xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision ca460cc2)
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 qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
612 {
613 	int i;
614 	for (i = 0; i < dev->caps.num_ports; i++) {
615 		if (qpn == dev->caps.qp0_proxy[i])
616 			return !!dev->caps.qp0_qkey[i];
617 	}
618 	return 0;
619 }
620 
621 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
622 			    struct ib_qp_init_attr *init_attr,
623 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp,
624 			    gfp_t gfp)
625 {
626 	int qpn;
627 	int err;
628 	struct mlx4_ib_sqp *sqp;
629 	struct mlx4_ib_qp *qp;
630 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
631 
632 	/* When tunneling special qps, we use a plain UD qp */
633 	if (sqpn) {
634 		if (mlx4_is_mfunc(dev->dev) &&
635 		    (!mlx4_is_master(dev->dev) ||
636 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
637 			if (init_attr->qp_type == IB_QPT_GSI)
638 				qp_type = MLX4_IB_QPT_PROXY_GSI;
639 			else {
640 				if (mlx4_is_master(dev->dev) ||
641 				    qp0_enabled_vf(dev->dev, sqpn))
642 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
643 				else
644 					qp_type = MLX4_IB_QPT_PROXY_SMI;
645 			}
646 		}
647 		qpn = sqpn;
648 		/* add extra sg entry for tunneling */
649 		init_attr->cap.max_recv_sge++;
650 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
651 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
652 			container_of(init_attr,
653 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
654 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
655 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
656 		    !mlx4_is_master(dev->dev))
657 			return -EINVAL;
658 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
659 			qp_type = MLX4_IB_QPT_TUN_GSI;
660 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
661 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
662 					     tnl_init->port))
663 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
664 		else
665 			qp_type = MLX4_IB_QPT_TUN_SMI;
666 		/* we are definitely in the PPF here, since we are creating
667 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
668 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
669 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
670 		sqpn = qpn;
671 	}
672 
673 	if (!*caller_qp) {
674 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
675 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
676 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
677 			sqp = kzalloc(sizeof (struct mlx4_ib_sqp), gfp);
678 			if (!sqp)
679 				return -ENOMEM;
680 			qp = &sqp->qp;
681 			qp->pri.vid = 0xFFFF;
682 			qp->alt.vid = 0xFFFF;
683 		} else {
684 			qp = kzalloc(sizeof (struct mlx4_ib_qp), gfp);
685 			if (!qp)
686 				return -ENOMEM;
687 			qp->pri.vid = 0xFFFF;
688 			qp->alt.vid = 0xFFFF;
689 		}
690 	} else
691 		qp = *caller_qp;
692 
693 	qp->mlx4_ib_qp_type = qp_type;
694 
695 	mutex_init(&qp->mutex);
696 	spin_lock_init(&qp->sq.lock);
697 	spin_lock_init(&qp->rq.lock);
698 	INIT_LIST_HEAD(&qp->gid_list);
699 	INIT_LIST_HEAD(&qp->steering_rules);
700 
701 	qp->state	 = IB_QPS_RESET;
702 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
703 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
704 
705 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
706 	if (err)
707 		goto err;
708 
709 	if (pd->uobject) {
710 		struct mlx4_ib_create_qp ucmd;
711 
712 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
713 			err = -EFAULT;
714 			goto err;
715 		}
716 
717 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
718 
719 		err = set_user_sq_size(dev, qp, &ucmd);
720 		if (err)
721 			goto err;
722 
723 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
724 				       qp->buf_size, 0, 0);
725 		if (IS_ERR(qp->umem)) {
726 			err = PTR_ERR(qp->umem);
727 			goto err;
728 		}
729 
730 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
731 				    ilog2(qp->umem->page_size), &qp->mtt);
732 		if (err)
733 			goto err_buf;
734 
735 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
736 		if (err)
737 			goto err_mtt;
738 
739 		if (qp_has_rq(init_attr)) {
740 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
741 						  ucmd.db_addr, &qp->db);
742 			if (err)
743 				goto err_mtt;
744 		}
745 	} else {
746 		qp->sq_no_prefetch = 0;
747 
748 		if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
749 			qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
750 
751 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
752 			qp->flags |= MLX4_IB_QP_LSO;
753 
754 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
755 			if (dev->steering_support ==
756 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
757 				qp->flags |= MLX4_IB_QP_NETIF;
758 			else
759 				goto err;
760 		}
761 
762 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
763 		if (err)
764 			goto err;
765 
766 		if (qp_has_rq(init_attr)) {
767 			err = mlx4_db_alloc(dev->dev, &qp->db, 0, gfp);
768 			if (err)
769 				goto err;
770 
771 			*qp->db.db = 0;
772 		}
773 
774 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf, gfp)) {
775 			err = -ENOMEM;
776 			goto err_db;
777 		}
778 
779 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
780 				    &qp->mtt);
781 		if (err)
782 			goto err_buf;
783 
784 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf, gfp);
785 		if (err)
786 			goto err_mtt;
787 
788 		qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), gfp);
789 		qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), gfp);
790 		if (!qp->sq.wrid || !qp->rq.wrid) {
791 			err = -ENOMEM;
792 			goto err_wrid;
793 		}
794 	}
795 
796 	if (sqpn) {
797 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
798 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
799 			if (alloc_proxy_bufs(pd->device, qp)) {
800 				err = -ENOMEM;
801 				goto err_wrid;
802 			}
803 		}
804 	} else {
805 		/* Raw packet QPNs must be aligned to 8 bits. If not, the WQE
806 		 * BlueFlame setup flow wrongly causes VLAN insertion. */
807 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
808 			err = mlx4_qp_reserve_range(dev->dev, 1, 1 << 8, &qpn);
809 		else
810 			if (qp->flags & MLX4_IB_QP_NETIF)
811 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
812 			else
813 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
814 							    &qpn);
815 		if (err)
816 			goto err_proxy;
817 	}
818 
819 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp, gfp);
820 	if (err)
821 		goto err_qpn;
822 
823 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
824 		qp->mqp.qpn |= (1 << 23);
825 
826 	/*
827 	 * Hardware wants QPN written in big-endian order (after
828 	 * shifting) for send doorbell.  Precompute this value to save
829 	 * a little bit when posting sends.
830 	 */
831 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
832 
833 	qp->mqp.event = mlx4_ib_qp_event;
834 	if (!*caller_qp)
835 		*caller_qp = qp;
836 	return 0;
837 
838 err_qpn:
839 	if (!sqpn) {
840 		if (qp->flags & MLX4_IB_QP_NETIF)
841 			mlx4_ib_steer_qp_free(dev, qpn, 1);
842 		else
843 			mlx4_qp_release_range(dev->dev, qpn, 1);
844 	}
845 err_proxy:
846 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
847 		free_proxy_bufs(pd->device, qp);
848 err_wrid:
849 	if (pd->uobject) {
850 		if (qp_has_rq(init_attr))
851 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
852 	} else {
853 		kfree(qp->sq.wrid);
854 		kfree(qp->rq.wrid);
855 	}
856 
857 err_mtt:
858 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
859 
860 err_buf:
861 	if (pd->uobject)
862 		ib_umem_release(qp->umem);
863 	else
864 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
865 
866 err_db:
867 	if (!pd->uobject && qp_has_rq(init_attr))
868 		mlx4_db_free(dev->dev, &qp->db);
869 
870 err:
871 	if (!*caller_qp)
872 		kfree(qp);
873 	return err;
874 }
875 
876 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
877 {
878 	switch (state) {
879 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
880 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
881 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
882 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
883 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
884 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
885 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
886 	default:		return -1;
887 	}
888 }
889 
890 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
891 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
892 {
893 	if (send_cq == recv_cq) {
894 		spin_lock_irq(&send_cq->lock);
895 		__acquire(&recv_cq->lock);
896 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
897 		spin_lock_irq(&send_cq->lock);
898 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
899 	} else {
900 		spin_lock_irq(&recv_cq->lock);
901 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
902 	}
903 }
904 
905 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
906 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
907 {
908 	if (send_cq == recv_cq) {
909 		__release(&recv_cq->lock);
910 		spin_unlock_irq(&send_cq->lock);
911 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
912 		spin_unlock(&recv_cq->lock);
913 		spin_unlock_irq(&send_cq->lock);
914 	} else {
915 		spin_unlock(&send_cq->lock);
916 		spin_unlock_irq(&recv_cq->lock);
917 	}
918 }
919 
920 static void del_gid_entries(struct mlx4_ib_qp *qp)
921 {
922 	struct mlx4_ib_gid_entry *ge, *tmp;
923 
924 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
925 		list_del(&ge->list);
926 		kfree(ge);
927 	}
928 }
929 
930 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
931 {
932 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
933 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
934 	else
935 		return to_mpd(qp->ibqp.pd);
936 }
937 
938 static void get_cqs(struct mlx4_ib_qp *qp,
939 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
940 {
941 	switch (qp->ibqp.qp_type) {
942 	case IB_QPT_XRC_TGT:
943 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
944 		*recv_cq = *send_cq;
945 		break;
946 	case IB_QPT_XRC_INI:
947 		*send_cq = to_mcq(qp->ibqp.send_cq);
948 		*recv_cq = *send_cq;
949 		break;
950 	default:
951 		*send_cq = to_mcq(qp->ibqp.send_cq);
952 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
953 		break;
954 	}
955 }
956 
957 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
958 			      int is_user)
959 {
960 	struct mlx4_ib_cq *send_cq, *recv_cq;
961 
962 	if (qp->state != IB_QPS_RESET) {
963 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
964 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
965 			pr_warn("modify QP %06x to RESET failed.\n",
966 			       qp->mqp.qpn);
967 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
968 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
969 			qp->pri.smac = 0;
970 			qp->pri.smac_port = 0;
971 		}
972 		if (qp->alt.smac) {
973 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
974 			qp->alt.smac = 0;
975 		}
976 		if (qp->pri.vid < 0x1000) {
