xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision ee89bd6b)
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 static const __be32 mlx4_ib_opcode[] = {
94 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
95 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
96 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
97 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
98 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
99 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
100 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
101 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
102 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
103 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
104 	[IB_WR_FAST_REG_MR]			= cpu_to_be32(MLX4_OPCODE_FMR),
105 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
106 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
107 	[IB_WR_BIND_MW]				= cpu_to_be32(MLX4_OPCODE_BIND_MW),
108 };
109 
110 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
111 {
112 	return container_of(mqp, struct mlx4_ib_sqp, qp);
113 }
114 
115 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
116 {
117 	if (!mlx4_is_master(dev->dev))
118 		return 0;
119 
120 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
121 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
122 		8 * MLX4_MFUNC_MAX;
123 }
124 
125 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
126 {
127 	int proxy_sqp = 0;
128 	int real_sqp = 0;
129 	int i;
130 	/* PPF or Native -- real SQP */
131 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
132 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
133 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
134 	if (real_sqp)
135 		return 1;
136 	/* VF or PF -- proxy SQP */
137 	if (mlx4_is_mfunc(dev->dev)) {
138 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
139 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] ||
140 			    qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) {
141 				proxy_sqp = 1;
142 				break;
143 			}
144 		}
145 	}
146 	return proxy_sqp;
147 }
148 
149 /* used for INIT/CLOSE port logic */
150 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
151 {
152 	int proxy_qp0 = 0;
153 	int real_qp0 = 0;
154 	int i;
155 	/* PPF or Native -- real QP0 */
156 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
157 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
158 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
159 	if (real_qp0)
160 		return 1;
161 	/* VF or PF -- proxy QP0 */
162 	if (mlx4_is_mfunc(dev->dev)) {
163 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
164 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) {
165 				proxy_qp0 = 1;
166 				break;
167 			}
168 		}
169 	}
170 	return proxy_qp0;
171 }
172 
173 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
174 {
175 	return mlx4_buf_offset(&qp->buf, offset);
176 }
177 
178 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
179 {
180 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
181 }
182 
183 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
184 {
185 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
186 }
187 
188 /*
189  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
190  * first four bytes of every 64 byte chunk with
191  *     0x7FFFFFF | (invalid_ownership_value << 31).
192  *
193  * When the max work request size is less than or equal to the WQE
194  * basic block size, as an optimization, we can stamp all WQEs with
195  * 0xffffffff, and skip the very first chunk of each WQE.
196  */
197 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
198 {
199 	__be32 *wqe;
200 	int i;
201 	int s;
202 	int ind;
203 	void *buf;
204 	__be32 stamp;
205 	struct mlx4_wqe_ctrl_seg *ctrl;
206 
207 	if (qp->sq_max_wqes_per_wr > 1) {
208 		s = roundup(size, 1U << qp->sq.wqe_shift);
209 		for (i = 0; i < s; i += 64) {
210 			ind = (i >> qp->sq.wqe_shift) + n;
211 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
212 						       cpu_to_be32(0xffffffff);
213 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
214 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
215 			*wqe = stamp;
216 		}
217 	} else {
218 		ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
219 		s = (ctrl->fence_size & 0x3f) << 4;
220 		for (i = 64; i < s; i += 64) {
221 			wqe = buf + i;
222 			*wqe = cpu_to_be32(0xffffffff);
223 		}
224 	}
225 }
226 
227 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
228 {
229 	struct mlx4_wqe_ctrl_seg *ctrl;
230 	struct mlx4_wqe_inline_seg *inl;
231 	void *wqe;
232 	int s;
233 
234 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
235 	s = sizeof(struct mlx4_wqe_ctrl_seg);
236 
237 	if (qp->ibqp.qp_type == IB_QPT_UD) {
238 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
239 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
240 		memset(dgram, 0, sizeof *dgram);
241 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
242 		s += sizeof(struct mlx4_wqe_datagram_seg);
243 	}
244 
245 	/* Pad the remainder of the WQE with an inline data segment. */
246 	if (size > s) {
247 		inl = wqe + s;
248 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
249 	}
250 	ctrl->srcrb_flags = 0;
251 	ctrl->fence_size = size / 16;
252 	/*
253 	 * Make sure descriptor is fully written before setting ownership bit
254 	 * (because HW can start executing as soon as we do).
255 	 */
256 	wmb();
257 
258 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
259 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
260 
261 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
262 }
263 
264 /* Post NOP WQE to prevent wrap-around in the middle of WR */
265 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
266 {
267 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
268 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
269 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
270 		ind += s;
271 	}
272 	return ind;
273 }
274 
275 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
276 {
277 	struct ib_event event;
278 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
279 
280 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
281 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
282 
283 	if (ibqp->event_handler) {
284 		event.device     = ibqp->device;
285 		event.element.qp = ibqp;
286 		switch (type) {
287 		case MLX4_EVENT_TYPE_PATH_MIG:
288 			event.event = IB_EVENT_PATH_MIG;
289 			break;
290 		case MLX4_EVENT_TYPE_COMM_EST:
291 			event.event = IB_EVENT_COMM_EST;
292 			break;
293 		case MLX4_EVENT_TYPE_SQ_DRAINED:
294 			event.event = IB_EVENT_SQ_DRAINED;
295 			break;
296 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
297 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
298 			break;
299 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
300 			event.event = IB_EVENT_QP_FATAL;
301 			break;
302 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
303 			event.event = IB_EVENT_PATH_MIG_ERR;
304 			break;
305 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
306 			event.event = IB_EVENT_QP_REQ_ERR;
307 			break;
308 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
309 			event.event = IB_EVENT_QP_ACCESS_ERR;
310 			break;
311 		default:
312 			pr_warn("Unexpected event type %d "
313 			       "on QP %06x\n", type, qp->qpn);
314 			return;
315 		}
316 
317 		ibqp->event_handler(&event, ibqp->qp_context);
318 	}
319 }
320 
321 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
322 {
323 	/*
324 	 * UD WQEs must have a datagram segment.
325 	 * RC and UC WQEs might have a remote address segment.
326 	 * MLX WQEs need two extra inline data segments (for the UD
327 	 * header and space for the ICRC).
328 	 */
329 	switch (type) {
330 	case MLX4_IB_QPT_UD:
331 		return sizeof (struct mlx4_wqe_ctrl_seg) +
332 			sizeof (struct mlx4_wqe_datagram_seg) +
333 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
334 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
335 	case MLX4_IB_QPT_PROXY_SMI:
336 	case MLX4_IB_QPT_PROXY_GSI:
337 		return sizeof (struct mlx4_wqe_ctrl_seg) +
338 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
339 	case MLX4_IB_QPT_TUN_SMI_OWNER:
340 	case MLX4_IB_QPT_TUN_GSI:
341 		return sizeof (struct mlx4_wqe_ctrl_seg) +
342 			sizeof (struct mlx4_wqe_datagram_seg);
343 
344 	case MLX4_IB_QPT_UC:
345 		return sizeof (struct mlx4_wqe_ctrl_seg) +
346 			sizeof (struct mlx4_wqe_raddr_seg);
347 	case MLX4_IB_QPT_RC:
348 		return sizeof (struct mlx4_wqe_ctrl_seg) +
349 			sizeof (struct mlx4_wqe_atomic_seg) +
350 			sizeof (struct mlx4_wqe_raddr_seg);
351 	case MLX4_IB_QPT_SMI:
352 	case MLX4_IB_QPT_GSI:
353 		return sizeof (struct mlx4_wqe_ctrl_seg) +
354 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
355 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
356 					   MLX4_INLINE_ALIGN) *
357 			      sizeof (struct mlx4_wqe_inline_seg),
358 			      sizeof (struct mlx4_wqe_data_seg)) +
359 			ALIGN(4 +
360 			      sizeof (struct mlx4_wqe_inline_seg),
361 			      sizeof (struct mlx4_wqe_data_seg));
362 	default:
363 		return sizeof (struct mlx4_wqe_ctrl_seg);
364 	}
365 }
366 
367 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
368 		       int is_user, int has_rq, struct mlx4_ib_qp *qp)
369 {
370 	/* Sanity check RQ size before proceeding */
371 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
372 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
373 		return -EINVAL;
374 
375 	if (!has_rq) {
376 		if (cap->max_recv_wr)
377 			return -EINVAL;
378 
379 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
380 	} else {
381 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
382 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
383 			return -EINVAL;
384 
385 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
386 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
387 		qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
388 	}
389 
390 	/* leave userspace return values as they were, so as not to break ABI */
391 	if (is_user) {
392 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
393 		cap->max_recv_sge = qp->rq.max_gs;
394 	} else {
395 		cap->max_recv_wr  = qp->rq.max_post =
396 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
397 		cap->max_recv_sge = min(qp->rq.max_gs,
398 					min(dev->dev->caps.max_sq_sg,
399 					    dev->dev->caps.max_rq_sg));
400 	}
401 
402 	return 0;
403 }
404 
405 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
406 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
407 {
408 	int s;
409 
410 	/* Sanity check SQ size before proceeding */
411 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
412 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
413 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
414 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
415 		return -EINVAL;
416 
417 	/*
418 	 * For MLX transport we need 2 extra S/G entries:
419 	 * one for the header and one for the checksum at the end
420 	 */
421 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
422 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
423 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
424 		return -EINVAL;
425 
426 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
427 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
428 		send_wqe_overhead(type, qp->flags);
429 
430 	if (s > dev->dev->caps.max_sq_desc_sz)
431 		return -EINVAL;
432 
433 	/*
434 	 * Hermon supports shrinking WQEs, such that a single work
435 	 * request can include multiple units of 1 << wqe_shift.  This
436 	 * way, work requests can differ in size, and do not have to
437 	 * be a power of 2 in size, saving memory and speeding up send
438 	 * WR posting.  Unfortunately, if we do this then the
439 	 * wqe_index field in CQEs can't be used to look up the WR ID
440 	 * anymore, so we do this only if selective signaling is off.
