n",
1673 			       attr->path_mtu);
1674 			goto out;
1675 		}
1676 		context->mtu_msgmax = (attr->path_mtu << 5) |
1677 			ilog2(dev->dev->caps.max_msg_sz);
1678 	}
1679 
1680 	if (qp->rq.wqe_cnt)
1681 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1682 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
1683 
1684 	if (qp->sq.wqe_cnt)
1685 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1686 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
1687 
1688 	if (new_state == IB_QPS_RESET && qp->counter_index)
1689 		mlx4_ib_free_qp_counter(dev, qp);
1690 
1691 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1692 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1693 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1694 		if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1695 			context->param3 |= cpu_to_be32(1 << 30);
1696 	}
1697 
1698 	if (qp->ibqp.uobject)
1699 		context->usr_page = cpu_to_be32(
1700 			mlx4_to_hw_uar_index(dev->dev,
1701 					     to_mucontext(ibqp->uobject->context)->uar.index));
1702 	else
1703 		context->usr_page = cpu_to_be32(
1704 			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
1705 
1706 	if (attr_mask & IB_QP_DEST_QPN)
1707 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1708 
1709 	if (attr_mask & IB_QP_PORT) {
1710 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1711 		    !(attr_mask & IB_QP_AV)) {
1712 			mlx4_set_sched(&context->pri_path, attr->port_num);
1713 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1714 		}
1715 	}
1716 
1717 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1718 		err = create_qp_lb_counter(dev, qp);
1719 		if (err)
1720 			goto out;
1721 
1722 		counter_index =
1723 			dev->counters_table[qp->port - 1].default_counter;
1724 		if (qp->counter_index)
1725 			counter_index = qp->counter_index->index;
1726 
1727 		if (counter_index != -1) {
1728 			context->pri_path.counter_index = counter_index;
1729 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1730 			if (qp->counter_index) {
1731 				context->pri_path.fl |=
1732 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
1733 				context->pri_path.vlan_control |=
1734 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
1735 			}
1736 		} else
1737 			context->pri_path.counter_index =
1738 				MLX4_SINK_COUNTER_INDEX(dev->dev);
1739 
1740 		if (qp->flags & MLX4_IB_QP_NETIF) {
1741 			mlx4_ib_steer_qp_reg(dev, qp, 1);
1742 			steer_qp = 1;
1743 		}
1744 
1745 		if (ibqp->qp_type == IB_QPT_GSI) {
1746 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
1747 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
1748 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
1749 
1750 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
1751 		}
1752 	}
1753 
1754 	if (attr_mask & IB_QP_PKEY_INDEX) {
1755 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1756 			context->pri_path.disable_pkey_check = 0x40;
1757 		context->pri_path.pkey_index = attr->pkey_index;
1758 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1759 	}
1760 
1761 	if (attr_mask & IB_QP_AV) {
1762 		u8 port_num = mlx4_is_bonded(to_mdev(ibqp->device)->dev) ? 1 :
1763 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1764 		union ib_gid gid;
1765 		struct ib_gid_attr gid_attr = {.gid_type = IB_GID_TYPE_IB};
1766 		u16 vlan = 0xffff;
1767 		u8 smac[ETH_ALEN];
1768 		int status = 0;
1769 		int is_eth =
1770 			rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
1771 			rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
1772 
1773 		if (is_eth) {
1774 			int index =
1775 				rdma_ah_read_grh(&attr->ah_attr)->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 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
2425 				int index, union ib_gid *gid,
2426 				enum ib_gid_type *gid_type)
2427 {
2428 	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
2429 	struct mlx4_port_gid_table *port_gid_table;
2430 	unsigned long flags;
2431 
2432 	port_gid_table = &iboe->gids[port_num - 1];
2433 	spin_lock_irqsave(&iboe->lock, flags);
2434 	memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
2435 	*gid_type = port_gid_table->gids[index].gid_type;
2436 	spin_unlock_irqrestore(&iboe->lock, flags);
2437 	if (!memcmp(gid, &zgid, sizeof(*gid)))
2438 		return -ENOENT;
2439 
2440 	return 0;
2441 }
2442 
2443 #define MLX4_ROCEV2_QP1_SPORT 0xC000
2444 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_ud_wr *wr,
2445 			    void *wqe, unsigned *mlx_seg_len)
2446 {
2447 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
2448 	struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
2449 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2450 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2451 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2452 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2453 	union ib_gid sgid;
