xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision afb46f79)
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/slab.h>
36 #include <linux/netdevice.h>
37 
38 #include <rdma/ib_cache.h>
39 #include <rdma/ib_pack.h>
40 #include <rdma/ib_addr.h>
41 #include <rdma/ib_mad.h>
42 
43 #include <linux/mlx4/qp.h>
44 
45 #include "mlx4_ib.h"
46 #include "user.h"
47 
48 enum {
49 	MLX4_IB_ACK_REQ_FREQ	= 8,
50 };
51 
52 enum {
53 	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
54 	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
55 	MLX4_IB_LINK_TYPE_IB		= 0,
56 	MLX4_IB_LINK_TYPE_ETH		= 1
57 };
58 
59 enum {
60 	/*
61 	 * Largest possible UD header: send with GRH and immediate
62 	 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
63 	 * tag.  (LRH would only use 8 bytes, so Ethernet is the
64 	 * biggest case)
65 	 */
66 	MLX4_IB_UD_HEADER_SIZE		= 82,
67 	MLX4_IB_LSO_HEADER_SPARE	= 128,
68 };
69 
70 enum {
71 	MLX4_IB_IBOE_ETHERTYPE		= 0x8915
72 };
73 
74 struct mlx4_ib_sqp {
75 	struct mlx4_ib_qp	qp;
76 	int			pkey_index;
77 	u32			qkey;
78 	u32			send_psn;
79 	struct ib_ud_header	ud_header;
80 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
81 };
82 
83 enum {
84 	MLX4_IB_MIN_SQ_STRIDE	= 6,
85 	MLX4_IB_CACHE_LINE_SIZE	= 64,
86 };
87 
88 enum {
89 	MLX4_RAW_QP_MTU		= 7,
90 	MLX4_RAW_QP_MSGMAX	= 31,
91 };
92 
93 #ifndef ETH_ALEN
94 #define ETH_ALEN        6
95 #endif
96 static inline u64 mlx4_mac_to_u64(u8 *addr)
97 {
98 	u64 mac = 0;
99 	int i;
100 
101 	for (i = 0; i < ETH_ALEN; i++) {
102 		mac <<= 8;
103 		mac |= addr[i];
104 	}
105 	return mac;
106 }
107 
108 static const __be32 mlx4_ib_opcode[] = {
109 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
110 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
111 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
112 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
113 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
114 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
115 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
116 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
117 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
118 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
119 	[IB_WR_FAST_REG_MR]			= cpu_to_be32(MLX4_OPCODE_FMR),
120 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
121 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
122 	[IB_WR_BIND_MW]				= cpu_to_be32(MLX4_OPCODE_BIND_MW),
123 };
124 
125 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
126 {
127 	return container_of(mqp, struct mlx4_ib_sqp, qp);
128 }
129 
130 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
131 {
132 	if (!mlx4_is_master(dev->dev))
133 		return 0;
134 
135 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
136 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
137 		8 * MLX4_MFUNC_MAX;
138 }
139 
140 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
141 {
142 	int proxy_sqp = 0;
143 	int real_sqp = 0;
144 	int i;
145 	/* PPF or Native -- real SQP */
146 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
147 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
148 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
149 	if (real_sqp)
150 		return 1;
151 	/* VF or PF -- proxy SQP */
152 	if (mlx4_is_mfunc(dev->dev)) {
153 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
154 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] ||
155 			    qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) {
156 				proxy_sqp = 1;
157 				break;
158 			}
159 		}
160 	}
161 	return proxy_sqp;
162 }
163 
164 /* used for INIT/CLOSE port logic */
165 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
166 {
167 	int proxy_qp0 = 0;
168 	int real_qp0 = 0;
169 	int i;
170 	/* PPF or Native -- real QP0 */
171 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
172 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
173 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
174 	if (real_qp0)
175 		return 1;
176 	/* VF or PF -- proxy QP0 */
177 	if (mlx4_is_mfunc(dev->dev)) {
178 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
179 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) {
180 				proxy_qp0 = 1;
181 				break;
182 			}
183 		}
184 	}
185 	return proxy_qp0;
186 }
187 
188 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
189 {
190 	return mlx4_buf_offset(&qp->buf, offset);
191 }
192 
193 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
194 {
195 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
196 }
197 
198 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
199 {
200 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
201 }
202 
203 /*
204  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
205  * first four bytes of every 64 byte chunk with
206  *     0x7FFFFFF | (invalid_ownership_value << 31).
207  *
208  * When the max work request size is less than or equal to the WQE
209  * basic block size, as an optimization, we can stamp all WQEs with
210  * 0xffffffff, and skip the very first chunk of each WQE.
211  */
212 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
213 {
214 	__be32 *wqe;
215 	int i;
216 	int s;
217 	int ind;
218 	void *buf;
219 	__be32 stamp;
220 	struct mlx4_wqe_ctrl_seg *ctrl;
221 
222 	if (qp->sq_max_wqes_per_wr > 1) {
223 		s = roundup(size, 1U << qp->sq.wqe_shift);
224 		for (i = 0; i < s; i += 64) {
225 			ind = (i >> qp->sq.wqe_shift) + n;
226 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
227 						       cpu_to_be32(0xffffffff);
228 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
229 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
230 			*wqe = stamp;
231 		}
232 	} else {
233 		ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
234 		s = (ctrl->fence_size & 0x3f) << 4;
235 		for (i = 64; i < s; i += 64) {
236 			wqe = buf + i;
237 			*wqe = cpu_to_be32(0xffffffff);
238 		}
239 	}
240 }
241 
242 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
243 {
244 	struct mlx4_wqe_ctrl_seg *ctrl;
245 	struct mlx4_wqe_inline_seg *inl;
246 	void *wqe;
247 	int s;
248 
249 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
250 	s = sizeof(struct mlx4_wqe_ctrl_seg);
251 
252 	if (qp->ibqp.qp_type == IB_QPT_UD) {
253 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
254 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
255 		memset(dgram, 0, sizeof *dgram);
256 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
257 		s += sizeof(struct mlx4_wqe_datagram_seg);
258 	}
259 
260 	/* Pad the remainder of the WQE with an inline data segment. */
261 	if (size > s) {
262 		inl = wqe + s;
263 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
264 	}
265 	ctrl->srcrb_flags = 0;
266 	ctrl->fence_size = size / 16;
267 	/*
268 	 * Make sure descriptor is fully written before setting ownership bit
269 	 * (because HW can start executing as soon as we do).
270 	 */
271 	wmb();
272 
273 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
274 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
275 
276 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
277 }
278 
279 /* Post NOP WQE to prevent wrap-around in the middle of WR */
280 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
281 {
282 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
283 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
284 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
285 		ind += s;
286 	}
287 	return ind;
288 }
289 
290 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
291 {
292 	struct ib_event event;
293 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
294 
295 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
296 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
297 
298 	if (ibqp->event_handler) {
299 		event.device     = ibqp->device;
300 		event.element.qp = ibqp;
301 		switch (type) {
302 		case MLX4_EVENT_TYPE_PATH_MIG:
303 			event.event = IB_EVENT_PATH_MIG;
304 			break;
305 		case MLX4_EVENT_TYPE_COMM_EST:
306 			event.event = IB_EVENT_COMM_EST;
307 			break;
308 		case MLX4_EVENT_TYPE_SQ_DRAINED:
309 			event.event = IB_EVENT_SQ_DRAINED;
310 			break;
311 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
312 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
313 			break;
314 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
315 			event.event = IB_EVENT_QP_FATAL;
316 			break;
317 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
318 			event.event = IB_EVENT_PATH_MIG_ERR;
319 			break;
320 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
321 			event.event = IB_EVENT_QP_REQ_ERR;
322 			break;
323 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
324 			event.event = IB_EVENT_QP_ACCESS_ERR;
325 			break;
326 		default:
327 			pr_warn("Unexpected event type %d "
328 			       "on QP %06x\n", type, qp->qpn);
329 			return;
330 		}
331 
332 		ibqp->event_handler(&event, ibqp->qp_context);
333 	}
334 }
335 
336 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
337 {
338 	/*
339 	 * UD WQEs must have a datagram segment.
340 	 * RC and UC WQEs might have a remote address segment.
341 	 * MLX WQEs need two extra inline data segments (for the UD
342 	 * header and space for the ICRC).
343 	 */
344 	switch (type) {
345 	case MLX4_IB_QPT_UD:
346 		return sizeof (struct mlx4_wqe_ctrl_seg) +
347 			sizeof (struct mlx4_wqe_datagram_seg) +
348 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
349 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
350 	case MLX4_IB_QPT_PROXY_SMI:
351 	case MLX4_IB_QPT_PROXY_GSI:
352 		return sizeof (struct mlx4_wqe_ctrl_seg) +
353 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
354 	case MLX4_IB_QPT_TUN_SMI_OWNER:
355 	case MLX4_IB_QPT_TUN_GSI:
356 		return sizeof (struct mlx4_wqe_ctrl_seg) +
357 			sizeof (struct mlx4_wqe_datagram_seg);
358 
359 	case MLX4_IB_QPT_UC:
360 		return sizeof (struct mlx4_wqe_ctrl_seg) +
361 			sizeof (struct mlx4_wqe_raddr_seg);
362 	case MLX4_IB_QPT_RC:
363 		return sizeof (struct mlx4_wqe_ctrl_seg) +
364 			sizeof (struct mlx4_wqe_atomic_seg) +
365 			sizeof (struct mlx4_wqe_raddr_seg);
366 	case MLX4_IB_QPT_SMI:
367 	case MLX4_IB_QPT_GSI:
368 		return sizeof (struct mlx4_wqe_ctrl_seg) +
369 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
370 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
371 					   MLX4_INLINE_ALIGN) *
372 			      sizeof (struct mlx4_wqe_inline_seg),
373 			      sizeof (struct mlx4_wqe_data_seg)) +
374 			ALIGN(4 +
375 			      sizeof (struct mlx4_wqe_inline_seg),
376 			      sizeof (struct mlx4_wqe_data_seg));
377 	default:
378 		return sizeof (struct mlx4_wqe_ctrl_seg);
379 	}
380 }
381 
382 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
383 		       int is_user, int has_rq, struct mlx4_ib_qp *qp)
384 {
385 	/* Sanity check RQ size before proceeding */
386 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
387 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
388 		return -EINVAL;
389 
390 	if (!has_rq) {
391 		if (cap->max_recv_wr)
392 			return -EINVAL;
393 
394 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
395 	} else {
396 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
397 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
398 			return -EINVAL;
399 
400 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
401 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
402 		qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
403 	}
404 
405 	/* leave userspace return values as they were, so as not to break ABI */
406 	if (is_user) {
407 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
408 		cap->max_recv_sge = qp->rq.max_gs;
409 	} else {
410 		cap->max_recv_wr  = qp->rq.max_post =
411 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
412 		cap->max_recv_sge = min(qp->rq.max_gs,
413 					min(dev->dev->caps.max_sq_sg,
414 					    dev->dev->caps.max_rq_sg));
415 	}
416 
417 	return 0;
418 }
419 
420 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
421 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
422 {
423 	int s;
424 
425 	/* Sanity check SQ size before proceeding */
426 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
427 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
428 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
429 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
430 		return -EINVAL;
431 
432 	/*
433 	 * For MLX transport we need 2 extra S/G entries:
434 	 * one for the header and one for the checksum at the end
435 	 */
436 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
437 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
438 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
439 		return -EINVAL;
440 
441 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
442 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
443 		send_wqe_overhead(type, qp->flags);
444 
445 	if (s > dev->dev->caps.max_sq_desc_sz)
446 		return -EINVAL;
447 
448 	/*
449 	 * Hermon supports shrinking WQEs, such that a single work
450 	 * request can include multiple units of 1 << wqe_shift.  This
451 	 * way, work requests can differ in size, and do not have to
452 	 * be a power of 2 in size, saving memory and speeding up send
453 	 * WR posting.  Unfortunately, if we do this then the
454 	 * wqe_index field in CQEs can't be used to look up the WR ID
455 	 * anymore, so we do this only if selective signaling is off.