441 	 *
442 	 * Further, on 32-bit platforms, we can't use vmap() to make
443 	 * the QP buffer virtually contiguous.  Thus we have to use
444 	 * constant-sized WRs to make sure a WR is always fully within
445 	 * a single page-sized chunk.
446 	 *
447 	 * Finally, we use NOP work requests to pad the end of the
448 	 * work queue, to avoid wrap-around in the middle of WR.  We
449 	 * set NEC bit to avoid getting completions with error for
450 	 * these NOP WRs, but since NEC is only supported starting
451 	 * with firmware 2.2.232, we use constant-sized WRs for older
452 	 * firmware.
453 	 *
454 	 * And, since MLX QPs only support SEND, we use constant-sized
455 	 * WRs in this case.
456 	 *
457 	 * We look for the smallest value of wqe_shift such that the
458 	 * resulting number of wqes does not exceed device
459 	 * capabilities.
460 	 *
461 	 * We set WQE size to at least 64 bytes, this way stamping
462 	 * invalidates each WQE.
463 	 */
464 	if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
465 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
466 	    type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
467 	    !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
468 		      MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
469 		qp->sq.wqe_shift = ilog2(64);
470 	else
471 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
472 
473 	for (;;) {
474 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
475 
476 		/*
477 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
478 		 * allow HW to prefetch.
479 		 */
480 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
481 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
482 						    qp->sq_max_wqes_per_wr +
483 						    qp->sq_spare_wqes);
484 
485 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
486 			break;
487 
488 		if (qp->sq_max_wqes_per_wr <= 1)
489 			return -EINVAL;
490 
491 		++qp->sq.wqe_shift;
492 	}
493 
494 	qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
495 			     (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
496 			 send_wqe_overhead(type, qp->flags)) /
497 		sizeof (struct mlx4_wqe_data_seg);
498 
499 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
500 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
501 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
502 		qp->rq.offset = 0;
503 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
504 	} else {
505 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
506 		qp->sq.offset = 0;
507 	}
508 
509 	cap->max_send_wr  = qp->sq.max_post =
510 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
511 	cap->max_send_sge = min(qp->sq.max_gs,
512 				min(dev->dev->caps.max_sq_sg,
513 				    dev->dev->caps.max_rq_sg));
514 	/* We don't support inline sends for kernel QPs (yet) */
515 	cap->max_inline_data = 0;
516 
517 	return 0;
518 }
519 
520 static int set_user_sq_size(struct mlx4_ib_dev *dev,
521 			    struct mlx4_ib_qp *qp,
522 			    struct mlx4_ib_create_qp *ucmd)
523 {
524 	/* Sanity check SQ size before proceeding */
525 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
526 	    ucmd->log_sq_stride >
527 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
528 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
529 		return -EINVAL;
530 
531 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
532 	qp->sq.wqe_shift = ucmd->log_sq_stride;
533 
534 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
535 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
536 
537 	return 0;
538 }
539 
540 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
541 {
542 	int i;
543 
544 	qp->sqp_proxy_rcv =
545 		kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
546 			GFP_KERNEL);
547 	if (!qp->sqp_proxy_rcv)
548 		return -ENOMEM;
549 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
550 		qp->sqp_proxy_rcv[i].addr =
551 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
552 				GFP_KERNEL);
553 		if (!qp->sqp_proxy_rcv[i].addr)
554 			goto err;
555 		qp->sqp_proxy_rcv[i].map =
556 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
557 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
558 					  DMA_FROM_DEVICE);
559 	}
560 	return 0;
561 
562 err:
563 	while (i > 0) {
564 		--i;
565 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
566 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
567 				    DMA_FROM_DEVICE);
568 		kfree(qp->sqp_proxy_rcv[i].addr);
569 	}
570 	kfree(qp->sqp_proxy_rcv);
571 	qp->sqp_proxy_rcv = NULL;
572 	return -ENOMEM;
573 }
574 
575 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
576 {
577 	int i;
578 
579 	for (i = 0; i < qp->rq.wqe_cnt; 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 }
587 
588 static int qp_has_rq(struct ib_qp_init_attr *attr)
589 {
590 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
591 		return 0;
592 
593 	return !attr->srq;
594 }
595 
596 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
597 			    struct ib_qp_init_attr *init_attr,
598 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp)
599 {
600 	int qpn;
601 	int err;
602 	struct mlx4_ib_sqp *sqp;
603 	struct mlx4_ib_qp *qp;
604 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
605 
606 	/* When tunneling special qps, we use a plain UD qp */
607 	if (sqpn) {
608 		if (mlx4_is_mfunc(dev->dev) &&
609 		    (!mlx4_is_master(dev->dev) ||
610 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
611 			if (init_attr->qp_type == IB_QPT_GSI)
612 				qp_type = MLX4_IB_QPT_PROXY_GSI;
613 			else if (mlx4_is_master(dev->dev))
614 				qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
615 			else
616 				qp_type = MLX4_IB_QPT_PROXY_SMI;
617 		}
618 		qpn = sqpn;
619 		/* add extra sg entry for tunneling */
620 		init_attr->cap.max_recv_sge++;
621 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
622 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
623 			container_of(init_attr,
624 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
625 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
626 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
627 		    !mlx4_is_master(dev->dev))
628 			return -EINVAL;
629 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
630 			qp_type = MLX4_IB_QPT_TUN_GSI;
631 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev))
632 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
633 		else
634 			qp_type = MLX4_IB_QPT_TUN_SMI;
635 		/* we are definitely in the PPF here, since we are creating
636 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
637 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
638 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
639 		sqpn = qpn;
640 	}
641 
642 	if (!*caller_qp) {
643 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
644 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
645 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
646 			sqp = kzalloc(sizeof (struct mlx4_ib_sqp), GFP_KERNEL);
647 			if (!sqp)
648 				return -ENOMEM;
649 			qp = &sqp->qp;
650 		} else {
651 			qp = kzalloc(sizeof (struct mlx4_ib_qp), GFP_KERNEL);
652 			if (!qp)
653 				return -ENOMEM;
654 		}
655 	} else
656 		qp = *caller_qp;
657 
658 	qp->mlx4_ib_qp_type = qp_type;
659 
660 	mutex_init(&qp->mutex);
661 	spin_lock_init(&qp->sq.lock);
662 	spin_lock_init(&qp->rq.lock);
663 	INIT_LIST_HEAD(&qp->gid_list);
664 	INIT_LIST_HEAD(&qp->steering_rules);
665 
666 	qp->state	 = IB_QPS_RESET;
667 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
668 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
669 
670 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
671 	if (err)
672 		goto err;
673 
674 	if (pd->uobject) {
675 		struct mlx4_ib_create_qp ucmd;
676 
677 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
678 			err = -EFAULT;
679 			goto err;
680 		}
681 
682 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
683 
684 		err = set_user_sq_size(dev, qp, &ucmd);
685 		if (err)
686 			goto err;
687 
688 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
689 				       qp->buf_size, 0, 0);
690 		if (IS_ERR(qp->umem)) {
691 			err = PTR_ERR(qp->umem);
692 			goto err;
693 		}
694 
695 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
696 				    ilog2(qp->umem->page_size), &qp->mtt);
697 		if (err)
698 			goto err_buf;
699 
700 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
701 		if (err)
702 			goto err_mtt;
703 
704 		if (qp_has_rq(init_attr)) {
705 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
706 						  ucmd.db_addr, &qp->db);
707 			if (err)
708 				goto err_mtt;
709 		}
710 	} else {
711 		qp->sq_no_prefetch = 0;
712 
713 		if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
714 			qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
715 
716 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
717 			qp->flags |= MLX4_IB_QP_LSO;
718 
719 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
720 		if (err)
721 			goto err;
722 
723 		if (qp_has_rq(init_attr)) {
724 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
725 			if (err)
726 				goto err;
727 
728 			*qp->db.db = 0;
729 		}
730 
731 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf)) {
732 			err = -ENOMEM;
733 			goto err_db;
734 		}
735 
736 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
737 				    &qp->mtt);
738 		if (err)
739 			goto err_buf;
740 
741 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
742 		if (err)
743 			goto err_mtt;
744 
745 		qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), GFP_KERNEL);
746 		qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), GFP_KERNEL);
747 
748 		if (!qp->sq.wrid || !qp->rq.wrid) {
749 			err = -ENOMEM;
750 			goto err_wrid;
751 		}
752 	}
753 
754 	if (sqpn) {
755 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
756 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
757 			if (alloc_proxy_bufs(pd->device, qp)) {
758 				err = -ENOMEM;
759 				goto err_wrid;
760 			}
761 		}
762 	} else {
763 		/* Raw packet QPNs must be aligned to 8 bits. If not, the WQE
764 		 * BlueFlame setup flow wrongly causes VLAN insertion. */
765 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
766 			err = mlx4_qp_reserve_range(dev->dev, 1, 1 << 8, &qpn);
767 		else
768 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn);
769 		if (err)
770 			goto err_proxy;
771 	}
772 
773 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
774 	if (err)
775 		goto err_qpn;
776 
777 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
778 		qp->mqp.qpn |= (1 << 23);
779 
780 	/*
781 	 * Hardware wants QPN written in big-endian order (after
782 	 * shifting) for send doorbell.  Precompute this value to save
783 	 * a little bit when posting sends.