2454 	u16 pkey;
2455 	int send_size;
2456 	int header_size;
2457 	int spc;
2458 	int i;
2459 	int err = 0;
2460 	u16 vlan = 0xffff;
2461 	bool is_eth;
2462 	bool is_vlan = false;
2463 	bool is_grh;
2464 	bool is_udp = false;
2465 	int ip_version = 0;
2466 
2467 	send_size = 0;
2468 	for (i = 0; i < wr->wr.num_sge; ++i)
2469 		send_size += wr->wr.sg_list[i].length;
2470 
2471 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2472 	is_grh = mlx4_ib_ah_grh_present(ah);
2473 	if (is_eth) {
2474 		enum ib_gid_type gid_type;
2475 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2476 			/* When multi-function is enabled, the ib_core gid
2477 			 * indexes don't necessarily match the hw ones, so
2478 			 * we must use our own cache */
2479 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2480 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
2481 							   ah->av.ib.gid_index, &sgid.raw[0]);
2482 			if (err)
2483 				return err;
2484 		} else  {
2485 			err = fill_gid_by_hw_index(ibdev, sqp->qp.port,
2486 					    ah->av.ib.gid_index,
2487 					    &sgid, &gid_type);
2488 			if (!err) {
2489 				is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
2490 				if (is_udp) {
2491 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
2492 						ip_version = 4;
2493 					else
2494 						ip_version = 6;
2495 					is_grh = false;
2496 				}
2497 			} else {
2498 				return err;
2499 			}
2500 		}
2501 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
2502 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
2503 			is_vlan = 1;
2504 		}
2505 	}
2506 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
2507 			  ip_version, is_udp, 0, &sqp->ud_header);
2508 	if (err)
2509 		return err;
2510 
2511 	if (!is_eth) {
2512 		sqp->ud_header.lrh.service_level =
2513 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2514 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
2515 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2516 	}
2517 
2518 	if (is_grh || (ip_version == 6)) {
2519 		sqp->ud_header.grh.traffic_class =
2520 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2521 		sqp->ud_header.grh.flow_label    =
2522 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
2523 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
2524 		if (is_eth) {
2525 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
2526 		} else {
2527 			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2528 				/* When multi-function is enabled, the ib_core gid
2529 				 * indexes don't necessarily match the hw ones, so
2530 				 * we must use our own cache
2531 				 */
2532 				sqp->ud_header.grh.source_gid.global.subnet_prefix =
2533 					cpu_to_be64(atomic64_read(&(to_mdev(ib_dev)->sriov.
2534 								    demux[sqp->qp.port - 1].
2535 								    subnet_prefix)));
2536 				sqp->ud_header.grh.source_gid.global.interface_id =
2537 					to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2538 						       guid_cache[ah->av.ib.gid_index];
2539 			} else {
2540 				ib_get_cached_gid(ib_dev,
2541 						  be32_to_cpu(ah->av.ib.port_pd) >> 24,
2542 						  ah->av.ib.gid_index,
2543 						  &sqp->ud_header.grh.source_gid, NULL);
2544 			}
2545 		}
2546 		memcpy(sqp->ud_header.grh.destination_gid.raw,
2547 		       ah->av.ib.dgid, 16);
2548 	}
2549 
2550 	if (ip_version == 4) {
2551 		sqp->ud_header.ip4.tos =
2552 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2553 		sqp->ud_header.ip4.id = 0;
2554 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
2555 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
2556 
2557 		memcpy(&sqp->ud_header.ip4.saddr,
2558 		       sgid.raw + 12, 4);
2559 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
2560 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
2561 	}
2562 
2563 	if (is_udp) {
2564 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
2565 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
2566 		sqp->ud_header.udp.csum = 0;
2567 	}
2568 
2569 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2570 
2571 	if (!is_eth) {
2572 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
2573 					  (sqp->ud_header.lrh.destination_lid ==
2574 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
2575 					  (sqp->ud_header.lrh.service_level << 8));
2576 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
2577 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
2578 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
2579 	}
2580 
2581 	switch (wr->wr.opcode) {
2582 	case IB_WR_SEND:
2583 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
2584 		sqp->ud_header.immediate_present = 0;
2585 		break;
2586 	case IB_WR_SEND_WITH_IMM:
2587 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2588 		sqp->ud_header.immediate_present = 1;
2589 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
2590 		break;
2591 	default:
2592 		return -EINVAL;
2593 	}
2594 
2595 	if (is_eth) {
2596 		struct in6_addr in6;
2597 		u16 ether_type;
2598 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
2599 
2600 		ether_type = (!is_udp) ? ETH_P_IBOE:
2601 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
2602 
2603 		mlx->sched_prio = cpu_to_be16(pcp);
2604 
2605 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
2606 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
2607 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
2608 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
2609 		memcpy(&in6, sgid.raw, sizeof(in6));
2610 
2611 
2612 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
2613 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
2614 		if (!is_vlan) {
2615 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
2616 		} else {
2617 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
2618 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
2619 		}
2620 	} else {
2621 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 :
2622 							sl_to_vl(to_mdev(ib_dev),
2623 								 sqp->ud_header.lrh.service_level,
2624 								 sqp->qp.port);
2625 		if (sqp->qp.ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
2626 			return -EINVAL;
2627 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
2628 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
2629 	}
2630 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2631 	if (!sqp->qp.ibqp.qp_num)
2632 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
2633 	else
2634 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
2635 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2636 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2637 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2638 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
2639 					       sqp->qkey : wr->remote_qkey);
2640 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
2641 
2642 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2643 
2644 	if (0) {
2645 		pr_err("built UD header of size %d:\n", header_size);
2646 		for (i = 0; i < header_size / 4; ++i) {
2647 			if (i % 8 == 0)
2648 				pr_err("  [%02x] ", i * 4);
2649 			pr_cont(" %08x",
2650 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
2651 			if ((i + 1) % 8 == 0)
2652 				pr_cont("\n");
2653 		}
2654 		pr_err("\n");
2655 	}
2656 
2657 	/*
2658 	 * Inline data segments may not cross a 64 byte boundary.  If
2659 	 * our UD header is bigger than the space available up to the
2660 	 * next 64 byte boundary in the WQE, use two inline data
2661 	 * segments to hold the UD header.
2662 	 */
2663 	spc = MLX4_INLINE_ALIGN -
2664 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2665 	if (header_size <= spc) {
2666 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2667 		memcpy(inl + 1, sqp->header_buf, header_size);
2668 		i = 1;
2669 	} else {
2670 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2671 		memcpy(inl + 1, sqp->header_buf, spc);
2672 
2673 		inl = (void *) (inl + 1) + spc;
2674 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2675 		/*
2676 		 * Need a barrier here to make sure all the data is
2677 		 * visible before the byte_count field is set.
2678 		 * Otherwise the HCA prefetcher could grab the 64-byte
2679 		 * chunk with this inline segment and get a valid (!=
2680 		 * 0xffffffff) byte count but stale data, and end up
2681 		 * generating a packet with bad headers.
2682 		 *
2683 		 * The first inline segment's byte_count field doesn't
2684 		 * need a barrier, because it comes after a
2685 		 * control/MLX segment and therefore is at an offset
2686 		 * of 16 mod 64.
2687 		 */
2688 		wmb();
2689 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2690 		i = 2;
2691 	}
2692 
2693 	*mlx_seg_len =
2694 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2695 	return 0;
2696 }
2697 
2698 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2699 {
2700 	unsigned cur;
2701 	struct mlx4_ib_cq *cq;
2702 
2703 	cur = wq->head - wq->tail;
2704 	if (likely(cur + nreq < wq->max_post))
2705 		return 0;
2706 
2707 	cq = to_mcq(ib_cq);
2708 	spin_lock(&cq->lock);
2709 	cur = wq->head - wq->tail;
2710 	spin_unlock(&cq->lock);
2711 
2712 	return cur + nreq >= wq->max_post;
2713 }
2714 
2715 static __be32 convert_access(int acc)
2716 {
2717 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
2718 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
2719 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
2720 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
2721 	       (acc & IB_ACCESS_REMOTE_READ   ?