456 	 *
457 	 * Further, on 32-bit platforms, we can't use vmap() to make
458 	 * the QP buffer virtually contiguous.  Thus we have to use
459 	 * constant-sized WRs to make sure a WR is always fully within
460 	 * a single page-sized chunk.
461 	 *
462 	 * Finally, we use NOP work requests to pad the end of the
463 	 * work queue, to avoid wrap-around in the middle of WR.  We
464 	 * set NEC bit to avoid getting completions with error for
465 	 * these NOP WRs, but since NEC is only supported starting
466 	 * with firmware 2.2.232, we use constant-sized WRs for older
467 	 * firmware.
468 	 *
469 	 * And, since MLX QPs only support SEND, we use constant-sized
470 	 * WRs in this case.
471 	 *
472 	 * We look for the smallest value of wqe_shift such that the
473 	 * resulting number of wqes does not exceed device
474 	 * capabilities.
475 	 *
476 	 * We set WQE size to at least 64 bytes, this way stamping
477 	 * invalidates each WQE.
478 	 */
479 	if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
480 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
481 	    type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
482 	    !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
483 		      MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
484 		qp->sq.wqe_shift = ilog2(64);
485 	else
486 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
487 
488 	for (;;) {
489 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
490 
491 		/*
492 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
493 		 * allow HW to prefetch.
494 		 */
495 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
496 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
497 						    qp->sq_max_wqes_per_wr +
498 						    qp->sq_spare_wqes);
499 
500 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
501 			break;
502 
503 		if (qp->sq_max_wqes_per_wr <= 1)
504 			return -EINVAL;
505 
506 		++qp->sq.wqe_shift;
507 	}
508 
509 	qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
510 			     (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
511 			 send_wqe_overhead(type, qp->flags)) /
512 		sizeof (struct mlx4_wqe_data_seg);
513 
514 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
515 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
516 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
517 		qp->rq.offset = 0;
518 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
519 	} else {
520 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
521 		qp->sq.offset = 0;
522 	}
523 
524 	cap->max_send_wr  = qp->sq.max_post =
525 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
526 	cap->max_send_sge = min(qp->sq.max_gs,
527 				min(dev->dev->caps.max_sq_sg,
528 				    dev->dev->caps.max_rq_sg));
529 	/* We don't support inline sends for kernel QPs (yet) */
530 	cap->max_inline_data = 0;
531 
532 	return 0;
533 }
534 
535 static int set_user_sq_size(struct mlx4_ib_dev *dev,
536 			    struct mlx4_ib_qp *qp,
537 			    struct mlx4_ib_create_qp *ucmd)
538 {
539 	/* Sanity check SQ size before proceeding */
540 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
541 	    ucmd->log_sq_stride >
542 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
543 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
544 		return -EINVAL;
545 
546 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
547 	qp->sq.wqe_shift = ucmd->log_sq_stride;
548 
549 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
550 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
551 
552 	return 0;
553 }
554 
555 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
556 {
557 	int i;
558 
559 	qp->sqp_proxy_rcv =
560 		kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
561 			GFP_KERNEL);
562 	if (!qp->sqp_proxy_rcv)
563 		return -ENOMEM;
564 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
565 		qp->sqp_proxy_rcv[i].addr =
566 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
567 				GFP_KERNEL);
568 		if (!qp->sqp_proxy_rcv[i].addr)
569 			goto err;
570 		qp->sqp_proxy_rcv[i].map =
571 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
572 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
573 					  DMA_FROM_DEVICE);
574 	}
575 	return 0;
576 
577 err:
578 	while (i > 0) {
579 		--i;
580 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
581 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
582 				    DMA_FROM_DEVICE);
583 		kfree(qp->sqp_proxy_rcv[i].addr);
584 	}
585 	kfree(qp->sqp_proxy_rcv);
586 	qp->sqp_proxy_rcv = NULL;
587 	return -ENOMEM;
588 }
589 
590 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
591 {
592 	int i;
593 
594 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
595 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
596 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
597 				    DMA_FROM_DEVICE);
598 		kfree(qp->sqp_proxy_rcv[i].addr);
599 	}
600 	kfree(qp->sqp_proxy_rcv);
601 }
602 
603 static int qp_has_rq(struct ib_qp_init_attr *attr)
604 {
605 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
606 		return 0;
607 
608 	return !attr->srq;
609 }
610 
611 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
612 			    struct ib_qp_init_attr *init_attr,
613 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp)
614 {
615 	int qpn;
616 	int err;
617 	struct mlx4_ib_sqp *sqp;
618 	struct mlx4_ib_qp *qp;
619 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
620 
621 	/* When tunneling special qps, we use a plain UD qp */
622 	if (sqpn) {
623 		if (mlx4_is_mfunc(dev->dev) &&
624 		    (!mlx4_is_master(dev->dev) ||
625 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
626 			if (init_attr->qp_type == IB_QPT_GSI)
627 				qp_type = MLX4_IB_QPT_PROXY_GSI;
628 			else if (mlx4_is_master(dev->dev))
629 				qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
630 			else
631 				qp_type = MLX4_IB_QPT_PROXY_SMI;
632 		}
633 		qpn = sqpn;
634 		/* add extra sg entry for tunneling */
635 		init_attr->cap.max_recv_sge++;
636 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
637 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
638 			container_of(init_attr,
639 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
640 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
641 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
642 		    !mlx4_is_master(dev->dev))
643 			return -EINVAL;
644 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
645 			qp_type = MLX4_IB_QPT_TUN_GSI;
646 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev))
647 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
648 		else
649 			qp_type = MLX4_IB_QPT_TUN_SMI;
650 		/* we are definitely in the PPF here, since we are creating
651 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
652 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
653 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
654 		sqpn = qpn;
655 	}
656 
657 	if (!*caller_qp) {
658 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
659 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
660 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
661 			sqp = kzalloc(sizeof (struct mlx4_ib_sqp), GFP_KERNEL);
662 			if (!sqp)
663 				return -ENOMEM;
664 			qp = &sqp->qp;
665 			qp->pri.vid = 0xFFFF;
666 			qp->alt.vid = 0xFFFF;
667 		} else {
668 			qp = kzalloc(sizeof (struct mlx4_ib_qp), GFP_KERNEL);
669 			if (!qp)
670 				return -ENOMEM;
671 			qp->pri.vid = 0xFFFF;
672 			qp->alt.vid = 0xFFFF;
673 		}
674 	} else
675 		qp = *caller_qp;
676 
677 	qp->mlx4_ib_qp_type = qp_type;
678 
679 	mutex_init(&qp->mutex);
680 	spin_lock_init(&qp->sq.lock);
681 	spin_lock_init(&qp->rq.lock);
682 	INIT_LIST_HEAD(&qp->gid_list);
683 	INIT_LIST_HEAD(&qp->steering_rules);
684 
685 	qp->state	 = IB_QPS_RESET;
686 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
687 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
688 
689 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
690 	if (err)
691 		goto err;
692 
693 	if (pd->uobject) {
694 		struct mlx4_ib_create_qp ucmd;
695 
696 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
697 			err = -EFAULT;
698 			goto err;
699 		}
700 
701 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
702 
703 		err = set_user_sq_size(dev, qp, &ucmd);
704 		if (err)
705 			goto err;
706 
707 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
708 				       qp->buf_size, 0, 0);
709 		if (IS_ERR(qp->umem)) {
710 			err = PTR_ERR(qp->umem);
711 			goto err;
712 		}
713 
714 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
715 				    ilog2(qp->umem->page_size), &qp->mtt);
716 		if (err)
717 			goto err_buf;
718 
719 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
720 		if (err)
721 			goto err_mtt;
722 
723 		if (qp_has_rq(init_attr)) {
724 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
725 						  ucmd.db_addr, &qp->db);
726 			if (err)
727 				goto err_mtt;
728 		}
729 	} else {
730 		qp->sq_no_prefetch = 0;
731 
732 		if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
733 			qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
734 
735 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
736 			qp->flags |= MLX4_IB_QP_LSO;
737 
738 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
739 			if (dev->steering_support ==
740 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
741 				qp->flags |= MLX4_IB_QP_NETIF;
742 			else
743 				goto err;
744 		}
745 
746 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
747 		if (err)
748 			goto err;
749 
750 		if (qp_has_rq(init_attr)) {
751 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
752 			if (err)
753 				goto err;
754 
755 			*qp->db.db = 0;
756 		}
757 
758 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf)) {
759 			err = -ENOMEM;
760 			goto err_db;
761 		}
762 
763 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
764 				    &qp->mtt);
765 		if (err)
766 			goto err_buf;
767 
768 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
769 		if (err)
770 			goto err_mtt;
771 
772 		qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), GFP_KERNEL);
773 		qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), GFP_KERNEL);
774 
775 		if (!qp->sq.wrid || !qp->rq.wrid) {
776 			err = -ENOMEM;
777 			goto err_wrid;
778 		}
779 	}
780 
781 	if (sqpn) {
782 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
783 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
784 			if (alloc_proxy_bufs(pd->device, qp)) {
785 				err = -ENOMEM;
786 				goto err_wrid;
787 			}
788 		}
789 	} else {
790 		/* Raw packet QPNs must be aligned to 8 bits. If not, the WQE
791 		 * BlueFlame setup flow wrongly causes VLAN insertion. */
792 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
793 			err = mlx4_qp_reserve_range(dev->dev, 1, 1 << 8, &qpn);
794 		else
795 			if (qp->flags & MLX4_IB_QP_NETIF)
796 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
797 			else
798 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
799 							    &qpn);
800 		if (err)
801 			goto err_proxy;
802 	}
803 
804 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
805 	if (err)
806 		goto err_qpn;
807 
808 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
809 		qp->mqp.qpn |= (1 << 23);
810 
811 	/*
812 	 * Hardware wants QPN written in big-endian order (after
813 	 * shifting) for send doorbell.  Precompute this value to save
814 	 * a little bit when posting sends.
