xref: /linux/drivers/infiniband/hw/mlx4/qp.c (revision 0d456bad36d42d16022be045c8a53ddbb59ee478)
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 };
108 
109 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
110 {
111 	return container_of(mqp, struct mlx4_ib_sqp, qp);
112 }
113 
114 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
115 {
116 	if (!mlx4_is_master(dev->dev))
117 		return 0;
118 
119 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
120 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
121 		8 * MLX4_MFUNC_MAX;
122 }
123 
124 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
125 {
126 	int proxy_sqp = 0;
127 	int real_sqp = 0;
128 	int i;
129 	/* PPF or Native -- real SQP */
130 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
131 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
132 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
133 	if (real_sqp)
134 		return 1;
135 	/* VF or PF -- proxy SQP */
136 	if (mlx4_is_mfunc(dev->dev)) {
137 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
138 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] ||
139 			    qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) {
140 				proxy_sqp = 1;
141 				break;
142 			}
143 		}
144 	}
145 	return proxy_sqp;
146 }
147 
148 /* used for INIT/CLOSE port logic */
149 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
150 {
151 	int proxy_qp0 = 0;
152 	int real_qp0 = 0;
153 	int i;
154 	/* PPF or Native -- real QP0 */
155 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
156 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
157 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
158 	if (real_qp0)
159 		return 1;
160 	/* VF or PF -- proxy QP0 */
161 	if (mlx4_is_mfunc(dev->dev)) {
162 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
163 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) {
164 				proxy_qp0 = 1;
165 				break;
166 			}
167 		}
168 	}
169 	return proxy_qp0;
170 }
171 
172 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
173 {
174 	return mlx4_buf_offset(&qp->buf, offset);
175 }
176 
177 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
178 {
179 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
180 }
181 
182 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
183 {
184 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
185 }
186 
187 /*
188  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
189  * first four bytes of every 64 byte chunk with
190  *     0x7FFFFFF | (invalid_ownership_value << 31).
191  *
192  * When the max work request size is less than or equal to the WQE
193  * basic block size, as an optimization, we can stamp all WQEs with
194  * 0xffffffff, and skip the very first chunk of each WQE.
195  */
196 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
197 {
198 	__be32 *wqe;
199 	int i;
200 	int s;
201 	int ind;
202 	void *buf;
203 	__be32 stamp;
204 	struct mlx4_wqe_ctrl_seg *ctrl;
205 
206 	if (qp->sq_max_wqes_per_wr > 1) {
207 		s = roundup(size, 1U << qp->sq.wqe_shift);
208 		for (i = 0; i < s; i += 64) {
209 			ind = (i >> qp->sq.wqe_shift) + n;
210 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
211 						       cpu_to_be32(0xffffffff);
212 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
213 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
214 			*wqe = stamp;
215 		}
216 	} else {
217 		ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
218 		s = (ctrl->fence_size & 0x3f) << 4;
219 		for (i = 64; i < s; i += 64) {
220 			wqe = buf + i;
221 			*wqe = cpu_to_be32(0xffffffff);
222 		}
223 	}
224 }
225 
226 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
227 {
228 	struct mlx4_wqe_ctrl_seg *ctrl;
229 	struct mlx4_wqe_inline_seg *inl;
230 	void *wqe;
231 	int s;
232 
233 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
234 	s = sizeof(struct mlx4_wqe_ctrl_seg);
235 
236 	if (qp->ibqp.qp_type == IB_QPT_UD) {
237 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
238 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
239 		memset(dgram, 0, sizeof *dgram);
240 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
241 		s += sizeof(struct mlx4_wqe_datagram_seg);
242 	}
243 
244 	/* Pad the remainder of the WQE with an inline data segment. */
245 	if (size > s) {
246 		inl = wqe + s;
247 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
248 	}
249 	ctrl->srcrb_flags = 0;
250 	ctrl->fence_size = size / 16;
251 	/*
252 	 * Make sure descriptor is fully written before setting ownership bit
253 	 * (because HW can start executing as soon as we do).
254 	 */
255 	wmb();
256 
257 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
258 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
259 
260 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
261 }
262 
263 /* Post NOP WQE to prevent wrap-around in the middle of WR */
264 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
265 {
266 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
267 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
268 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
269 		ind += s;
270 	}
271 	return ind;
272 }
273 
274 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
275 {
276 	struct ib_event event;
277 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
278 
279 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
280 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
281 
282 	if (ibqp->event_handler) {
283 		event.device     = ibqp->device;
284 		event.element.qp = ibqp;
285 		switch (type) {
286 		case MLX4_EVENT_TYPE_PATH_MIG:
287 			event.event = IB_EVENT_PATH_MIG;
288 			break;
289 		case MLX4_EVENT_TYPE_COMM_EST:
290 			event.event = IB_EVENT_COMM_EST;
291 			break;
292 		case MLX4_EVENT_TYPE_SQ_DRAINED:
293 			event.event = IB_EVENT_SQ_DRAINED;
294 			break;
295 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
296 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
297 			break;
298 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
299 			event.event = IB_EVENT_QP_FATAL;
300 			break;
301 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
302 			event.event = IB_EVENT_PATH_MIG_ERR;
303 			break;
304 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
305 			event.event = IB_EVENT_QP_REQ_ERR;
306 			break;
307 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
308 			event.event = IB_EVENT_QP_ACCESS_ERR;
309 			break;
310 		default:
311 			pr_warn("Unexpected event type %d "
312 			       "on QP %06x\n", type, qp->qpn);
313 			return;
314 		}
315 
316 		ibqp->event_handler(&event, ibqp->qp_context);
317 	}
318 }
319 
320 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
321 {
322 	/*
323 	 * UD WQEs must have a datagram segment.
324 	 * RC and UC WQEs might have a remote address segment.
325 	 * MLX WQEs need two extra inline data segments (for the UD
326 	 * header and space for the ICRC).
327 	 */
328 	switch (type) {
329 	case MLX4_IB_QPT_UD:
330 		return sizeof (struct mlx4_wqe_ctrl_seg) +
331 			sizeof (struct mlx4_wqe_datagram_seg) +
332 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
333 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
334 	case MLX4_IB_QPT_PROXY_SMI:
335 	case MLX4_IB_QPT_PROXY_GSI:
336 		return sizeof (struct mlx4_wqe_ctrl_seg) +
337 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
338 	case MLX4_IB_QPT_TUN_SMI_OWNER:
339 	case MLX4_IB_QPT_TUN_GSI:
340 		return sizeof (struct mlx4_wqe_ctrl_seg) +
341 			sizeof (struct mlx4_wqe_datagram_seg);
342 
343 	case MLX4_IB_QPT_UC:
344 		return sizeof (struct mlx4_wqe_ctrl_seg) +
345 			sizeof (struct mlx4_wqe_raddr_seg);
346 	case MLX4_IB_QPT_RC:
347 		return sizeof (struct mlx4_wqe_ctrl_seg) +
348 			sizeof (struct mlx4_wqe_atomic_seg) +
349 			sizeof (struct mlx4_wqe_raddr_seg);
350 	case MLX4_IB_QPT_SMI:
351 	case MLX4_IB_QPT_GSI:
352 		return sizeof (struct mlx4_wqe_ctrl_seg) +
353 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
354 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
355 					   MLX4_INLINE_ALIGN) *
356 			      sizeof (struct mlx4_wqe_inline_seg),
357 			      sizeof (struct mlx4_wqe_data_seg)) +
358 			ALIGN(4 +
359 			      sizeof (struct mlx4_wqe_inline_seg),
360 			      sizeof (struct mlx4_wqe_data_seg));
361 	default:
362 		return sizeof (struct mlx4_wqe_ctrl_seg);
363 	}
364 }
365 
366 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
367 		       int is_user, int has_rq, struct mlx4_ib_qp *qp)
368 {
369 	/* Sanity check RQ size before proceeding */
370 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
371 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
372 		return -EINVAL;
373 
374 	if (!has_rq) {
375 		if (cap->max_recv_wr)
376 			return -EINVAL;
377 
378 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
379 	} else {
380 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
381 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
382 			return -EINVAL;
383 
384 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
385 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
386 		qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
387 	}
388 
389 	/* leave userspace return values as they were, so as not to break ABI */
390 	if (is_user) {
391 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
392 		cap->max_recv_sge = qp->rq.max_gs;
393 	} else {
394 		cap->max_recv_wr  = qp->rq.max_post =
395 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
396 		cap->max_recv_sge = min(qp->rq.max_gs,
397 					min(dev->dev->caps.max_sq_sg,
398 					    dev->dev->caps.max_rq_sg));
399 	}
400 
401 	return 0;
402 }
403 
404 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
405 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
406 {
407 	int s;
408 
409 	/* Sanity check SQ size before proceeding */
410 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
411 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
412 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
413 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
414 		return -EINVAL;
415 
416 	/*
417 	 * For MLX transport we need 2 extra S/G entries:
418 	 * one for the header and one for the checksum at the end
419 	 */
420 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
421 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
422 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
423 		return -EINVAL;
424 
425 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
426 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
427 		send_wqe_overhead(type, qp->flags);
428 
429 	if (s > dev->dev->caps.max_sq_desc_sz)
430 		return -EINVAL;
431 
432 	/*
433 	 * Hermon supports shrinking WQEs, such that a single work
434 	 * request can include multiple units of 1 << wqe_shift.  This
435 	 * way, work requests can differ in size, and do not have to
436 	 * be a power of 2 in size, saving memory and speeding up send
437 	 * WR posting.  Unfortunately, if we do this then the
438 	 * wqe_index field in CQEs can't be used to look up the WR ID
439 	 * anymore, so we do this only if selective signaling is off.