784 	 */
785 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
786 
787 	qp->mqp.event = mlx4_ib_qp_event;
788 	if (!*caller_qp)
789 		*caller_qp = qp;
790 	return 0;
791 
792 err_qpn:
793 	if (!sqpn)
794 		mlx4_qp_release_range(dev->dev, qpn, 1);
795 err_proxy:
796 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
797 		free_proxy_bufs(pd->device, qp);
798 err_wrid:
799 	if (pd->uobject) {
800 		if (qp_has_rq(init_attr))
801 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
802 	} else {
803 		kfree(qp->sq.wrid);
804 		kfree(qp->rq.wrid);
805 	}
806 
807 err_mtt:
808 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
809 
810 err_buf:
811 	if (pd->uobject)
812 		ib_umem_release(qp->umem);
813 	else
814 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
815 
816 err_db:
817 	if (!pd->uobject && qp_has_rq(init_attr))
818 		mlx4_db_free(dev->dev, &qp->db);
819 
820 err:
821 	if (!*caller_qp)
822 		kfree(qp);
823 	return err;
824 }
825 
826 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
827 {
828 	switch (state) {
829 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
830 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
831 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
832 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
833 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
834 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
835 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
836 	default:		return -1;
837 	}
838 }
839 
840 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
841 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
842 {
843 	if (send_cq == recv_cq) {
844 		spin_lock_irq(&send_cq->lock);
845 		__acquire(&recv_cq->lock);
846 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
847 		spin_lock_irq(&send_cq->lock);
848 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
849 	} else {
850 		spin_lock_irq(&recv_cq->lock);
851 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
852 	}
853 }
854 
855 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
856 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
857 {
858 	if (send_cq == recv_cq) {
859 		__release(&recv_cq->lock);
860 		spin_unlock_irq(&send_cq->lock);
861 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
862 		spin_unlock(&recv_cq->lock);
863 		spin_unlock_irq(&send_cq->lock);
864 	} else {
865 		spin_unlock(&send_cq->lock);
866 		spin_unlock_irq(&recv_cq->lock);
867 	}
868 }
869 
870 static void del_gid_entries(struct mlx4_ib_qp *qp)
871 {
872 	struct mlx4_ib_gid_entry *ge, *tmp;
873 
874 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
875 		list_del(&ge->list);
876 		kfree(ge);
877 	}
878 }
879 
880 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
881 {
882 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
883 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
884 	else
885 		return to_mpd(qp->ibqp.pd);
886 }
887 
888 static void get_cqs(struct mlx4_ib_qp *qp,
889 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
890 {
891 	switch (qp->ibqp.qp_type) {
892 	case IB_QPT_XRC_TGT:
893 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
894 		*recv_cq = *send_cq;
895 		break;
896 	case IB_QPT_XRC_INI:
897 		*send_cq = to_mcq(qp->ibqp.send_cq);
898 		*recv_cq = *send_cq;
899 		break;
900 	default:
901 		*send_cq = to_mcq(qp->ibqp.send_cq);
902 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
903 		break;
904 	}
905 }
906 
907 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
908 			      int is_user)
909 {
910 	struct mlx4_ib_cq *send_cq, *recv_cq;
911 
912 	if (qp->state != IB_QPS_RESET)
913 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
914 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
915 			pr_warn("modify QP %06x to RESET failed.\n",
916 			       qp->mqp.qpn);
917 
918 	get_cqs(qp, &send_cq, &recv_cq);
919 
920 	mlx4_ib_lock_cqs(send_cq, recv_cq);
921 
922 	if (!is_user) {
923 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
924 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
925 		if (send_cq != recv_cq)
926 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
927 	}
928 
929 	mlx4_qp_remove(dev->dev, &qp->mqp);
930 
931 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
932 
933 	mlx4_qp_free(dev->dev, &qp->mqp);
934 
935 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp))
936 		mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
937 
938 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
939 
940 	if (is_user) {
941 		if (qp->rq.wqe_cnt)
942 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
943 					      &qp->db);
944 		ib_umem_release(qp->umem);
945 	} else {
946 		kfree(qp->sq.wrid);
947 		kfree(qp->rq.wrid);
948 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
949 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
950 			free_proxy_bufs(&dev->ib_dev, qp);
951 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
952 		if (qp->rq.wqe_cnt)
953 			mlx4_db_free(dev->dev, &qp->db);
954 	}
955 
956 	del_gid_entries(qp);
957 }
958 
959 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
960 {
961 	/* Native or PPF */
962 	if (!mlx4_is_mfunc(dev->dev) ||
963 	    (mlx4_is_master(dev->dev) &&
964 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
965 		return  dev->dev->phys_caps.base_sqpn +
966 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
967 			attr->port_num - 1;
968 	}
969 	/* PF or VF -- creating proxies */
970 	if (attr->qp_type == IB_QPT_SMI)
971 		return dev->dev->caps.qp0_proxy[attr->port_num - 1];
972 	else
973 		return dev->dev->caps.qp1_proxy[attr->port_num - 1];
974 }
975 
976 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
977 				struct ib_qp_init_attr *init_attr,
978 				struct ib_udata *udata)
979 {
980 	struct mlx4_ib_qp *qp = NULL;
981 	int err;
982 	u16 xrcdn = 0;
983 
984 	/*
985 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
986 	 * and only for kernel UD QPs.
987 	 */
988 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
989 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
990 					MLX4_IB_SRIOV_TUNNEL_QP | MLX4_IB_SRIOV_SQP))
991 		return ERR_PTR(-EINVAL);
992 
993 	if (init_attr->create_flags &&
994 	    (udata ||
995 	     ((init_attr->create_flags & ~MLX4_IB_SRIOV_SQP) &&
996 	      init_attr->qp_type != IB_QPT_UD) ||
997 	     ((init_attr->create_flags & MLX4_IB_SRIOV_SQP) &&
998 	      init_attr->qp_type > IB_QPT_GSI)))
999 		return ERR_PTR(-EINVAL);
1000 
1001 	switch (init_attr->qp_type) {
1002 	case IB_QPT_XRC_TGT:
1003 		pd = to_mxrcd(init_attr->xrcd)->pd;
1004 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1005 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1006 		/* fall through */
1007 	case IB_QPT_XRC_INI:
1008 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1009 			return ERR_PTR(-ENOSYS);
1010 		init_attr->recv_cq = init_attr->send_cq;
1011 		/* fall through */
1012 	case IB_QPT_RC:
1013 	case IB_QPT_UC:
1014 	case IB_QPT_RAW_PACKET:
1015 		qp = kzalloc(sizeof *qp, GFP_KERNEL);
1016 		if (!qp)
1017 			return ERR_PTR(-ENOMEM);
1018 		/* fall through */
1019 	case IB_QPT_UD:
1020 	{
1021 		err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1022 				       udata, 0, &qp);
1023 		if (err)
1024 			return ERR_PTR(err);
1025 
1026 		qp->ibqp.qp_num = qp->mqp.qpn;
1027 		qp->xrcdn = xrcdn;
1028 
1029 		break;
1030 	}
1031 	case IB_QPT_SMI:
1032 	case IB_QPT_GSI:
1033 	{
1034 		/* Userspace is not allowed to create special QPs: */
1035 		if (udata)
1036 			return ERR_PTR(-EINVAL);
1037 
1038 		err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1039 				       get_sqp_num(to_mdev(pd->device), init_attr),
1040 				       &qp);
1041 		if (err)
1042 			return ERR_PTR(err);
1043 
1044 		qp->port	= init_attr->port_num;
1045 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
1046 
1047 		break;
1048 	}
1049 	default:
1050 		/* Don't support raw QPs */
1051 		return ERR_PTR(-EINVAL);
1052 	}
1053 
1054 	return &qp->ibqp;
1055 }
1056 
1057 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1058 {
1059 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1060 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1061 	struct mlx4_ib_pd *pd;
1062 
1063 	if (is_qp0(dev, mqp))
1064 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1065 
1066 	pd = get_pd(mqp);
1067 	destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1068 
1069 	if (is_sqp(dev, mqp))
1070 		kfree(to_msqp(mqp));
1071 	else
1072 		kfree(mqp);
1073 
1074 	return 0;
1075 }
1076 
1077 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1078 {
1079 	switch (type) {
1080 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1081 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1082 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1083 	case MLX4_IB_QPT_XRC_INI:
1084 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1085 	case MLX4_IB_QPT_SMI:
1086 	case MLX4_IB_QPT_GSI:
1087 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1088 
1089 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1090 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1091 						MLX4_QP_ST_MLX : -1);
1092 	case MLX4_IB_QPT_PROXY_SMI:
1093 	case MLX4_IB_QPT_TUN_SMI:
1094 	case MLX4_IB_QPT_PROXY_GSI:
1095 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1096 						MLX4_QP_ST_UD : -1);
1097 	default:			return -1;
1098 	}
1099 }
1100 
1101 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1102 				   int attr_mask)
1103 {
1104 	u8 dest_rd_atomic;
1105 	u32 access_flags;
1106 	u32 hw_access_flags = 0;
1107 
1108 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1109 		dest_rd_atomic = attr->max_dest_rd_atomic;
1110 	else
1111 		dest_rd_atomic = qp->resp_depth;
1112 
1113 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1114 		access_flags = attr->qp_access_flags;
1115 	else
1116 		access_flags = qp->atomic_rd_en;
1117 
1118 	if (!dest_rd_atomic)
1119 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1120 
1121 	if (access_flags & IB_ACCESS_REMOTE_READ)
1122 		hw_access_flags |= MLX4_QP_BIT_RRE;
1123 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1124 		hw_access_flags |= MLX4_QP_BIT_RAE;
1125 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1126 		hw_access_flags |= MLX4_QP_BIT_RWE;
1127 
1128 	return cpu_to_be32(hw_access_flags);
1129 }
1130 
1131 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1132 			    int attr_mask)
1133 {
1134 	if (attr_mask & IB_QP_PKEY_INDEX)
1135 		sqp->pkey_index = attr->pkey_index;
1136 	if (attr_mask & IB_QP_QKEY)
1137 		sqp->qkey = attr->qkey;
1138 	if (attr_mask & IB_QP_SQ_PSN)
1139 		sqp->send_psn = attr->sq_psn;
1140 }
1141 
1142 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1143 {
1144 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1145 }
1146 
1147 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
1148 			 struct mlx4_qp_path *path, u8 port)
1149 {
1150 	int err;
1151 	int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
1152 		IB_LINK_LAYER_ETHERNET;
1153 	u8 mac[6];
1154 	int is_mcast;
1155 	u16 vlan_tag;
1156 	int vidx;
1157 
1158 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
1159 	path->rlid	    = cpu_to_be16(ah->dlid);
1160 	if (ah->static_rate) {
1161 		path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