815 	 */
816 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
817 
818 	qp->mqp.event = mlx4_ib_qp_event;
819 	if (!*caller_qp)
820 		*caller_qp = qp;
821 	return 0;
822 
823 err_qpn:
824 	if (!sqpn) {
825 		if (qp->flags & MLX4_IB_QP_NETIF)
826 			mlx4_ib_steer_qp_free(dev, qpn, 1);
827 		else
828 			mlx4_qp_release_range(dev->dev, qpn, 1);
829 	}
830 err_proxy:
831 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
832 		free_proxy_bufs(pd->device, qp);
833 err_wrid:
834 	if (pd->uobject) {
835 		if (qp_has_rq(init_attr))
836 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
837 	} else {
838 		kfree(qp->sq.wrid);
839 		kfree(qp->rq.wrid);
840 	}
841 
842 err_mtt:
843 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
844 
845 err_buf:
846 	if (pd->uobject)
847 		ib_umem_release(qp->umem);
848 	else
849 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
850 
851 err_db:
852 	if (!pd->uobject && qp_has_rq(init_attr))
853 		mlx4_db_free(dev->dev, &qp->db);
854 
855 err:
856 	if (!*caller_qp)
857 		kfree(qp);
858 	return err;
859 }
860 
861 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
862 {
863 	switch (state) {
864 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
865 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
866 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
867 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
868 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
869 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
870 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
871 	default:		return -1;
872 	}
873 }
874 
875 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
876 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
877 {
878 	if (send_cq == recv_cq) {
879 		spin_lock_irq(&send_cq->lock);
880 		__acquire(&recv_cq->lock);
881 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
882 		spin_lock_irq(&send_cq->lock);
883 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
884 	} else {
885 		spin_lock_irq(&recv_cq->lock);
886 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
887 	}
888 }
889 
890 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
891 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
892 {
893 	if (send_cq == recv_cq) {
894 		__release(&recv_cq->lock);
895 		spin_unlock_irq(&send_cq->lock);
896 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
897 		spin_unlock(&recv_cq->lock);
898 		spin_unlock_irq(&send_cq->lock);
899 	} else {
900 		spin_unlock(&send_cq->lock);
901 		spin_unlock_irq(&recv_cq->lock);
902 	}
903 }
904 
905 static void del_gid_entries(struct mlx4_ib_qp *qp)
906 {
907 	struct mlx4_ib_gid_entry *ge, *tmp;
908 
909 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
910 		list_del(&ge->list);
911 		kfree(ge);
912 	}
913 }
914 
915 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
916 {
917 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
918 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
919 	else
920 		return to_mpd(qp->ibqp.pd);
921 }
922 
923 static void get_cqs(struct mlx4_ib_qp *qp,
924 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
925 {
926 	switch (qp->ibqp.qp_type) {
927 	case IB_QPT_XRC_TGT:
928 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
929 		*recv_cq = *send_cq;
930 		break;
931 	case IB_QPT_XRC_INI:
932 		*send_cq = to_mcq(qp->ibqp.send_cq);
933 		*recv_cq = *send_cq;
934 		break;
935 	default:
936 		*send_cq = to_mcq(qp->ibqp.send_cq);
937 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
938 		break;
939 	}
940 }
941 
942 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
943 			      int is_user)
944 {
945 	struct mlx4_ib_cq *send_cq, *recv_cq;
946 
947 	if (qp->state != IB_QPS_RESET) {
948 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
949 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
950 			pr_warn("modify QP %06x to RESET failed.\n",
951 			       qp->mqp.qpn);
952 		if (qp->pri.smac) {
953 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
954 			qp->pri.smac = 0;
955 		}
956 		if (qp->alt.smac) {
957 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
958 			qp->alt.smac = 0;
959 		}
960 		if (qp->pri.vid < 0x1000) {
961 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
962 			qp->pri.vid = 0xFFFF;
963 			qp->pri.candidate_vid = 0xFFFF;
964 			qp->pri.update_vid = 0;
965 		}
966 		if (qp->alt.vid < 0x1000) {
967 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
968 			qp->alt.vid = 0xFFFF;
969 			qp->alt.candidate_vid = 0xFFFF;
970 			qp->alt.update_vid = 0;
971 		}
972 	}
973 
974 	get_cqs(qp, &send_cq, &recv_cq);
975 
976 	mlx4_ib_lock_cqs(send_cq, recv_cq);
977 
978 	if (!is_user) {
979 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
980 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
981 		if (send_cq != recv_cq)
982 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
983 	}
984 
985 	mlx4_qp_remove(dev->dev, &qp->mqp);
986 
987 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
988 
989 	mlx4_qp_free(dev->dev, &qp->mqp);
990 
991 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
992 		if (qp->flags & MLX4_IB_QP_NETIF)
993 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
994 		else
995 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
996 	}
997 
998 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
999 
1000 	if (is_user) {
1001 		if (qp->rq.wqe_cnt)
1002 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
1003 					      &qp->db);
1004 		ib_umem_release(qp->umem);
1005 	} else {
1006 		kfree(qp->sq.wrid);
1007 		kfree(qp->rq.wrid);
1008 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1009 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1010 			free_proxy_bufs(&dev->ib_dev, qp);
1011 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1012 		if (qp->rq.wqe_cnt)
1013 			mlx4_db_free(dev->dev, &qp->db);
1014 	}
1015 
1016 	del_gid_entries(qp);
1017 }
1018 
1019 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1020 {
1021 	/* Native or PPF */
1022 	if (!mlx4_is_mfunc(dev->dev) ||
1023 	    (mlx4_is_master(dev->dev) &&
1024 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1025 		return  dev->dev->phys_caps.base_sqpn +
1026 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1027 			attr->port_num - 1;
1028 	}
1029 	/* PF or VF -- creating proxies */
1030 	if (attr->qp_type == IB_QPT_SMI)
1031 		return dev->dev->caps.qp0_proxy[attr->port_num - 1];
1032 	else
1033 		return dev->dev->caps.qp1_proxy[attr->port_num - 1];
1034 }
1035 
1036 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1037 				struct ib_qp_init_attr *init_attr,
1038 				struct ib_udata *udata)
1039 {
1040 	struct mlx4_ib_qp *qp = NULL;
1041 	int err;
1042 	u16 xrcdn = 0;
1043 
1044 	/*
1045 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1046 	 * and only for kernel UD QPs.
1047 	 */
1048 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1049 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1050 					MLX4_IB_SRIOV_TUNNEL_QP |
1051 					MLX4_IB_SRIOV_SQP |
1052 					MLX4_IB_QP_NETIF))
1053 		return ERR_PTR(-EINVAL);
1054 
1055 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1056 		if (init_attr->qp_type != IB_QPT_UD)
1057 			return ERR_PTR(-EINVAL);
1058 	}
1059 
1060 	if (init_attr->create_flags &&
1061 	    (udata ||
1062 	     ((init_attr->create_flags & ~MLX4_IB_SRIOV_SQP) &&
1063 	      init_attr->qp_type != IB_QPT_UD) ||
1064 	     ((init_attr->create_flags & MLX4_IB_SRIOV_SQP) &&
1065 	      init_attr->qp_type > IB_QPT_GSI)))
1066 		return ERR_PTR(-EINVAL);
1067 
1068 	switch (init_attr->qp_type) {
1069 	case IB_QPT_XRC_TGT:
1070 		pd = to_mxrcd(init_attr->xrcd)->pd;
1071 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1072 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1073 		/* fall through */
1074 	case IB_QPT_XRC_INI:
1075 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1076 			return ERR_PTR(-ENOSYS);
1077 		init_attr->recv_cq = init_attr->send_cq;
1078 		/* fall through */
1079 	case IB_QPT_RC:
1080 	case IB_QPT_UC:
1081 	case IB_QPT_RAW_PACKET:
1082 		qp = kzalloc(sizeof *qp, GFP_KERNEL);
1083 		if (!qp)
1084 			return ERR_PTR(-ENOMEM);
1085 		qp->pri.vid = 0xFFFF;
1086 		qp->alt.vid = 0xFFFF;
1087 		/* fall through */
1088 	case IB_QPT_UD:
1089 	{
1090 		err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1091 				       udata, 0, &qp);
1092 		if (err)
1093 			return ERR_PTR(err);
1094 
1095 		qp->ibqp.qp_num = qp->mqp.qpn;
1096 		qp->xrcdn = xrcdn;
1097 
1098 		break;
1099 	}
1100 	case IB_QPT_SMI:
1101 	case IB_QPT_GSI:
1102 	{
1103 		/* Userspace is not allowed to create special QPs: */
1104 		if (udata)
1105 			return ERR_PTR(-EINVAL);
1106 
1107 		err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1108 				       get_sqp_num(to_mdev(pd->device), init_attr),
1109 				       &qp);
1110 		if (err)
1111 			return ERR_PTR(err);
1112 
1113 		qp->port	= init_attr->port_num;
1114 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
1115 
1116 		break;
1117 	}
1118 	default:
1119 		/* Don't support raw QPs */
1120 		return ERR_PTR(-EINVAL);
1121 	}
1122 
1123 	return &qp->ibqp;
1124 }
1125 
1126 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1127 {
1128 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1129 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1130 	struct mlx4_ib_pd *pd;
1131 
1132 	if (is_qp0(dev, mqp))
1133 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1134 
1135 	pd = get_pd(mqp);
1136 	destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1137 
1138 	if (is_sqp(dev, mqp))
1139 		kfree(to_msqp(mqp));
1140 	else
1141 		kfree(mqp);
1142 
1143 	return 0;
1144 }
1145 
1146 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1147 {
1148 	switch (type) {
1149 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1150 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1151 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1152 	case MLX4_IB_QPT_XRC_INI:
1153 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1154 	case MLX4_IB_QPT_SMI:
1155 	case MLX4_IB_QPT_GSI:
1156 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1157 
1158 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1159 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1160 						MLX4_QP_ST_MLX : -1);
1161 	case MLX4_IB_QPT_PROXY_SMI:
1162 	case MLX4_IB_QPT_TUN_SMI:
1163 	case MLX4_IB_QPT_PROXY_GSI:
1164 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1165 						MLX4_QP_ST_UD : -1);
1166 	default:			return -1;
1167 	}
1168 }
1169 
1170 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1171 				   int attr_mask)
1172 {
1173 	u8 dest_rd_atomic;
1174 	u32 access_flags;
1175 	u32 hw_access_flags = 0;
1176 
1177 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1178 		dest_rd_atomic = attr->max_dest_rd_atomic;
1179 	else
1180 		dest_rd_atomic = qp->resp_depth;
1181 
1182 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1183 		access_flags = attr->qp_access_flags;
1184 	else
1185 		access_flags = qp->atomic_rd_en;
1186 
1187 	if (!dest_rd_atomic)
1188 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1189 
1190 	if (access_flags & IB_ACCESS_REMOTE_READ)
1191 		hw_access_flags |= MLX4_QP_BIT_RRE;
1192 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1193 		hw_access_flags |= MLX4_QP_BIT_RAE;
1194 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1195 		hw_access_flags |= MLX4_QP_BIT_RWE;
1196 
1197 	return cpu_to_be32(hw_access_flags);
1198 }
1199 
1200 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1201 			    int attr_mask)
1202 {
1203 	if (attr_mask & IB_QP_PKEY_INDEX)
1204 		sqp->pkey_index = attr->pkey_index;
1205 	if (attr_mask & IB_QP_QKEY)
1206 		sqp->qkey = attr->qkey;
1207 	if (attr_mask & IB_QP_SQ_PSN)
1208 		sqp->send_psn = attr->sq_psn;
1209 }
1210 
1211 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1212 {
1213 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1214 }
1215 
1216 static int _mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
1217 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1218 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1219 {
1220 	int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
1221 		IB_LINK_LAYER_ETHERNET;
1222 	int vidx;
1223 	int smac_index;
1224 	int err;
1225 
1226 
1227 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
1228 	path->rlid	    = cpu_to_be16(ah->dlid);
1229 	if (ah->static_rate) {
1230 		path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
1231 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1232 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1233 			--path->static_rate;
1234 	} else
1235 		path->static_rate = 0;
1236 
1237 	if (ah->ah_flags & IB_AH_GRH) {
1238 		if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
1239 			pr_err("sgid_index (%u) too large. max is %d\n",
1240 			       ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1241 			return -1;
1242 		}
1243 
1244 		path->grh_mylmc |= 1 << 7;
1245 		path->mgid_index = ah->grh.sgid_index;
1246 		path->hop_limit  = ah->grh.hop_limit;
1247 		path->tclass_flowlabel =
1248 			cpu_to_be32((ah->grh.traffic_class << 20) |
1249 				    (ah->grh.flow_label));
1250 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1251 	}
1252 
1253 	if (is_eth) {
1254 		if (!(ah->ah_flags & IB_AH_GRH))
1255 			return -1;
1256 
1257 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1258 			((port - 1) << 6) | ((ah->sl & 7) << 3);
1259 
1260 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1261 		if (vlan_tag < 0x1000) {
1262 			if (smac_info->vid < 0x1000) {
1263 				/* both valid vlan ids */
1264 				if (smac_info->vid != vlan_tag) {
1265 					/* different VIDs.  unreg old and reg new */
1266 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1267 					if (err)
1268 						return err;
1269 					smac_info->candidate_vid = vlan_tag;
1270 					smac_info->candidate_vlan_index = vidx;
1271 					smac_info->candidate_vlan_port = port;
1272 					smac_info->update_vid = 1;
1273 					path->vlan_index = vidx;
1274 				} else {
1275 					path->vlan_index = smac_info->vlan_index;
1276 				}
1277 			} else {
1278 				/* no current vlan tag in qp */
1279 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1280 				if (err)
1281 					return err;
1282 				smac_info->candidate_vid = vlan_tag;
1283 				smac_info->candidate_vlan_index = vidx;
1284 				smac_info->candidate_vlan_port = port;
1285 				smac_info->update_vid = 1;
1286 				path->vlan_index = vidx;
1287 			}
1288 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1289 			path->fl = 1 << 6;
1290 		} else {
1291 			/* have current vlan tag. unregister it at modify-qp success */
1292 			if (smac_info->vid < 0x1000) {
1293 				smac_info->candidate_vid = 0xFFFF;
1294 				smac_info->update_vid = 1;
1295 			}
1296 		}
1297 
1298 		/* get smac_index for RoCE use.