440 	 *
441 	 * Further, on 32-bit platforms, we can't use vmap() to make
442 	 * the QP buffer virtually contiguous.  Thus we have to use
443 	 * constant-sized WRs to make sure a WR is always fully within
444 	 * a single page-sized chunk.
445 	 *
446 	 * Finally, we use NOP work requests to pad the end of the
447 	 * work queue, to avoid wrap-around in the middle of WR.  We
448 	 * set NEC bit to avoid getting completions with error for
449 	 * these NOP WRs, but since NEC is only supported starting
450 	 * with firmware 2.2.232, we use constant-sized WRs for older
451 	 * firmware.
452 	 *
453 	 * And, since MLX QPs only support SEND, we use constant-sized
454 	 * WRs in this case.
455 	 *
456 	 * We look for the smallest value of wqe_shift such that the
457 	 * resulting number of wqes does not exceed device
458 	 * capabilities.
459 	 *
460 	 * We set WQE size to at least 64 bytes, this way stamping
461 	 * invalidates each WQE.
462 	 */
463 	if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
464 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
465 	    type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
466 	    !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
467 		      MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
468 		qp->sq.wqe_shift = ilog2(64);
469 	else
470 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
471 
472 	for (;;) {
473 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
474 
475 		/*
476 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
477 		 * allow HW to prefetch.
478 		 */
479 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
480 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
481 						    qp->sq_max_wqes_per_wr +
482 						    qp->sq_spare_wqes);
483 
484 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
485 			break;
486 
487 		if (qp->sq_max_wqes_per_wr <= 1)
488 			return -EINVAL;
489 
490 		++qp->sq.wqe_shift;
491 	}
492 
493 	qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
494 			     (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
495 			 send_wqe_overhead(type, qp->flags)) /
496 		sizeof (struct mlx4_wqe_data_seg);
497 
498 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
499 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
500 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
501 		qp->rq.offset = 0;
502 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
503 	} else {
504 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
505 		qp->sq.offset = 0;
506 	}
507 
508 	cap->max_send_wr  = qp->sq.max_post =
509 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
510 	cap->max_send_sge = min(qp->sq.max_gs,
511 				min(dev->dev->caps.max_sq_sg,
512 				    dev->dev->caps.max_rq_sg));
513 	/* We don't support inline sends for kernel QPs (yet) */
514 	cap->max_inline_data = 0;
515 
516 	return 0;
517 }
518 
519 static int set_user_sq_size(struct mlx4_ib_dev *dev,
520 			    struct mlx4_ib_qp *qp,
521 			    struct mlx4_ib_create_qp *ucmd)
522 {
523 	/* Sanity check SQ size before proceeding */
524 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
525 	    ucmd->log_sq_stride >
526 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
527 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
528 		return -EINVAL;
529 
530 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
531 	qp->sq.wqe_shift = ucmd->log_sq_stride;
532 
533 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
534 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
535 
536 	return 0;
537 }
538 
539 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
540 {
541 	int i;
542 
543 	qp->sqp_proxy_rcv =
544 		kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
545 			GFP_KERNEL);
546 	if (!qp->sqp_proxy_rcv)
547 		return -ENOMEM;
548 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
549 		qp->sqp_proxy_rcv[i].addr =
550 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
551 				GFP_KERNEL);
552 		if (!qp->sqp_proxy_rcv[i].addr)
553 			goto err;
554 		qp->sqp_proxy_rcv[i].map =
555 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
556 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
557 					  DMA_FROM_DEVICE);
558 	}
559 	return 0;
560 
561 err:
562 	while (i > 0) {
563 		--i;
564 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
565 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
566 				    DMA_FROM_DEVICE);
567 		kfree(qp->sqp_proxy_rcv[i].addr);
568 	}
569 	kfree(qp->sqp_proxy_rcv);
570 	qp->sqp_proxy_rcv = NULL;
571 	return -ENOMEM;
572 }
573 
574 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
575 {
576 	int i;
577 
578 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
579 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
580 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
581 				    DMA_FROM_DEVICE);
582 		kfree(qp->sqp_proxy_rcv[i].addr);
583 	}
584 	kfree(qp->sqp_proxy_rcv);
585 }
586 
587 static int qp_has_rq(struct ib_qp_init_attr *attr)
588 {
589 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
590 		return 0;
591 
592 	return !attr->srq;
593 }
594 
595 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
596 			    struct ib_qp_init_attr *init_attr,
597 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp)
598 {
599 	int qpn;
600 	int err;
601 	struct mlx4_ib_sqp *sqp;
602 	struct mlx4_ib_qp *qp;
603 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
604 
605 	/* When tunneling special qps, we use a plain UD qp */
606 	if (sqpn) {
607 		if (mlx4_is_mfunc(dev->dev) &&
608 		    (!mlx4_is_master(dev->dev) ||
609 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
610 			if (init_attr->qp_type == IB_QPT_GSI)
611 				qp_type = MLX4_IB_QPT_PROXY_GSI;
612 			else if (mlx4_is_master(dev->dev))
613 				qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
614 			else
615 				qp_type = MLX4_IB_QPT_PROXY_SMI;
616 		}
617 		qpn = sqpn;
618 		/* add extra sg entry for tunneling */
619 		init_attr->cap.max_recv_sge++;
620 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
621 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
622 			container_of(init_attr,
623 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
624 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
625 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
626 		    !mlx4_is_master(dev->dev))
627 			return -EINVAL;
628 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
629 			qp_type = MLX4_IB_QPT_TUN_GSI;
630 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev))
631 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
632 		else
633 			qp_type = MLX4_IB_QPT_TUN_SMI;
634 		/* we are definitely in the PPF here, since we are creating
635 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
636 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
637 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
638 		sqpn = qpn;
639 	}
640 
641 	if (!*caller_qp) {
642 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
643 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
644 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
645 			sqp = kzalloc(sizeof (struct mlx4_ib_sqp), GFP_KERNEL);
646 			if (!sqp)
647 				return -ENOMEM;
648 			qp = &sqp->qp;
649 		} else {
650 			qp = kzalloc(sizeof (struct mlx4_ib_qp), GFP_KERNEL);
651 			if (!qp)
652 				return -ENOMEM;
653 		}
654 	} else
655 		qp = *caller_qp;
656 
657 	qp->mlx4_ib_qp_type = qp_type;
658 
659 	mutex_init(&qp->mutex);
660 	spin_lock_init(&qp->sq.lock);
661 	spin_lock_init(&qp->rq.lock);
662 	INIT_LIST_HEAD(&qp->gid_list);
663 	INIT_LIST_HEAD(&qp->steering_rules);
664 
665 	qp->state	 = IB_QPS_RESET;
666 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
667 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
668 
669 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
670 	if (err)
671 		goto err;
672 
673 	if (pd->uobject) {
674 		struct mlx4_ib_create_qp ucmd;
675 
676 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
677 			err = -EFAULT;
678 			goto err;
679 		}
680 
681 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
682 
683 		err = set_user_sq_size(dev, qp, &ucmd);
684 		if (err)
685 			goto err;
686 
687 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
688 				       qp->buf_size, 0, 0);
689 		if (IS_ERR(qp->umem)) {
690 			err = PTR_ERR(qp->umem);
691 			goto err;
692 		}
693 
694 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
695 				    ilog2(qp->umem->page_size), &qp->mtt);
696 		if (err)
697 			goto err_buf;
698 
699 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
700 		if (err)
701 			goto err_mtt;
702 
703 		if (qp_has_rq(init_attr)) {
704 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
705 						  ucmd.db_addr, &qp->db);
706 			if (err)
707 				goto err_mtt;
708 		}
709 	} else {
710 		qp->sq_no_prefetch = 0;
711 
712 		if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
713 			qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
714 
715 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
716 			qp->flags |= MLX4_IB_QP_LSO;
717 
718 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
719 		if (err)
720 			goto err;
721 
722 		if (qp_has_rq(init_attr)) {
723 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
724 			if (err)
725 				goto err;
726 
727 			*qp->db.db = 0;
728 		}
729 
730 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf)) {
731 			err = -ENOMEM;
732 			goto err_db;
733 		}
734 
735 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
736 				    &qp->mtt);
737 		if (err)
738 			goto err_buf;
739 
740 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
741 		if (err)
742 			goto err_mtt;
743 
744 		qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), GFP_KERNEL);
745 		qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), GFP_KERNEL);
746 
747 		if (!qp->sq.wrid || !qp->rq.wrid) {
748 			err = -ENOMEM;
749 			goto err_wrid;
750 		}
751 	}
752 
753 	if (sqpn) {
754 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
755 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
756 			if (alloc_proxy_bufs(pd->device, qp)) {
757 				err = -ENOMEM;
758 				goto err_wrid;
759 			}
760 		}
761 	} else {
762 		/* Raw packet QPNs must be aligned to 8 bits. If not, the WQE
763 		 * BlueFlame setup flow wrongly causes VLAN insertion. */
764 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
765 			err = mlx4_qp_reserve_range(dev->dev, 1, 1 << 8, &qpn);
766 		else
767 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn);
768 		if (err)
769 			goto err_proxy;
770 	}
771 
772 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
773 	if (err)
774 		goto err_qpn;
775 
776 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
777 		qp->mqp.qpn |= (1 << 23);
778 
779 	/*
780 	 * Hardware wants QPN written in big-endian order (after
781 	 * shifting) for send doorbell.  Precompute this value to save
782 	 * a little bit when posting sends.