977 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
978 			qp->pri.vid = 0xFFFF;
979 			qp->pri.candidate_vid = 0xFFFF;
980 			qp->pri.update_vid = 0;
981 		}
982 		if (qp->alt.vid < 0x1000) {
983 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
984 			qp->alt.vid = 0xFFFF;
985 			qp->alt.candidate_vid = 0xFFFF;
986 			qp->alt.update_vid = 0;
987 		}
988 	}
989 
990 	get_cqs(qp, &send_cq, &recv_cq);
991 
992 	mlx4_ib_lock_cqs(send_cq, recv_cq);
993 
994 	if (!is_user) {
995 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
996 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
997 		if (send_cq != recv_cq)
998 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
999 	}
1000 
1001 	mlx4_qp_remove(dev->dev, &qp->mqp);
1002 
1003 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1004 
1005 	mlx4_qp_free(dev->dev, &qp->mqp);
1006 
1007 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1008 		if (qp->flags & MLX4_IB_QP_NETIF)
1009 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1010 		else
1011 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1012 	}
1013 
1014 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1015 
1016 	if (is_user) {
1017 		if (qp->rq.wqe_cnt)
1018 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
1019 					      &qp->db);
1020 		ib_umem_release(qp->umem);
1021 	} else {
1022 		kfree(qp->sq.wrid);
1023 		kfree(qp->rq.wrid);
1024 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1025 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1026 			free_proxy_bufs(&dev->ib_dev, qp);
1027 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1028 		if (qp->rq.wqe_cnt)
1029 			mlx4_db_free(dev->dev, &qp->db);
1030 	}
1031 
1032 	del_gid_entries(qp);
1033 }
1034 
1035 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1036 {
1037 	/* Native or PPF */
1038 	if (!mlx4_is_mfunc(dev->dev) ||
1039 	    (mlx4_is_master(dev->dev) &&
1040 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1041 		return  dev->dev->phys_caps.base_sqpn +
1042 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1043 			attr->port_num - 1;
1044 	}
1045 	/* PF or VF -- creating proxies */
1046 	if (attr->qp_type == IB_QPT_SMI)
1047 		return dev->dev->caps.qp0_proxy[attr->port_num - 1];
1048 	else
1049 		return dev->dev->caps.qp1_proxy[attr->port_num - 1];
1050 }
1051 
1052 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1053 				struct ib_qp_init_attr *init_attr,
1054 				struct ib_udata *udata)
1055 {
1056 	struct mlx4_ib_qp *qp = NULL;
1057 	int err;
1058 	u16 xrcdn = 0;
1059 	gfp_t gfp;
1060 
1061 	gfp = (init_attr->create_flags & MLX4_IB_QP_CREATE_USE_GFP_NOIO) ?
1062 		GFP_NOIO : GFP_KERNEL;
1063 	/*
1064 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1065 	 * and only for kernel UD QPs.
1066 	 */
1067 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1068 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1069 					MLX4_IB_SRIOV_TUNNEL_QP |
1070 					MLX4_IB_SRIOV_SQP |
1071 					MLX4_IB_QP_NETIF |
1072 					MLX4_IB_QP_CREATE_USE_GFP_NOIO))
1073 		return ERR_PTR(-EINVAL);
1074 
1075 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1076 		if (init_attr->qp_type != IB_QPT_UD)
1077 			return ERR_PTR(-EINVAL);
1078 	}
1079 
1080 	if (init_attr->create_flags &&
1081 	    (udata ||
1082 	     ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP | MLX4_IB_QP_CREATE_USE_GFP_NOIO)) &&
1083 	      init_attr->qp_type != IB_QPT_UD) ||
1084 	     ((init_attr->create_flags & MLX4_IB_SRIOV_SQP) &&
1085 	      init_attr->qp_type > IB_QPT_GSI)))
1086 		return ERR_PTR(-EINVAL);
1087 
1088 	switch (init_attr->qp_type) {
1089 	case IB_QPT_XRC_TGT:
1090 		pd = to_mxrcd(init_attr->xrcd)->pd;
1091 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1092 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1093 		/* fall through */
1094 	case IB_QPT_XRC_INI:
1095 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1096 			return ERR_PTR(-ENOSYS);
1097 		init_attr->recv_cq = init_attr->send_cq;
1098 		/* fall through */
1099 	case IB_QPT_RC:
1100 	case IB_QPT_UC:
1101 	case IB_QPT_RAW_PACKET:
1102 		qp = kzalloc(sizeof *qp, gfp);
1103 		if (!qp)
1104 			return ERR_PTR(-ENOMEM);
1105 		qp->pri.vid = 0xFFFF;
1106 		qp->alt.vid = 0xFFFF;
1107 		/* fall through */
1108 	case IB_QPT_UD:
1109 	{
1110 		err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1111 				       udata, 0, &qp, gfp);
1112 		if (err)
1113 			return ERR_PTR(err);
1114 
1115 		qp->ibqp.qp_num = qp->mqp.qpn;
1116 		qp->xrcdn = xrcdn;
1117 
1118 		break;
1119 	}
1120 	case IB_QPT_SMI:
1121 	case IB_QPT_GSI:
1122 	{
1123 		/* Userspace is not allowed to create special QPs: */
1124 		if (udata)
1125 			return ERR_PTR(-EINVAL);
1126 
1127 		err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1128 				       get_sqp_num(to_mdev(pd->device), init_attr),
1129 				       &qp, gfp);
1130 		if (err)
1131 			return ERR_PTR(err);
1132 
1133 		qp->port	= init_attr->port_num;
1134 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
1135 
1136 		break;
1137 	}
1138 	default:
1139 		/* Don't support raw QPs */
1140 		return ERR_PTR(-EINVAL);
1141 	}
1142 
1143 	return &qp->ibqp;
1144 }
1145 
1146 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1147 {
1148 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1149 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1150 	struct mlx4_ib_pd *pd;
1151 
1152 	if (is_qp0(dev, mqp))
1153 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1154 
1155 	if (dev->qp1_proxy[mqp->port - 1] == mqp) {
1156 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1157 		dev->qp1_proxy[mqp->port - 1] = NULL;
1158 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1159 	}
1160 
1161 	pd = get_pd(mqp);
1162 	destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1163 
1164 	if (is_sqp(dev, mqp))
1165 		kfree(to_msqp(mqp));
1166 	else
1167 		kfree(mqp);
1168 
1169 	return 0;
1170 }
1171 
1172 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1173 {
1174 	switch (type) {
1175 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1176 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1177 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1178 	case MLX4_IB_QPT_XRC_INI:
1179 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1180 	case MLX4_IB_QPT_SMI:
1181 	case MLX4_IB_QPT_GSI:
1182 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1183 
1184 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1185 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1186 						MLX4_QP_ST_MLX : -1);
1187 	case MLX4_IB_QPT_PROXY_SMI:
1188 	case MLX4_IB_QPT_TUN_SMI:
1189 	case MLX4_IB_QPT_PROXY_GSI:
1190 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1191 						MLX4_QP_ST_UD : -1);
1192 	default:			return -1;
1193 	}
1194 }
1195 
1196 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1197 				   int attr_mask)
1198 {
1199 	u8 dest_rd_atomic;
1200 	u32 access_flags;
1201 	u32 hw_access_flags = 0;
1202 
1203 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1204 		dest_rd_atomic = attr->max_dest_rd_atomic;
1205 	else
1206 		dest_rd_atomic = qp->resp_depth;
1207 
1208 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1209 		access_flags = attr->qp_access_flags;
1210 	else
1211 		access_flags = qp->atomic_rd_en;
1212 
1213 	if (!dest_rd_atomic)
1214 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1215 
1216 	if (access_flags & IB_ACCESS_REMOTE_READ)
1217 		hw_access_flags |= MLX4_QP_BIT_RRE;
1218 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1219 		hw_access_flags |= MLX4_QP_BIT_RAE;
1220 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1221 		hw_access_flags |= MLX4_QP_BIT_RWE;
1222 
1223 	return cpu_to_be32(hw_access_flags);
1224 }
1225 
1226 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1227 			    int attr_mask)
1228 {
1229 	if (attr_mask & IB_QP_PKEY_INDEX)
1230 		sqp->pkey_index = attr->pkey_index;
1231 	if (attr_mask & IB_QP_QKEY)
1232 		sqp->qkey = attr->qkey;
1233 	if (attr_mask & IB_QP_SQ_PSN)
1234 		sqp->send_psn = attr->sq_psn;
1235 }
1236 
1237 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1238 {
1239 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1240 }
1241 
1242 static int _mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
1243 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1244 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1245 {
1246 	int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
1247 		IB_LINK_LAYER_ETHERNET;
1248 	int vidx;
1249 	int smac_index;
1250 	int err;
1251 
1252 
1253 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
1254 	path->rlid	    = cpu_to_be16(ah->dlid);
1255 	if (ah->static_rate) {
1256 		path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
1257 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1258 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1259 			--path->static_rate;
1260 	} else
1261 		path->static_rate = 0;
1262 
1263 	if (ah->ah_flags & IB_AH_GRH) {
1264 		if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
1265 			pr_err("sgid_index (%u) too large. max is %d\n",
1266 			       ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1267 			return -1;
1268 		}
1269 
1270 		path->grh_mylmc |= 1 << 7;
1271 		path->mgid_index = ah->grh.sgid_index;
1272 		path->hop_limit  = ah->grh.hop_limit;
1273 		path->tclass_flowlabel =
1274 			cpu_to_be32((ah->grh.traffic_class << 20) |
1275 				    (ah->grh.flow_label));
1276 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1277 	}
1278 
1279 	if (is_eth) {
1280 		if (!(ah->ah_flags & IB_AH_GRH))
1281 			return -1;
1282 
1283 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1284 			((port - 1) << 6) | ((ah->sl & 7) << 3);
1285 
1286 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1287 		if (vlan_tag < 0x1000) {
1288 			if (smac_info->vid < 0x1000) {
1289 				/* both valid vlan ids */
1290 				if (smac_info->vid != vlan_tag) {
1291 					/* different VIDs.  unreg old and reg new */
1292 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1293 					if (err)
1294 						return err;
1295 					smac_info->candidate_vid = vlan_tag;
1296 					smac_info->candidate_vlan_index = vidx;
1297 					smac_info->candidate_vlan_port = port;
1298 					smac_info->update_vid = 1;
1299 					path->vlan_index = vidx;
1300 				} else {
1301 					path->vlan_index = smac_info->vlan_index;
1302 				}
1303 			} else {
1304 				/* no current vlan tag in qp */
1305 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1306 				if (err)
1307 					return err;
1308 				smac_info->candidate_vid = vlan_tag;
1309 				smac_info->candidate_vlan_index = vidx;
1310 				smac_info->candidate_vlan_port = port;
1311 				smac_info->update_vid = 1;
1312 				path->vlan_index = vidx;
1313 			}
1314 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1315 			path->fl = 1 << 6;
1316 		} else {
1317 			/* have current vlan tag. unregister it at modify-qp success */
1318 			if (smac_info->vid < 0x1000) {
1319 				smac_info->candidate_vid = 0xFFFF;
1320 				smac_info->update_vid = 1;
1321 			}
1322 		}
1323 
1324 		/* get smac_index for RoCE use.