1162 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1163 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1164 			--path->static_rate;
1165 	} else
1166 		path->static_rate = 0;
1167 
1168 	if (ah->ah_flags & IB_AH_GRH) {
1169 		if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
1170 			pr_err("sgid_index (%u) too large. max is %d\n",
1171 			       ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1172 			return -1;
1173 		}
1174 
1175 		path->grh_mylmc |= 1 << 7;
1176 		path->mgid_index = ah->grh.sgid_index;
1177 		path->hop_limit  = ah->grh.hop_limit;
1178 		path->tclass_flowlabel =
1179 			cpu_to_be32((ah->grh.traffic_class << 20) |
1180 				    (ah->grh.flow_label));
1181 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1182 	}
1183 
1184 	if (is_eth) {
1185 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1186 			((port - 1) << 6) | ((ah->sl & 7) << 3);
1187 
1188 		if (!(ah->ah_flags & IB_AH_GRH))
1189 			return -1;
1190 
1191 		err = mlx4_ib_resolve_grh(dev, ah, mac, &is_mcast, port);
1192 		if (err)
1193 			return err;
1194 
1195 		memcpy(path->dmac, mac, 6);
1196 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1197 		/* use index 0 into MAC table for IBoE */
1198 		path->grh_mylmc &= 0x80;
1199 
1200 		vlan_tag = rdma_get_vlan_id(&dev->iboe.gid_table[port - 1][ah->grh.sgid_index]);
1201 		if (vlan_tag < 0x1000) {
1202 			if (mlx4_find_cached_vlan(dev->dev, port, vlan_tag, &vidx))
1203 				return -ENOENT;
1204 
1205 			path->vlan_index = vidx;
1206 			path->fl = 1 << 6;
1207 		}
1208 	} else
1209 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1210 			((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1211 
1212 	return 0;
1213 }
1214 
1215 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1216 {
1217 	struct mlx4_ib_gid_entry *ge, *tmp;
1218 
1219 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1220 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1221 			ge->added = 1;
1222 			ge->port = qp->port;
1223 		}
1224 	}
1225 }
1226 
1227 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1228 			       const struct ib_qp_attr *attr, int attr_mask,
1229 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1230 {
1231 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1232 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1233 	struct mlx4_ib_pd *pd;
1234 	struct mlx4_ib_cq *send_cq, *recv_cq;
1235 	struct mlx4_qp_context *context;
1236 	enum mlx4_qp_optpar optpar = 0;
1237 	int sqd_event;
1238 	int err = -EINVAL;
1239 
1240 	context = kzalloc(sizeof *context, GFP_KERNEL);
1241 	if (!context)
1242 		return -ENOMEM;
1243 
1244 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1245 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1246 
1247 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1248 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1249 	else {
1250 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
1251 		switch (attr->path_mig_state) {
1252 		case IB_MIG_MIGRATED:
1253 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1254 			break;
1255 		case IB_MIG_REARM:
1256 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1257 			break;
1258 		case IB_MIG_ARMED:
1259 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1260 			break;
1261 		}
1262 	}
1263 
1264 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1265 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1266 	else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1267 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1268 	else if (ibqp->qp_type == IB_QPT_UD) {
1269 		if (qp->flags & MLX4_IB_QP_LSO)
1270 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
1271 					      ilog2(dev->dev->caps.max_gso_sz);
1272 		else
1273 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1274 	} else if (attr_mask & IB_QP_PATH_MTU) {
1275 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1276 			pr_err("path MTU (%u) is invalid\n",
1277 			       attr->path_mtu);
1278 			goto out;
1279 		}
1280 		context->mtu_msgmax = (attr->path_mtu << 5) |
1281 			ilog2(dev->dev->caps.max_msg_sz);
1282 	}
1283 
1284 	if (qp->rq.wqe_cnt)
1285 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1286 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
1287 
1288 	if (qp->sq.wqe_cnt)
1289 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1290 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
1291 
1292 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1293 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1294 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1295 		if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1296 			context->param3 |= cpu_to_be32(1 << 30);
1297 	}
1298 
1299 	if (qp->ibqp.uobject)
1300 		context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
1301 	else
1302 		context->usr_page = cpu_to_be32(dev->priv_uar.index);
1303 
1304 	if (attr_mask & IB_QP_DEST_QPN)
1305 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1306 
1307 	if (attr_mask & IB_QP_PORT) {
1308 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1309 		    !(attr_mask & IB_QP_AV)) {
1310 			mlx4_set_sched(&context->pri_path, attr->port_num);
1311 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1312 		}
1313 	}
1314 
1315 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1316 		if (dev->counters[qp->port - 1] != -1) {
1317 			context->pri_path.counter_index =
1318 						dev->counters[qp->port - 1];
1319 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1320 		} else
1321 			context->pri_path.counter_index = 0xff;
1322 	}
1323 
1324 	if (attr_mask & IB_QP_PKEY_INDEX) {
1325 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1326 			context->pri_path.disable_pkey_check = 0x40;
1327 		context->pri_path.pkey_index = attr->pkey_index;
1328 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1329 	}
1330 
1331 	if (attr_mask & IB_QP_AV) {
1332 		if (mlx4_set_path(dev, &attr->ah_attr, &context->pri_path,
1333 				  attr_mask & IB_QP_PORT ?
1334 				  attr->port_num : qp->port))
1335 			goto out;
1336 
1337 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1338 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
1339 	}
1340 
1341 	if (attr_mask & IB_QP_TIMEOUT) {
1342 		context->pri_path.ackto |= attr->timeout << 3;
1343 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1344 	}
1345 
1346 	if (attr_mask & IB_QP_ALT_PATH) {
1347 		if (attr->alt_port_num == 0 ||
1348 		    attr->alt_port_num > dev->dev->caps.num_ports)
1349 			goto out;
1350 
1351 		if (attr->alt_pkey_index >=
1352 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
1353 			goto out;
1354 
1355 		if (mlx4_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
1356 				  attr->alt_port_num))
1357 			goto out;
1358 
1359 		context->alt_path.pkey_index = attr->alt_pkey_index;
1360 		context->alt_path.ackto = attr->alt_timeout << 3;
1361 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1362 	}
1363 
1364 	pd = get_pd(qp);
1365 	get_cqs(qp, &send_cq, &recv_cq);
1366 	context->pd       = cpu_to_be32(pd->pdn);
1367 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1368 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1369 	context->params1  = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1370 
1371 	/* Set "fast registration enabled" for all kernel QPs */
1372 	if (!qp->ibqp.uobject)
1373 		context->params1 |= cpu_to_be32(1 << 11);
1374 
1375 	if (attr_mask & IB_QP_RNR_RETRY) {
1376 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1377 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1378 	}
1379 
1380 	if (attr_mask & IB_QP_RETRY_CNT) {
1381 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1382 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1383 	}
1384 
1385 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1386 		if (attr->max_rd_atomic)
1387 			context->params1 |=
1388 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1389 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1390 	}
1391 
1392 	if (attr_mask & IB_QP_SQ_PSN)
1393 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1394 
1395 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1396 		if (attr->max_dest_rd_atomic)
1397 			context->params2 |=
1398 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1399 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1400 	}
1401 
1402 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1403 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1404 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1405 	}
1406 
1407 	if (ibqp->srq)
1408 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1409 
1410 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1411 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1412 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1413 	}
1414 	if (attr_mask & IB_QP_RQ_PSN)
1415 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1416 
1417 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1418 	if (attr_mask & IB_QP_QKEY) {
1419 		if (qp->mlx4_ib_qp_type &
1420 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1421 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1422 		else {
1423 			if (mlx4_is_mfunc(dev->dev) &&
1424 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1425 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1426 			    MLX4_RESERVED_QKEY_BASE) {
1427 				pr_err("Cannot use reserved QKEY"
1428 				       " 0x%x (range 0xffff0000..0xffffffff"
1429 				       " is reserved)\n", attr->qkey);
1430 				err = -EINVAL;
1431 				goto out;
1432 			}
1433 			context->qkey = cpu_to_be32(attr->qkey);
1434 		}
1435 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
1436 	}
1437 
1438 	if (ibqp->srq)
1439 		context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1440 
1441 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1442 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1443 
1444 	if (cur_state == IB_QPS_INIT &&
1445 	    new_state == IB_QPS_RTR  &&
1446 	    (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1447 	     ibqp->qp_type == IB_QPT_UD ||
1448 	     ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1449 		context->pri_path.sched_queue = (qp->port - 1) << 6;
1450 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1451 		    qp->mlx4_ib_qp_type &
1452 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1453 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1454 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1455 				context->pri_path.fl = 0x80;
1456 		} else {
1457 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1458 				context->pri_path.fl = 0x80;
1459 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1460 		}
1461 	}
1462 
1463 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET)
1464 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
1465 					MLX4_IB_LINK_TYPE_ETH;
1466 
1467 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1468 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1469 		sqd_event = 1;
1470 	else
1471 		sqd_event = 0;
1472 
1473 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1474 		context->rlkey |= (1 << 4);
1475 
1476 	/*
1477 	 * Before passing a kernel QP to the HW, make sure that the
1478 	 * ownership bits of the send queue are set and the SQ
1479 	 * headroom is stamped so that the hardware doesn't start
1480 	 * processing stale work requests.