783 	 */
784 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
785 
786 	qp->mqp.event = mlx4_ib_qp_event;
787 	if (!*caller_qp)
788 		*caller_qp = qp;
789 	return 0;
790 
791 err_qpn:
792 	if (!sqpn)
793 		mlx4_qp_release_range(dev->dev, qpn, 1);
794 err_proxy:
795 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
796 		free_proxy_bufs(pd->device, qp);
797 err_wrid:
798 	if (pd->uobject) {
799 		if (qp_has_rq(init_attr))
800 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
801 	} else {
802 		kfree(qp->sq.wrid);
803 		kfree(qp->rq.wrid);
804 	}
805 
806 err_mtt:
807 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
808 
809 err_buf:
810 	if (pd->uobject)
811 		ib_umem_release(qp->umem);
812 	else
813 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
814 
815 err_db:
816 	if (!pd->uobject && qp_has_rq(init_attr))
817 		mlx4_db_free(dev->dev, &qp->db);
818 
819 err:
820 	if (!*caller_qp)
821 		kfree(qp);
822 	return err;
823 }
824 
825 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
826 {
827 	switch (state) {
828 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
829 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
830 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
831 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
832 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
833 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
834 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
835 	default:		return -1;
836 	}
837 }
838 
839 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
840 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
841 {
842 	if (send_cq == recv_cq) {
843 		spin_lock_irq(&send_cq->lock);
844 		__acquire(&recv_cq->lock);
845 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
846 		spin_lock_irq(&send_cq->lock);
847 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
848 	} else {
849 		spin_lock_irq(&recv_cq->lock);
850 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
851 	}
852 }
853 
854 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
855 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
856 {
857 	if (send_cq == recv_cq) {
858 		__release(&recv_cq->lock);
859 		spin_unlock_irq(&send_cq->lock);
860 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
861 		spin_unlock(&recv_cq->lock);
862 		spin_unlock_irq(&send_cq->lock);
863 	} else {
864 		spin_unlock(&send_cq->lock);
865 		spin_unlock_irq(&recv_cq->lock);
866 	}
867 }
868 
869 static void del_gid_entries(struct mlx4_ib_qp *qp)
870 {
871 	struct mlx4_ib_gid_entry *ge, *tmp;
872 
873 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
874 		list_del(&ge->list);
875 		kfree(ge);
876 	}
877 }
878 
879 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
880 {
881 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
882 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
883 	else
884 		return to_mpd(qp->ibqp.pd);
885 }
886 
887 static void get_cqs(struct mlx4_ib_qp *qp,
888 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
889 {
890 	switch (qp->ibqp.qp_type) {
891 	case IB_QPT_XRC_TGT:
892 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
893 		*recv_cq = *send_cq;
894 		break;
895 	case IB_QPT_XRC_INI:
896 		*send_cq = to_mcq(qp->ibqp.send_cq);
897 		*recv_cq = *send_cq;
898 		break;
899 	default:
900 		*send_cq = to_mcq(qp->ibqp.send_cq);
901 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
902 		break;
903 	}
904 }
905 
906 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
907 			      int is_user)
908 {
909 	struct mlx4_ib_cq *send_cq, *recv_cq;
910 
911 	if (qp->state != IB_QPS_RESET)
912 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
913 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
914 			pr_warn("modify QP %06x to RESET failed.\n",
915 			       qp->mqp.qpn);
916 
917 	get_cqs(qp, &send_cq, &recv_cq);
918 
919 	mlx4_ib_lock_cqs(send_cq, recv_cq);
920 
921 	if (!is_user) {
922 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
923 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
924 		if (send_cq != recv_cq)
925 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
926 	}
927 
928 	mlx4_qp_remove(dev->dev, &qp->mqp);
929 
930 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
931 
932 	mlx4_qp_free(dev->dev, &qp->mqp);
933 
934 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp))
935 		mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
936 
937 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
938 
939 	if (is_user) {
940 		if (qp->rq.wqe_cnt)
941 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
942 					      &qp->db);
943 		ib_umem_release(qp->umem);
944 	} else {
945 		kfree(qp->sq.wrid);
946 		kfree(qp->rq.wrid);
947 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
948 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
949 			free_proxy_bufs(&dev->ib_dev, qp);
950 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
951 		if (qp->rq.wqe_cnt)
952 			mlx4_db_free(dev->dev, &qp->db);
953 	}
954 
955 	del_gid_entries(qp);
956 }
957 
958 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
959 {
960 	/* Native or PPF */
961 	if (!mlx4_is_mfunc(dev->dev) ||
962 	    (mlx4_is_master(dev->dev) &&
963 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
964 		return  dev->dev->phys_caps.base_sqpn +
965 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
966 			attr->port_num - 1;
967 	}
968 	/* PF or VF -- creating proxies */
969 	if (attr->qp_type == IB_QPT_SMI)
970 		return dev->dev->caps.qp0_proxy[attr->port_num - 1];
971 	else
972 		return dev->dev->caps.qp1_proxy[attr->port_num - 1];
973 }
974 
975 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
976 				struct ib_qp_init_attr *init_attr,
977 				struct ib_udata *udata)
978 {
979 	struct mlx4_ib_qp *qp = NULL;
980 	int err;
981 	u16 xrcdn = 0;
982 
983 	/*
984 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
985 	 * and only for kernel UD QPs.
986 	 */
987 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
988 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
989 					MLX4_IB_SRIOV_TUNNEL_QP | MLX4_IB_SRIOV_SQP))
990 		return ERR_PTR(-EINVAL);
991 
992 	if (init_attr->create_flags &&
993 	    (udata ||
994 	     ((init_attr->create_flags & ~MLX4_IB_SRIOV_SQP) &&
995 	      init_attr->qp_type != IB_QPT_UD) ||
996 	     ((init_attr->create_flags & MLX4_IB_SRIOV_SQP) &&
997 	      init_attr->qp_type > IB_QPT_GSI)))
998 		return ERR_PTR(-EINVAL);
999 
1000 	switch (init_attr->qp_type) {
1001 	case IB_QPT_XRC_TGT:
1002 		pd = to_mxrcd(init_attr->xrcd)->pd;
1003 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1004 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1005 		/* fall through */
1006 	case IB_QPT_XRC_INI:
1007 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1008 			return ERR_PTR(-ENOSYS);
1009 		init_attr->recv_cq = init_attr->send_cq;
1010 		/* fall through */
1011 	case IB_QPT_RC:
1012 	case IB_QPT_UC:
1013 	case IB_QPT_RAW_PACKET:
1014 		qp = kzalloc(sizeof *qp, GFP_KERNEL);
1015 		if (!qp)
1016 			return ERR_PTR(-ENOMEM);
1017 		/* fall through */
1018 	case IB_QPT_UD:
1019 	{
1020 		err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1021 				       udata, 0, &qp);
1022 		if (err)
1023 			return ERR_PTR(err);
1024 
1025 		qp->ibqp.qp_num = qp->mqp.qpn;
1026 		qp->xrcdn = xrcdn;
1027 
1028 		break;
1029 	}
1030 	case IB_QPT_SMI:
1031 	case IB_QPT_GSI:
1032 	{
1033 		/* Userspace is not allowed to create special QPs: */
1034 		if (udata)
1035 			return ERR_PTR(-EINVAL);
1036 
1037 		err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1038 				       get_sqp_num(to_mdev(pd->device), init_attr),
1039 				       &qp);
1040 		if (err)
1041 			return ERR_PTR(err);
1042 
1043 		qp->port	= init_attr->port_num;
1044 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
1045 
1046 		break;
1047 	}
1048 	default:
1049 		/* Don't support raw QPs */
1050 		return ERR_PTR(-EINVAL);
1051 	}
1052 
1053 	return &qp->ibqp;
1054 }
1055 
1056 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1057 {
1058 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1059 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1060 	struct mlx4_ib_pd *pd;
1061 
1062 	if (is_qp0(dev, mqp))
1063 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1064 
1065 	pd = get_pd(mqp);
1066 	destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1067 
1068 	if (is_sqp(dev, mqp))
1069 		kfree(to_msqp(mqp));
1070 	else
1071 		kfree(mqp);
1072 
1073 	return 0;
1074 }
1075 
1076 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1077 {
1078 	switch (type) {
1079 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1080 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1081 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1082 	case MLX4_IB_QPT_XRC_INI:
1083 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1084 	case MLX4_IB_QPT_SMI:
1085 	case MLX4_IB_QPT_GSI:
1086 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1087 
1088 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1089 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1090 						MLX4_QP_ST_MLX : -1);
1091 	case MLX4_IB_QPT_PROXY_SMI:
1092 	case MLX4_IB_QPT_TUN_SMI:
1093 	case MLX4_IB_QPT_PROXY_GSI:
1094 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1095 						MLX4_QP_ST_UD : -1);
1096 	default:			return -1;
1097 	}
1098 }
1099 
1100 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1101 				   int attr_mask)
1102 {
1103 	u8 dest_rd_atomic;
1104 	u32 access_flags;
1105 	u32 hw_access_flags = 0;
1106 
1107 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1108 		dest_rd_atomic = attr->max_dest_rd_atomic;
1109 	else
1110 		dest_rd_atomic = qp->resp_depth;
1111 
1112 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1113 		access_flags = attr->qp_access_flags;
1114 	else
1115 		access_flags = qp->atomic_rd_en;
1116 
1117 	if (!dest_rd_atomic)
1118 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1119 
1120 	if (access_flags & IB_ACCESS_REMOTE_READ)
1121 		hw_access_flags |= MLX4_QP_BIT_RRE;
1122 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1123 		hw_access_flags |= MLX4_QP_BIT_RAE;
1124 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1125 		hw_access_flags |= MLX4_QP_BIT_RWE;
1126 
1127 	return cpu_to_be32(hw_access_flags);
1128 }
1129 
1130 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1131 			    int attr_mask)
1132 {
1133 	if (attr_mask & IB_QP_PKEY_INDEX)
1134 		sqp->pkey_index = attr->pkey_index;
1135 	if (attr_mask & IB_QP_QKEY)
1136 		sqp->qkey = attr->qkey;
1137 	if (attr_mask & IB_QP_SQ_PSN)
1138 		sqp->send_psn = attr->sq_psn;
1139 }
1140 
1141 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1142 {
1143 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1144 }
1145 
1146 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
1147 			 struct mlx4_qp_path *path, u8 port)
1148 {
1149 	int err;
1150 	int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
1151 		IB_LINK_LAYER_ETHERNET;
1152 	u8 mac[6];
1153 	int is_mcast;
1154 	u16 vlan_tag;
1155 	int vidx;
1156 
1157 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
1158 	path->rlid	    = cpu_to_be16(ah->dlid);