1325 		 * If no smac was yet assigned, register one.
1326 		 * If one was already assigned, but the new mac differs,
1327 		 * unregister the old one and register the new one.
1328 		*/
1329 		if ((!smac_info->smac && !smac_info->smac_port) ||
1330 		    smac_info->smac != smac) {
1331 			/* register candidate now, unreg if needed, after success */
1332 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1333 			if (smac_index >= 0) {
1334 				smac_info->candidate_smac_index = smac_index;
1335 				smac_info->candidate_smac = smac;
1336 				smac_info->candidate_smac_port = port;
1337 			} else {
1338 				return -EINVAL;
1339 			}
1340 		} else {
1341 			smac_index = smac_info->smac_index;
1342 		}
1343 
1344 		memcpy(path->dmac, ah->dmac, 6);
1345 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1346 		/* put MAC table smac index for IBoE */
1347 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1348 	} else {
1349 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1350 			((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1351 	}
1352 
1353 	return 0;
1354 }
1355 
1356 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1357 			 enum ib_qp_attr_mask qp_attr_mask,
1358 			 struct mlx4_ib_qp *mqp,
1359 			 struct mlx4_qp_path *path, u8 port)
1360 {
1361 	return _mlx4_set_path(dev, &qp->ah_attr,
1362 			      mlx4_mac_to_u64((u8 *)qp->smac),
1363 			      (qp_attr_mask & IB_QP_VID) ? qp->vlan_id : 0xffff,
1364 			      path, &mqp->pri, port);
1365 }
1366 
1367 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1368 			     const struct ib_qp_attr *qp,
1369 			     enum ib_qp_attr_mask qp_attr_mask,
1370 			     struct mlx4_ib_qp *mqp,
1371 			     struct mlx4_qp_path *path, u8 port)
1372 {
1373 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1374 			      mlx4_mac_to_u64((u8 *)qp->alt_smac),
1375 			      (qp_attr_mask & IB_QP_ALT_VID) ?
1376 			      qp->alt_vlan_id : 0xffff,
1377 			      path, &mqp->alt, port);
1378 }
1379 
1380 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1381 {
1382 	struct mlx4_ib_gid_entry *ge, *tmp;
1383 
1384 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1385 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1386 			ge->added = 1;
1387 			ge->port = qp->port;
1388 		}
1389 	}
1390 }
1391 
1392 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp, u8 *smac,
1393 				    struct mlx4_qp_context *context)
1394 {
1395 	u64 u64_mac;
1396 	int smac_index;
1397 
1398 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1399 
1400 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1401 	if (!qp->pri.smac && !qp->pri.smac_port) {
1402 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1403 		if (smac_index >= 0) {
1404 			qp->pri.candidate_smac_index = smac_index;
1405 			qp->pri.candidate_smac = u64_mac;
1406 			qp->pri.candidate_smac_port = qp->port;
1407 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1408 		} else {
1409 			return -ENOENT;
1410 		}
1411 	}
1412 	return 0;
1413 }
1414 
1415 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1416 			       const struct ib_qp_attr *attr, int attr_mask,
1417 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1418 {
1419 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1420 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1421 	struct mlx4_ib_pd *pd;
1422 	struct mlx4_ib_cq *send_cq, *recv_cq;
1423 	struct mlx4_qp_context *context;
1424 	enum mlx4_qp_optpar optpar = 0;
1425 	int sqd_event;
1426 	int steer_qp = 0;
1427 	int err = -EINVAL;
1428 
1429 	/* APM is not supported under RoCE */
1430 	if (attr_mask & IB_QP_ALT_PATH &&
1431 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1432 	    IB_LINK_LAYER_ETHERNET)
1433 		return -ENOTSUPP;
1434 
1435 	context = kzalloc(sizeof *context, GFP_KERNEL);
1436 	if (!context)
1437 		return -ENOMEM;
1438 
1439 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1440 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1441 
1442 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1443 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1444 	else {
1445 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
1446 		switch (attr->path_mig_state) {
1447 		case IB_MIG_MIGRATED:
1448 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1449 			break;
1450 		case IB_MIG_REARM:
1451 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1452 			break;
1453 		case IB_MIG_ARMED:
1454 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1455 			break;
1456 		}
1457 	}
1458 
1459 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1460 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1461 	else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1462 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1463 	else if (ibqp->qp_type == IB_QPT_UD) {
1464 		if (qp->flags & MLX4_IB_QP_LSO)
1465 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
1466 					      ilog2(dev->dev->caps.max_gso_sz);
1467 		else
1468 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1469 	} else if (attr_mask & IB_QP_PATH_MTU) {
1470 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1471 			pr_err("path MTU (%u) is invalid\n",
1472 			       attr->path_mtu);
1473 			goto out;
1474 		}
1475 		context->mtu_msgmax = (attr->path_mtu << 5) |
1476 			ilog2(dev->dev->caps.max_msg_sz);
1477 	}
1478 
1479 	if (qp->rq.wqe_cnt)
1480 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1481 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
1482 
1483 	if (qp->sq.wqe_cnt)
1484 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1485 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
1486 
1487 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1488 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1489 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1490 		if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1491 			context->param3 |= cpu_to_be32(1 << 30);
1492 	}
1493 
1494 	if (qp->ibqp.uobject)
1495 		context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
1496 	else
1497 		context->usr_page = cpu_to_be32(dev->priv_uar.index);
1498 
1499 	if (attr_mask & IB_QP_DEST_QPN)
1500 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1501 
1502 	if (attr_mask & IB_QP_PORT) {
1503 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1504 		    !(attr_mask & IB_QP_AV)) {
1505 			mlx4_set_sched(&context->pri_path, attr->port_num);
1506 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1507 		}
1508 	}
1509 
1510 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1511 		if (dev->counters[qp->port - 1] != -1) {
1512 			context->pri_path.counter_index =
1513 						dev->counters[qp->port - 1];
1514 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1515 		} else
1516 			context->pri_path.counter_index = 0xff;
1517 
1518 		if (qp->flags & MLX4_IB_QP_NETIF) {
1519 			mlx4_ib_steer_qp_reg(dev, qp, 1);
1520 			steer_qp = 1;
1521 		}
1522 	}
1523 
1524 	if (attr_mask & IB_QP_PKEY_INDEX) {
1525 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1526 			context->pri_path.disable_pkey_check = 0x40;
1527 		context->pri_path.pkey_index = attr->pkey_index;
1528 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1529 	}
1530 
1531 	if (attr_mask & IB_QP_AV) {
1532 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
1533 				  attr_mask & IB_QP_PORT ?