1481 	 */
1482 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1483 		struct mlx4_wqe_ctrl_seg *ctrl;
1484 		int i;
1485 
1486 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1487 			ctrl = get_send_wqe(qp, i);
1488 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
1489 			if (qp->sq_max_wqes_per_wr == 1)
1490 				ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1491 
1492 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1493 		}
1494 	}
1495 
1496 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1497 			     to_mlx4_state(new_state), context, optpar,
1498 			     sqd_event, &qp->mqp);
1499 	if (err)
1500 		goto out;
1501 
1502 	qp->state = new_state;
1503 
1504 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1505 		qp->atomic_rd_en = attr->qp_access_flags;
1506 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1507 		qp->resp_depth = attr->max_dest_rd_atomic;
1508 	if (attr_mask & IB_QP_PORT) {
1509 		qp->port = attr->port_num;
1510 		update_mcg_macs(dev, qp);
1511 	}
1512 	if (attr_mask & IB_QP_ALT_PATH)
1513 		qp->alt_port = attr->alt_port_num;
1514 
1515 	if (is_sqp(dev, qp))
1516 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1517 
1518 	/*
1519 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
1520 	 * QP0 to RESET or ERROR, bring the link back down.
1521 	 */
1522 	if (is_qp0(dev, qp)) {
1523 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1524 			if (mlx4_INIT_PORT(dev->dev, qp->port))
1525 				pr_warn("INIT_PORT failed for port %d\n",
1526 				       qp->port);
1527 
1528 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1529 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1530 			mlx4_CLOSE_PORT(dev->dev, qp->port);
1531 	}
1532 
1533 	/*
1534 	 * If we moved a kernel QP to RESET, clean up all old CQ
1535 	 * entries and reinitialize the QP.
1536 	 */
1537 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1538 		mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1539 				 ibqp->srq ? to_msrq(ibqp->srq): NULL);
1540 		if (send_cq != recv_cq)
1541 			mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1542 
1543 		qp->rq.head = 0;
1544 		qp->rq.tail = 0;
1545 		qp->sq.head = 0;
1546 		qp->sq.tail = 0;
1547 		qp->sq_next_wqe = 0;
1548 		if (qp->rq.wqe_cnt)
1549 			*qp->db.db  = 0;
1550 	}
1551 
1552 out:
1553 	kfree(context);
1554 	return err;
1555 }
1556 
1557 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1558 		      int attr_mask, struct ib_udata *udata)
1559 {
1560 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1561 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1562 	enum ib_qp_state cur_state, new_state;
1563 	int err = -EINVAL;
1564 
1565 	mutex_lock(&qp->mutex);
1566 
1567 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1568 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1569 
1570 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
1571 		pr_debug("qpn 0x%x: invalid attribute mask specified "
1572 			 "for transition %d to %d. qp_type %d,"
1573 			 " attr_mask 0x%x\n",
1574 			 ibqp->qp_num, cur_state, new_state,
1575 			 ibqp->qp_type, attr_mask);
1576 		goto out;
1577 	}
1578 
1579 	if ((attr_mask & IB_QP_PORT) &&
1580 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
1581 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
1582 			 "for transition %d to %d. qp_type %d\n",
1583 			 ibqp->qp_num, attr->port_num, cur_state,
1584 			 new_state, ibqp->qp_type);
1585 		goto out;
1586 	}
1587 
1588 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
1589 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
1590 	     IB_LINK_LAYER_ETHERNET))
1591 		goto out;
1592 
1593 	if (attr_mask & IB_QP_PKEY_INDEX) {
1594 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1595 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
1596 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
1597 				 "for transition %d to %d. qp_type %d\n",
1598 				 ibqp->qp_num, attr->pkey_index, cur_state,
1599 				 new_state, ibqp->qp_type);
1600 			goto out;
1601 		}
1602 	}
1603 
1604 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1605 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
1606 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
1607 			 "Transition %d to %d. qp_type %d\n",
1608 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
1609 			 new_state, ibqp->qp_type);
1610 		goto out;
1611 	}
1612 
1613 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1614 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
1615 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
1616 			 "Transition %d to %d. qp_type %d\n",
1617 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
1618 			 new_state, ibqp->qp_type);
1619 		goto out;
1620 	}
1621 
1622 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1623 		err = 0;
1624 		goto out;
1625 	}
1626 
1627 	err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1628 
1629 out:
1630 	mutex_unlock(&qp->mutex);
1631 	return err;
1632 }
1633 
1634 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
1635 				  struct ib_send_wr *wr,
1636 				  void *wqe, unsigned *mlx_seg_len)
1637 {
1638 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
1639 	struct ib_device *ib_dev = &mdev->ib_dev;
1640 	struct mlx4_wqe_mlx_seg *mlx = wqe;
1641 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
1642 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
1643 	u16 pkey;
1644 	u32 qkey;
1645 	int send_size;
1646 	int header_size;
1647 	int spc;
1648 	int i;
1649 
1650 	if (wr->opcode != IB_WR_SEND)
1651 		return -EINVAL;
1652 
1653 	send_size = 0;
1654 
1655 	for (i = 0; i < wr->num_sge; ++i)
1656 		send_size += wr->sg_list[i].length;
1657 
1658 	/* for proxy-qp0 sends, need to add in size of tunnel header */
1659 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
1660 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
1661 		send_size += sizeof (struct mlx4_ib_tunnel_header);
1662 
1663 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, &sqp->ud_header);
1664 
1665 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
1666 		sqp->ud_header.lrh.service_level =
1667 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
1668 		sqp->ud_header.lrh.destination_lid =
1669 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
1670 		sqp->ud_header.lrh.source_lid =
1671 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
1672 	}
1673 
1674 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1675 
1676 	/* force loopback */
1677 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
1678 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
1679 
1680 	sqp->ud_header.lrh.virtual_lane    = 0;
1681 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1682 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
1683 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1684 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
1685 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1686 	else
1687 		sqp->ud_header.bth.destination_qpn =
1688 			cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
1689 
1690 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1691 	if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
1692 		return -EINVAL;
1693 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
1694 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
1695 
1696 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
1697 	sqp->ud_header.immediate_present = 0;
1698 
1699 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
1700 
1701 	/*
1702 	 * Inline data segments may not cross a 64 byte boundary.  If
1703 	 * our UD header is bigger than the space available up to the
1704 	 * next 64 byte boundary in the WQE, use two inline data
1705 	 * segments to hold the UD header.
1706 	 */
1707 	spc = MLX4_INLINE_ALIGN -
1708 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
1709 	if (header_size <= spc) {
1710 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
1711 		memcpy(inl + 1, sqp->header_buf, header_size);
1712 		i = 1;
1713 	} else {
1714 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
1715 		memcpy(inl + 1, sqp->header_buf, spc);
1716 
1717 		inl = (void *) (inl + 1) + spc;
1718 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
1719 		/*
1720 		 * Need a barrier here to make sure all the data is
1721 		 * visible before the byte_count field is set.
1722 		 * Otherwise the HCA prefetcher could grab the 64-byte
1723 		 * chunk with this inline segment and get a valid (!=
1724 		 * 0xffffffff) byte count but stale data, and end up
1725 		 * generating a packet with bad headers.
1726 		 *
1727 		 * The first inline segment's byte_count field doesn't
1728 		 * need a barrier, because it comes after a
1729 		 * control/MLX segment and therefore is at an offset
1730 		 * of 16 mod 64.