2722 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
2723 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
2724 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
2725 }
2726 
2727 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
2728 			struct ib_reg_wr *wr)
2729 {
2730 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
2731 
2732 	fseg->flags		= convert_access(wr->access);
2733 	fseg->mem_key		= cpu_to_be32(wr->key);
2734 	fseg->buf_list		= cpu_to_be64(mr->page_map);
2735 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
2736 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
2737 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
2738 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
2739 	fseg->reserved[0]	= 0;
2740 	fseg->reserved[1]	= 0;
2741 }
2742 
2743 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2744 {
2745 	memset(iseg, 0, sizeof(*iseg));
2746 	iseg->mem_key = cpu_to_be32(rkey);
2747 }
2748 
2749 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2750 					  u64 remote_addr, u32 rkey)
2751 {
2752 	rseg->raddr    = cpu_to_be64(remote_addr);
2753 	rseg->rkey     = cpu_to_be32(rkey);
2754 	rseg->reserved = 0;
2755 }
2756 
2757 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
2758 		struct ib_atomic_wr *wr)
2759 {
2760 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2761 		aseg->swap_add = cpu_to_be64(wr->swap);
2762 		aseg->compare  = cpu_to_be64(wr->compare_add);
2763 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2764 		aseg->swap_add = cpu_to_be64(wr->compare_add);
2765 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
2766 	} else {
2767 		aseg->swap_add = cpu_to_be64(wr->compare_add);
2768 		aseg->compare  = 0;
2769 	}
2770 
2771 }
2772 
2773 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2774 				  struct ib_atomic_wr *wr)
2775 {
2776 	aseg->swap_add		= cpu_to_be64(wr->swap);
2777 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
2778 	aseg->compare		= cpu_to_be64(wr->compare_add);
2779 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
2780 }
2781 
2782 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2783 			     struct ib_ud_wr *wr)
2784 {
2785 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
2786 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
2787 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
2788 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
2789 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
2790 }
2791 
2792 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2793 				    struct mlx4_wqe_datagram_seg *dseg,
2794 				    struct ib_ud_wr *wr,
2795 				    enum mlx4_ib_qp_type qpt)
2796 {
2797 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
2798 	struct mlx4_av sqp_av = {0};
2799 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
2800 
2801 	/* force loopback */
2802 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2803 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2804 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2805 			cpu_to_be32(0xf0000000);
2806 
2807 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2808 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
2809 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2810 	else
2811 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp0_tunnel[port - 1]);
2812 	/* Use QKEY from the QP context, which is set by master */
2813 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2814 }
2815 
2816 static void build_tunnel_header(struct ib_ud_wr *wr, void *wqe, unsigned *mlx_seg_len)
2817 {
2818 	struct mlx4_wqe_inline_seg *inl = wqe;
2819 	struct mlx4_ib_tunnel_header hdr;
2820 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2821 	int spc;
2822 	int i;
2823 
2824 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2825 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
2826 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
2827 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
2828 	memcpy(hdr.mac, ah->av.eth.mac, 6);
2829 	hdr.vlan = ah->av.eth.vlan;
2830 
2831 	spc = MLX4_INLINE_ALIGN -
2832 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2833 	if (sizeof (hdr) <= spc) {
2834 		memcpy(inl + 1, &hdr, sizeof (hdr));
2835 		wmb();
2836 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
2837 		i = 1;
2838 	} else {
2839 		memcpy(inl + 1, &hdr, spc);
2840 		wmb();
2841 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2842 
2843 		inl = (void *) (inl + 1) + spc;
2844 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2845 		wmb();
2846 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
2847 		i = 2;
2848 	}
2849 
2850 	*mlx_seg_len =
2851 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2852 }
2853 
2854 static void set_mlx_icrc_seg(void *dseg)
2855 {
2856 	u32 *t = dseg;
2857 	struct mlx4_wqe_inline_seg *iseg = dseg;
2858 
2859 	t[1] = 0;
2860 
2861 	/*
2862 	 * Need a barrier here before writing the byte_count field to
2863 	 * make sure that all the data is visible before the
2864 	 * byte_count field is set.  Otherwise, if the segment begins
2865 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2866 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2867 	 * stale data, and end up sending the wrong data.
2868 	 */
2869 	wmb();
2870 
2871 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2872 }
2873 
2874 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2875 {
2876 	dseg->lkey       = cpu_to_be32(sg->lkey);
2877 	dseg->addr       = cpu_to_be64(sg->addr);
2878 
2879 	/*
2880 	 * Need a barrier here before writing the byte_count field to
2881 	 * make sure that all the data is visible before the
2882 	 * byte_count field is set.  Otherwise, if the segment begins
2883 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2884 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2885 	 * stale data, and end up sending the wrong data.