1299 		 * If no smac was yet assigned, register one.
1300 		 * If one was already assigned, but the new mac differs,
1301 		 * unregister the old one and register the new one.
1302 		*/
1303 		if (!smac_info->smac || smac_info->smac != smac) {
1304 			/* register candidate now, unreg if needed, after success */
1305 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1306 			if (smac_index >= 0) {
1307 				smac_info->candidate_smac_index = smac_index;
1308 				smac_info->candidate_smac = smac;
1309 				smac_info->candidate_smac_port = port;
1310 			} else {
1311 				return -EINVAL;
1312 			}
1313 		} else {
1314 			smac_index = smac_info->smac_index;
1315 		}
1316 
1317 		memcpy(path->dmac, ah->dmac, 6);
1318 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1319 		/* put MAC table smac index for IBoE */
1320 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1321 	} else {
1322 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1323 			((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1324 	}
1325 
1326 	return 0;
1327 }
1328 
1329 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1330 			 enum ib_qp_attr_mask qp_attr_mask,
1331 			 struct mlx4_ib_qp *mqp,
1332 			 struct mlx4_qp_path *path, u8 port)
1333 {
1334 	return _mlx4_set_path(dev, &qp->ah_attr,
1335 			      mlx4_mac_to_u64((u8 *)qp->smac),
1336 			      (qp_attr_mask & IB_QP_VID) ? qp->vlan_id : 0xffff,
1337 			      path, &mqp->pri, port);
1338 }
1339 
1340 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1341 			     const struct ib_qp_attr *qp,
1342 			     enum ib_qp_attr_mask qp_attr_mask,
1343 			     struct mlx4_ib_qp *mqp,
1344 			     struct mlx4_qp_path *path, u8 port)
1345 {
1346 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1347 			      mlx4_mac_to_u64((u8 *)qp->alt_smac),
1348 			      (qp_attr_mask & IB_QP_ALT_VID) ?
1349 			      qp->alt_vlan_id : 0xffff,
1350 			      path, &mqp->alt, port);
1351 }
1352 
1353 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1354 {
1355 	struct mlx4_ib_gid_entry *ge, *tmp;
1356 
1357 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1358 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1359 			ge->added = 1;
1360 			ge->port = qp->port;
1361 		}
1362 	}
1363 }
1364 
1365 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp, u8 *smac,
1366 				    struct mlx4_qp_context *context)
1367 {
1368 	struct net_device *ndev;
1369 	u64 u64_mac;
1370 	int smac_index;
1371 
1372 
1373 	ndev = dev->iboe.netdevs[qp->port - 1];
1374 	if (ndev) {
1375 		smac = ndev->dev_addr;
1376 		u64_mac = mlx4_mac_to_u64(smac);
1377 	} else {
1378 		u64_mac = dev->dev->caps.def_mac[qp->port];
1379 	}
1380 
1381 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1382 	if (!qp->pri.smac) {
1383 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1384 		if (smac_index >= 0) {
1385 			qp->pri.candidate_smac_index = smac_index;
1386 			qp->pri.candidate_smac = u64_mac;
1387 			qp->pri.candidate_smac_port = qp->port;
1388 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1389 		} else {
1390 			return -ENOENT;
1391 		}
1392 	}
1393 	return 0;
1394 }
1395 
1396 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1397 			       const struct ib_qp_attr *attr, int attr_mask,
1398 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1399 {
1400 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1401 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1402 	struct mlx4_ib_pd *pd;
1403 	struct mlx4_ib_cq *send_cq, *recv_cq;
1404 	struct mlx4_qp_context *context;
1405 	enum mlx4_qp_optpar optpar = 0;
1406 	int sqd_event;
1407 	int steer_qp = 0;
1408 	int err = -EINVAL;
1409 
1410 	context = kzalloc(sizeof *context, GFP_KERNEL);
1411 	if (!context)
1412 		return -ENOMEM;
1413 
1414 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1415 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1416 
1417 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1418 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1419 	else {
1420 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
1421 		switch (attr->path_mig_state) {
1422 		case IB_MIG_MIGRATED:
1423 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1424 			break;
1425 		case IB_MIG_REARM:
1426 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1427 			break;
1428 		case IB_MIG_ARMED:
1429 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1430 			break;
1431 		}
1432 	}
1433 
1434 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1435 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1436 	else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1437 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1438 	else if (ibqp->qp_type == IB_QPT_UD) {
1439 		if (qp->flags & MLX4_IB_QP_LSO)
1440 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
1441 					      ilog2(dev->dev->caps.max_gso_sz);
1442 		else
1443 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1444 	} else if (attr_mask & IB_QP_PATH_MTU) {
1445 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1446 			pr_err("path MTU (%u) is invalid\n",
1447 			       attr->path_mtu);
1448 			goto out;
1449 		}
1450 		context->mtu_msgmax = (attr->path_mtu << 5) |
1451 			ilog2(dev->dev->caps.max_msg_sz);
1452 	}
1453 
1454 	if (qp->rq.wqe_cnt)
1455 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1456 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
1457 
1458 	if (qp->sq.wqe_cnt)
1459 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1460 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
1461 
1462 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1463 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1464 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1465 		if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1466 			context->param3 |= cpu_to_be32(1 << 30);
1467 	}
1468 
1469 	if (qp->ibqp.uobject)
1470 		context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
1471 	else
1472 		context->usr_page = cpu_to_be32(dev->priv_uar.index);
1473 
1474 	if (attr_mask & IB_QP_DEST_QPN)
1475 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1476 
1477 	if (attr_mask & IB_QP_PORT) {
1478 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1479 		    !(attr_mask & IB_QP_AV)) {
1480 			mlx4_set_sched(&context->pri_path, attr->port_num);
1481 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1482 		}
1483 	}
1484 
1485 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1486 		if (dev->counters[qp->port - 1] != -1) {
1487 			context->pri_path.counter_index =
1488 						dev->counters[qp->port - 1];
1489 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1490 		} else
1491 			context->pri_path.counter_index = 0xff;
1492 
1493 		if (qp->flags & MLX4_IB_QP_NETIF) {
1494 			mlx4_ib_steer_qp_reg(dev, qp, 1);
1495 			steer_qp = 1;
1496 		}
1497 	}
1498 
1499 	if (attr_mask & IB_QP_PKEY_INDEX) {
1500 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1501 			context->pri_path.disable_pkey_check = 0x40;
1502 		context->pri_path.pkey_index = attr->pkey_index;
1503 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1504 	}
1505 
1506 	if (attr_mask & IB_QP_AV) {
1507 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
1508 				  attr_mask & IB_QP_PORT ?