1731 		 */
1732 		wmb();
1733 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
1734 		i = 2;
1735 	}
1736 
1737 	*mlx_seg_len =
1738 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
1739 	return 0;
1740 }
1741 
1742 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
1743 			    void *wqe, unsigned *mlx_seg_len)
1744 {
1745 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
1746 	struct mlx4_wqe_mlx_seg *mlx = wqe;
1747 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
1748 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
1749 	struct net_device *ndev;
1750 	union ib_gid sgid;
1751 	u16 pkey;
1752 	int send_size;
1753 	int header_size;
1754 	int spc;
1755 	int i;
1756 	int err = 0;
1757 	u16 vlan = 0xffff;
1758 	bool is_eth;
1759 	bool is_vlan = false;
1760 	bool is_grh;
1761 
1762 	send_size = 0;
1763 	for (i = 0; i < wr->num_sge; ++i)
1764 		send_size += wr->sg_list[i].length;
1765 
1766 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
1767 	is_grh = mlx4_ib_ah_grh_present(ah);
1768 	if (is_eth) {
1769 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
1770 			/* When multi-function is enabled, the ib_core gid
1771 			 * indexes don't necessarily match the hw ones, so
1772 			 * we must use our own cache */
1773 			sgid.global.subnet_prefix =
1774 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
1775 				subnet_prefix;
1776 			sgid.global.interface_id =
1777 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
1778 				guid_cache[ah->av.ib.gid_index];
1779 		} else  {
1780 			err = ib_get_cached_gid(ib_dev,
1781 						be32_to_cpu(ah->av.ib.port_pd) >> 24,
1782 						ah->av.ib.gid_index, &sgid);
1783 			if (err)
1784 				return err;
1785 		}
1786 
1787 		vlan = rdma_get_vlan_id(&sgid);
1788 		is_vlan = vlan < 0x1000;
1789 	}
1790 	ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh, 0, &sqp->ud_header);
1791 
1792 	if (!is_eth) {
1793 		sqp->ud_header.lrh.service_level =
1794 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
1795 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
1796 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
1797 	}
1798 
1799 	if (is_grh) {
1800 		sqp->ud_header.grh.traffic_class =
1801 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
1802 		sqp->ud_header.grh.flow_label    =
1803 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
1804 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
1805 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
1806 			/* When multi-function is enabled, the ib_core gid
1807 			 * indexes don't necessarily match the hw ones, so
1808 			 * we must use our own cache */
1809 			sqp->ud_header.grh.source_gid.global.subnet_prefix =
1810 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
1811 						       subnet_prefix;
1812 			sqp->ud_header.grh.source_gid.global.interface_id =
1813 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
1814 					       guid_cache[ah->av.ib.gid_index];
1815 		} else
1816 			ib_get_cached_gid(ib_dev,
1817 					  be32_to_cpu(ah->av.ib.port_pd) >> 24,
1818 					  ah->av.ib.gid_index,
1819 					  &sqp->ud_header.grh.source_gid);
1820 		memcpy(sqp->ud_header.grh.destination_gid.raw,
1821 		       ah->av.ib.dgid, 16);
1822 	}
1823 
1824 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1825 
1826 	if (!is_eth) {
1827 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
1828 					  (sqp->ud_header.lrh.destination_lid ==
1829 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
1830 					  (sqp->ud_header.lrh.service_level << 8));
1831 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
1832 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
1833 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
1834 	}
1835 
1836 	switch (wr->opcode) {
1837 	case IB_WR_SEND:
1838 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
1839 		sqp->ud_header.immediate_present = 0;
1840 		break;
1841 	case IB_WR_SEND_WITH_IMM:
1842 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1843 		sqp->ud_header.immediate_present = 1;
1844 		sqp->ud_header.immediate_data    = wr->ex.imm_data;
1845 		break;
1846 	default:
1847 		return -EINVAL;
1848 	}
1849 
1850 	if (is_eth) {
1851 		u8 *smac;
1852 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
1853 
1854 		mlx->sched_prio = cpu_to_be16(pcp);
1855 
1856 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
1857 		/* FIXME: cache smac value? */
1858 		ndev = to_mdev(sqp->qp.ibqp.device)->iboe.netdevs[sqp->qp.port - 1];
1859 		if (!ndev)
1860 			return -ENODEV;
1861 		smac = ndev->dev_addr;
1862 		memcpy(sqp->ud_header.eth.smac_h, smac, 6);
1863 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
1864 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
1865 		if (!is_vlan) {
1866 			sqp->ud_header.eth.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
1867 		} else {
1868 			sqp->ud_header.vlan.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
1869 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
1870 		}
1871 	} else {
1872 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1873 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1874 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1875 	}
1876 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1877 	if (!sqp->qp.ibqp.qp_num)
1878 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
1879 	else
1880 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
1881 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1882 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1883 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1884 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1885 					       sqp->qkey : wr->wr.ud.remote_qkey);
1886 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1887 
1888 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
1889 
1890 	if (0) {
1891 		pr_err("built UD header of size %d:\n", header_size);
1892 		for (i = 0; i < header_size / 4; ++i) {
1893 			if (i % 8 == 0)
1894 				pr_err("  [%02x] ", i * 4);
1895 			pr_cont(" %08x",
1896 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
1897 			if ((i + 1) % 8 == 0)
1898 				pr_cont("\n");
1899 		}
1900 		pr_err("\n");
1901 	}
1902 
1903 	/*
1904 	 * Inline data segments may not cross a 64 byte boundary.  If
1905 	 * our UD header is bigger than the space available up to the
1906 	 * next 64 byte boundary in the WQE, use two inline data
1907 	 * segments to hold the UD header.
1908 	 */
1909 	spc = MLX4_INLINE_ALIGN -
1910 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
1911 	if (header_size <= spc) {
1912 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
1913 		memcpy(inl + 1, sqp->header_buf, header_size);
1914 		i = 1;
1915 	} else {
1916 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
1917 		memcpy(inl + 1, sqp->header_buf, spc);
1918 
1919 		inl = (void *) (inl + 1) + spc;
1920 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
1921 		/*
1922 		 * Need a barrier here to make sure all the data is
1923 		 * visible before the byte_count field is set.
1924 		 * Otherwise the HCA prefetcher could grab the 64-byte
1925 		 * chunk with this inline segment and get a valid (!=
1926 		 * 0xffffffff) byte count but stale data, and end up
1927 		 * generating a packet with bad headers.
1928 		 *
1929 		 * The first inline segment's byte_count field doesn't
1930 		 * need a barrier, because it comes after a
1931 		 * control/MLX segment and therefore is at an offset
1932 		 * of 16 mod 64.
1933 		 */
1934 		wmb();
1935 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
1936 		i = 2;
1937 	}
1938 
1939 	*mlx_seg_len =
1940 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
1941 	return 0;
1942 }
1943 
1944 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1945 {
1946 	unsigned cur;
1947 	struct mlx4_ib_cq *cq;
1948 
1949 	cur = wq->head - wq->tail;
1950 	if (likely(cur + nreq < wq->max_post))
1951 		return 0;
1952 
1953 	cq = to_mcq(ib_cq);
1954 	spin_lock(&cq->lock);
1955 	cur = wq->head - wq->tail;
1956 	spin_unlock(&cq->lock);
1957 
1958 	return cur + nreq >= wq->max_post;
1959 }
1960 
1961 static __be32 convert_access(int acc)
1962 {
1963 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
1964 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
1965 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
1966 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
1967 	       (acc & IB_ACCESS_REMOTE_READ   ?
1968 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
1969 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
1970 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
1971 }
1972 
1973 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr)
1974 {
1975 	struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
1976 	int i;
1977 
1978 	for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i)
1979 		mfrpl->mapped_page_list[i] =
1980 			cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] |
1981 				    MLX4_MTT_FLAG_PRESENT);
1982 
1983 	fseg->flags		= convert_access(wr->wr.fast_reg.access_flags);
1984 	fseg->mem_key		= cpu_to_be32(wr->wr.fast_reg.rkey);
1985 	fseg->buf_list		= cpu_to_be64(mfrpl->map);
1986 	fseg->start_addr	= cpu_to_be64(wr->wr.fast_reg.iova_start);
1987 	fseg->reg_len		= cpu_to_be64(wr->wr.fast_reg.length);
1988 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
1989 	fseg->page_size		= cpu_to_be32(wr->wr.fast_reg.page_shift);
1990 	fseg->reserved[0]	= 0;
1991 	fseg->reserved[1]	= 0;
1992 }
1993 
1994 static void set_bind_seg(struct mlx4_wqe_bind_seg *bseg, struct ib_send_wr *wr)
1995 {
1996 	bseg->flags1 =
1997 		convert_access(wr->wr.bind_mw.bind_info.mw_access_flags) &
1998 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ  |
1999 			    MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE |
2000 			    MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC);
2001 	bseg->flags2 = 0;
2002 	if (wr->wr.bind_mw.mw->type == IB_MW_TYPE_2)
2003 		bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_TYPE_2);
2004 	if (wr->wr.bind_mw.bind_info.mw_access_flags & IB_ZERO_BASED)
2005 		bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_ZERO_BASED);
2006 	bseg->new_rkey = cpu_to_be32(wr->wr.bind_mw.rkey);
2007 	bseg->lkey = cpu_to_be32(wr->wr.bind_mw.bind_info.mr->lkey);
2008 	bseg->addr = cpu_to_be64(wr->wr.bind_mw.bind_info.addr);
2009 	bseg->length = cpu_to_be64(wr->wr.bind_mw.bind_info.length);
2010 }
2011 
2012 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2013 {
2014 	memset(iseg, 0, sizeof(*iseg));
2015 	iseg->mem_key = cpu_to_be32(rkey);
2016 }
2017 
2018 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2019 					  u64 remote_addr, u32 rkey)
2020 {
2021 	rseg->raddr    = cpu_to_be64(remote_addr);
2022 	rseg->rkey     = cpu_to_be32(rkey);
2023 	rseg->reserved = 0;
2024 }
2025 
2026 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
2027 {
2028 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2029 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
2030 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
2031 	} else if (wr->opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2032 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2033 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add_mask);
2034 	} else {
2035 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2036 		aseg->compare  = 0;
2037 	}
2038 
2039 }
2040 
2041 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2042 				  struct ib_send_wr *wr)
2043 {
2044 	aseg->swap_add		= cpu_to_be64(wr->wr.atomic.swap);
2045 	aseg->swap_add_mask	= cpu_to_be64(wr->wr.atomic.swap_mask);
2046 	aseg->compare		= cpu_to_be64(wr->wr.atomic.compare_add);
2047 	aseg->compare_mask	= cpu_to_be64(wr->wr.atomic.compare_add_mask);
2048 }
2049 
2050 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2051 			     struct ib_send_wr *wr)
2052 {
2053 	memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
2054 	dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2055 	dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2056 	dseg->vlan = to_mah(wr->wr.ud.ah)->av.eth.vlan;
2057 	memcpy(dseg->mac, to_mah(wr->wr.ud.ah)->av.eth.mac, 6);
2058 }
2059 
2060 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2061 				    struct mlx4_wqe_datagram_seg *dseg,
2062 				    struct ib_send_wr *wr, enum ib_qp_type qpt)
2063 {
2064 	union mlx4_ext_av *av = &to_mah(wr->wr.ud.ah)->av;
2065 	struct mlx4_av sqp_av = {0};
2066 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
2067 
2068 	/* force loopback */
2069 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2070 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2071 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2072 			cpu_to_be32(0xf0000000);
2073 
2074 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2075 	/* This function used only for sending on QP1 proxies */
2076 	dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2077 	/* Use QKEY from the QP context, which is set by master */
2078 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2079 }
2080 
2081 static void build_tunnel_header(struct ib_send_wr *wr, void *wqe, unsigned *mlx_seg_len)
2082 {
2083 	struct mlx4_wqe_inline_seg *inl = wqe;
2084 	struct mlx4_ib_tunnel_header hdr;
2085 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2086 	int spc;
2087 	int i;
2088 
2089 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2090 	hdr.remote_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2091 	hdr.pkey_index = cpu_to_be16(wr->wr.ud.pkey_index);
2092 	hdr.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2093 
2094 	spc = MLX4_INLINE_ALIGN -
2095 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2096 	if (sizeof (hdr) <= spc) {
2097 		memcpy(inl + 1, &hdr, sizeof (hdr));
2098 		wmb();
2099 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
2100 		i = 1;
2101 	} else {
2102 		memcpy(inl + 1, &hdr, spc);
2103 		wmb();
2104 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2105 
2106 		inl = (void *) (inl + 1) + spc;
2107 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2108 		wmb();
2109 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
2110 		i = 2;
2111 	}
2112 
2113 	*mlx_seg_len =
2114 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2115 }
2116 
2117 static void set_mlx_icrc_seg(void *dseg)
2118 {
2119 	u32 *t = dseg;
2120 	struct mlx4_wqe_inline_seg *iseg = dseg;
2121 
2122 	t[1] = 0;
2123 
2124 	/*
2125 	 * Need a barrier here before writing the byte_count field to
2126 	 * make sure that all the data is visible before the
2127 	 * byte_count field is set.  Otherwise, if the segment begins
2128 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2129 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2130 	 * stale data, and end up sending the wrong data.