1159 	if (ah->static_rate) {
1160 		path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
1161 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1162 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1163 			--path->static_rate;
1164 	} else
1165 		path->static_rate = 0;
1166 
1167 	if (ah->ah_flags & IB_AH_GRH) {
1168 		if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
1169 			pr_err("sgid_index (%u) too large. max is %d\n",
1170 			       ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1171 			return -1;
1172 		}
1173 
1174 		path->grh_mylmc |= 1 << 7;
1175 		path->mgid_index = ah->grh.sgid_index;
1176 		path->hop_limit  = ah->grh.hop_limit;
1177 		path->tclass_flowlabel =
1178 			cpu_to_be32((ah->grh.traffic_class << 20) |
1179 				    (ah->grh.flow_label));
1180 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1181 	}
1182 
1183 	if (is_eth) {
1184 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1185 			((port - 1) << 6) | ((ah->sl & 7) << 3);
1186 
1187 		if (!(ah->ah_flags & IB_AH_GRH))
1188 			return -1;
1189 
1190 		err = mlx4_ib_resolve_grh(dev, ah, mac, &is_mcast, port);
1191 		if (err)
1192 			return err;
1193 
1194 		memcpy(path->dmac, mac, 6);
1195 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1196 		/* use index 0 into MAC table for IBoE */
1197 		path->grh_mylmc &= 0x80;
1198 
1199 		vlan_tag = rdma_get_vlan_id(&dev->iboe.gid_table[port - 1][ah->grh.sgid_index]);
1200 		if (vlan_tag < 0x1000) {
1201 			if (mlx4_find_cached_vlan(dev->dev, port, vlan_tag, &vidx))
1202 				return -ENOENT;
1203 
1204 			path->vlan_index = vidx;
1205 			path->fl = 1 << 6;
1206 		}
1207 	} else
1208 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1209 			((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1210 
1211 	return 0;
1212 }
1213 
1214 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1215 {
1216 	struct mlx4_ib_gid_entry *ge, *tmp;
1217 
1218 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1219 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1220 			ge->added = 1;
1221 			ge->port = qp->port;
1222 		}
1223 	}
1224 }
1225 
1226 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1227 			       const struct ib_qp_attr *attr, int attr_mask,
1228 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1229 {
1230 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1231 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1232 	struct mlx4_ib_pd *pd;
1233 	struct mlx4_ib_cq *send_cq, *recv_cq;
1234 	struct mlx4_qp_context *context;
1235 	enum mlx4_qp_optpar optpar = 0;
1236 	int sqd_event;
1237 	int err = -EINVAL;
1238 
1239 	context = kzalloc(sizeof *context, GFP_KERNEL);
1240 	if (!context)
1241 		return -ENOMEM;
1242 
1243 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1244 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1245 
1246 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1247 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1248 	else {
1249 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
1250 		switch (attr->path_mig_state) {
1251 		case IB_MIG_MIGRATED:
1252 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1253 			break;
1254 		case IB_MIG_REARM:
1255 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1256 			break;
1257 		case IB_MIG_ARMED:
1258 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1259 			break;
1260 		}
1261 	}
1262 
1263 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1264 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1265 	else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1266 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1267 	else if (ibqp->qp_type == IB_QPT_UD) {
1268 		if (qp->flags & MLX4_IB_QP_LSO)
1269 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
1270 					      ilog2(dev->dev->caps.max_gso_sz);
1271 		else
1272 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1273 	} else if (attr_mask & IB_QP_PATH_MTU) {
1274 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1275 			pr_err("path MTU (%u) is invalid\n",
1276 			       attr->path_mtu);
1277 			goto out;
1278 		}
1279 		context->mtu_msgmax = (attr->path_mtu << 5) |
1280 			ilog2(dev->dev->caps.max_msg_sz);
1281 	}
1282 
1283 	if (qp->rq.wqe_cnt)
1284 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1285 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
1286 
1287 	if (qp->sq.wqe_cnt)
1288 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1289 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
1290 
1291 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1292 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1293 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1294 	}
1295 
1296 	if (qp->ibqp.uobject)
1297 		context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
1298 	else
1299 		context->usr_page = cpu_to_be32(dev->priv_uar.index);
1300 
1301 	if (attr_mask & IB_QP_DEST_QPN)
1302 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1303 
1304 	if (attr_mask & IB_QP_PORT) {
1305 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1306 		    !(attr_mask & IB_QP_AV)) {
1307 			mlx4_set_sched(&context->pri_path, attr->port_num);
1308 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1309 		}
1310 	}
1311 
1312 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1313 		if (dev->counters[qp->port - 1] != -1) {
1314 			context->pri_path.counter_index =
1315 						dev->counters[qp->port - 1];
1316 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1317 		} else
1318 			context->pri_path.counter_index = 0xff;
1319 	}
1320 
1321 	if (attr_mask & IB_QP_PKEY_INDEX) {
1322 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1323 			context->pri_path.disable_pkey_check = 0x40;
1324 		context->pri_path.pkey_index = attr->pkey_index;
1325 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1326 	}
1327 
1328 	if (attr_mask & IB_QP_AV) {
1329 		if (mlx4_set_path(dev, &attr->ah_attr, &context->pri_path,
1330 				  attr_mask & IB_QP_PORT ?
1331 				  attr->port_num : qp->port))
1332 			goto out;
1333 
1334 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1335 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
1336 	}
1337 
1338 	if (attr_mask & IB_QP_TIMEOUT) {
1339 		context->pri_path.ackto |= attr->timeout << 3;
1340 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1341 	}
1342 
1343 	if (attr_mask & IB_QP_ALT_PATH) {
1344 		if (attr->alt_port_num == 0 ||
1345 		    attr->alt_port_num > dev->dev->caps.num_ports)
1346 			goto out;
1347 
1348 		if (attr->alt_pkey_index >=
1349 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
1350 			goto out;
1351 
1352 		if (mlx4_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
1353 				  attr->alt_port_num))
1354 			goto out;
1355 
1356 		context->alt_path.pkey_index = attr->alt_pkey_index;
1357 		context->alt_path.ackto = attr->alt_timeout << 3;
1358 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1359 	}
1360 
1361 	pd = get_pd(qp);
1362 	get_cqs(qp, &send_cq, &recv_cq);
1363 	context->pd       = cpu_to_be32(pd->pdn);
1364 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1365 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1366 	context->params1  = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1367 
1368 	/* Set "fast registration enabled" for all kernel QPs */
1369 	if (!qp->ibqp.uobject)
1370 		context->params1 |= cpu_to_be32(1 << 11);
1371 
1372 	if (attr_mask & IB_QP_RNR_RETRY) {
1373 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1374 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1375 	}
1376 
1377 	if (attr_mask & IB_QP_RETRY_CNT) {
1378 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1379 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1380 	}
1381 
1382 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1383 		if (attr->max_rd_atomic)
1384 			context->params1 |=
1385 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1386 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1387 	}
1388 
1389 	if (attr_mask & IB_QP_SQ_PSN)
1390 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1391 
1392 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1393 		if (attr->max_dest_rd_atomic)
1394 			context->params2 |=
1395 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1396 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1397 	}
1398 
1399 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1400 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1401 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1402 	}
1403 
1404 	if (ibqp->srq)
1405 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1406 
1407 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1408 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1409 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1410 	}
1411 	if (attr_mask & IB_QP_RQ_PSN)
1412 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1413 
1414 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1415 	if (attr_mask & IB_QP_QKEY) {
1416 		if (qp->mlx4_ib_qp_type &
1417 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1418 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1419 		else {
1420 			if (mlx4_is_mfunc(dev->dev) &&
1421 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1422 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1423 			    MLX4_RESERVED_QKEY_BASE) {
1424 				pr_err("Cannot use reserved QKEY"
1425 				       " 0x%x (range 0xffff0000..0xffffffff"
1426 				       " is reserved)\n", attr->qkey);
1427 				err = -EINVAL;
1428 				goto out;
1429 			}
1430 			context->qkey = cpu_to_be32(attr->qkey);
1431 		}
1432 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
1433 	}
1434 
1435 	if (ibqp->srq)
1436 		context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1437 
1438 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1439 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1440 
1441 	if (cur_state == IB_QPS_INIT &&
1442 	    new_state == IB_QPS_RTR  &&
1443 	    (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1444 	     ibqp->qp_type == IB_QPT_UD ||
1445 	     ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1446 		context->pri_path.sched_queue = (qp->port - 1) << 6;
1447 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1448 		    qp->mlx4_ib_qp_type &
1449 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1450 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1451 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1452 				context->pri_path.fl = 0x80;
1453 		} else {
1454 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1455 				context->pri_path.fl = 0x80;
1456 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1457 		}
1458 	}
1459 
1460 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1461 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1462 		sqd_event = 1;
1463 	else
1464 		sqd_event = 0;
1465 
1466 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1467 		context->rlkey |= (1 << 4);
1468 
1469 	/*
1470 	 * Before passing a kernel QP to the HW, make sure that the
1471 	 * ownership bits of the send queue are set and the SQ
1472 	 * headroom is stamped so that the hardware doesn't start
1473 	 * processing stale work requests.