1509 				  attr->port_num : qp->port))
1510 			goto out;
1511 
1512 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1513 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
1514 	}
1515 
1516 	if (attr_mask & IB_QP_TIMEOUT) {
1517 		context->pri_path.ackto |= attr->timeout << 3;
1518 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1519 	}
1520 
1521 	if (attr_mask & IB_QP_ALT_PATH) {
1522 		if (attr->alt_port_num == 0 ||
1523 		    attr->alt_port_num > dev->dev->caps.num_ports)
1524 			goto out;
1525 
1526 		if (attr->alt_pkey_index >=
1527 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
1528 			goto out;
1529 
1530 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
1531 				      &context->alt_path,
1532 				      attr->alt_port_num))
1533 			goto out;
1534 
1535 		context->alt_path.pkey_index = attr->alt_pkey_index;
1536 		context->alt_path.ackto = attr->alt_timeout << 3;
1537 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1538 	}
1539 
1540 	pd = get_pd(qp);
1541 	get_cqs(qp, &send_cq, &recv_cq);
1542 	context->pd       = cpu_to_be32(pd->pdn);
1543 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1544 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1545 	context->params1  = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1546 
1547 	/* Set "fast registration enabled" for all kernel QPs */
1548 	if (!qp->ibqp.uobject)
1549 		context->params1 |= cpu_to_be32(1 << 11);
1550 
1551 	if (attr_mask & IB_QP_RNR_RETRY) {
1552 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1553 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1554 	}
1555 
1556 	if (attr_mask & IB_QP_RETRY_CNT) {
1557 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1558 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1559 	}
1560 
1561 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1562 		if (attr->max_rd_atomic)
1563 			context->params1 |=
1564 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1565 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1566 	}
1567 
1568 	if (attr_mask & IB_QP_SQ_PSN)
1569 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1570 
1571 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1572 		if (attr->max_dest_rd_atomic)
1573 			context->params2 |=
1574 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1575 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1576 	}
1577 
1578 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1579 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1580 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1581 	}
1582 
1583 	if (ibqp->srq)
1584 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1585 
1586 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1587 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1588 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1589 	}
1590 	if (attr_mask & IB_QP_RQ_PSN)
1591 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1592 
1593 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1594 	if (attr_mask & IB_QP_QKEY) {
1595 		if (qp->mlx4_ib_qp_type &
1596 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1597 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1598 		else {
1599 			if (mlx4_is_mfunc(dev->dev) &&
1600 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1601 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1602 			    MLX4_RESERVED_QKEY_BASE) {
1603 				pr_err("Cannot use reserved QKEY"
1604 				       " 0x%x (range 0xffff0000..0xffffffff"
1605 				       " is reserved)\n", attr->qkey);
1606 				err = -EINVAL;
1607 				goto out;
1608 			}
1609 			context->qkey = cpu_to_be32(attr->qkey);
1610 		}
1611 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
1612 	}
1613 
1614 	if (ibqp->srq)
1615 		context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1616 
1617 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1618 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1619 
1620 	if (cur_state == IB_QPS_INIT &&
1621 	    new_state == IB_QPS_RTR  &&
1622 	    (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1623 	     ibqp->qp_type == IB_QPT_UD ||
1624 	     ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1625 		context->pri_path.sched_queue = (qp->port - 1) << 6;
1626 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1627 		    qp->mlx4_ib_qp_type &
1628 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1629 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1630 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1631 				context->pri_path.fl = 0x80;
1632 		} else {
1633 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1634 				context->pri_path.fl = 0x80;
1635 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1636 		}
1637 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1638 		    IB_LINK_LAYER_ETHERNET) {
1639 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
1640 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
1641 				context->pri_path.feup = 1 << 7; /* don't fsm */
1642 			/* handle smac_index */
1643 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
1644 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
1645 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
1646 				err = handle_eth_ud_smac_index(dev, qp, (u8 *)attr->smac, context);
1647 				if (err)
1648 					return -EINVAL;
1649 			}
1650 		}
1651 	}
1652 
1653 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET)
1654 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
1655 					MLX4_IB_LINK_TYPE_ETH;
1656 
1657 	if (ibqp->qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
1658 		int is_eth = rdma_port_get_link_layer(
1659 				&dev->ib_dev, qp->port) ==
1660 				IB_LINK_LAYER_ETHERNET;
1661 		if (is_eth) {
1662 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
1663 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
1664 		}
1665 	}
1666 
1667 
1668 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1669 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1670 		sqd_event = 1;
1671 	else
1672 		sqd_event = 0;
1673 
1674 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1675 		context->rlkey |= (1 << 4);
1676 
1677 	/*
1678 	 * Before passing a kernel QP to the HW, make sure that the
1679 	 * ownership bits of the send queue are set and the SQ
1680 	 * headroom is stamped so that the hardware doesn't start
1681 	 * processing stale work requests.
1682 	 */
1683 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1684 		struct mlx4_wqe_ctrl_seg *ctrl;
1685 		int i;
1686 
1687 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1688 			ctrl = get_send_wqe(qp, i);
1689 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
1690 			if (qp->sq_max_wqes_per_wr == 1)
1691 				ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1692 
1693 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1694 		}
1695 	}
1696 
1697 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1698 			     to_mlx4_state(new_state), context, optpar,
1699 			     sqd_event, &qp->mqp);
1700 	if (err)
1701 		goto out;
1702 
1703 	qp->state = new_state;
1704 
1705 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1706 		qp->atomic_rd_en = attr->qp_access_flags;
1707 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1708 		qp->resp_depth = attr->max_dest_rd_atomic;
1709 	if (attr_mask & IB_QP_PORT) {
1710 		qp->port = attr->port_num;
1711 		update_mcg_macs(dev, qp);
1712 	}
1713 	if (attr_mask & IB_QP_ALT_PATH)
1714 		qp->alt_port = attr->alt_port_num;
1715 
1716 	if (is_sqp(dev, qp))
1717 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1718 
1719 	/*
1720 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
1721 	 * QP0 to RESET or ERROR, bring the link back down.
1722 	 */
1723 	if (is_qp0(dev, qp)) {
1724 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1725 			if (mlx4_INIT_PORT(dev->dev, qp->port))
1726 				pr_warn("INIT_PORT failed for port %d\n",
1727 				       qp->port);
1728 
1729 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1730 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1731 			mlx4_CLOSE_PORT(dev->dev, qp->port);
1732 	}
1733 
1734 	/*
1735 	 * If we moved a kernel QP to RESET, clean up all old CQ
1736 	 * entries and reinitialize the QP.
1737 	 */
1738 	if (new_state == IB_QPS_RESET) {
1739 		if (!ibqp->uobject) {
1740 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1741 					 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
1742 			if (send_cq != recv_cq)
1743 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1744 
1745 			qp->rq.head = 0;
1746 			qp->rq.tail = 0;
1747 			qp->sq.head = 0;
1748 			qp->sq.tail = 0;
1749 			qp->sq_next_wqe = 0;
1750 			if (qp->rq.wqe_cnt)
1751 				*qp->db.db  = 0;
1752 
1753 			if (qp->flags & MLX4_IB_QP_NETIF)
1754 				mlx4_ib_steer_qp_reg(dev, qp, 0);
1755 		}
1756 		if (qp->pri.smac) {
1757 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1758 			qp->pri.smac = 0;
1759 		}
1760 		if (qp->alt.smac) {
1761 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1762 			qp->alt.smac = 0;
1763 		}
1764 		if (qp->pri.vid < 0x1000) {
1765 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1766 			qp->pri.vid = 0xFFFF;
1767 			qp->pri.candidate_vid = 0xFFFF;
1768 			qp->pri.update_vid = 0;
1769 		}
1770 
1771 		if (qp->alt.vid < 0x1000) {
1772 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1773 			qp->alt.vid = 0xFFFF;
1774 			qp->alt.candidate_vid = 0xFFFF;
1775 			qp->alt.update_vid = 0;
1776 		}
1777 	}
1778 out:
1779 	if (err && steer_qp)
1780 		mlx4_ib_steer_qp_reg(dev, qp, 0);
1781 	kfree(context);
1782 	if (qp->pri.candidate_smac) {
1783 		if (err) {
1784 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
1785 		} else {
1786 			if (qp->pri.smac)
1787 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1788 			qp->pri.smac = qp->pri.candidate_smac;
1789 			qp->pri.smac_index = qp->pri.candidate_smac_index;
1790 			qp->pri.smac_port = qp->pri.candidate_smac_port;
1791 		}
1792 		qp->pri.candidate_smac = 0;
1793 		qp->pri.candidate_smac_index = 0;
1794 		qp->pri.candidate_smac_port = 0;
1795 	}
1796 	if (qp->alt.candidate_smac) {
1797 		if (err) {
1798 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
1799 		} else {
1800 			if (qp->alt.smac)
1801 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1802 			qp->alt.smac = qp->alt.candidate_smac;
1803 			qp->alt.smac_index = qp->alt.candidate_smac_index;
1804 			qp->alt.smac_port = qp->alt.candidate_smac_port;
1805 		}
1806 		qp->alt.candidate_smac = 0;
1807 		qp->alt.candidate_smac_index = 0;
1808 		qp->alt.candidate_smac_port = 0;
1809 	}
1810 
1811 	if (qp->pri.update_vid) {
1812 		if (err) {
1813 			if (qp->pri.candidate_vid < 0x1000)
1814 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
1815 						     qp->pri.candidate_vid);
1816 		} else {
1817 			if (qp->pri.vid < 0x1000)
1818 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
1819 						     qp->pri.vid);
1820 			qp->pri.vid = qp->pri.candidate_vid;
1821 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
1822 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
1823 		}
1824 		qp->pri.candidate_vid = 0xFFFF;
1825 		qp->pri.update_vid = 0;
1826 	}
1827 
1828 	if (qp->alt.update_vid) {
1829 		if (err) {
1830 			if (qp->alt.candidate_vid < 0x1000)
1831 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
1832 						     qp->alt.candidate_vid);
1833 		} else {
1834 			if (qp->alt.vid < 0x1000)
1835 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
1836 						     qp->alt.vid);
1837 			qp->alt.vid = qp->alt.candidate_vid;
1838 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
1839 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
1840 		}
1841 		qp->alt.candidate_vid = 0xFFFF;
1842 		qp->alt.update_vid = 0;
1843 	}
1844 
1845 	return err;
1846 }
1847 
1848 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1849 		      int attr_mask, struct ib_udata *udata)
1850 {
1851 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1852 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1853 	enum ib_qp_state cur_state, new_state;
1854 	int err = -EINVAL;
1855 	int ll;
1856 	mutex_lock(&qp->mutex);
1857 
1858 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1859 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1860 
1861 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1862 		ll = IB_LINK_LAYER_UNSPECIFIED;
1863 	} else {
1864 		int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1865 		ll = rdma_port_get_link_layer(&dev->ib_dev, port);
1866 	}
1867 
1868 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1869 				attr_mask, ll)) {
1870 		pr_debug("qpn 0x%x: invalid attribute mask specified "
1871 			 "for transition %d to %d. qp_type %d,"
1872 			 " attr_mask 0x%x\n",
1873 			 ibqp->qp_num, cur_state, new_state,
1874 			 ibqp->qp_type, attr_mask);
1875 		goto out;
1876 	}
1877 
1878 	if ((attr_mask & IB_QP_PORT) &&
1879 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
1880 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
1881 			 "for transition %d to %d. qp_type %d\n",
1882 			 ibqp->qp_num, attr->port_num, cur_state,
1883 			 new_state, ibqp->qp_type);
1884 		goto out;
1885 	}
1886 
1887 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
1888 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
1889 	     IB_LINK_LAYER_ETHERNET))
1890 		goto out;
1891 
1892 	if (attr_mask & IB_QP_PKEY_INDEX) {
1893 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1894 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
1895 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
1896 				 "for transition %d to %d. qp_type %d\n",
1897 				 ibqp->qp_num, attr->pkey_index, cur_state,
1898 				 new_state, ibqp->qp_type);
1899 			goto out;
1900 		}
1901 	}
1902 
1903 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1904 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
1905 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
1906 			 "Transition %d to %d. qp_type %d\n",
1907 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
1908 			 new_state, ibqp->qp_type);
1909 		goto out;
1910 	}
1911 
1912 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1913 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
1914 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
1915 			 "Transition %d to %d. qp_type %d\n",
1916 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
1917 			 new_state, ibqp->qp_type);
1918 		goto out;
1919 	}
1920 
1921 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1922 		err = 0;
1923 		goto out;
1924 	}
1925 
1926 	err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1927 
1928 out:
1929 	mutex_unlock(&qp->mutex);
1930 	return err;
1931 }
1932 
1933 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
1934 				  struct ib_send_wr *wr,
1935 				  void *wqe, unsigned *mlx_seg_len)
1936 {
1937 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
1938 	struct ib_device *ib_dev = &mdev->ib_dev;
1939 	struct mlx4_wqe_mlx_seg *mlx = wqe;
1940 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
1941 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
1942 	u16 pkey;
1943 	u32 qkey;
1944 	int send_size;
1945 	int header_size;
1946 	int spc;
1947 	int i;
1948 
1949 	if (wr->opcode != IB_WR_SEND)
1950 		return -EINVAL;
1951 
1952 	send_size = 0;
1953 
1954 	for (i = 0; i < wr->num_sge; ++i)
1955 		send_size += wr->sg_list[i].length;
1956 
1957 	/* for proxy-qp0 sends, need to add in size of tunnel header */
1958 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
1959 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
1960 		send_size += sizeof (struct mlx4_ib_tunnel_header);
1961 
1962 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, &sqp->ud_header);
1963 
1964 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
1965 		sqp->ud_header.lrh.service_level =
1966 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
1967 		sqp->ud_header.lrh.destination_lid =
1968 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
1969 		sqp->ud_header.lrh.source_lid =
1970 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
1971 	}
1972 
1973 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1974 
1975 	/* force loopback */
1976 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
1977 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
1978 
1979 	sqp->ud_header.lrh.virtual_lane    = 0;
1980 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1981 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
1982 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1983 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
1984 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1985 	else
1986 		sqp->ud_header.bth.destination_qpn =
1987 			cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
1988 
1989 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1990 	if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
1991 		return -EINVAL;
1992 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
1993 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
1994 
1995 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
1996 	sqp->ud_header.immediate_present = 0;
1997 
1998 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
1999 
2000 	/*
2001 	 * Inline data segments may not cross a 64 byte boundary.  If
2002 	 * our UD header is bigger than the space available up to the
2003 	 * next 64 byte boundary in the WQE, use two inline data
2004 	 * segments to hold the UD header.