2886 	 */
2887 	wmb();
2888 
2889 	dseg->byte_count = cpu_to_be32(sg->length);
2890 }
2891 
2892 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2893 {
2894 	dseg->byte_count = cpu_to_be32(sg->length);
2895 	dseg->lkey       = cpu_to_be32(sg->lkey);
2896 	dseg->addr       = cpu_to_be64(sg->addr);
2897 }
2898 
2899 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_ud_wr *wr,
2900 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2901 			 __be32 *lso_hdr_sz, __be32 *blh)
2902 {
2903 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
2904 
2905 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2906 		*blh = cpu_to_be32(1 << 6);
2907 
2908 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2909 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
2910 		return -EINVAL;
2911 
2912 	memcpy(wqe->header, wr->header, wr->hlen);
2913 
2914 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
2915 	*lso_seg_len = halign;
2916 	return 0;
2917 }
2918 
2919 static __be32 send_ieth(struct ib_send_wr *wr)
2920 {
2921 	switch (wr->opcode) {
2922 	case IB_WR_SEND_WITH_IMM:
2923 	case IB_WR_RDMA_WRITE_WITH_IMM:
2924 		return wr->ex.imm_data;
2925 
2926 	case IB_WR_SEND_WITH_INV:
2927 		return cpu_to_be32(wr->ex.invalidate_rkey);
2928 
2929 	default:
2930 		return 0;
2931 	}
2932 }
2933 
2934 static void add_zero_len_inline(void *wqe)
2935 {
2936 	struct mlx4_wqe_inline_seg *inl = wqe;
2937 	memset(wqe, 0, 16);
2938 	inl->byte_count = cpu_to_be32(1 << 31);
2939 }
2940 
2941 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2942 		      struct ib_send_wr **bad_wr)
2943 {
2944 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2945 	void *wqe;
2946 	struct mlx4_wqe_ctrl_seg *ctrl;
2947 	struct mlx4_wqe_data_seg *dseg;
2948 	unsigned long flags;
2949 	int nreq;
2950 	int err = 0;
2951 	unsigned ind;
2952 	int uninitialized_var(stamp);
2953 	int uninitialized_var(size);
2954 	unsigned uninitialized_var(seglen);
2955 	__be32 dummy;
2956 	__be32 *lso_wqe;
2957 	__be32 uninitialized_var(lso_hdr_sz);
2958 	__be32 blh;
2959 	int i;
2960 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
2961 
2962 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2963 		struct mlx4_ib_sqp *sqp = to_msqp(qp);
2964 
2965 		if (sqp->roce_v2_gsi) {
2966 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
2967 			enum ib_gid_type gid_type;
2968 			union ib_gid gid;
2969 
2970 			if (!fill_gid_by_hw_index(mdev, sqp->qp.port,
2971 					   ah->av.ib.gid_index,
2972 					   &gid, &gid_type))
2973 				qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
2974 						to_mqp(sqp->roce_v2_gsi) : qp;
2975 			else
2976 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
2977 				       ah->av.ib.gid_index);
2978 		}
2979 	}
2980 
2981 	spin_lock_irqsave(&qp->sq.lock, flags);
2982 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
2983 		err = -EIO;
2984 		*bad_wr = wr;
2985 		nreq = 0;
2986 		goto out;
2987 	}
2988 
2989 	ind = qp->sq_next_wqe;
2990 
2991 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2992 		lso_wqe = &dummy;
2993 		blh = 0;
2994 
2995 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2996 			err = -ENOMEM;
2997 			*bad_wr = wr;
2998 			goto out;
2999 		}
3000 
3001 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3002 			err = -EINVAL;
3003 			*bad_wr = wr;
3004 			goto out;
3005 		}
3006 
3007 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3008 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3009 
3010 		ctrl->srcrb_flags =
3011 			(wr->send_flags & IB_SEND_SIGNALED ?
3012 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3013 			(wr->send_flags & IB_SEND_SOLICITED ?
3014 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3015 			((wr->send_flags & IB_SEND_IP_CSUM) ?