1474 	 */
1475 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1476 		struct mlx4_wqe_ctrl_seg *ctrl;
1477 		int i;
1478 
1479 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1480 			ctrl = get_send_wqe(qp, i);
1481 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
1482 			if (qp->sq_max_wqes_per_wr == 1)
1483 				ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1484 
1485 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1486 		}
1487 	}
1488 
1489 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1490 			     to_mlx4_state(new_state), context, optpar,
1491 			     sqd_event, &qp->mqp);
1492 	if (err)
1493 		goto out;
1494 
1495 	qp->state = new_state;
1496 
1497 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1498 		qp->atomic_rd_en = attr->qp_access_flags;
1499 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1500 		qp->resp_depth = attr->max_dest_rd_atomic;
1501 	if (attr_mask & IB_QP_PORT) {
1502 		qp->port = attr->port_num;
1503 		update_mcg_macs(dev, qp);
1504 	}
1505 	if (attr_mask & IB_QP_ALT_PATH)
1506 		qp->alt_port = attr->alt_port_num;
1507 
1508 	if (is_sqp(dev, qp))
1509 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1510 
1511 	/*
1512 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
1513 	 * QP0 to RESET or ERROR, bring the link back down.
1514 	 */
1515 	if (is_qp0(dev, qp)) {
1516 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1517 			if (mlx4_INIT_PORT(dev->dev, qp->port))
1518 				pr_warn("INIT_PORT failed for port %d\n",
1519 				       qp->port);
1520 
1521 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1522 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1523 			mlx4_CLOSE_PORT(dev->dev, qp->port);
1524 	}
1525 
1526 	/*
1527 	 * If we moved a kernel QP to RESET, clean up all old CQ
1528 	 * entries and reinitialize the QP.
1529 	 */
1530 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1531 		mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1532 				 ibqp->srq ? to_msrq(ibqp->srq): NULL);
1533 		if (send_cq != recv_cq)
1534 			mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1535 
1536 		qp->rq.head = 0;
1537 		qp->rq.tail = 0;
1538 		qp->sq.head = 0;
1539 		qp->sq.tail = 0;
1540 		qp->sq_next_wqe = 0;
1541 		if (qp->rq.wqe_cnt)
1542 			*qp->db.db  = 0;
1543 	}
1544 
1545 out:
1546 	kfree(context);
1547 	return err;
1548 }
1549 
1550 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1551 		      int attr_mask, struct ib_udata *udata)
1552 {
1553 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1554 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1555 	enum ib_qp_state cur_state, new_state;
1556 	int err = -EINVAL;
1557 
1558 	mutex_lock(&qp->mutex);
1559 
1560 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1561 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1562 
1563 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
1564 		pr_debug("qpn 0x%x: invalid attribute mask specified "
1565 			 "for transition %d to %d. qp_type %d,"
1566 			 " attr_mask 0x%x\n",
1567 			 ibqp->qp_num, cur_state, new_state,
1568 			 ibqp->qp_type, attr_mask);
1569 		goto out;
1570 	}
1571 
1572 	if ((attr_mask & IB_QP_PORT) &&
1573 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
1574 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
1575 			 "for transition %d to %d. qp_type %d\n",
1576 			 ibqp->qp_num, attr->port_num, cur_state,
1577 			 new_state, ibqp->qp_type);
1578 		goto out;
1579 	}
1580 
1581 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
1582 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
1583 	     IB_LINK_LAYER_ETHERNET))
1584 		goto out;
1585 
1586 	if (attr_mask & IB_QP_PKEY_INDEX) {
1587 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1588 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
1589 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
1590 				 "for transition %d to %d. qp_type %d\n",
1591 				 ibqp->qp_num, attr->pkey_index, cur_state,
1592 				 new_state, ibqp->qp_type);
1593 			goto out;
1594 		}
1595 	}
1596 
1597 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1598 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
1599 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
1600 			 "Transition %d to %d. qp_type %d\n",
1601 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
1602 			 new_state, ibqp->qp_type);
1603 		goto out;
1604 	}
1605 
1606 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1607 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
1608 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
1609 			 "Transition %d to %d. qp_type %d\n",
1610 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
1611 			 new_state, ibqp->qp_type);
1612 		goto out;
1613 	}
1614 
1615 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1616 		err = 0;
1617 		goto out;
1618 	}
1619 
1620 	err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1621 
1622 out:
1623 	mutex_unlock(&qp->mutex);
1624 	return err;
1625 }
1626 
1627 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
1628 				  struct ib_send_wr *wr,
1629 				  void *wqe, unsigned *mlx_seg_len)
1630 {
1631 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
1632 	struct ib_device *ib_dev = &mdev->ib_dev;
1633 	struct mlx4_wqe_mlx_seg *mlx = wqe;
1634 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
1635 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
1636 	u16 pkey;
1637 	u32 qkey;
1638 	int send_size;
1639 	int header_size;
1640 	int spc;
1641 	int i;
1642 
1643 	if (wr->opcode != IB_WR_SEND)
1644 		return -EINVAL;
1645 
1646 	send_size = 0;
1647 
1648 	for (i = 0; i < wr->num_sge; ++i)
1649 		send_size += wr->sg_list[i].length;
1650 
1651 	/* for proxy-qp0 sends, need to add in size of tunnel header */
1652 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
1653 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
1654 		send_size += sizeof (struct mlx4_ib_tunnel_header);
1655 
1656 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, &sqp->ud_header);
1657 
1658 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
1659 		sqp->ud_header.lrh.service_level =
1660 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
1661 		sqp->ud_header.lrh.destination_lid =
1662 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
1663 		sqp->ud_header.lrh.source_lid =
1664 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
1665 	}
1666 
1667 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1668 
1669 	/* force loopback */
1670 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
1671 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
1672 
1673 	sqp->ud_header.lrh.virtual_lane    = 0;
1674 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1675 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
1676 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1677 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
1678 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1679 	else
1680 		sqp->ud_header.bth.destination_qpn =
1681 			cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
1682 
1683 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1684 	if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
1685 		return -EINVAL;
1686 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
1687 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
1688 
1689 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
1690 	sqp->ud_header.immediate_present = 0;
1691 
1692 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
1693 
1694 	/*
1695 	 * Inline data segments may not cross a 64 byte boundary.  If
1696 	 * our UD header is bigger than the space available up to the
1697 	 * next 64 byte boundary in the WQE, use two inline data
1698 	 * segments to hold the UD header.
1699 	 */
1700 	spc = MLX4_INLINE_ALIGN -
1701 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
1702 	if (header_size <= spc) {
1703 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
1704 		memcpy(inl + 1, sqp->header_buf, header_size);
1705 		i = 1;
1706 	} else {
1707 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
1708 		memcpy(inl + 1, sqp->header_buf, spc);
1709 
1710 		inl = (void *) (inl + 1) + spc;
1711 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
1712 		/*
1713 		 * Need a barrier here to make sure all the data is
1714 		 * visible before the byte_count field is set.
1715 		 * Otherwise the HCA prefetcher could grab the 64-byte
1716 		 * chunk with this inline segment and get a valid (!=
1717 		 * 0xffffffff) byte count but stale data, and end up
1718 		 * generating a packet with bad headers.
1719 		 *
1720 		 * The first inline segment's byte_count field doesn't
1721 		 * need a barrier, because it comes after a
1722 		 * control/MLX segment and therefore is at an offset
1723 		 * of 16 mod 64.