2005 	 */
2006 	spc = MLX4_INLINE_ALIGN -
2007 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2008 	if (header_size <= spc) {
2009 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2010 		memcpy(inl + 1, sqp->header_buf, header_size);
2011 		i = 1;
2012 	} else {
2013 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2014 		memcpy(inl + 1, sqp->header_buf, spc);
2015 
2016 		inl = (void *) (inl + 1) + spc;
2017 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2018 		/*
2019 		 * Need a barrier here to make sure all the data is
2020 		 * visible before the byte_count field is set.
2021 		 * Otherwise the HCA prefetcher could grab the 64-byte
2022 		 * chunk with this inline segment and get a valid (!=
2023 		 * 0xffffffff) byte count but stale data, and end up
2024 		 * generating a packet with bad headers.
2025 		 *
2026 		 * The first inline segment's byte_count field doesn't
2027 		 * need a barrier, because it comes after a
2028 		 * control/MLX segment and therefore is at an offset
2029 		 * of 16 mod 64.
2030 		 */
2031 		wmb();
2032 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2033 		i = 2;
2034 	}
2035 
2036 	*mlx_seg_len =
2037 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2038 	return 0;
2039 }
2040 
2041 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
2042 			    void *wqe, unsigned *mlx_seg_len)
2043 {
2044 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
2045 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2046 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2047 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2048 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2049 	union ib_gid sgid;
2050 	u16 pkey;
2051 	int send_size;
2052 	int header_size;
2053 	int spc;
2054 	int i;
2055 	int err = 0;
2056 	u16 vlan = 0xffff;
2057 	bool is_eth;
2058 	bool is_vlan = false;
2059 	bool is_grh;
2060 
2061 	send_size = 0;
2062 	for (i = 0; i < wr->num_sge; ++i)
2063 		send_size += wr->sg_list[i].length;
2064 
2065 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2066 	is_grh = mlx4_ib_ah_grh_present(ah);
2067 	if (is_eth) {
2068 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2069 			/* When multi-function is enabled, the ib_core gid
2070 			 * indexes don't necessarily match the hw ones, so
2071 			 * we must use our own cache */
2072 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2073 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
2074 							   ah->av.ib.gid_index, &sgid.raw[0]);
2075 			if (err)
2076 				return err;
2077 		} else  {
2078 			err = ib_get_cached_gid(ib_dev,
2079 						be32_to_cpu(ah->av.ib.port_pd) >> 24,
2080 						ah->av.ib.gid_index, &sgid);
2081 			if (err)
2082 				return err;
2083 		}
2084 
2085 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
2086 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
2087 			is_vlan = 1;
2088 		}
2089 	}
2090 	ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh, 0, &sqp->ud_header);
2091 
2092 	if (!is_eth) {
2093 		sqp->ud_header.lrh.service_level =
2094 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2095 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
2096 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2097 	}
2098 
2099 	if (is_grh) {
2100 		sqp->ud_header.grh.traffic_class =
2101 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2102 		sqp->ud_header.grh.flow_label    =
2103 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
2104 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
2105 		if (is_eth)
2106 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
2107 		else {
2108 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2109 			/* When multi-function is enabled, the ib_core gid
2110 			 * indexes don't necessarily match the hw ones, so
2111 			 * we must use our own cache */
2112 			sqp->ud_header.grh.source_gid.global.subnet_prefix =
2113 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2114 						       subnet_prefix;
2115 			sqp->ud_header.grh.source_gid.global.interface_id =
2116 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2117 					       guid_cache[ah->av.ib.gid_index];
2118 		} else
2119 			ib_get_cached_gid(ib_dev,
2120 					  be32_to_cpu(ah->av.ib.port_pd) >> 24,
2121 					  ah->av.ib.gid_index,
2122 					  &sqp->ud_header.grh.source_gid);
2123 		}
2124 		memcpy(sqp->ud_header.grh.destination_gid.raw,
2125 		       ah->av.ib.dgid, 16);
2126 	}
2127 
2128 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2129 
2130 	if (!is_eth) {
2131 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
2132 					  (sqp->ud_header.lrh.destination_lid ==
2133 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
2134 					  (sqp->ud_header.lrh.service_level << 8));
2135 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
2136 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
2137 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
2138 	}
2139 
2140 	switch (wr->opcode) {
2141 	case IB_WR_SEND:
2142 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
2143 		sqp->ud_header.immediate_present = 0;
2144 		break;
2145 	case IB_WR_SEND_WITH_IMM:
2146 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2147 		sqp->ud_header.immediate_present = 1;
2148 		sqp->ud_header.immediate_data    = wr->ex.imm_data;
2149 		break;
2150 	default:
2151 		return -EINVAL;
2152 	}
2153 
2154 	if (is_eth) {
2155 		u8 *smac;
2156 		struct in6_addr in6;
2157 
2158 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
2159 
2160 		mlx->sched_prio = cpu_to_be16(pcp);
2161 
2162 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
2163 		/* FIXME: cache smac value? */
2164 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
2165 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
2166 		memcpy(&in6, sgid.raw, sizeof(in6));
2167 
2168 		if (!mlx4_is_mfunc(to_mdev(ib_dev)->dev))
2169 			smac = to_mdev(sqp->qp.ibqp.device)->
2170 				iboe.netdevs[sqp->qp.port - 1]->dev_addr;
2171 		else	/* use the src mac of the tunnel */
2172 			smac = ah->av.eth.s_mac;
2173 		memcpy(sqp->ud_header.eth.smac_h, smac, 6);
2174 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
2175 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
2176 		if (!is_vlan) {
2177 			sqp->ud_header.eth.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2178 		} else {
2179 			sqp->ud_header.vlan.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2180 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
2181 		}
2182 	} else {
2183 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
2184 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
2185 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
2186 	}
2187 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
2188 	if (!sqp->qp.ibqp.qp_num)
2189 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
2190 	else
2191 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
2192 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2193 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2194 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2195 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
2196 					       sqp->qkey : wr->wr.ud.remote_qkey);
2197 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
2198 
2199 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2200 
2201 	if (0) {
2202 		pr_err("built UD header of size %d:\n", header_size);
2203 		for (i = 0; i < header_size / 4; ++i) {
2204 			if (i % 8 == 0)
2205 				pr_err("  [%02x] ", i * 4);
2206 			pr_cont(" %08x",
2207 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
2208 			if ((i + 1) % 8 == 0)
2209 				pr_cont("\n");
2210 		}
2211 		pr_err("\n");
2212 	}
2213 
2214 	/*
2215 	 * Inline data segments may not cross a 64 byte boundary.  If
2216 	 * our UD header is bigger than the space available up to the
2217 	 * next 64 byte boundary in the WQE, use two inline data
2218 	 * segments to hold the UD header.
2219 	 */
2220 	spc = MLX4_INLINE_ALIGN -
2221 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2222 	if (header_size <= spc) {
2223 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2224 		memcpy(inl + 1, sqp->header_buf, header_size);
2225 		i = 1;
2226 	} else {
2227 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2228 		memcpy(inl + 1, sqp->header_buf, spc);
2229 
2230 		inl = (void *) (inl + 1) + spc;
2231 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2232 		/*
2233 		 * Need a barrier here to make sure all the data is
2234 		 * visible before the byte_count field is set.
2235 		 * Otherwise the HCA prefetcher could grab the 64-byte
2236 		 * chunk with this inline segment and get a valid (!=
2237 		 * 0xffffffff) byte count but stale data, and end up
2238 		 * generating a packet with bad headers.
2239 		 *
2240 		 * The first inline segment's byte_count field doesn't
2241 		 * need a barrier, because it comes after a
2242 		 * control/MLX segment and therefore is at an offset
2243 		 * of 16 mod 64.
2244 		 */
2245 		wmb();
2246 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2247 		i = 2;
2248 	}
2249 
2250 	*mlx_seg_len =
2251 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2252 	return 0;
2253 }
2254 
2255 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2256 {
2257 	unsigned cur;
2258 	struct mlx4_ib_cq *cq;
2259 
2260 	cur = wq->head - wq->tail;
2261 	if (likely(cur + nreq < wq->max_post))
2262 		return 0;
2263 
2264 	cq = to_mcq(ib_cq);
2265 	spin_lock(&cq->lock);
2266 	cur = wq->head - wq->tail;
2267 	spin_unlock(&cq->lock);
2268 
2269 	return cur + nreq >= wq->max_post;
2270 }
2271 
2272 static __be32 convert_access(int acc)
2273 {
2274 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
2275 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
2276 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
2277 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
2278 	       (acc & IB_ACCESS_REMOTE_READ   ?