1534 				  attr->port_num : qp->port))
1535 			goto out;
1536 
1537 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1538 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
1539 	}
1540 
1541 	if (attr_mask & IB_QP_TIMEOUT) {
1542 		context->pri_path.ackto |= attr->timeout << 3;
1543 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1544 	}
1545 
1546 	if (attr_mask & IB_QP_ALT_PATH) {
1547 		if (attr->alt_port_num == 0 ||
1548 		    attr->alt_port_num > dev->dev->caps.num_ports)
1549 			goto out;
1550 
1551 		if (attr->alt_pkey_index >=
1552 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
1553 			goto out;
1554 
1555 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
1556 				      &context->alt_path,
1557 				      attr->alt_port_num))
1558 			goto out;
1559 
1560 		context->alt_path.pkey_index = attr->alt_pkey_index;
1561 		context->alt_path.ackto = attr->alt_timeout << 3;
1562 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1563 	}
1564 
1565 	pd = get_pd(qp);
1566 	get_cqs(qp, &send_cq, &recv_cq);
1567 	context->pd       = cpu_to_be32(pd->pdn);
1568 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1569 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1570 	context->params1  = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1571 
1572 	/* Set "fast registration enabled" for all kernel QPs */
1573 	if (!qp->ibqp.uobject)
1574 		context->params1 |= cpu_to_be32(1 << 11);
1575 
1576 	if (attr_mask & IB_QP_RNR_RETRY) {
1577 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1578 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1579 	}
1580 
1581 	if (attr_mask & IB_QP_RETRY_CNT) {
1582 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1583 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1584 	}
1585 
1586 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1587 		if (attr->max_rd_atomic)
1588 			context->params1 |=
1589 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1590 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1591 	}
1592 
1593 	if (attr_mask & IB_QP_SQ_PSN)
1594 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1595 
1596 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1597 		if (attr->max_dest_rd_atomic)
1598 			context->params2 |=
1599 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1600 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1601 	}
1602 
1603 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1604 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1605 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1606 	}
1607 
1608 	if (ibqp->srq)
1609 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1610 
1611 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1612 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1613 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1614 	}
1615 	if (attr_mask & IB_QP_RQ_PSN)
1616 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1617 
1618 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1619 	if (attr_mask & IB_QP_QKEY) {
1620 		if (qp->mlx4_ib_qp_type &
1621 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1622 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1623 		else {
1624 			if (mlx4_is_mfunc(dev->dev) &&
1625 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1626 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1627 			    MLX4_RESERVED_QKEY_BASE) {
1628 				pr_err("Cannot use reserved QKEY"
1629 				       " 0x%x (range 0xffff0000..0xffffffff"
1630 				       " is reserved)\n", attr->qkey);
1631 				err = -EINVAL;
1632 				goto out;
1633 			}
1634 			context->qkey = cpu_to_be32(attr->qkey);
1635 		}
1636 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
1637 	}
1638 
1639 	if (ibqp->srq)
1640 		context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1641 
1642 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1643 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1644 
1645 	if (cur_state == IB_QPS_INIT &&
1646 	    new_state == IB_QPS_RTR  &&
1647 	    (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1648 	     ibqp->qp_type == IB_QPT_UD ||
1649 	     ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1650 		context->pri_path.sched_queue = (qp->port - 1) << 6;
1651 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1652 		    qp->mlx4_ib_qp_type &
1653 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1654 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1655 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1656 				context->pri_path.fl = 0x80;
1657 		} else {
1658 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1659 				context->pri_path.fl = 0x80;
1660 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1661 		}
1662 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1663 		    IB_LINK_LAYER_ETHERNET) {
1664 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
1665 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
1666 				context->pri_path.feup = 1 << 7; /* don't fsm */
1667 			/* handle smac_index */
1668 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
1669 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
1670 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
1671 				err = handle_eth_ud_smac_index(dev, qp, (u8 *)attr->smac, context);
1672 				if (err)
1673 					return -EINVAL;
1674 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1675 					dev->qp1_proxy[qp->port - 1] = qp;
1676 			}
1677 		}
1678 	}
1679 
1680 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1681 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
1682 					MLX4_IB_LINK_TYPE_ETH;
1683 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1684 			/* set QP to receive both tunneled & non-tunneled packets */
1685 			if (!(context->flags & cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET)))
1686 				context->srqn = cpu_to_be32(7 << 28);
1687 		}
1688 	}
1689 
1690 	if (ibqp->qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
1691 		int is_eth = rdma_port_get_link_layer(
1692 				&dev->ib_dev, qp->port) ==
1693 				IB_LINK_LAYER_ETHERNET;
1694 		if (is_eth) {
1695 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
1696 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
1697 		}
1698 	}
1699 
1700 
1701 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1702 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1703 		sqd_event = 1;
1704 	else
1705 		sqd_event = 0;
1706 
1707 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1708 		context->rlkey |= (1 << 4);
1709 
1710 	/*
1711 	 * Before passing a kernel QP to the HW, make sure that the
1712 	 * ownership bits of the send queue are set and the SQ
1713 	 * headroom is stamped so that the hardware doesn't start
1714 	 * processing stale work requests.
1715 	 */
1716 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1717 		struct mlx4_wqe_ctrl_seg *ctrl;
1718 		int i;
1719 
1720 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1721 			ctrl = get_send_wqe(qp, i);
1722 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
1723 			if (qp->sq_max_wqes_per_wr == 1)
1724 				ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1725 
1726 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1727 		}
1728 	}
1729 
1730 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1731 			     to_mlx4_state(new_state), context, optpar,
1732 			     sqd_event, &qp->mqp);
1733 	if (err)
1734 		goto out;
1735 
1736 	qp->state = new_state;
1737 
1738 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1739 		qp->atomic_rd_en = attr->qp_access_flags;
1740 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1741 		qp->resp_depth = attr->max_dest_rd_atomic;
1742 	if (attr_mask & IB_QP_PORT) {
1743 		qp->port = attr->port_num;
1744 		update_mcg_macs(dev, qp);
1745 	}
1746 	if (attr_mask & IB_QP_ALT_PATH)
1747 		qp->alt_port = attr->alt_port_num;
1748 
1749 	if (is_sqp(dev, qp))
1750 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1751 
1752 	/*
1753 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
1754 	 * QP0 to RESET or ERROR, bring the link back down.
1755 	 */
1756 	if (is_qp0(dev, qp)) {
1757 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1758 			if (mlx4_INIT_PORT(dev->dev, qp->port))
1759 				pr_warn("INIT_PORT failed for port %d\n",
1760 				       qp->port);
1761 
1762 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1763 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1764 			mlx4_CLOSE_PORT(dev->dev, qp->port);
1765 	}
1766 
1767 	/*
1768 	 * If we moved a kernel QP to RESET, clean up all old CQ
1769 	 * entries and reinitialize the QP.
1770 	 */
1771 	if (new_state == IB_QPS_RESET) {
1772 		if (!ibqp->uobject) {
1773 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1774 					 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
1775 			if (send_cq != recv_cq)
1776 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1777 
1778 			qp->rq.head = 0;
1779 			qp->rq.tail = 0;
1780 			qp->sq.head = 0;
1781 			qp->sq.tail = 0;
1782 			qp->sq_next_wqe = 0;
1783 			if (qp->rq.wqe_cnt)
1784 				*qp->db.db  = 0;
1785 
1786 			if (qp->flags & MLX4_IB_QP_NETIF)
1787 				mlx4_ib_steer_qp_reg(dev, qp, 0);
1788 		}
1789 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1790 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1791 			qp->pri.smac = 0;
1792 			qp->pri.smac_port = 0;
1793 		}
1794 		if (qp->alt.smac) {
1795 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1796 			qp->alt.smac = 0;
1797 		}
1798 		if (qp->pri.vid < 0x1000) {
1799 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1800 			qp->pri.vid = 0xFFFF;
1801 			qp->pri.candidate_vid = 0xFFFF;
1802 			qp->pri.update_vid = 0;
1803 		}
1804 
1805 		if (qp->alt.vid < 0x1000) {
1806 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1807 			qp->alt.vid = 0xFFFF;
1808 			qp->alt.candidate_vid = 0xFFFF;
1809 			qp->alt.update_vid = 0;
1810 		}
1811 	}
1812 out:
1813 	if (err && steer_qp)
1814 		mlx4_ib_steer_qp_reg(dev, qp, 0);
1815 	kfree(context);
1816 	if (qp->pri.candidate_smac ||
1817 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
1818 		if (err) {
1819 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
1820 		} else {
1821 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
1822 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1823 			qp->pri.smac = qp->pri.candidate_smac;
1824 			qp->pri.smac_index = qp->pri.candidate_smac_index;
1825 			qp->pri.smac_port = qp->pri.candidate_smac_port;
1826 		}
1827 		qp->pri.candidate_smac = 0;
1828 		qp->pri.candidate_smac_index = 0;
1829 		qp->pri.candidate_smac_port = 0;
1830 	}
1831 	if (qp->alt.candidate_smac) {
1832 		if (err) {
1833 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
1834 		} else {
1835 			if (qp->alt.smac)
1836 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1837 			qp->alt.smac = qp->alt.candidate_smac;
1838 			qp->alt.smac_index = qp->alt.candidate_smac_index;
1839 			qp->alt.smac_port = qp->alt.candidate_smac_port;
1840 		}
1841 		qp->alt.candidate_smac = 0;
1842 		qp->alt.candidate_smac_index = 0;
1843 		qp->alt.candidate_smac_port = 0;
1844 	}
1845 
1846 	if (qp->pri.update_vid) {
1847 		if (err) {
1848 			if (qp->pri.candidate_vid < 0x1000)
1849 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
1850 						     qp->pri.candidate_vid);
1851 		} else {
1852 			if (qp->pri.vid < 0x1000)
1853 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
1854 						     qp->pri.vid);
1855 			qp->pri.vid = qp->pri.candidate_vid;
1856 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
1857 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
1858 		}
1859 		qp->pri.candidate_vid = 0xFFFF;
1860 		qp->pri.update_vid = 0;
1861 	}
1862 
1863 	if (qp->alt.update_vid) {
1864 		if (err) {
1865 			if (qp->alt.candidate_vid < 0x1000)
1866 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
1867 						     qp->alt.candidate_vid);
1868 		} else {
1869 			if (qp->alt.vid < 0x1000)
1870 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
1871 						     qp->alt.vid);
1872 			qp->alt.vid = qp->alt.candidate_vid;
1873 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
1874 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
1875 		}
1876 		qp->alt.candidate_vid = 0xFFFF;
1877 		qp->alt.update_vid = 0;
1878 	}
1879 
1880 	return err;
1881 }
1882 
1883 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1884 		      int attr_mask, struct ib_udata *udata)
1885 {
1886 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1887 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1888 	enum ib_qp_state cur_state, new_state;
1889 	int err = -EINVAL;
1890 	int ll;
1891 	mutex_lock(&qp->mutex);
1892 
1893 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1894 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1895 
1896 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1897 		ll = IB_LINK_LAYER_UNSPECIFIED;
1898 	} else {
1899 		int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1900 		ll = rdma_port_get_link_layer(&dev->ib_dev, port);
1901 	}
1902 
1903 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1904 				attr_mask, ll)) {
1905 		pr_debug("qpn 0x%x: invalid attribute mask specified "
1906 			 "for transition %d to %d. qp_type %d,"
1907 			 " attr_mask 0x%x\n",
1908 			 ibqp->qp_num, cur_state, new_state,
1909 			 ibqp->qp_type, attr_mask);