1724 		 */
1725 		wmb();
1726 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
1727 		i = 2;
1728 	}
1729 
1730 	*mlx_seg_len =
1731 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
1732 	return 0;
1733 }
1734 
1735 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
1736 			    void *wqe, unsigned *mlx_seg_len)
1737 {
1738 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
1739 	struct mlx4_wqe_mlx_seg *mlx = wqe;
1740 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
1741 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
1742 	struct net_device *ndev;
1743 	union ib_gid sgid;
1744 	u16 pkey;
1745 	int send_size;
1746 	int header_size;
1747 	int spc;
1748 	int i;
1749 	int is_eth;
1750 	int is_vlan = 0;
1751 	int is_grh;
1752 	u16 vlan;
1753 	int err = 0;
1754 
1755 	send_size = 0;
1756 	for (i = 0; i < wr->num_sge; ++i)
1757 		send_size += wr->sg_list[i].length;
1758 
1759 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
1760 	is_grh = mlx4_ib_ah_grh_present(ah);
1761 	if (is_eth) {
1762 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
1763 			/* When multi-function is enabled, the ib_core gid
1764 			 * indexes don't necessarily match the hw ones, so
1765 			 * we must use our own cache */
1766 			sgid.global.subnet_prefix =
1767 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
1768 				subnet_prefix;
1769 			sgid.global.interface_id =
1770 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
1771 				guid_cache[ah->av.ib.gid_index];
1772 		} else  {
1773 			err = ib_get_cached_gid(ib_dev,
1774 						be32_to_cpu(ah->av.ib.port_pd) >> 24,
1775 						ah->av.ib.gid_index, &sgid);
1776 			if (err)
1777 				return err;
1778 		}
1779 
1780 		vlan = rdma_get_vlan_id(&sgid);
1781 		is_vlan = vlan < 0x1000;
1782 	}
1783 	ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh, 0, &sqp->ud_header);
1784 
1785 	if (!is_eth) {
1786 		sqp->ud_header.lrh.service_level =
1787 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
1788 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
1789 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
1790 	}
1791 
1792 	if (is_grh) {
1793 		sqp->ud_header.grh.traffic_class =
1794 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
1795 		sqp->ud_header.grh.flow_label    =
1796 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
1797 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
1798 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
1799 			/* When multi-function is enabled, the ib_core gid
1800 			 * indexes don't necessarily match the hw ones, so
1801 			 * we must use our own cache */
1802 			sqp->ud_header.grh.source_gid.global.subnet_prefix =
1803 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
1804 						       subnet_prefix;
1805 			sqp->ud_header.grh.source_gid.global.interface_id =
1806 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
1807 					       guid_cache[ah->av.ib.gid_index];
1808 		} else
1809 			ib_get_cached_gid(ib_dev,
1810 					  be32_to_cpu(ah->av.ib.port_pd) >> 24,
1811 					  ah->av.ib.gid_index,
1812 					  &sqp->ud_header.grh.source_gid);
1813 		memcpy(sqp->ud_header.grh.destination_gid.raw,
1814 		       ah->av.ib.dgid, 16);
1815 	}
1816 
1817 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1818 
1819 	if (!is_eth) {
1820 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
1821 					  (sqp->ud_header.lrh.destination_lid ==
1822 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
1823 					  (sqp->ud_header.lrh.service_level << 8));
1824 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
1825 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
1826 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
1827 	}
1828 
1829 	switch (wr->opcode) {
1830 	case IB_WR_SEND:
1831 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
1832 		sqp->ud_header.immediate_present = 0;
1833 		break;
1834 	case IB_WR_SEND_WITH_IMM:
1835 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1836 		sqp->ud_header.immediate_present = 1;
1837 		sqp->ud_header.immediate_data    = wr->ex.imm_data;
1838 		break;
1839 	default:
1840 		return -EINVAL;
1841 	}
1842 
1843 	if (is_eth) {
1844 		u8 *smac;
1845 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
1846 
1847 		mlx->sched_prio = cpu_to_be16(pcp);
1848 
1849 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
1850 		/* FIXME: cache smac value? */
1851 		ndev = to_mdev(sqp->qp.ibqp.device)->iboe.netdevs[sqp->qp.port - 1];
1852 		if (!ndev)
1853 			return -ENODEV;
1854 		smac = ndev->dev_addr;
1855 		memcpy(sqp->ud_header.eth.smac_h, smac, 6);
1856 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
1857 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
1858 		if (!is_vlan) {
1859 			sqp->ud_header.eth.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
1860 		} else {
1861 			sqp->ud_header.vlan.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
1862 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
1863 		}
1864 	} else {
1865 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1866 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1867 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1868 	}
1869 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1870 	if (!sqp->qp.ibqp.qp_num)
1871 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
1872 	else
1873 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
1874 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1875 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1876 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1877 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1878 					       sqp->qkey : wr->wr.ud.remote_qkey);
1879 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1880 
1881 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
1882 
1883 	if (0) {
1884 		pr_err("built UD header of size %d:\n", header_size);
1885 		for (i = 0; i < header_size / 4; ++i) {
1886 			if (i % 8 == 0)
1887 				pr_err("  [%02x] ", i * 4);
1888 			pr_cont(" %08x",
1889 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
1890 			if ((i + 1) % 8 == 0)
1891 				pr_cont("\n");
1892 		}
1893 		pr_err("\n");
1894 	}
1895 
1896 	/*
1897 	 * Inline data segments may not cross a 64 byte boundary.  If
1898 	 * our UD header is bigger than the space available up to the
1899 	 * next 64 byte boundary in the WQE, use two inline data
1900 	 * segments to hold the UD header.
1901 	 */
1902 	spc = MLX4_INLINE_ALIGN -
1903 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
1904 	if (header_size <= spc) {
1905 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
1906 		memcpy(inl + 1, sqp->header_buf, header_size);
1907 		i = 1;
1908 	} else {
1909 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
1910 		memcpy(inl + 1, sqp->header_buf, spc);
1911 
1912 		inl = (void *) (inl + 1) + spc;
1913 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
1914 		/*
1915 		 * Need a barrier here to make sure all the data is
1916 		 * visible before the byte_count field is set.
1917 		 * Otherwise the HCA prefetcher could grab the 64-byte
1918 		 * chunk with this inline segment and get a valid (!=
1919 		 * 0xffffffff) byte count but stale data, and end up
1920 		 * generating a packet with bad headers.
1921 		 *
1922 		 * The first inline segment's byte_count field doesn't
1923 		 * need a barrier, because it comes after a
1924 		 * control/MLX segment and therefore is at an offset
1925 		 * of 16 mod 64.
1926 		 */
1927 		wmb();
1928 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
1929 		i = 2;
1930 	}
1931 
1932 	*mlx_seg_len =
1933 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
1934 	return 0;
1935 }
1936 
1937 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1938 {
1939 	unsigned cur;
1940 	struct mlx4_ib_cq *cq;
1941 
1942 	cur = wq->head - wq->tail;
1943 	if (likely(cur + nreq < wq->max_post))
1944 		return 0;
1945 
1946 	cq = to_mcq(ib_cq);
1947 	spin_lock(&cq->lock);
1948 	cur = wq->head - wq->tail;
1949 	spin_unlock(&cq->lock);
1950 
1951 	return cur + nreq >= wq->max_post;
1952 }
1953 
1954 static __be32 convert_access(int acc)
1955 {
1956 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? cpu_to_be32(MLX4_WQE_FMR_PERM_ATOMIC)       : 0) |
1957 	       (acc & IB_ACCESS_REMOTE_WRITE  ? cpu_to_be32(MLX4_WQE_FMR_PERM_REMOTE_WRITE) : 0) |
1958 	       (acc & IB_ACCESS_REMOTE_READ   ? cpu_to_be32(MLX4_WQE_FMR_PERM_REMOTE_READ)  : 0) |
1959 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
1960 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
1961 }
1962 
1963 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr)
1964 {
1965 	struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
1966 	int i;
1967 
1968 	for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i)
1969 		mfrpl->mapped_page_list[i] =
1970 			cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] |
1971 				    MLX4_MTT_FLAG_PRESENT);
1972 
1973 	fseg->flags		= convert_access(wr->wr.fast_reg.access_flags);
1974 	fseg->mem_key		= cpu_to_be32(wr->wr.fast_reg.rkey);
1975 	fseg->buf_list		= cpu_to_be64(mfrpl->map);
1976 	fseg->start_addr	= cpu_to_be64(wr->wr.fast_reg.iova_start);
1977 	fseg->reg_len		= cpu_to_be64(wr->wr.fast_reg.length);
1978 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
1979 	fseg->page_size		= cpu_to_be32(wr->wr.fast_reg.page_shift);
1980 	fseg->reserved[0]	= 0;
1981 	fseg->reserved[1]	= 0;
1982 }
1983 
1984 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
1985 {
1986 	iseg->flags	= 0;
1987 	iseg->mem_key	= cpu_to_be32(rkey);
1988 	iseg->guest_id	= 0;
1989 	iseg->pa	= 0;
1990 }
1991 
1992 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
1993 					  u64 remote_addr, u32 rkey)
1994 {
1995 	rseg->raddr    = cpu_to_be64(remote_addr);
1996 	rseg->rkey     = cpu_to_be32(rkey);
1997 	rseg->reserved = 0;
1998 }
1999 
2000 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
2001 {
2002 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2003 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
2004 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
2005 	} else if (wr->opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2006 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2007 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add_mask);
2008 	} else {
2009 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2010 		aseg->compare  = 0;
2011 	}
2012 
2013 }
2014 
2015 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2016 				  struct ib_send_wr *wr)
2017 {
2018 	aseg->swap_add		= cpu_to_be64(wr->wr.atomic.swap);
2019 	aseg->swap_add_mask	= cpu_to_be64(wr->wr.atomic.swap_mask);
2020 	aseg->compare		= cpu_to_be64(wr->wr.atomic.compare_add);
2021 	aseg->compare_mask	= cpu_to_be64(wr->wr.atomic.compare_add_mask);
2022 }
2023 
2024 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2025 			     struct ib_send_wr *wr)
2026 {
2027 	memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
2028 	dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2029 	dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2030 	dseg->vlan = to_mah(wr->wr.ud.ah)->av.eth.vlan;
2031 	memcpy(dseg->mac, to_mah(wr->wr.ud.ah)->av.eth.mac, 6);
2032 }
2033 
2034 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2035 				    struct mlx4_wqe_datagram_seg *dseg,
2036 				    struct ib_send_wr *wr, enum ib_qp_type qpt)
2037 {
2038 	union mlx4_ext_av *av = &to_mah(wr->wr.ud.ah)->av;
2039 	struct mlx4_av sqp_av = {0};
2040 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
2041 
2042 	/* force loopback */
2043 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2044 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2045 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2046 			cpu_to_be32(0xf0000000);
2047 
2048 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2049 	/* This function used only for sending on QP1 proxies */
2050 	dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2051 	/* Use QKEY from the QP context, which is set by master */
2052 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2053 }
2054 
2055 static void build_tunnel_header(struct ib_send_wr *wr, void *wqe, unsigned *mlx_seg_len)
2056 {
2057 	struct mlx4_wqe_inline_seg *inl = wqe;
2058 	struct mlx4_ib_tunnel_header hdr;
2059 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2060 	int spc;
2061 	int i;
2062 
2063 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2064 	hdr.remote_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2065 	hdr.pkey_index = cpu_to_be16(wr->wr.ud.pkey_index);
2066 	hdr.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2067 
2068 	spc = MLX4_INLINE_ALIGN -
2069 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2070 	if (sizeof (hdr) <= spc) {
2071 		memcpy(inl + 1, &hdr, sizeof (hdr));
2072 		wmb();
2073 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
2074 		i = 1;
2075 	} else {
2076 		memcpy(inl + 1, &hdr, spc);
2077 		wmb();
2078 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2079 
2080 		inl = (void *) (inl + 1) + spc;
2081 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2082 		wmb();
2083 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
2084 		i = 2;
2085 	}
2086 
2087 	*mlx_seg_len =
2088 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2089 }
2090 
2091 static void set_mlx_icrc_seg(void *dseg)
2092 {
2093 	u32 *t = dseg;
2094 	struct mlx4_wqe_inline_seg *iseg = dseg;
2095 
2096 	t[1] = 0;
2097 
2098 	/*
2099 	 * Need a barrier here before writing the byte_count field to
2100 	 * make sure that all the data is visible before the
2101 	 * byte_count field is set.  Otherwise, if the segment begins
2102 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2103 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2104 	 * stale data, and end up sending the wrong data.