2131 	 */
2132 	wmb();
2133 
2134 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2135 }
2136 
2137 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2138 {
2139 	dseg->lkey       = cpu_to_be32(sg->lkey);
2140 	dseg->addr       = cpu_to_be64(sg->addr);
2141 
2142 	/*
2143 	 * Need a barrier here before writing the byte_count field to
2144 	 * make sure that all the data is visible before the
2145 	 * byte_count field is set.  Otherwise, if the segment begins
2146 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2147 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2148 	 * stale data, and end up sending the wrong data.
2149 	 */
2150 	wmb();
2151 
2152 	dseg->byte_count = cpu_to_be32(sg->length);
2153 }
2154 
2155 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2156 {
2157 	dseg->byte_count = cpu_to_be32(sg->length);
2158 	dseg->lkey       = cpu_to_be32(sg->lkey);
2159 	dseg->addr       = cpu_to_be64(sg->addr);
2160 }
2161 
2162 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr,
2163 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2164 			 __be32 *lso_hdr_sz, __be32 *blh)
2165 {
2166 	unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
2167 
2168 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2169 		*blh = cpu_to_be32(1 << 6);
2170 
2171 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2172 		     wr->num_sge > qp->sq.max_gs - (halign >> 4)))
2173 		return -EINVAL;
2174 
2175 	memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
2176 
2177 	*lso_hdr_sz  = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 |
2178 				   wr->wr.ud.hlen);
2179 	*lso_seg_len = halign;
2180 	return 0;
2181 }
2182 
2183 static __be32 send_ieth(struct ib_send_wr *wr)
2184 {
2185 	switch (wr->opcode) {
2186 	case IB_WR_SEND_WITH_IMM:
2187 	case IB_WR_RDMA_WRITE_WITH_IMM:
2188 		return wr->ex.imm_data;
2189 
2190 	case IB_WR_SEND_WITH_INV:
2191 		return cpu_to_be32(wr->ex.invalidate_rkey);
2192 
2193 	default:
2194 		return 0;
2195 	}
2196 }
2197 
2198 static void add_zero_len_inline(void *wqe)
2199 {
2200 	struct mlx4_wqe_inline_seg *inl = wqe;
2201 	memset(wqe, 0, 16);
2202 	inl->byte_count = cpu_to_be32(1 << 31);
2203 }
2204 
2205 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2206 		      struct ib_send_wr **bad_wr)
2207 {
2208 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2209 	void *wqe;
2210 	struct mlx4_wqe_ctrl_seg *ctrl;
2211 	struct mlx4_wqe_data_seg *dseg;
2212 	unsigned long flags;
2213 	int nreq;
2214 	int err = 0;
2215 	unsigned ind;
2216 	int uninitialized_var(stamp);
2217 	int uninitialized_var(size);
2218 	unsigned uninitialized_var(seglen);
2219 	__be32 dummy;
2220 	__be32 *lso_wqe;
2221 	__be32 uninitialized_var(lso_hdr_sz);
2222 	__be32 blh;
2223 	int i;
2224 
2225 	spin_lock_irqsave(&qp->sq.lock, flags);
2226 
2227 	ind = qp->sq_next_wqe;
2228 
2229 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2230 		lso_wqe = &dummy;
2231 		blh = 0;
2232 
2233 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2234 			err = -ENOMEM;
2235 			*bad_wr = wr;
2236 			goto out;
2237 		}
2238 
2239 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
2240 			err = -EINVAL;
2241 			*bad_wr = wr;
2242 			goto out;
2243 		}
2244 
2245 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
2246 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
2247 
2248 		ctrl->srcrb_flags =
2249 			(wr->send_flags & IB_SEND_SIGNALED ?
2250 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
2251 			(wr->send_flags & IB_SEND_SOLICITED ?
2252 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
2253 			((wr->send_flags & IB_SEND_IP_CSUM) ?
2254 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
2255 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
2256 			qp->sq_signal_bits;
2257 
2258 		ctrl->imm = send_ieth(wr);
2259 
2260 		wqe += sizeof *ctrl;
2261 		size = sizeof *ctrl / 16;
2262 
2263 		switch (qp->mlx4_ib_qp_type) {
2264 		case MLX4_IB_QPT_RC:
2265 		case MLX4_IB_QPT_UC:
2266 			switch (wr->opcode) {
2267 			case IB_WR_ATOMIC_CMP_AND_SWP:
2268 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2269 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
2270 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2271 					      wr->wr.atomic.rkey);
2272 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2273 
2274 				set_atomic_seg(wqe, wr);
2275 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
2276 
2277 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2278 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
2279 
2280 				break;
2281 
2282 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2283 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2284 					      wr->wr.atomic.rkey);
2285 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2286 
2287 				set_masked_atomic_seg(wqe, wr);
2288 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
2289 
2290 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2291 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
2292 
2293 				break;
2294 
2295 			case IB_WR_RDMA_READ:
2296 			case IB_WR_RDMA_WRITE:
2297 			case IB_WR_RDMA_WRITE_WITH_IMM:
2298 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2299 					      wr->wr.rdma.rkey);
2300 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2301 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
2302 				break;
2303 
2304 			case IB_WR_LOCAL_INV:
2305 				ctrl->srcrb_flags |=
2306 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2307 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
2308 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
2309 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
2310 				break;
2311 
2312 			case IB_WR_FAST_REG_MR:
2313 				ctrl->srcrb_flags |=
2314 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2315 				set_fmr_seg(wqe, wr);
2316 				wqe  += sizeof (struct mlx4_wqe_fmr_seg);
2317 				size += sizeof (struct mlx4_wqe_fmr_seg) / 16;
2318 				break;
2319 
2320 			case IB_WR_BIND_MW:
2321 				ctrl->srcrb_flags |=
2322 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2323 				set_bind_seg(wqe, wr);
2324 				wqe  += sizeof(struct mlx4_wqe_bind_seg);
2325 				size += sizeof(struct mlx4_wqe_bind_seg) / 16;
2326 				break;
2327 			default:
2328 				/* No extra segments required for sends */
2329 				break;
2330 			}
2331 			break;
2332 
2333 		case MLX4_IB_QPT_TUN_SMI_OWNER:
2334 			err =  build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2335 			if (unlikely(err)) {
2336 				*bad_wr = wr;
2337 				goto out;
2338 			}
2339 			wqe  += seglen;
2340 			size += seglen / 16;
2341 			break;
2342 		case MLX4_IB_QPT_TUN_SMI:
2343 		case MLX4_IB_QPT_TUN_GSI:
2344 			/* this is a UD qp used in MAD responses to slaves. */
2345 			set_datagram_seg(wqe, wr);
2346 			/* set the forced-loopback bit in the data seg av */
2347 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
2348 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2349 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2350 			break;
2351 		case MLX4_IB_QPT_UD:
2352 			set_datagram_seg(wqe, wr);
2353 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2354 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2355 
2356 			if (wr->opcode == IB_WR_LSO) {
2357 				err = build_lso_seg(wqe, wr, qp, &seglen, &lso_hdr_sz, &blh);
2358 				if (unlikely(err)) {
2359 					*bad_wr = wr;
2360 					goto out;
2361 				}
2362 				lso_wqe = (__be32 *) wqe;
2363 				wqe  += seglen;
2364 				size += seglen / 16;
2365 			}
2366 			break;
2367 
2368 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
2369 			if (unlikely(!mlx4_is_master(to_mdev(ibqp->device)->dev))) {
2370 				err = -ENOSYS;
2371 				*bad_wr = wr;
2372 				goto out;
2373 			}
2374 			err = build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2375 			if (unlikely(err)) {
2376 				*bad_wr = wr;
2377 				goto out;
2378 			}
2379 			wqe  += seglen;
2380 			size += seglen / 16;
2381 			/* to start tunnel header on a cache-line boundary */
2382 			add_zero_len_inline(wqe);
2383 			wqe += 16;
2384 			size++;
2385 			build_tunnel_header(wr, wqe, &seglen);
2386 			wqe  += seglen;
2387 			size += seglen / 16;
2388 			break;
2389 		case MLX4_IB_QPT_PROXY_SMI:
2390 			/* don't allow QP0 sends on guests */
2391 			err = -ENOSYS;
2392 			*bad_wr = wr;
2393 			goto out;
2394 		case MLX4_IB_QPT_PROXY_GSI:
2395 			/* If we are tunneling special qps, this is a UD qp.