3016 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3017 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3018 			qp->sq_signal_bits;
3019 
3020 		ctrl->imm = send_ieth(wr);
3021 
3022 		wqe += sizeof *ctrl;
3023 		size = sizeof *ctrl / 16;
3024 
3025 		switch (qp->mlx4_ib_qp_type) {
3026 		case MLX4_IB_QPT_RC:
3027 		case MLX4_IB_QPT_UC:
3028 			switch (wr->opcode) {
3029 			case IB_WR_ATOMIC_CMP_AND_SWP:
3030 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3031 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3032 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3033 					      atomic_wr(wr)->rkey);
3034 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3035 
3036 				set_atomic_seg(wqe, atomic_wr(wr));
3037 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
3038 
3039 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3040 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3041 
3042 				break;
3043 
3044 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3045 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3046 					      atomic_wr(wr)->rkey);
3047 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3048 
3049 				set_masked_atomic_seg(wqe, atomic_wr(wr));
3050 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
3051 
3052 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3053 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3054 
3055 				break;
3056 
3057 			case IB_WR_RDMA_READ:
3058 			case IB_WR_RDMA_WRITE:
3059 			case IB_WR_RDMA_WRITE_WITH_IMM:
3060 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3061 					      rdma_wr(wr)->rkey);
3062 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3063 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3064 				break;
3065 
3066 			case IB_WR_LOCAL_INV:
3067 				ctrl->srcrb_flags |=
3068 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3069 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3070 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3071 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3072 				break;
3073 
3074 			case IB_WR_REG_MR:
3075 				ctrl->srcrb_flags |=
3076 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3077 				set_reg_seg(wqe, reg_wr(wr));
3078 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3079 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3080 				break;
3081 
3082 			default:
3083 				/* No extra segments required for sends */
3084 				break;
3085 			}
3086 			break;
3087 
3088 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3089 			err =  build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3090 					ctrl, &seglen);
3091 			if (unlikely(err)) {
3092 				*bad_wr = wr;
3093 				goto out;
3094 			}
3095 			wqe  += seglen;
3096 			size += seglen / 16;
3097 			break;
3098 		case MLX4_IB_QPT_TUN_SMI:
3099 		case MLX4_IB_QPT_TUN_GSI:
3100 			/* this is a UD qp used in MAD responses to slaves. */
3101 			set_datagram_seg(wqe, ud_wr(wr));
3102 			/* set the forced-loopback bit in the data seg av */
3103 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3104 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3105 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3106 			break;
3107 		case MLX4_IB_QPT_UD:
3108 			set_datagram_seg(wqe, ud_wr(wr));
3109 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3110 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3111 
3112 			if (wr->opcode == IB_WR_LSO) {
3113 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3114 						&lso_hdr_sz, &blh);
3115 				if (unlikely(err)) {
3116 					*bad_wr = wr;
3117 					goto out;
3118 				}
3119 				lso_wqe = (__be32 *) wqe;
3120 				wqe  += seglen;
3121 				size += seglen / 16;
3122 			}
3123 			break;
3124 
3125 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3126 			err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3127 					ctrl, &seglen);
3128 			if (unlikely(err)) {
3129 				*bad_wr = wr;
3130 				goto out;
3131 			}
3132 			wqe  += seglen;
3133 			size += seglen / 16;
3134 			/* to start tunnel header on a cache-line boundary */
3135 			add_zero_len_inline(wqe);
3136 			wqe += 16;
3137 			size++;
3138 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3139 			wqe  += seglen;
3140 			size += seglen / 16;
3141 			break;
3142 		case MLX4_IB_QPT_PROXY_SMI:
3143 		case MLX4_IB_QPT_PROXY_GSI:
3144 			/* If we are tunneling special qps, this is a UD qp.
3145 			 * In this case we first add a UD segment targeting
3146 			 * the tunnel qp, and then add a header with address
3147 			 * information */
3148 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3149 						ud_wr(wr),
3150 						qp->mlx4_ib_qp_type);
3151 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3152 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3153 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3154 			wqe  += seglen;
3155 			size += seglen / 16;
3156 			break;
3157 
3158 		case MLX4_IB_QPT_SMI:
3159 		case MLX4_IB_QPT_GSI:
3160 			err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
3161 					&seglen);
3162 			if (unlikely(err)) {
3163 				*bad_wr = wr;
3164 				goto out;
3165 			}
3166 			wqe  += seglen;
3167 			size += seglen / 16;
3168 			break;
3169 
3170 		default:
3171 			break;
3172 		}
3173 
3174 		/*
3175 		 * Write data segments in reverse order, so as to
3176 		 * overwrite cacheline stamp last within each
3177 		 * cacheline.  This avoids issues with WQE
3178 		 * prefetching.
3179 		 */
3180 
3181 		dseg = wqe;
3182 		dseg += wr->num_sge - 1;
3183 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3184 
3185 		/* Add one more inline data segment for ICRC for MLX sends */
3186 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3187 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3188 			     qp->mlx4_ib_qp_type &
3189 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3190 			set_mlx_icrc_seg(dseg + 1);
3191 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3192 		}
3193 
3194 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3195 			set_data_seg(dseg, wr->sg_list + i);
3196 
3197 		/*
3198 		 * Possibly overwrite stamping in cacheline with LSO
3199 		 * segment only after making sure all data segments
3200 		 * are written.