2105 	 */
2106 	wmb();
2107 
2108 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2109 }
2110 
2111 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2112 {
2113 	dseg->lkey       = cpu_to_be32(sg->lkey);
2114 	dseg->addr       = cpu_to_be64(sg->addr);
2115 
2116 	/*
2117 	 * Need a barrier here before writing the byte_count field to
2118 	 * make sure that all the data is visible before the
2119 	 * byte_count field is set.  Otherwise, if the segment begins
2120 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2121 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2122 	 * stale data, and end up sending the wrong data.
2123 	 */
2124 	wmb();
2125 
2126 	dseg->byte_count = cpu_to_be32(sg->length);
2127 }
2128 
2129 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2130 {
2131 	dseg->byte_count = cpu_to_be32(sg->length);
2132 	dseg->lkey       = cpu_to_be32(sg->lkey);
2133 	dseg->addr       = cpu_to_be64(sg->addr);
2134 }
2135 
2136 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr,
2137 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2138 			 __be32 *lso_hdr_sz, __be32 *blh)
2139 {
2140 	unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
2141 
2142 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2143 		*blh = cpu_to_be32(1 << 6);
2144 
2145 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2146 		     wr->num_sge > qp->sq.max_gs - (halign >> 4)))
2147 		return -EINVAL;
2148 
2149 	memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
2150 
2151 	*lso_hdr_sz  = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 |
2152 				   wr->wr.ud.hlen);
2153 	*lso_seg_len = halign;
2154 	return 0;
2155 }
2156 
2157 static __be32 send_ieth(struct ib_send_wr *wr)
2158 {
2159 	switch (wr->opcode) {
2160 	case IB_WR_SEND_WITH_IMM:
2161 	case IB_WR_RDMA_WRITE_WITH_IMM:
2162 		return wr->ex.imm_data;
2163 
2164 	case IB_WR_SEND_WITH_INV:
2165 		return cpu_to_be32(wr->ex.invalidate_rkey);
2166 
2167 	default:
2168 		return 0;
2169 	}
2170 }
2171 
2172 static void add_zero_len_inline(void *wqe)
2173 {
2174 	struct mlx4_wqe_inline_seg *inl = wqe;
2175 	memset(wqe, 0, 16);
2176 	inl->byte_count = cpu_to_be32(1 << 31);
2177 }
2178 
2179 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2180 		      struct ib_send_wr **bad_wr)
2181 {
2182 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2183 	void *wqe;
2184 	struct mlx4_wqe_ctrl_seg *ctrl;
2185 	struct mlx4_wqe_data_seg *dseg;
2186 	unsigned long flags;
2187 	int nreq;
2188 	int err = 0;
2189 	unsigned ind;
2190 	int uninitialized_var(stamp);
2191 	int uninitialized_var(size);
2192 	unsigned uninitialized_var(seglen);
2193 	__be32 dummy;
2194 	__be32 *lso_wqe;
2195 	__be32 uninitialized_var(lso_hdr_sz);
2196 	__be32 blh;
2197 	int i;
2198 
2199 	spin_lock_irqsave(&qp->sq.lock, flags);
2200 
2201 	ind = qp->sq_next_wqe;
2202 
2203 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2204 		lso_wqe = &dummy;
2205 		blh = 0;
2206 
2207 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2208 			err = -ENOMEM;
2209 			*bad_wr = wr;
2210 			goto out;
2211 		}
2212 
2213 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
2214 			err = -EINVAL;
2215 			*bad_wr = wr;
2216 			goto out;
2217 		}
2218 
2219 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
2220 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
2221 
2222 		ctrl->srcrb_flags =
2223 			(wr->send_flags & IB_SEND_SIGNALED ?
2224 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
2225 			(wr->send_flags & IB_SEND_SOLICITED ?
2226 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
2227 			((wr->send_flags & IB_SEND_IP_CSUM) ?
2228 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
2229 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
2230 			qp->sq_signal_bits;
2231 
2232 		ctrl->imm = send_ieth(wr);
2233 
2234 		wqe += sizeof *ctrl;
2235 		size = sizeof *ctrl / 16;
2236 
2237 		switch (qp->mlx4_ib_qp_type) {
2238 		case MLX4_IB_QPT_RC:
2239 		case MLX4_IB_QPT_UC:
2240 			switch (wr->opcode) {
2241 			case IB_WR_ATOMIC_CMP_AND_SWP:
2242 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2243 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
2244 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2245 					      wr->wr.atomic.rkey);
2246 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2247 
2248 				set_atomic_seg(wqe, wr);
2249 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
2250 
2251 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2252 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
2253 
2254 				break;
2255 
2256 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2257 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2258 					      wr->wr.atomic.rkey);
2259 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2260 
2261 				set_masked_atomic_seg(wqe, wr);
2262 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
2263 
2264 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2265 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
2266 
2267 				break;
2268 
2269 			case IB_WR_RDMA_READ:
2270 			case IB_WR_RDMA_WRITE:
2271 			case IB_WR_RDMA_WRITE_WITH_IMM:
2272 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2273 					      wr->wr.rdma.rkey);
2274 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2275 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
2276 				break;
2277 
2278 			case IB_WR_LOCAL_INV:
2279 				ctrl->srcrb_flags |=
2280 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2281 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
2282 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
2283 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
2284 				break;
2285 
2286 			case IB_WR_FAST_REG_MR:
2287 				ctrl->srcrb_flags |=
2288 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2289 				set_fmr_seg(wqe, wr);
2290 				wqe  += sizeof (struct mlx4_wqe_fmr_seg);
2291 				size += sizeof (struct mlx4_wqe_fmr_seg) / 16;
2292 				break;
2293 
2294 			default:
2295 				/* No extra segments required for sends */
2296 				break;
2297 			}
2298 			break;
2299 
2300 		case MLX4_IB_QPT_TUN_SMI_OWNER:
2301 			err =  build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2302 			if (unlikely(err)) {
2303 				*bad_wr = wr;
2304 				goto out;
2305 			}
2306 			wqe  += seglen;
2307 			size += seglen / 16;
2308 			break;
2309 		case MLX4_IB_QPT_TUN_SMI:
2310 		case MLX4_IB_QPT_TUN_GSI:
2311 			/* this is a UD qp used in MAD responses to slaves. */
2312 			set_datagram_seg(wqe, wr);
2313 			/* set the forced-loopback bit in the data seg av */
2314 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
2315 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2316 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2317 			break;
2318 		case MLX4_IB_QPT_UD:
2319 			set_datagram_seg(wqe, wr);
2320 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2321 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2322 
2323 			if (wr->opcode == IB_WR_LSO) {
2324 				err = build_lso_seg(wqe, wr, qp, &seglen, &lso_hdr_sz, &blh);
2325 				if (unlikely(err)) {
2326 					*bad_wr = wr;
2327 					goto out;
2328 				}
2329 				lso_wqe = (__be32 *) wqe;
2330 				wqe  += seglen;
2331 				size += seglen / 16;
2332 			}
2333 			break;
2334 
2335 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
2336 			if (unlikely(!mlx4_is_master(to_mdev(ibqp->device)->dev))) {
2337 				err = -ENOSYS;
2338 				*bad_wr = wr;
2339 				goto out;
2340 			}
2341 			err = build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2342 			if (unlikely(err)) {
2343 				*bad_wr = wr;
2344 				goto out;
2345 			}
2346 			wqe  += seglen;
2347 			size += seglen / 16;
2348 			/* to start tunnel header on a cache-line boundary */
2349 			add_zero_len_inline(wqe);
2350 			wqe += 16;
2351 			size++;
2352 			build_tunnel_header(wr, wqe, &seglen);
2353 			wqe  += seglen;
2354 			size += seglen / 16;
2355 			break;
2356 		case MLX4_IB_QPT_PROXY_SMI:
2357 			/* don't allow QP0 sends on guests */
2358 			err = -ENOSYS;
2359 			*bad_wr = wr;
2360 			goto out;
2361 		case MLX4_IB_QPT_PROXY_GSI:
2362 			/* If we are tunneling special qps, this is a UD qp.