1910 		goto out;
1911 	}
1912 
1913 	if ((attr_mask & IB_QP_PORT) &&
1914 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
1915 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
1916 			 "for transition %d to %d. qp_type %d\n",
1917 			 ibqp->qp_num, attr->port_num, cur_state,
1918 			 new_state, ibqp->qp_type);
1919 		goto out;
1920 	}
1921 
1922 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
1923 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
1924 	     IB_LINK_LAYER_ETHERNET))
1925 		goto out;
1926 
1927 	if (attr_mask & IB_QP_PKEY_INDEX) {
1928 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1929 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
1930 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
1931 				 "for transition %d to %d. qp_type %d\n",
1932 				 ibqp->qp_num, attr->pkey_index, cur_state,
1933 				 new_state, ibqp->qp_type);
1934 			goto out;
1935 		}
1936 	}
1937 
1938 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1939 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
1940 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
1941 			 "Transition %d to %d. qp_type %d\n",
1942 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
1943 			 new_state, ibqp->qp_type);
1944 		goto out;
1945 	}
1946 
1947 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1948 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
1949 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
1950 			 "Transition %d to %d. qp_type %d\n",
1951 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
1952 			 new_state, ibqp->qp_type);
1953 		goto out;
1954 	}
1955 
1956 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1957 		err = 0;
1958 		goto out;
1959 	}
1960 
1961 	err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1962 
1963 out:
1964 	mutex_unlock(&qp->mutex);
1965 	return err;
1966 }
1967 
1968 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
1969 {
1970 	int i;
1971 	for (i = 0; i < dev->caps.num_ports; i++) {
1972 		if (qpn == dev->caps.qp0_proxy[i] ||
1973 		    qpn == dev->caps.qp0_tunnel[i]) {
1974 			*qkey = dev->caps.qp0_qkey[i];
1975 			return 0;
1976 		}
1977 	}
1978 	return -EINVAL;
1979 }
1980 
1981 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
1982 				  struct ib_send_wr *wr,
1983 				  void *wqe, unsigned *mlx_seg_len)
1984 {
1985 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
1986 	struct ib_device *ib_dev = &mdev->ib_dev;
1987 	struct mlx4_wqe_mlx_seg *mlx = wqe;
1988 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
1989 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
1990 	u16 pkey;
1991 	u32 qkey;
1992 	int send_size;
1993 	int header_size;
1994 	int spc;
1995 	int i;
1996 
1997 	if (wr->opcode != IB_WR_SEND)
1998 		return -EINVAL;
1999 
2000 	send_size = 0;
2001 
2002 	for (i = 0; i < wr->num_sge; ++i)
2003 		send_size += wr->sg_list[i].length;
2004 
2005 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2006 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2007 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2008 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2009 
2010 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, &sqp->ud_header);
2011 
2012 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2013 		sqp->ud_header.lrh.service_level =
2014 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2015 		sqp->ud_header.lrh.destination_lid =
2016 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2017 		sqp->ud_header.lrh.source_lid =
2018 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2019 	}
2020 
2021 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2022 
2023 	/* force loopback */
2024 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2025 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2026 
2027 	sqp->ud_header.lrh.virtual_lane    = 0;
2028 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
2029 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2030 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2031 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2032 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2033 	else
2034 		sqp->ud_header.bth.destination_qpn =
2035 			cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
2036 
2037 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2038 	if (mlx4_is_master(mdev->dev)) {
2039 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2040 			return -EINVAL;
2041 	} else {
2042 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2043 			return -EINVAL;
2044 	}
2045 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2046 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2047 
2048 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2049 	sqp->ud_header.immediate_present = 0;
2050 
2051 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2052 
2053 	/*
2054 	 * Inline data segments may not cross a 64 byte boundary.  If
2055 	 * our UD header is bigger than the space available up to the
2056 	 * next 64 byte boundary in the WQE, use two inline data
2057 	 * segments to hold the UD header.
2058 	 */
2059 	spc = MLX4_INLINE_ALIGN -
2060 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2061 	if (header_size <= spc) {
2062 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2063 		memcpy(inl + 1, sqp->header_buf, header_size);
2064 		i = 1;
2065 	} else {
2066 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2067 		memcpy(inl + 1, sqp->header_buf, spc);
2068 
2069 		inl = (void *) (inl + 1) + spc;
2070 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2071 		/*
2072 		 * Need a barrier here to make sure all the data is
2073 		 * visible before the byte_count field is set.
2074 		 * Otherwise the HCA prefetcher could grab the 64-byte
2075 		 * chunk with this inline segment and get a valid (!=
2076 		 * 0xffffffff) byte count but stale data, and end up
2077 		 * generating a packet with bad headers.
2078 		 *
2079 		 * The first inline segment's byte_count field doesn't
2080 		 * need a barrier, because it comes after a
2081 		 * control/MLX segment and therefore is at an offset
2082 		 * of 16 mod 64.
2083 		 */
2084 		wmb();
2085 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2086 		i = 2;
2087 	}
2088 
2089 	*mlx_seg_len =
2090 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2091 	return 0;
2092 }
2093 
2094 static void mlx4_u64_to_smac(u8 *dst_mac, u64 src_mac)
2095 {
2096 	int i;
2097 
2098 	for (i = ETH_ALEN; i; i--) {
2099 		dst_mac[i - 1] = src_mac & 0xff;
2100 		src_mac >>= 8;
2101 	}
2102 }
2103 
2104 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
2105 			    void *wqe, unsigned *mlx_seg_len)
2106 {
2107 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
2108 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2109 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2110 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2111 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2112 	union ib_gid sgid;
2113 	u16 pkey;
2114 	int send_size;
2115 	int header_size;
2116 	int spc;
2117 	int i;
2118 	int err = 0;
2119 	u16 vlan = 0xffff;
2120 	bool is_eth;
2121 	bool is_vlan = false;
2122 	bool is_grh;
2123 
2124 	send_size = 0;
2125 	for (i = 0; i < wr->num_sge; ++i)
2126 		send_size += wr->sg_list[i].length;
2127 
2128 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2129 	is_grh = mlx4_ib_ah_grh_present(ah);
2130 	if (is_eth) {
2131 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2132 			/* When multi-function is enabled, the ib_core gid
2133 			 * indexes don't necessarily match the hw ones, so
2134 			 * we must use our own cache */
2135 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2136 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
2137 							   ah->av.ib.gid_index, &sgid.raw[0]);
2138 			if (err)
2139 				return err;
2140 		} else  {
2141 			err = ib_get_cached_gid(ib_dev,
2142 						be32_to_cpu(ah->av.ib.port_pd) >> 24,
2143 						ah->av.ib.gid_index, &sgid);
2144 			if (err)
2145 				return err;
2146 		}
2147 
2148 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
2149 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
2150 			is_vlan = 1;
2151 		}
2152 	}
2153 	ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh, 0, &sqp->ud_header);
2154 
2155 	if (!is_eth) {
2156 		sqp->ud_header.lrh.service_level =
2157 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2158 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
2159 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2160 	}
2161 
2162 	if (is_grh) {
2163 		sqp->ud_header.grh.traffic_class =
2164 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2165 		sqp->ud_header.grh.flow_label    =
2166 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
2167 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
2168 		if (is_eth)
2169 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
2170 		else {
2171 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2172 			/* When multi-function is enabled, the ib_core gid
2173 			 * indexes don't necessarily match the hw ones, so
2174 			 * we must use our own cache */
2175 			sqp->ud_header.grh.source_gid.global.subnet_prefix =
2176 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2177 						       subnet_prefix;
2178 			sqp->ud_header.grh.source_gid.global.interface_id =
2179 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2180 					       guid_cache[ah->av.ib.gid_index];
2181 		} else
2182 			ib_get_cached_gid(ib_dev,
2183 					  be32_to_cpu(ah->av.ib.port_pd) >> 24,
2184 					  ah->av.ib.gid_index,
2185 					  &sqp->ud_header.grh.source_gid);
2186 		}
2187 		memcpy(sqp->ud_header.grh.destination_gid.raw,
2188 		       ah->av.ib.dgid, 16);
2189 	}
2190 
2191 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2192 
2193 	if (!is_eth) {
2194 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
2195 					  (sqp->ud_header.lrh.destination_lid ==
2196 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
2197 					  (sqp->ud_header.lrh.service_level << 8));
2198 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
2199 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
2200 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
2201 	}
2202 
2203 	switch (wr->opcode) {
2204 	case IB_WR_SEND:
2205 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
2206 		sqp->ud_header.immediate_present = 0;
2207 		break;
2208 	case IB_WR_SEND_WITH_IMM:
2209 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2210 		sqp->ud_header.immediate_present = 1;
2211 		sqp->ud_header.immediate_data    = wr->ex.imm_data;
2212 		break;
2213 	default:
2214 		return -EINVAL;
2215 	}
2216 
2217 	if (is_eth) {
2218 		struct in6_addr in6;
2219 
2220 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
2221 
2222 		mlx->sched_prio = cpu_to_be16(pcp);
2223 
2224 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
2225 		/* FIXME: cache smac value? */
2226 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
2227 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
2228 		memcpy(&in6, sgid.raw, sizeof(in6));
2229 
2230 		if (!mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2231 			u64 mac = atomic64_read(&to_mdev(ib_dev)->iboe.mac[sqp->qp.port - 1]);
2232 			u8 smac[ETH_ALEN];
2233 
2234 			mlx4_u64_to_smac(smac, mac);
2235 			memcpy(sqp->ud_header.eth.smac_h, smac, ETH_ALEN);
2236 		} else {
2237 			/* use the src mac of the tunnel */
2238 			memcpy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac, ETH_ALEN);
2239 		}
2240 
2241 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
2242 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
2243 		if (!is_vlan) {
2244 			sqp->ud_header.eth.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2245 		} else {
2246 			sqp->ud_header.vlan.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2247 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
2248 		}
2249 	} else {
2250 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
2251 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
2252 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
2253 	}
2254 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
2255 	if (!sqp->qp.ibqp.qp_num)
2256 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
2257 	else
2258 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
2259 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2260 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2261 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2262 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
2263 					       sqp->qkey : wr->wr.ud.remote_qkey);
2264 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
2265 
2266 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2267 
2268 	if (0) {
2269 		pr_err("built UD header of size %d:\n", header_size);
2270 		for (i = 0; i < header_size / 4; ++i) {
2271 			if (i % 8 == 0)
2272 				pr_err("  [%02x] ", i * 4);
2273 			pr_cont(" %08x",
2274 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
2275 			if ((i + 1) % 8 == 0)
2276 				pr_cont("\n");
2277 		}
2278 		pr_err("\n");
2279 	}
2280 
2281 	/*
2282 	 * Inline data segments may not cross a 64 byte boundary.  If
2283 	 * our UD header is bigger than the space available up to the
2284 	 * next 64 byte boundary in the WQE, use two inline data
2285 	 * segments to hold the UD header.