2396 			 * In this case we first add a UD segment targeting
2397 			 * the tunnel qp, and then add a header with address
2398 			 * information */
2399 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe, wr, ibqp->qp_type);
2400 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2401 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2402 			build_tunnel_header(wr, wqe, &seglen);
2403 			wqe  += seglen;
2404 			size += seglen / 16;
2405 			break;
2406 
2407 		case MLX4_IB_QPT_SMI:
2408 		case MLX4_IB_QPT_GSI:
2409 			err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
2410 			if (unlikely(err)) {
2411 				*bad_wr = wr;
2412 				goto out;
2413 			}
2414 			wqe  += seglen;
2415 			size += seglen / 16;
2416 			break;
2417 
2418 		default:
2419 			break;
2420 		}
2421 
2422 		/*
2423 		 * Write data segments in reverse order, so as to
2424 		 * overwrite cacheline stamp last within each
2425 		 * cacheline.  This avoids issues with WQE
2426 		 * prefetching.
2427 		 */
2428 
2429 		dseg = wqe;
2430 		dseg += wr->num_sge - 1;
2431 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
2432 
2433 		/* Add one more inline data segment for ICRC for MLX sends */
2434 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2435 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
2436 			     qp->mlx4_ib_qp_type &
2437 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
2438 			set_mlx_icrc_seg(dseg + 1);
2439 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
2440 		}
2441 
2442 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
2443 			set_data_seg(dseg, wr->sg_list + i);
2444 
2445 		/*
2446 		 * Possibly overwrite stamping in cacheline with LSO
2447 		 * segment only after making sure all data segments
2448 		 * are written.
2449 		 */
2450 		wmb();
2451 		*lso_wqe = lso_hdr_sz;
2452 
2453 		ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
2454 				    MLX4_WQE_CTRL_FENCE : 0) | size;
2455 
2456 		/*
2457 		 * Make sure descriptor is fully written before
2458 		 * setting ownership bit (because HW can start
2459 		 * executing as soon as we do).
2460 		 */
2461 		wmb();
2462 
2463 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
2464 			*bad_wr = wr;
2465 			err = -EINVAL;
2466 			goto out;
2467 		}
2468 
2469 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
2470 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
2471 
2472 		stamp = ind + qp->sq_spare_wqes;
2473 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
2474 
2475 		/*
2476 		 * We can improve latency by not stamping the last
2477 		 * send queue WQE until after ringing the doorbell, so
2478 		 * only stamp here if there are still more WQEs to post.
2479 		 *
2480 		 * Same optimization applies to padding with NOP wqe
2481 		 * in case of WQE shrinking (used to prevent wrap-around
2482 		 * in the middle of WR).
2483 		 */
2484 		if (wr->next) {
2485 			stamp_send_wqe(qp, stamp, size * 16);
2486 			ind = pad_wraparound(qp, ind);
2487 		}
2488 	}
2489 
2490 out:
2491 	if (likely(nreq)) {
2492 		qp->sq.head += nreq;
2493 
2494 		/*
2495 		 * Make sure that descriptors are written before
2496 		 * doorbell record.
2497 		 */
2498 		wmb();
2499 
2500 		writel(qp->doorbell_qpn,
2501 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
2502 
2503 		/*
2504 		 * Make sure doorbells don't leak out of SQ spinlock
2505 		 * and reach the HCA out of order.
2506 		 */
2507 		mmiowb();
2508 
2509 		stamp_send_wqe(qp, stamp, size * 16);
2510 
2511 		ind = pad_wraparound(qp, ind);
2512 		qp->sq_next_wqe = ind;
2513 	}
2514 
2515 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2516 
2517 	return err;
2518 }
2519 
2520 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2521 		      struct ib_recv_wr **bad_wr)
2522 {
2523 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2524 	struct mlx4_wqe_data_seg *scat;
2525 	unsigned long flags;
2526 	int err = 0;
2527 	int nreq;
2528 	int ind;
2529 	int max_gs;
2530 	int i;
2531 
2532 	max_gs = qp->rq.max_gs;
2533 	spin_lock_irqsave(&qp->rq.lock, flags);
2534 
2535 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
2536 
2537 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2538 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2539 			err = -ENOMEM;
2540 			*bad_wr = wr;
2541 			goto out;
2542 		}
2543 
2544 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2545 			err = -EINVAL;
2546 			*bad_wr = wr;
2547 			goto out;
2548 		}
2549 
2550 		scat = get_recv_wqe(qp, ind);
2551 
2552 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
2553 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
2554 			ib_dma_sync_single_for_device(ibqp->device,
2555 						      qp->sqp_proxy_rcv[ind].map,
2556 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
2557 						      DMA_FROM_DEVICE);
2558 			scat->byte_count =
2559 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
2560 			/* use dma lkey from upper layer entry */
2561 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
2562 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
2563 			scat++;
2564 			max_gs--;
2565 		}
2566 
2567 		for (i = 0; i < wr->num_sge; ++i)
2568 			__set_data_seg(scat + i, wr->sg_list + i);
2569 
2570 		if (i < max_gs) {
2571 			scat[i].byte_count = 0;
2572 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
2573 			scat[i].addr       = 0;
2574 		}
2575 
2576 		qp->rq.wrid[ind] = wr->wr_id;
2577 
2578 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
2579 	}
2580 
2581 out:
2582 	if (likely(nreq)) {
2583 		qp->rq.head += nreq;
2584 
2585 		/*
2586 		 * Make sure that descriptors are written before
2587 		 * doorbell record.
2588 		 */
2589 		wmb();
2590 
2591 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
2592 	}
2593 
2594 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2595 
2596 	return err;
2597 }
2598 
2599 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
2600 {
2601 	switch (mlx4_state) {
2602 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
2603 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
2604 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
2605 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
2606 	case MLX4_QP_STATE_SQ_DRAINING:
2607 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
2608 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
2609 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
2610 	default:		     return -1;
2611 	}
2612 }
2613 
2614 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
2615 {
2616 	switch (mlx4_mig_state) {
2617 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
2618 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
2619 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
2620 	default: return -1;
2621 	}
2622 }
2623 
2624 static int to_ib_qp_access_flags(int mlx4_flags)
2625 {
2626 	int ib_flags = 0;
2627 
2628 	if (mlx4_flags & MLX4_QP_BIT_RRE)
2629 		ib_flags |= IB_ACCESS_REMOTE_READ;
2630 	if (mlx4_flags & MLX4_QP_BIT_RWE)
2631 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
2632 	if (mlx4_flags & MLX4_QP_BIT_RAE)
2633 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
2634 
2635 	return ib_flags;
2636 }
2637 
2638 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
2639 				struct mlx4_qp_path *path)
2640 {
2641 	struct mlx4_dev *dev = ibdev->dev;
2642 	int is_eth;
2643 
2644 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
2645 	ib_ah_attr->port_num	  = path->sched_queue & 0x40 ? 2 : 1;
2646 
2647 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
2648 		return;
2649 
2650 	is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
2651 		IB_LINK_LAYER_ETHERNET;
2652 	if (is_eth)
2653 		ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
2654 		((path->sched_queue & 4) << 1);
2655 	else
2656 		ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
2657 
2658 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
2659 	ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
2660 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
2661 	ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
2662 	if (ib_ah_attr->ah_flags) {
2663 		ib_ah_attr->grh.sgid_index = path->mgid_index;
2664 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
2665 		ib_ah_attr->grh.traffic_class =
2666 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
2667 		ib_ah_attr->grh.flow_label =
2668 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
2669 		memcpy(ib_ah_attr->grh.dgid.raw,
2670 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
2671 	}
2672 }
2673 
2674 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
2675 		     struct ib_qp_init_attr *qp_init_attr)
2676 {
2677 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2678 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2679 	struct mlx4_qp_context context;
2680 	int mlx4_state;
2681 	int err = 0;
2682 
2683 	mutex_lock(&qp->mutex);
2684 
2685 	if (qp->state == IB_QPS_RESET) {
2686 		qp_attr->qp_state = IB_QPS_RESET;
2687 		goto done;
2688 	}
2689 
2690 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
2691 	if (err) {
2692 		err = -EINVAL;
2693 		goto out;
2694 	}
2695 
2696 	mlx4_state = be32_to_cpu(context.flags) >> 28;
2697 
2698 	qp->state		     = to_ib_qp_state(mlx4_state);
2699 	qp_attr->qp_state	     = qp->state;
2700 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
2701 	qp_attr->path_mig_state	     =
2702 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
2703 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
2704 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
2705 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
2706 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
2707 	qp_attr->qp_access_flags     =
2708 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
2709 
2710 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
2711 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
2712 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
2713 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
2714 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
2715 	}
2716 
2717 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
2718 	if (qp_attr->qp_state == IB_QPS_INIT)
2719 		qp_attr->port_num = qp->port;
2720 	else
2721 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
2722 
2723 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
2724 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
2725 
2726 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
2727 
2728 	qp_attr->max_dest_rd_atomic =
2729 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
2730 	qp_attr->min_rnr_timer	    =
2731 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
2732 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
2733 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
2734 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
2735 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
2736 
2737 done:
2738 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
2739 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
2740 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
2741 
2742 	if (!ibqp->uobject) {
2743 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
2744 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
2745 	} else {
2746 		qp_attr->cap.max_send_wr  = 0;
2747 		qp_attr->cap.max_send_sge = 0;
2748 	}
2749 
2750 	/*
2751 	 * We don't support inline sends for kernel QPs (yet), and we
2752 	 * don't know what userspace's value should be.
2753 	 */
2754 	qp_attr->cap.max_inline_data = 0;
2755 
2756 	qp_init_attr->cap	     = qp_attr->cap;
2757 
2758 	qp_init_attr->create_flags = 0;
2759 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
2760 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
2761 
2762 	if (qp->flags & MLX4_IB_QP_LSO)
2763 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
2764 
2765 	qp_init_attr->sq_sig_type =
2766 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
2767 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
2768 
2769 out:
2770 	mutex_unlock(&qp->mutex);
2771 	return err;
2772 }
2773 
2774