2363 			 * In this case we first add a UD segment targeting
2364 			 * the tunnel qp, and then add a header with address
2365 			 * information */
2366 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe, wr, ibqp->qp_type);
2367 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2368 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2369 			build_tunnel_header(wr, wqe, &seglen);
2370 			wqe  += seglen;
2371 			size += seglen / 16;
2372 			break;
2373 
2374 		case MLX4_IB_QPT_SMI:
2375 		case MLX4_IB_QPT_GSI:
2376 			err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
2377 			if (unlikely(err)) {
2378 				*bad_wr = wr;
2379 				goto out;
2380 			}
2381 			wqe  += seglen;
2382 			size += seglen / 16;
2383 			break;
2384 
2385 		default:
2386 			break;
2387 		}
2388 
2389 		/*
2390 		 * Write data segments in reverse order, so as to
2391 		 * overwrite cacheline stamp last within each
2392 		 * cacheline.  This avoids issues with WQE
2393 		 * prefetching.
2394 		 */
2395 
2396 		dseg = wqe;
2397 		dseg += wr->num_sge - 1;
2398 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
2399 
2400 		/* Add one more inline data segment for ICRC for MLX sends */
2401 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2402 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
2403 			     qp->mlx4_ib_qp_type &
2404 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
2405 			set_mlx_icrc_seg(dseg + 1);
2406 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
2407 		}
2408 
2409 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
2410 			set_data_seg(dseg, wr->sg_list + i);
2411 
2412 		/*
2413 		 * Possibly overwrite stamping in cacheline with LSO
2414 		 * segment only after making sure all data segments
2415 		 * are written.
2416 		 */
2417 		wmb();
2418 		*lso_wqe = lso_hdr_sz;
2419 
2420 		ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
2421 				    MLX4_WQE_CTRL_FENCE : 0) | size;
2422 
2423 		/*
2424 		 * Make sure descriptor is fully written before
2425 		 * setting ownership bit (because HW can start
2426 		 * executing as soon as we do).
2427 		 */
2428 		wmb();
2429 
2430 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
2431 			*bad_wr = wr;
2432 			err = -EINVAL;
2433 			goto out;
2434 		}
2435 
2436 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
2437 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
2438 
2439 		stamp = ind + qp->sq_spare_wqes;
2440 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
2441 
2442 		/*
2443 		 * We can improve latency by not stamping the last
2444 		 * send queue WQE until after ringing the doorbell, so
2445 		 * only stamp here if there are still more WQEs to post.
2446 		 *
2447 		 * Same optimization applies to padding with NOP wqe
2448 		 * in case of WQE shrinking (used to prevent wrap-around
2449 		 * in the middle of WR).
2450 		 */
2451 		if (wr->next) {
2452 			stamp_send_wqe(qp, stamp, size * 16);
2453 			ind = pad_wraparound(qp, ind);
2454 		}
2455 	}
2456 
2457 out:
2458 	if (likely(nreq)) {
2459 		qp->sq.head += nreq;
2460 
2461 		/*
2462 		 * Make sure that descriptors are written before
2463 		 * doorbell record.
2464 		 */
2465 		wmb();
2466 
2467 		writel(qp->doorbell_qpn,
2468 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
2469 
2470 		/*
2471 		 * Make sure doorbells don't leak out of SQ spinlock
2472 		 * and reach the HCA out of order.
2473 		 */
2474 		mmiowb();
2475 
2476 		stamp_send_wqe(qp, stamp, size * 16);
2477 
2478 		ind = pad_wraparound(qp, ind);
2479 		qp->sq_next_wqe = ind;
2480 	}
2481 
2482 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2483 
2484 	return err;
2485 }
2486 
2487 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2488 		      struct ib_recv_wr **bad_wr)
2489 {
2490 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2491 	struct mlx4_wqe_data_seg *scat;
2492 	unsigned long flags;
2493 	int err = 0;
2494 	int nreq;
2495 	int ind;
2496 	int max_gs;
2497 	int i;
2498 
2499 	max_gs = qp->rq.max_gs;
2500 	spin_lock_irqsave(&qp->rq.lock, flags);
2501 
2502 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
2503 
2504 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2505 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2506 			err = -ENOMEM;
2507 			*bad_wr = wr;
2508 			goto out;
2509 		}
2510 
2511 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2512 			err = -EINVAL;
2513 			*bad_wr = wr;
2514 			goto out;
2515 		}
2516 
2517 		scat = get_recv_wqe(qp, ind);
2518 
2519 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
2520 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
2521 			ib_dma_sync_single_for_device(ibqp->device,
2522 						      qp->sqp_proxy_rcv[ind].map,
2523 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
2524 						      DMA_FROM_DEVICE);
2525 			scat->byte_count =
2526 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
2527 			/* use dma lkey from upper layer entry */
2528 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
2529 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
2530 			scat++;
2531 			max_gs--;
2532 		}
2533 
2534 		for (i = 0; i < wr->num_sge; ++i)
2535 			__set_data_seg(scat + i, wr->sg_list + i);
2536 
2537 		if (i < max_gs) {
2538 			scat[i].byte_count = 0;
2539 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
2540 			scat[i].addr       = 0;
2541 		}
2542 
2543 		qp->rq.wrid[ind] = wr->wr_id;
2544 
2545 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
2546 	}
2547 
2548 out:
2549 	if (likely(nreq)) {
2550 		qp->rq.head += nreq;
2551 
2552 		/*
2553 		 * Make sure that descriptors are written before
2554 		 * doorbell record.
2555 		 */
2556 		wmb();
2557 
2558 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
2559 	}
2560 
2561 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2562 
2563 	return err;
2564 }
2565 
2566 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
2567 {
2568 	switch (mlx4_state) {
2569 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
2570 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
2571 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
2572 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
2573 	case MLX4_QP_STATE_SQ_DRAINING:
2574 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
2575 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
2576 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
2577 	default:		     return -1;
2578 	}
2579 }
2580 
2581 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
2582 {
2583 	switch (mlx4_mig_state) {
2584 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
2585 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
2586 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
2587 	default: return -1;
2588 	}
2589 }
2590 
2591 static int to_ib_qp_access_flags(int mlx4_flags)
2592 {
2593 	int ib_flags = 0;
2594 
2595 	if (mlx4_flags & MLX4_QP_BIT_RRE)
2596 		ib_flags |= IB_ACCESS_REMOTE_READ;
2597 	if (mlx4_flags & MLX4_QP_BIT_RWE)
2598 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
2599 	if (mlx4_flags & MLX4_QP_BIT_RAE)
2600 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
2601 
2602 	return ib_flags;
2603 }
2604 
2605 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
2606 				struct mlx4_qp_path *path)
2607 {
2608 	struct mlx4_dev *dev = ibdev->dev;
2609 	int is_eth;
2610 
2611 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
2612 	ib_ah_attr->port_num	  = path->sched_queue & 0x40 ? 2 : 1;
2613 
2614 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
2615 		return;
2616 
2617 	is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
2618 		IB_LINK_LAYER_ETHERNET;
2619 	if (is_eth)
2620 		ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
2621 		((path->sched_queue & 4) << 1);
2622 	else
2623 		ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
2624 
2625 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
2626 	ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
2627 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
2628 	ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
2629 	if (ib_ah_attr->ah_flags) {
2630 		ib_ah_attr->grh.sgid_index = path->mgid_index;
2631 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
2632 		ib_ah_attr->grh.traffic_class =
2633 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
2634 		ib_ah_attr->grh.flow_label =
2635 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
2636 		memcpy(ib_ah_attr->grh.dgid.raw,
2637 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
2638 	}
2639 }
2640 
2641 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
2642 		     struct ib_qp_init_attr *qp_init_attr)
2643 {
2644 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2645 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2646 	struct mlx4_qp_context context;
2647 	int mlx4_state;
2648 	int err = 0;
2649 
2650 	mutex_lock(&qp->mutex);
2651 
2652 	if (qp->state == IB_QPS_RESET) {
2653 		qp_attr->qp_state = IB_QPS_RESET;
2654 		goto done;
2655 	}
2656 
2657 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
2658 	if (err) {
2659 		err = -EINVAL;
2660 		goto out;
2661 	}
2662 
2663 	mlx4_state = be32_to_cpu(context.flags) >> 28;
2664 
2665 	qp->state		     = to_ib_qp_state(mlx4_state);
2666 	qp_attr->qp_state	     = qp->state;
2667 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
2668 	qp_attr->path_mig_state	     =
2669 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
2670 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
2671 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
2672 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
2673 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
2674 	qp_attr->qp_access_flags     =
2675 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
2676 
2677 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
2678 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
2679 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
2680 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
2681 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
2682 	}
2683 
2684 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
2685 	if (qp_attr->qp_state == IB_QPS_INIT)
2686 		qp_attr->port_num = qp->port;
2687 	else
2688 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
2689 
2690 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
2691 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
2692 
2693 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
2694 
2695 	qp_attr->max_dest_rd_atomic =
2696 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
2697 	qp_attr->min_rnr_timer	    =
2698 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
2699 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
2700 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
2701 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
2702 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
2703 
2704 done:
2705 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
2706 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
2707 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
2708 
2709 	if (!ibqp->uobject) {
2710 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
2711 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
2712 	} else {
2713 		qp_attr->cap.max_send_wr  = 0;
2714 		qp_attr->cap.max_send_sge = 0;
2715 	}
2716 
2717 	/*
2718 	 * We don't support inline sends for kernel QPs (yet), and we
2719 	 * don't know what userspace's value should be.
2720 	 */
2721 	qp_attr->cap.max_inline_data = 0;
2722 
2723 	qp_init_attr->cap	     = qp_attr->cap;
2724 
2725 	qp_init_attr->create_flags = 0;
2726 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
2727 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
2728 
2729 	if (qp->flags & MLX4_IB_QP_LSO)
2730 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
2731 
2732 	qp_init_attr->sq_sig_type =
2733 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
2734 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
2735 
2736 out:
2737 	mutex_unlock(&qp->mutex);
2738 	return err;
2739 }
2740 
2741