2286 	 */
2287 	spc = MLX4_INLINE_ALIGN -
2288 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2289 	if (header_size <= spc) {
2290 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2291 		memcpy(inl + 1, sqp->header_buf, header_size);
2292 		i = 1;
2293 	} else {
2294 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2295 		memcpy(inl + 1, sqp->header_buf, spc);
2296 
2297 		inl = (void *) (inl + 1) + spc;
2298 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2299 		/*
2300 		 * Need a barrier here to make sure all the data is
2301 		 * visible before the byte_count field is set.
2302 		 * Otherwise the HCA prefetcher could grab the 64-byte
2303 		 * chunk with this inline segment and get a valid (!=
2304 		 * 0xffffffff) byte count but stale data, and end up
2305 		 * generating a packet with bad headers.
2306 		 *
2307 		 * The first inline segment's byte_count field doesn't
2308 		 * need a barrier, because it comes after a
2309 		 * control/MLX segment and therefore is at an offset
2310 		 * of 16 mod 64.
2311 		 */
2312 		wmb();
2313 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2314 		i = 2;
2315 	}
2316 
2317 	*mlx_seg_len =
2318 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2319 	return 0;
2320 }
2321 
2322 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2323 {
2324 	unsigned cur;
2325 	struct mlx4_ib_cq *cq;
2326 
2327 	cur = wq->head - wq->tail;
2328 	if (likely(cur + nreq < wq->max_post))
2329 		return 0;
2330 
2331 	cq = to_mcq(ib_cq);
2332 	spin_lock(&cq->lock);
2333 	cur = wq->head - wq->tail;
2334 	spin_unlock(&cq->lock);
2335 
2336 	return cur + nreq >= wq->max_post;
2337 }
2338 
2339 static __be32 convert_access(int acc)
2340 {
2341 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
2342 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
2343 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
2344 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
2345 	       (acc & IB_ACCESS_REMOTE_READ   ?
2346 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
2347 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
2348 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
2349 }
2350 
2351 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr)
2352 {
2353 	struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
2354 	int i;
2355 
2356 	for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i)
2357 		mfrpl->mapped_page_list[i] =
2358 			cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] |
2359 				    MLX4_MTT_FLAG_PRESENT);
2360 
2361 	fseg->flags		= convert_access(wr->wr.fast_reg.access_flags);
2362 	fseg->mem_key		= cpu_to_be32(wr->wr.fast_reg.rkey);
2363 	fseg->buf_list		= cpu_to_be64(mfrpl->map);
2364 	fseg->start_addr	= cpu_to_be64(wr->wr.fast_reg.iova_start);
2365 	fseg->reg_len		= cpu_to_be64(wr->wr.fast_reg.length);
2366 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
2367 	fseg->page_size		= cpu_to_be32(wr->wr.fast_reg.page_shift);
2368 	fseg->reserved[0]	= 0;
2369 	fseg->reserved[1]	= 0;
2370 }
2371 
2372 static void set_bind_seg(struct mlx4_wqe_bind_seg *bseg, struct ib_send_wr *wr)
2373 {
2374 	bseg->flags1 =
2375 		convert_access(wr->wr.bind_mw.bind_info.mw_access_flags) &
2376 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ  |
2377 			    MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE |
2378 			    MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC);
2379 	bseg->flags2 = 0;
2380 	if (wr->wr.bind_mw.mw->type == IB_MW_TYPE_2)
2381 		bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_TYPE_2);
2382 	if (wr->wr.bind_mw.bind_info.mw_access_flags & IB_ZERO_BASED)
2383 		bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_ZERO_BASED);
2384 	bseg->new_rkey = cpu_to_be32(wr->wr.bind_mw.rkey);
2385 	bseg->lkey = cpu_to_be32(wr->wr.bind_mw.bind_info.mr->lkey);
2386 	bseg->addr = cpu_to_be64(wr->wr.bind_mw.bind_info.addr);
2387 	bseg->length = cpu_to_be64(wr->wr.bind_mw.bind_info.length);
2388 }
2389 
2390 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2391 {
2392 	memset(iseg, 0, sizeof(*iseg));
2393 	iseg->mem_key = cpu_to_be32(rkey);
2394 }
2395 
2396 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2397 					  u64 remote_addr, u32 rkey)
2398 {
2399 	rseg->raddr    = cpu_to_be64(remote_addr);
2400 	rseg->rkey     = cpu_to_be32(rkey);
2401 	rseg->reserved = 0;
2402 }
2403 
2404 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
2405 {
2406 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2407 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
2408 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
2409 	} else if (wr->opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2410 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2411 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add_mask);
2412 	} else {
2413 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2414 		aseg->compare  = 0;
2415 	}
2416 
2417 }
2418 
2419 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2420 				  struct ib_send_wr *wr)
2421 {
2422 	aseg->swap_add		= cpu_to_be64(wr->wr.atomic.swap);
2423 	aseg->swap_add_mask	= cpu_to_be64(wr->wr.atomic.swap_mask);
2424 	aseg->compare		= cpu_to_be64(wr->wr.atomic.compare_add);
2425 	aseg->compare_mask	= cpu_to_be64(wr->wr.atomic.compare_add_mask);
2426 }
2427 
2428 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2429 			     struct ib_send_wr *wr)
2430 {
2431 	memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
2432 	dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2433 	dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2434 	dseg->vlan = to_mah(wr->wr.ud.ah)->av.eth.vlan;
2435 	memcpy(dseg->mac, to_mah(wr->wr.ud.ah)->av.eth.mac, 6);
2436 }
2437 
2438 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2439 				    struct mlx4_wqe_datagram_seg *dseg,
2440 				    struct ib_send_wr *wr,
2441 				    enum mlx4_ib_qp_type qpt)
2442 {
2443 	union mlx4_ext_av *av = &to_mah(wr->wr.ud.ah)->av;
2444 	struct mlx4_av sqp_av = {0};
2445 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
2446 
2447 	/* force loopback */
2448 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2449 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2450 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2451 			cpu_to_be32(0xf0000000);
2452 
2453 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2454 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
2455 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2456 	else
2457 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp0_tunnel[port - 1]);
2458 	/* Use QKEY from the QP context, which is set by master */
2459 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2460 }
2461 
2462 static void build_tunnel_header(struct ib_send_wr *wr, void *wqe, unsigned *mlx_seg_len)
2463 {
2464 	struct mlx4_wqe_inline_seg *inl = wqe;
2465 	struct mlx4_ib_tunnel_header hdr;
2466 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2467 	int spc;
2468 	int i;
2469 
2470 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2471 	hdr.remote_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2472 	hdr.pkey_index = cpu_to_be16(wr->wr.ud.pkey_index);
2473 	hdr.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2474 	memcpy(hdr.mac, ah->av.eth.mac, 6);
2475 	hdr.vlan = ah->av.eth.vlan;
2476 
2477 	spc = MLX4_INLINE_ALIGN -
2478 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2479 	if (sizeof (hdr) <= spc) {
2480 		memcpy(inl + 1, &hdr, sizeof (hdr));
2481 		wmb();
2482 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
2483 		i = 1;
2484 	} else {
2485 		memcpy(inl + 1, &hdr, spc);
2486 		wmb();
2487 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2488 
2489 		inl = (void *) (inl + 1) + spc;
2490 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2491 		wmb();
2492 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
2493 		i = 2;
2494 	}
2495 
2496 	*mlx_seg_len =
2497 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2498 }
2499 
2500 static void set_mlx_icrc_seg(void *dseg)
2501 {
2502 	u32 *t = dseg;
2503 	struct mlx4_wqe_inline_seg *iseg = dseg;
2504 
2505 	t[1] = 0;
2506 
2507 	/*
2508 	 * Need a barrier here before writing the byte_count field to
2509 	 * make sure that all the data is visible before the
2510 	 * byte_count field is set.  Otherwise, if the segment begins
2511 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2512 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2513 	 * stale data, and end up sending the wrong data.