3201 		 */
3202 		wmb();
3203 		*lso_wqe = lso_hdr_sz;
3204 
3205 		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3206 					     MLX4_WQE_CTRL_FENCE : 0) | size;
3207 
3208 		/*
3209 		 * Make sure descriptor is fully written before
3210 		 * setting ownership bit (because HW can start
3211 		 * executing as soon as we do).
3212 		 */
3213 		wmb();
3214 
3215 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3216 			*bad_wr = wr;
3217 			err = -EINVAL;
3218 			goto out;
3219 		}
3220 
3221 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3222 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3223 
3224 		stamp = ind + qp->sq_spare_wqes;
3225 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
3226 
3227 		/*
3228 		 * We can improve latency by not stamping the last
3229 		 * send queue WQE until after ringing the doorbell, so
3230 		 * only stamp here if there are still more WQEs to post.
3231 		 *
3232 		 * Same optimization applies to padding with NOP wqe
3233 		 * in case of WQE shrinking (used to prevent wrap-around
3234 		 * in the middle of WR).
3235 		 */
3236 		if (wr->next) {
3237 			stamp_send_wqe(qp, stamp, size * 16);
3238 			ind = pad_wraparound(qp, ind);
3239 		}
3240 	}
3241 
3242 out:
3243 	if (likely(nreq)) {
3244 		qp->sq.head += nreq;
3245 
3246 		/*
3247 		 * Make sure that descriptors are written before
3248 		 * doorbell record.
3249 		 */
3250 		wmb();
3251 
3252 		writel(qp->doorbell_qpn,
3253 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3254 
3255 		/*
3256 		 * Make sure doorbells don't leak out of SQ spinlock
3257 		 * and reach the HCA out of order.
3258 		 */
3259 		mmiowb();
3260 
3261 		stamp_send_wqe(qp, stamp, size * 16);
3262 
3263 		ind = pad_wraparound(qp, ind);
3264 		qp->sq_next_wqe = ind;
3265 	}
3266 
3267 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3268 
3269 	return err;
3270 }
3271 
3272 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3273 		      struct ib_recv_wr **bad_wr)
3274 {
3275 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3276 	struct mlx4_wqe_data_seg *scat;
3277 	unsigned long flags;
3278 	int err = 0;
3279 	int nreq;
3280 	int ind;
3281 	int max_gs;
3282 	int i;
3283 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3284 
3285 	max_gs = qp->rq.max_gs;
3286 	spin_lock_irqsave(&qp->rq.lock, flags);
3287 
3288 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3289 		err = -EIO;
3290 		*bad_wr = wr;
3291 		nreq = 0;
3292 		goto out;
3293 	}
3294 
3295 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3296 
3297 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3298 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3299 			err = -ENOMEM;
3300 			*bad_wr = wr;
3301 			goto out;
3302 		}
3303 
3304 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3305 			err = -EINVAL;
3306 			*bad_wr = wr;
3307 			goto out;
3308 		}
3309 
3310 		scat = get_recv_wqe(qp, ind);
3311 
3312 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3313 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3314 			ib_dma_sync_single_for_device(ibqp->device,
3315 						      qp->sqp_proxy_rcv[ind].map,
3316 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3317 						      DMA_FROM_DEVICE);
3318 			scat->byte_count =
3319 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3320 			/* use dma lkey from upper layer entry */
3321 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3322 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3323 			scat++;
3324 			max_gs--;
3325 		}
3326 
3327 		for (i = 0; i < wr->num_sge; ++i)
3328 			__set_data_seg(scat + i, wr->sg_list + i);
3329 
3330 		if (i < max_gs) {
3331 			scat[i].byte_count = 0;
3332 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3333 			scat[i].addr       = 0;
3334 		}
3335 
3336 		qp->rq.wrid[ind] = wr->wr_id;
3337 
3338 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3339 	}
3340 
3341 out:
3342 	if (likely(nreq)) {
3343 		qp->rq.head += nreq;
3344 
3345 		/*
3346 		 * Make sure that descriptors are written before
3347 		 * doorbell record.