2279 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
2280 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
2281 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
2282 }
2283 
2284 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr)
2285 {
2286 	struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
2287 	int i;
2288 
2289 	for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i)
2290 		mfrpl->mapped_page_list[i] =
2291 			cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] |
2292 				    MLX4_MTT_FLAG_PRESENT);
2293 
2294 	fseg->flags		= convert_access(wr->wr.fast_reg.access_flags);
2295 	fseg->mem_key		= cpu_to_be32(wr->wr.fast_reg.rkey);
2296 	fseg->buf_list		= cpu_to_be64(mfrpl->map);
2297 	fseg->start_addr	= cpu_to_be64(wr->wr.fast_reg.iova_start);
2298 	fseg->reg_len		= cpu_to_be64(wr->wr.fast_reg.length);
2299 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
2300 	fseg->page_size		= cpu_to_be32(wr->wr.fast_reg.page_shift);
2301 	fseg->reserved[0]	= 0;
2302 	fseg->reserved[1]	= 0;
2303 }
2304 
2305 static void set_bind_seg(struct mlx4_wqe_bind_seg *bseg, struct ib_send_wr *wr)
2306 {
2307 	bseg->flags1 =
2308 		convert_access(wr->wr.bind_mw.bind_info.mw_access_flags) &
2309 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ  |
2310 			    MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE |
2311 			    MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC);
2312 	bseg->flags2 = 0;
2313 	if (wr->wr.bind_mw.mw->type == IB_MW_TYPE_2)
2314 		bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_TYPE_2);
2315 	if (wr->wr.bind_mw.bind_info.mw_access_flags & IB_ZERO_BASED)
2316 		bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_ZERO_BASED);
2317 	bseg->new_rkey = cpu_to_be32(wr->wr.bind_mw.rkey);
2318 	bseg->lkey = cpu_to_be32(wr->wr.bind_mw.bind_info.mr->lkey);
2319 	bseg->addr = cpu_to_be64(wr->wr.bind_mw.bind_info.addr);
2320 	bseg->length = cpu_to_be64(wr->wr.bind_mw.bind_info.length);
2321 }
2322 
2323 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2324 {
2325 	memset(iseg, 0, sizeof(*iseg));
2326 	iseg->mem_key = cpu_to_be32(rkey);
2327 }
2328 
2329 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2330 					  u64 remote_addr, u32 rkey)
2331 {
2332 	rseg->raddr    = cpu_to_be64(remote_addr);
2333 	rseg->rkey     = cpu_to_be32(rkey);
2334 	rseg->reserved = 0;
2335 }
2336 
2337 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
2338 {
2339 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2340 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
2341 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
2342 	} else if (wr->opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2343 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2344 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add_mask);
2345 	} else {
2346 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2347 		aseg->compare  = 0;
2348 	}
2349 
2350 }
2351 
2352 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2353 				  struct ib_send_wr *wr)
2354 {
2355 	aseg->swap_add		= cpu_to_be64(wr->wr.atomic.swap);
2356 	aseg->swap_add_mask	= cpu_to_be64(wr->wr.atomic.swap_mask);
2357 	aseg->compare		= cpu_to_be64(wr->wr.atomic.compare_add);
2358 	aseg->compare_mask	= cpu_to_be64(wr->wr.atomic.compare_add_mask);
2359 }
2360 
2361 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2362 			     struct ib_send_wr *wr)
2363 {
2364 	memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
2365 	dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2366 	dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2367 	dseg->vlan = to_mah(wr->wr.ud.ah)->av.eth.vlan;
2368 	memcpy(dseg->mac, to_mah(wr->wr.ud.ah)->av.eth.mac, 6);
2369 }
2370 
2371 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2372 				    struct mlx4_wqe_datagram_seg *dseg,
2373 				    struct ib_send_wr *wr, enum ib_qp_type qpt)
2374 {
2375 	union mlx4_ext_av *av = &to_mah(wr->wr.ud.ah)->av;
2376 	struct mlx4_av sqp_av = {0};
2377 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
2378 
2379 	/* force loopback */
2380 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2381 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2382 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2383 			cpu_to_be32(0xf0000000);
2384 
2385 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2386 	/* This function used only for sending on QP1 proxies */
2387 	dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2388 	/* Use QKEY from the QP context, which is set by master */
2389 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2390 }
2391 
2392 static void build_tunnel_header(struct ib_send_wr *wr, void *wqe, unsigned *mlx_seg_len)
2393 {
2394 	struct mlx4_wqe_inline_seg *inl = wqe;
2395 	struct mlx4_ib_tunnel_header hdr;
2396 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2397 	int spc;
2398 	int i;
2399 
2400 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2401 	hdr.remote_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2402 	hdr.pkey_index = cpu_to_be16(wr->wr.ud.pkey_index);
2403 	hdr.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2404 	memcpy(hdr.mac, ah->av.eth.mac, 6);
2405 	hdr.vlan = ah->av.eth.vlan;
2406 
2407 	spc = MLX4_INLINE_ALIGN -
2408 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2409 	if (sizeof (hdr) <= spc) {
2410 		memcpy(inl + 1, &hdr, sizeof (hdr));
2411 		wmb();
2412 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
2413 		i = 1;
2414 	} else {
2415 		memcpy(inl + 1, &hdr, spc);
2416 		wmb();
2417 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2418 
2419 		inl = (void *) (inl + 1) + spc;
2420 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2421 		wmb();
2422 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
2423 		i = 2;
2424 	}
2425 
2426 	*mlx_seg_len =
2427 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2428 }
2429 
2430 static void set_mlx_icrc_seg(void *dseg)
2431 {
2432 	u32 *t = dseg;
2433 	struct mlx4_wqe_inline_seg *iseg = dseg;
2434 
2435 	t[1] = 0;
2436 
2437 	/*
2438 	 * Need a barrier here before writing the byte_count field to
2439 	 * make sure that all the data is visible before the
2440 	 * byte_count field is set.  Otherwise, if the segment begins
2441 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2442 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2443 	 * stale data, and end up sending the wrong data.
2444 	 */
2445 	wmb();
2446 
2447 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2448 }
2449 
2450 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2451 {
2452 	dseg->lkey       = cpu_to_be32(sg->lkey);
2453 	dseg->addr       = cpu_to_be64(sg->addr);
2454 
2455 	/*
2456 	 * Need a barrier here before writing the byte_count field to
2457 	 * make sure that all the data is visible before the
2458 	 * byte_count field is set.  Otherwise, if the segment begins
2459 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2460 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2461 	 * stale data, and end up sending the wrong data.
2462 	 */
2463 	wmb();
2464 
2465 	dseg->byte_count = cpu_to_be32(sg->length);
2466 }
2467 
2468 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2469 {
2470 	dseg->byte_count = cpu_to_be32(sg->length);
2471 	dseg->lkey       = cpu_to_be32(sg->lkey);
2472 	dseg->addr       = cpu_to_be64(sg->addr);
2473 }
2474 
2475 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr,
2476 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2477 			 __be32 *lso_hdr_sz, __be32 *blh)
2478 {
2479 	unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
2480 
2481 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2482 		*blh = cpu_to_be32(1 << 6);
2483 
2484 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2485 		     wr->num_sge > qp->sq.max_gs - (halign >> 4)))
2486 		return -EINVAL;
2487 
2488 	memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
2489 
2490 	*lso_hdr_sz  = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 |
2491 				   wr->wr.ud.hlen);
2492 	*lso_seg_len = halign;
2493 	return 0;
2494 }
2495 
2496 static __be32 send_ieth(struct ib_send_wr *wr)
2497 {
2498 	switch (wr->opcode) {
2499 	case IB_WR_SEND_WITH_IMM:
2500 	case IB_WR_RDMA_WRITE_WITH_IMM:
2501 		return wr->ex.imm_data;
2502 
2503 	case IB_WR_SEND_WITH_INV:
2504 		return cpu_to_be32(wr->ex.invalidate_rkey);
2505 
2506 	default:
2507 		return 0;
2508 	}
2509 }
2510 
2511 static void add_zero_len_inline(void *wqe)
2512 {
2513 	struct mlx4_wqe_inline_seg *inl = wqe;
2514 	memset(wqe, 0, 16);
2515 	inl->byte_count = cpu_to_be32(1 << 31);
2516 }
2517 
2518 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2519 		      struct ib_send_wr **bad_wr)
2520 {
2521 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2522 	void *wqe;
2523 	struct mlx4_wqe_ctrl_seg *ctrl;
2524 	struct mlx4_wqe_data_seg *dseg;
2525 	unsigned long flags;
2526 	int nreq;
2527 	int err = 0;
2528 	unsigned ind;
2529 	int uninitialized_var(stamp);
2530 	int uninitialized_var(size);
2531 	unsigned uninitialized_var(seglen);
2532 	__be32 dummy;
2533 	__be32 *lso_wqe;
2534 	__be32 uninitialized_var(lso_hdr_sz);
2535 	__be32 blh;
2536 	int i;
2537 
2538 	spin_lock_irqsave(&qp->sq.lock, flags);
2539 
2540 	ind = qp->sq_next_wqe;
2541 
2542 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2543 		lso_wqe = &dummy;
2544 		blh = 0;
2545 
2546 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2547 			err = -ENOMEM;
2548 			*bad_wr = wr;
2549 			goto out;
2550 		}
2551 
2552 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
2553 			err = -EINVAL;
2554 			*bad_wr = wr;
2555 			goto out;
2556 		}
2557 
2558 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
2559 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
2560 
2561 		ctrl->srcrb_flags =
2562 			(wr->send_flags & IB_SEND_SIGNALED ?
2563 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
2564 			(wr->send_flags & IB_SEND_SOLICITED ?
2565 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
2566 			((wr->send_flags & IB_SEND_IP_CSUM) ?
2567 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
2568 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
2569 			qp->sq_signal_bits;
2570 
2571 		ctrl->imm = send_ieth(wr);
2572 
2573 		wqe += sizeof *ctrl;
2574 		size = sizeof *ctrl / 16;
2575 
2576 		switch (qp->mlx4_ib_qp_type) {
2577 		case MLX4_IB_QPT_RC:
2578 		case MLX4_IB_QPT_UC:
2579 			switch (wr->opcode) {
2580 			case IB_WR_ATOMIC_CMP_AND_SWP:
2581 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2582 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
2583 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2584 					      wr->wr.atomic.rkey);
2585 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2586 
2587 				set_atomic_seg(wqe, wr);
2588 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
2589 
2590 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2591 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
2592 
2593 				break;
2594 
2595 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2596 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2597 					      wr->wr.atomic.rkey);
2598 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2599 
2600 				set_masked_atomic_seg(wqe, wr);
2601 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
2602 
2603 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2604 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
2605 
2606 				break;
2607 
2608 			case IB_WR_RDMA_READ:
2609 			case IB_WR_RDMA_WRITE:
2610 			case IB_WR_RDMA_WRITE_WITH_IMM:
2611 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2612 					      wr->wr.rdma.rkey);
2613 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2614 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
2615 				break;
2616 
2617 			case IB_WR_LOCAL_INV:
2618 				ctrl->srcrb_flags |=
2619 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2620 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
2621 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
2622 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
2623 				break;
2624 
2625 			case IB_WR_FAST_REG_MR:
2626 				ctrl->srcrb_flags |=
2627 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2628 				set_fmr_seg(wqe, wr);
2629 				wqe  += sizeof (struct mlx4_wqe_fmr_seg);
2630 				size += sizeof (struct mlx4_wqe_fmr_seg) / 16;
2631 				break;
2632 
2633 			case IB_WR_BIND_MW:
2634 				ctrl->srcrb_flags |=
2635 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2636 				set_bind_seg(wqe, wr);
2637 				wqe  += sizeof(struct mlx4_wqe_bind_seg);
2638 				size += sizeof(struct mlx4_wqe_bind_seg) / 16;
2639 				break;
2640 			default:
2641 				/* No extra segments required for sends */
2642 				break;
2643 			}
2644 			break;
2645 
2646 		case MLX4_IB_QPT_TUN_SMI_OWNER:
2647 			err =  build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2648 			if (unlikely(err)) {
2649 				*bad_wr = wr;
2650 				goto out;
2651 			}
2652 			wqe  += seglen;
2653 			size += seglen / 16;
2654 			break;
2655 		case MLX4_IB_QPT_TUN_SMI:
2656 		case MLX4_IB_QPT_TUN_GSI:
2657 			/* this is a UD qp used in MAD responses to slaves. */
2658 			set_datagram_seg(wqe, wr);
2659 			/* set the forced-loopback bit in the data seg av */
2660 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
2661 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2662 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2663 			break;
2664 		case MLX4_IB_QPT_UD:
2665 			set_datagram_seg(wqe, wr);
2666 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2667 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2668 
2669 			if (wr->opcode == IB_WR_LSO) {
2670 				err = build_lso_seg(wqe, wr, qp, &seglen, &lso_hdr_sz, &blh);
2671 				if (unlikely(err)) {
2672 					*bad_wr = wr;
2673 					goto out;
2674 				}
2675 				lso_wqe = (__be32 *) wqe;
2676 				wqe  += seglen;
2677 				size += seglen / 16;
2678 			}
2679 			break;
2680 
2681 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
2682 			if (unlikely(!mlx4_is_master(to_mdev(ibqp->device)->dev))) {
2683 				err = -ENOSYS;
2684 				*bad_wr = wr;
2685 				goto out;
2686 			}
2687 			err = build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2688 			if (unlikely(err)) {
2689 				*bad_wr = wr;
2690 				goto out;
2691 			}
2692 			wqe  += seglen;
2693 			size += seglen / 16;
2694 			/* to start tunnel header on a cache-line boundary */
2695 			add_zero_len_inline(wqe);
2696 			wqe += 16;
2697 			size++;
2698 			build_tunnel_header(wr, wqe, &seglen);
2699 			wqe  += seglen;
2700 			size += seglen / 16;
2701 			break;
2702 		case MLX4_IB_QPT_PROXY_SMI:
2703 			/* don't allow QP0 sends on guests */
2704 			err = -ENOSYS;
2705 			*bad_wr = wr;
2706 			goto out;
2707 		case MLX4_IB_QPT_PROXY_GSI:
2708 			/* If we are tunneling special qps, this is a UD qp.