2514 	 */
2515 	wmb();
2516 
2517 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2518 }
2519 
2520 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2521 {
2522 	dseg->lkey       = cpu_to_be32(sg->lkey);
2523 	dseg->addr       = cpu_to_be64(sg->addr);
2524 
2525 	/*
2526 	 * Need a barrier here before writing the byte_count field to
2527 	 * make sure that all the data is visible before the
2528 	 * byte_count field is set.  Otherwise, if the segment begins
2529 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2530 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2531 	 * stale data, and end up sending the wrong data.
2532 	 */
2533 	wmb();
2534 
2535 	dseg->byte_count = cpu_to_be32(sg->length);
2536 }
2537 
2538 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2539 {
2540 	dseg->byte_count = cpu_to_be32(sg->length);
2541 	dseg->lkey       = cpu_to_be32(sg->lkey);
2542 	dseg->addr       = cpu_to_be64(sg->addr);
2543 }
2544 
2545 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr,
2546 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2547 			 __be32 *lso_hdr_sz, __be32 *blh)
2548 {
2549 	unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
2550 
2551 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2552 		*blh = cpu_to_be32(1 << 6);
2553 
2554 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2555 		     wr->num_sge > qp->sq.max_gs - (halign >> 4)))
2556 		return -EINVAL;
2557 
2558 	memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
2559 
2560 	*lso_hdr_sz  = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 |
2561 				   wr->wr.ud.hlen);
2562 	*lso_seg_len = halign;
2563 	return 0;
2564 }
2565 
2566 static __be32 send_ieth(struct ib_send_wr *wr)
2567 {
2568 	switch (wr->opcode) {
2569 	case IB_WR_SEND_WITH_IMM:
2570 	case IB_WR_RDMA_WRITE_WITH_IMM:
2571 		return wr->ex.imm_data;
2572 
2573 	case IB_WR_SEND_WITH_INV:
2574 		return cpu_to_be32(wr->ex.invalidate_rkey);
2575 
2576 	default:
2577 		return 0;
2578 	}
2579 }
2580 
2581 static void add_zero_len_inline(void *wqe)
2582 {
2583 	struct mlx4_wqe_inline_seg *inl = wqe;
2584 	memset(wqe, 0, 16);
2585 	inl->byte_count = cpu_to_be32(1 << 31);
2586 }
2587 
2588 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2589 		      struct ib_send_wr **bad_wr)
2590 {
2591 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2592 	void *wqe;
2593 	struct mlx4_wqe_ctrl_seg *ctrl;
2594 	struct mlx4_wqe_data_seg *dseg;
2595 	unsigned long flags;
2596 	int nreq;
2597 	int err = 0;
2598 	unsigned ind;
2599 	int uninitialized_var(stamp);
2600 	int uninitialized_var(size);
2601 	unsigned uninitialized_var(seglen);
2602 	__be32 dummy;
2603 	__be32 *lso_wqe;
2604 	__be32 uninitialized_var(lso_hdr_sz);
2605 	__be32 blh;
2606 	int i;
2607 
2608 	spin_lock_irqsave(&qp->sq.lock, flags);
2609 
2610 	ind = qp->sq_next_wqe;
2611 
2612 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2613 		lso_wqe = &dummy;
2614 		blh = 0;
2615 
2616 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2617 			err = -ENOMEM;
2618 			*bad_wr = wr;
2619 			goto out;
2620 		}
2621 
2622 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
2623 			err = -EINVAL;
2624 			*bad_wr = wr;
2625 			goto out;
2626 		}
2627 
2628 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
2629 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
2630 
2631 		ctrl->srcrb_flags =
2632 			(wr->send_flags & IB_SEND_SIGNALED ?
2633 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
2634 			(wr->send_flags & IB_SEND_SOLICITED ?
2635 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
2636 			((wr->send_flags & IB_SEND_IP_CSUM) ?
2637 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
2638 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
2639 			qp->sq_signal_bits;
2640 
2641 		ctrl->imm = send_ieth(wr);
2642 
2643 		wqe += sizeof *ctrl;
2644 		size = sizeof *ctrl / 16;
2645 
2646 		switch (qp->mlx4_ib_qp_type) {
2647 		case MLX4_IB_QPT_RC:
2648 		case MLX4_IB_QPT_UC:
2649 			switch (wr->opcode) {
2650 			case IB_WR_ATOMIC_CMP_AND_SWP:
2651 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2652 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
2653 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2654 					      wr->wr.atomic.rkey);
2655 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2656 
2657 				set_atomic_seg(wqe, wr);
2658 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
2659 
2660 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2661 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
2662 
2663 				break;
2664 
2665 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2666 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2667 					      wr->wr.atomic.rkey);
2668 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2669 
2670 				set_masked_atomic_seg(wqe, wr);
2671 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
2672 
2673 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2674 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
2675 
2676 				break;
2677 
2678 			case IB_WR_RDMA_READ:
2679 			case IB_WR_RDMA_WRITE:
2680 			case IB_WR_RDMA_WRITE_WITH_IMM:
2681 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2682 					      wr->wr.rdma.rkey);
2683 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2684 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
2685 				break;
2686 
2687 			case IB_WR_LOCAL_INV:
2688 				ctrl->srcrb_flags |=
2689 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2690 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
2691 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
2692 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
2693 				break;
2694 
2695 			case IB_WR_FAST_REG_MR:
2696 				ctrl->srcrb_flags |=
2697 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2698 				set_fmr_seg(wqe, wr);
2699 				wqe  += sizeof (struct mlx4_wqe_fmr_seg);
2700 				size += sizeof (struct mlx4_wqe_fmr_seg) / 16;
2701 				break;
2702 
2703 			case IB_WR_BIND_MW:
2704 				ctrl->srcrb_flags |=
2705 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2706 				set_bind_seg(wqe, wr);
2707 				wqe  += sizeof(struct mlx4_wqe_bind_seg);
2708 				size += sizeof(struct mlx4_wqe_bind_seg) / 16;
2709 				break;
2710 			default:
2711 				/* No extra segments required for sends */
2712 				break;
2713 			}
2714 			break;
2715 
2716 		case MLX4_IB_QPT_TUN_SMI_OWNER:
2717 			err =  build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2718 			if (unlikely(err)) {
2719 				*bad_wr = wr;
2720 				goto out;
2721 			}
2722 			wqe  += seglen;
2723 			size += seglen / 16;
2724 			break;
2725 		case MLX4_IB_QPT_TUN_SMI:
2726 		case MLX4_IB_QPT_TUN_GSI:
2727 			/* this is a UD qp used in MAD responses to slaves. */
2728 			set_datagram_seg(wqe, wr);
2729 			/* set the forced-loopback bit in the data seg av */
2730 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
2731 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2732 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2733 			break;
2734 		case MLX4_IB_QPT_UD:
2735 			set_datagram_seg(wqe, wr);
2736 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2737 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2738 
2739 			if (wr->opcode == IB_WR_LSO) {
2740 				err = build_lso_seg(wqe, wr, qp, &seglen, &lso_hdr_sz, &blh);
2741 				if (unlikely(err)) {
2742 					*bad_wr = wr;
2743 					goto out;
2744 				}
2745 				lso_wqe = (__be32 *) wqe;
2746 				wqe  += seglen;
2747 				size += seglen / 16;
2748 			}
2749 			break;
2750 
2751 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
2752 			err = build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2753 			if (unlikely(err)) {
2754 				*bad_wr = wr;
2755 				goto out;
2756 			}
2757 			wqe  += seglen;
2758 			size += seglen / 16;
2759 			/* to start tunnel header on a cache-line boundary */
2760 			add_zero_len_inline(wqe);
2761 			wqe += 16;
2762 			size++;
2763 			build_tunnel_header(wr, wqe, &seglen);
2764 			wqe  += seglen;
2765 			size += seglen / 16;
2766 			break;
2767 		case MLX4_IB_QPT_PROXY_SMI:
2768 		case MLX4_IB_QPT_PROXY_GSI:
2769 			/* If we are tunneling special qps, this is a UD qp.