3348 		 */
3349 		wmb();
3350 
3351 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3352 	}
3353 
3354 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3355 
3356 	return err;
3357 }
3358 
3359 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3360 {
3361 	switch (mlx4_state) {
3362 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3363 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3364 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3365 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3366 	case MLX4_QP_STATE_SQ_DRAINING:
3367 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3368 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3369 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3370 	default:		     return -1;
3371 	}
3372 }
3373 
3374 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3375 {
3376 	switch (mlx4_mig_state) {
3377 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3378 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3379 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3380 	default: return -1;
3381 	}
3382 }
3383 
3384 static int to_ib_qp_access_flags(int mlx4_flags)
3385 {
3386 	int ib_flags = 0;
3387 
3388 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3389 		ib_flags |= IB_ACCESS_REMOTE_READ;
3390 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3391 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3392 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3393 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3394 
3395 	return ib_flags;
3396 }
3397 
3398 static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
3399 			    struct rdma_ah_attr *ah_attr,
3400 			    struct mlx4_qp_path *path)
3401 {
3402 	struct mlx4_dev *dev = ibdev->dev;
3403 	u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
3404 
3405 	memset(ah_attr, 0, sizeof(*ah_attr));
3406 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
3407 	if (port_num == 0 || port_num > dev->caps.num_ports)
3408 		return;
3409 
3410 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
3411 		rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
3412 			       ((path->sched_queue & 4) << 1));
3413 	else
3414 		rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
3415 	rdma_ah_set_port_num(ah_attr, port_num);
3416 
3417 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
3418 	rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
3419 	rdma_ah_set_static_rate(ah_attr,
3420 				path->static_rate ? path->static_rate - 5 : 0);
3421 	if (path->grh_mylmc & (1 << 7)) {
3422 		rdma_ah_set_grh(ah_attr, NULL,
3423 				be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
3424 				path->mgid_index,
3425 				path->hop_limit,
3426 				(be32_to_cpu(path->tclass_flowlabel)
3427 				 >> 20) & 0xff);
3428 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
3429 	}
3430 }
3431 
3432 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3433 		     struct ib_qp_init_attr *qp_init_attr)
3434 {
3435 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3436 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3437 	struct mlx4_qp_context context;
3438 	int mlx4_state;
3439 	int err = 0;
3440 
3441 	mutex_lock(&qp->mutex);
3442 
3443 	if (qp->state == IB_QPS_RESET) {
3444 		qp_attr->qp_state = IB_QPS_RESET;
3445 		goto done;
3446 	}
3447 
3448 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3449 	if (err) {
3450 		err = -EINVAL;
3451 		goto out;
3452 	}
3453 
3454 	mlx4_state = be32_to_cpu(context.flags) >> 28;
3455 
3456 	qp->state		     = to_ib_qp_state(mlx4_state);
3457 	qp_attr->qp_state	     = qp->state;
3458 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
3459 	qp_attr->path_mig_state	     =
3460 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3461 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
3462 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3463 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
3464 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
3465 	qp_attr->qp_access_flags     =
3466 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
3467 
3468 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3469 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3470 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3471 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3472 		qp_attr->alt_port_num	=
3473 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
3474 	}
3475 
3476 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3477 	if (qp_attr->qp_state == IB_QPS_INIT)
3478 		qp_attr->port_num = qp->port;
3479 	else
3480 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3481 
3482 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3483 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3484 
3485 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3486 
3487 	qp_attr->max_dest_rd_atomic =
3488 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3489 	qp_attr->min_rnr_timer	    =
3490 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3491 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
3492 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
3493 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
3494 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
3495 
3496 done:
3497 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
3498 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
3499 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
3500 
3501 	if (!ibqp->uobject) {
3502 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
3503 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
3504 	} else {
3505 		qp_attr->cap.max_send_wr  = 0;
3506 		qp_attr->cap.max_send_sge = 0;
3507 	}
3508 
3509 	/*
3510 	 * We don't support inline sends for kernel QPs (yet), and we
3511 	 * don't know what userspace's value should be.
3512 	 */
3513 	qp_attr->cap.max_inline_data = 0;
3514 
3515 	qp_init_attr->cap	     = qp_attr->cap;
3516 
3517 	qp_init_attr->create_flags = 0;
3518 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3519 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3520 
3521 	if (qp->flags & MLX4_IB_QP_LSO)
3522 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
3523 
3524 	if (qp->flags & MLX4_IB_QP_NETIF)
3525 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
3526 
3527 	qp_init_attr->sq_sig_type =
3528 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
3529 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3530 
3531 out:
3532 	mutex_unlock(&qp->mutex);
3533 	return err;
3534 }
3535 
3536