2709 			 * In this case we first add a UD segment targeting
2710 			 * the tunnel qp, and then add a header with address
2711 			 * information */
2712 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe, wr, ibqp->qp_type);
2713 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2714 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2715 			build_tunnel_header(wr, wqe, &seglen);
2716 			wqe  += seglen;
2717 			size += seglen / 16;
2718 			break;
2719 
2720 		case MLX4_IB_QPT_SMI:
2721 		case MLX4_IB_QPT_GSI:
2722 			err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
2723 			if (unlikely(err)) {
2724 				*bad_wr = wr;
2725 				goto out;
2726 			}
2727 			wqe  += seglen;
2728 			size += seglen / 16;
2729 			break;
2730 
2731 		default:
2732 			break;
2733 		}
2734 
2735 		/*
2736 		 * Write data segments in reverse order, so as to
2737 		 * overwrite cacheline stamp last within each
2738 		 * cacheline.  This avoids issues with WQE
2739 		 * prefetching.
2740 		 */
2741 
2742 		dseg = wqe;
2743 		dseg += wr->num_sge - 1;
2744 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
2745 
2746 		/* Add one more inline data segment for ICRC for MLX sends */
2747 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2748 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
2749 			     qp->mlx4_ib_qp_type &
2750 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
2751 			set_mlx_icrc_seg(dseg + 1);
2752 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
2753 		}
2754 
2755 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
2756 			set_data_seg(dseg, wr->sg_list + i);
2757 
2758 		/*
2759 		 * Possibly overwrite stamping in cacheline with LSO
2760 		 * segment only after making sure all data segments
2761 		 * are written.
2762 		 */
2763 		wmb();
2764 		*lso_wqe = lso_hdr_sz;
2765 
2766 		ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
2767 				    MLX4_WQE_CTRL_FENCE : 0) | size;
2768 
2769 		/*
2770 		 * Make sure descriptor is fully written before
2771 		 * setting ownership bit (because HW can start
2772 		 * executing as soon as we do).
2773 		 */
2774 		wmb();
2775 
2776 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
2777 			*bad_wr = wr;
2778 			err = -EINVAL;
2779 			goto out;
2780 		}
2781 
2782 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
2783 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
2784 
2785 		stamp = ind + qp->sq_spare_wqes;
2786 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
2787 
2788 		/*
2789 		 * We can improve latency by not stamping the last
2790 		 * send queue WQE until after ringing the doorbell, so
2791 		 * only stamp here if there are still more WQEs to post.
2792 		 *
2793 		 * Same optimization applies to padding with NOP wqe
2794 		 * in case of WQE shrinking (used to prevent wrap-around
2795 		 * in the middle of WR).
2796 		 */
2797 		if (wr->next) {
2798 			stamp_send_wqe(qp, stamp, size * 16);
2799 			ind = pad_wraparound(qp, ind);
2800 		}
2801 	}
2802 
2803 out:
2804 	if (likely(nreq)) {
2805 		qp->sq.head += nreq;
2806 
2807 		/*
2808 		 * Make sure that descriptors are written before
2809 		 * doorbell record.
2810 		 */
2811 		wmb();
2812 
2813 		writel(qp->doorbell_qpn,
2814 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
2815 
2816 		/*
2817 		 * Make sure doorbells don't leak out of SQ spinlock
2818 		 * and reach the HCA out of order.
2819 		 */
2820 		mmiowb();
2821 
2822 		stamp_send_wqe(qp, stamp, size * 16);
2823 
2824 		ind = pad_wraparound(qp, ind);
2825 		qp->sq_next_wqe = ind;
2826 	}
2827 
2828 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2829 
2830 	return err;
2831 }
2832 
2833 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2834 		      struct ib_recv_wr **bad_wr)
2835 {
2836 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2837 	struct mlx4_wqe_data_seg *scat;
2838 	unsigned long flags;
2839 	int err = 0;
2840 	int nreq;
2841 	int ind;
2842 	int max_gs;
2843 	int i;
2844 
2845 	max_gs = qp->rq.max_gs;
2846 	spin_lock_irqsave(&qp->rq.lock, flags);
2847 
2848 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
2849 
2850 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2851 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2852 			err = -ENOMEM;
2853 			*bad_wr = wr;
2854 			goto out;
2855 		}
2856 
2857 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2858 			err = -EINVAL;
2859 			*bad_wr = wr;
2860 			goto out;
2861 		}
2862 
2863 		scat = get_recv_wqe(qp, ind);
2864 
2865 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
2866 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
2867 			ib_dma_sync_single_for_device(ibqp->device,
2868 						      qp->sqp_proxy_rcv[ind].map,
2869 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
2870 						      DMA_FROM_DEVICE);
2871 			scat->byte_count =
2872 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
2873 			/* use dma lkey from upper layer entry */
2874 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
2875 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
2876 			scat++;
2877 			max_gs--;
2878 		}
2879 
2880 		for (i = 0; i < wr->num_sge; ++i)
2881 			__set_data_seg(scat + i, wr->sg_list + i);
2882 
2883 		if (i < max_gs) {
2884 			scat[i].byte_count = 0;
2885 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
2886 			scat[i].addr       = 0;
2887 		}
2888 
2889 		qp->rq.wrid[ind] = wr->wr_id;
2890 
2891 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
2892 	}
2893 
2894 out:
2895 	if (likely(nreq)) {
2896 		qp->rq.head += nreq;
2897 
2898 		/*
2899 		 * Make sure that descriptors are written before
2900 		 * doorbell record.
2901 		 */
2902 		wmb();
2903 
2904 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
2905 	}
2906 
2907 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2908 
2909 	return err;
2910 }
2911 
2912 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
2913 {
2914 	switch (mlx4_state) {
2915 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
2916 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
2917 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
2918 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
2919 	case MLX4_QP_STATE_SQ_DRAINING:
2920 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
2921 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
2922 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
2923 	default:		     return -1;
2924 	}
2925 }
2926 
2927 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
2928 {
2929 	switch (mlx4_mig_state) {
2930 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
2931 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
2932 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
2933 	default: return -1;
2934 	}
2935 }
2936 
2937 static int to_ib_qp_access_flags(int mlx4_flags)
2938 {
2939 	int ib_flags = 0;
2940 
2941 	if (mlx4_flags & MLX4_QP_BIT_RRE)
2942 		ib_flags |= IB_ACCESS_REMOTE_READ;
2943 	if (mlx4_flags & MLX4_QP_BIT_RWE)
2944 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
2945 	if (mlx4_flags & MLX4_QP_BIT_RAE)
2946 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
2947 
2948 	return ib_flags;
2949 }
2950 
2951 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
2952 				struct mlx4_qp_path *path)
2953 {
2954 	struct mlx4_dev *dev = ibdev->dev;
2955 	int is_eth;
2956 
2957 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
2958 	ib_ah_attr->port_num	  = path->sched_queue & 0x40 ? 2 : 1;
2959 
2960 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
2961 		return;
2962 
2963 	is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
2964 		IB_LINK_LAYER_ETHERNET;
2965 	if (is_eth)
2966 		ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
2967 		((path->sched_queue & 4) << 1);
2968 	else
2969 		ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
2970 
2971 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
2972 	ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
2973 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
2974 	ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
2975 	if (ib_ah_attr->ah_flags) {
2976 		ib_ah_attr->grh.sgid_index = path->mgid_index;
2977 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
2978 		ib_ah_attr->grh.traffic_class =
2979 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
2980 		ib_ah_attr->grh.flow_label =
2981 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
2982 		memcpy(ib_ah_attr->grh.dgid.raw,
2983 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
2984 	}
2985 }
2986 
2987 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
2988 		     struct ib_qp_init_attr *qp_init_attr)
2989 {
2990 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2991 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2992 	struct mlx4_qp_context context;
2993 	int mlx4_state;
2994 	int err = 0;
2995 
2996 	mutex_lock(&qp->mutex);
2997 
2998 	if (qp->state == IB_QPS_RESET) {
2999 		qp_attr->qp_state = IB_QPS_RESET;
3000 		goto done;
3001 	}
3002 
3003 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3004 	if (err) {
3005 		err = -EINVAL;
3006 		goto out;
3007 	}
3008 
3009 	mlx4_state = be32_to_cpu(context.flags) >> 28;
3010 
3011 	qp->state		     = to_ib_qp_state(mlx4_state);
3012 	qp_attr->qp_state	     = qp->state;
3013 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
3014 	qp_attr->path_mig_state	     =
3015 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3016 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
3017 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3018 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
3019 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
3020 	qp_attr->qp_access_flags     =
3021 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
3022 
3023 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3024 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3025 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3026 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3027 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
3028 	}
3029 
3030 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3031 	if (qp_attr->qp_state == IB_QPS_INIT)
3032 		qp_attr->port_num = qp->port;
3033 	else
3034 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3035 
3036 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3037 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3038 
3039 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3040 
3041 	qp_attr->max_dest_rd_atomic =
3042 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3043 	qp_attr->min_rnr_timer	    =
3044 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3045 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
3046 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
3047 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
3048 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
3049 
3050 done:
3051 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
3052 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
3053 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
3054 
3055 	if (!ibqp->uobject) {
3056 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
3057 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
3058 	} else {
3059 		qp_attr->cap.max_send_wr  = 0;
3060 		qp_attr->cap.max_send_sge = 0;
3061 	}
3062 
3063 	/*
3064 	 * We don't support inline sends for kernel QPs (yet), and we
3065 	 * don't know what userspace's value should be.
3066 	 */
3067 	qp_attr->cap.max_inline_data = 0;
3068 
3069 	qp_init_attr->cap	     = qp_attr->cap;
3070 
3071 	qp_init_attr->create_flags = 0;
3072 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3073 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3074 
3075 	if (qp->flags & MLX4_IB_QP_LSO)
3076 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
3077 
3078 	if (qp->flags & MLX4_IB_QP_NETIF)
3079 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
3080 
3081 	qp_init_attr->sq_sig_type =
3082 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
3083 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3084 
3085 out:
3086 	mutex_unlock(&qp->mutex);
3087 	return err;
3088 }
3089 
3090