2770 			 * In this case we first add a UD segment targeting
2771 			 * the tunnel qp, and then add a header with address
2772 			 * information */
2773 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe, wr,
2774 						qp->mlx4_ib_qp_type);
2775 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2776 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2777 			build_tunnel_header(wr, wqe, &seglen);
2778 			wqe  += seglen;
2779 			size += seglen / 16;
2780 			break;
2781 
2782 		case MLX4_IB_QPT_SMI:
2783 		case MLX4_IB_QPT_GSI:
2784 			err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
2785 			if (unlikely(err)) {
2786 				*bad_wr = wr;
2787 				goto out;
2788 			}
2789 			wqe  += seglen;
2790 			size += seglen / 16;
2791 			break;
2792 
2793 		default:
2794 			break;
2795 		}
2796 
2797 		/*
2798 		 * Write data segments in reverse order, so as to
2799 		 * overwrite cacheline stamp last within each
2800 		 * cacheline.  This avoids issues with WQE
2801 		 * prefetching.
2802 		 */
2803 
2804 		dseg = wqe;
2805 		dseg += wr->num_sge - 1;
2806 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
2807 
2808 		/* Add one more inline data segment for ICRC for MLX sends */
2809 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2810 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
2811 			     qp->mlx4_ib_qp_type &
2812 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
2813 			set_mlx_icrc_seg(dseg + 1);
2814 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
2815 		}
2816 
2817 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
2818 			set_data_seg(dseg, wr->sg_list + i);
2819 
2820 		/*
2821 		 * Possibly overwrite stamping in cacheline with LSO
2822 		 * segment only after making sure all data segments
2823 		 * are written.
2824 		 */
2825 		wmb();
2826 		*lso_wqe = lso_hdr_sz;
2827 
2828 		ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
2829 				    MLX4_WQE_CTRL_FENCE : 0) | size;
2830 
2831 		/*
2832 		 * Make sure descriptor is fully written before
2833 		 * setting ownership bit (because HW can start
2834 		 * executing as soon as we do).
2835 		 */
2836 		wmb();
2837 
2838 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
2839 			*bad_wr = wr;
2840 			err = -EINVAL;
2841 			goto out;
2842 		}
2843 
2844 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
2845 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
2846 
2847 		stamp = ind + qp->sq_spare_wqes;
2848 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
2849 
2850 		/*
2851 		 * We can improve latency by not stamping the last
2852 		 * send queue WQE until after ringing the doorbell, so
2853 		 * only stamp here if there are still more WQEs to post.
2854 		 *
2855 		 * Same optimization applies to padding with NOP wqe
2856 		 * in case of WQE shrinking (used to prevent wrap-around
2857 		 * in the middle of WR).
2858 		 */
2859 		if (wr->next) {
2860 			stamp_send_wqe(qp, stamp, size * 16);
2861 			ind = pad_wraparound(qp, ind);
2862 		}
2863 	}
2864 
2865 out:
2866 	if (likely(nreq)) {
2867 		qp->sq.head += nreq;
2868 
2869 		/*
2870 		 * Make sure that descriptors are written before
2871 		 * doorbell record.
2872 		 */
2873 		wmb();
2874 
2875 		writel(qp->doorbell_qpn,
2876 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
2877 
2878 		/*
2879 		 * Make sure doorbells don't leak out of SQ spinlock
2880 		 * and reach the HCA out of order.
2881 		 */
2882 		mmiowb();
2883 
2884 		stamp_send_wqe(qp, stamp, size * 16);
2885 
2886 		ind = pad_wraparound(qp, ind);
2887 		qp->sq_next_wqe = ind;
2888 	}
2889 
2890 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2891 
2892 	return err;
2893 }
2894 
2895 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2896 		      struct ib_recv_wr **bad_wr)
2897 {
2898 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2899 	struct mlx4_wqe_data_seg *scat;
2900 	unsigned long flags;
2901 	int err = 0;
2902 	int nreq;
2903 	int ind;
2904 	int max_gs;
2905 	int i;
2906 
2907 	max_gs = qp->rq.max_gs;
2908 	spin_lock_irqsave(&qp->rq.lock, flags);
2909 
2910 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
2911 
2912 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2913 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2914 			err = -ENOMEM;
2915 			*bad_wr = wr;
2916 			goto out;
2917 		}
2918 
2919 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2920 			err = -EINVAL;
2921 			*bad_wr = wr;
2922 			goto out;
2923 		}
2924 
2925 		scat = get_recv_wqe(qp, ind);
2926 
2927 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
2928 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
2929 			ib_dma_sync_single_for_device(ibqp->device,
2930 						      qp->sqp_proxy_rcv[ind].map,
2931 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
2932 						      DMA_FROM_DEVICE);
2933 			scat->byte_count =
2934 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
2935 			/* use dma lkey from upper layer entry */
2936 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
2937 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
2938 			scat++;
2939 			max_gs--;
2940 		}
2941 
2942 		for (i = 0; i < wr->num_sge; ++i)
2943 			__set_data_seg(scat + i, wr->sg_list + i);
2944 
2945 		if (i < max_gs) {
2946 			scat[i].byte_count = 0;
2947 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
2948 			scat[i].addr       = 0;
2949 		}
2950 
2951 		qp->rq.wrid[ind] = wr->wr_id;
2952 
2953 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
2954 	}
2955 
2956 out:
2957 	if (likely(nreq)) {
2958 		qp->rq.head += nreq;
2959 
2960 		/*
2961 		 * Make sure that descriptors are written before
2962 		 * doorbell record.
2963 		 */
2964 		wmb();
2965 
2966 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
2967 	}
2968 
2969 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2970 
2971 	return err;
2972 }
2973 
2974 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
2975 {
2976 	switch (mlx4_state) {
2977 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
2978 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
2979 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
2980 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
2981 	case MLX4_QP_STATE_SQ_DRAINING:
2982 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
2983 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
2984 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
2985 	default:		     return -1;
2986 	}
2987 }
2988 
2989 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
2990 {
2991 	switch (mlx4_mig_state) {
2992 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
2993 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
2994 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
2995 	default: return -1;
2996 	}
2997 }
2998 
2999 static int to_ib_qp_access_flags(int mlx4_flags)
3000 {
3001 	int ib_flags = 0;
3002 
3003 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3004 		ib_flags |= IB_ACCESS_REMOTE_READ;
3005 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3006 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3007 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3008 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3009 
3010 	return ib_flags;
3011 }
3012 
3013 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
3014 				struct mlx4_qp_path *path)
3015 {
3016 	struct mlx4_dev *dev = ibdev->dev;
3017 	int is_eth;
3018 
3019 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
3020 	ib_ah_attr->port_num	  = path->sched_queue & 0x40 ? 2 : 1;
3021 
3022 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
3023 		return;
3024 
3025 	is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
3026 		IB_LINK_LAYER_ETHERNET;
3027 	if (is_eth)
3028 		ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
3029 		((path->sched_queue & 4) << 1);
3030 	else
3031 		ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
3032 
3033 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
3034 	ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
3035 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
3036 	ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
3037 	if (ib_ah_attr->ah_flags) {
3038 		ib_ah_attr->grh.sgid_index = path->mgid_index;
3039 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
3040 		ib_ah_attr->grh.traffic_class =
3041 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3042 		ib_ah_attr->grh.flow_label =
3043 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3044 		memcpy(ib_ah_attr->grh.dgid.raw,
3045 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
3046 	}
3047 }
3048 
3049 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3050 		     struct ib_qp_init_attr *qp_init_attr)
3051 {
3052 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3053 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3054 	struct mlx4_qp_context context;
3055 	int mlx4_state;
3056 	int err = 0;
3057 
3058 	mutex_lock(&qp->mutex);
3059 
3060 	if (qp->state == IB_QPS_RESET) {
3061 		qp_attr->qp_state = IB_QPS_RESET;
3062 		goto done;
3063 	}
3064 
3065 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3066 	if (err) {
3067 		err = -EINVAL;
3068 		goto out;
3069 	}
3070 
3071 	mlx4_state = be32_to_cpu(context.flags) >> 28;
3072 
3073 	qp->state		     = to_ib_qp_state(mlx4_state);
3074 	qp_attr->qp_state	     = qp->state;
3075 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
3076 	qp_attr->path_mig_state	     =
3077 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3078 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
3079 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3080 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
3081 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
3082 	qp_attr->qp_access_flags     =
3083 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
3084 
3085 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3086 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3087 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3088 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3089 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
3090 	}
3091 
3092 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3093 	if (qp_attr->qp_state == IB_QPS_INIT)
3094 		qp_attr->port_num = qp->port;
3095 	else
3096 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3097 
3098 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3099 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3100 
3101 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3102 
3103 	qp_attr->max_dest_rd_atomic =
3104 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3105 	qp_attr->min_rnr_timer	    =
3106 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3107 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
3108 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
3109 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
3110 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
3111 
3112 done:
3113 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
3114 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
3115 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
3116 
3117 	if (!ibqp->uobject) {
3118 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
3119 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
3120 	} else {
3121 		qp_attr->cap.max_send_wr  = 0;
3122 		qp_attr->cap.max_send_sge = 0;
3123 	}
3124 
3125 	/*
3126 	 * We don't support inline sends for kernel QPs (yet), and we
3127 	 * don't know what userspace's value should be.
3128 	 */
3129 	qp_attr->cap.max_inline_data = 0;
3130 
3131 	qp_init_attr->cap	     = qp_attr->cap;
3132 
3133 	qp_init_attr->create_flags = 0;
3134 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3135 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3136 
3137 	if (qp->flags & MLX4_IB_QP_LSO)
3138 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
3139 
3140 	if (qp->flags & MLX4_IB_QP_NETIF)
3141 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
3142 
3143 	qp_init_attr->sq_sig_type =
3144 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
3145 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3146 
3147 out:
3148 	mutex_unlock(&qp->mutex);
3149 	return err;
3150 }
3151 
3152