xref: /linux/drivers/infiniband/hw/mlx4/qp.c (revision 2ba9268dd603d23e17643437b2246acb6844953b)
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/driver.h>
44 #include <linux/mlx4/qp.h>
45 
46 #include "mlx4_ib.h"
47 #include "user.h"
48 
49 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
50 			     struct mlx4_ib_cq *recv_cq);
51 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
52 			       struct mlx4_ib_cq *recv_cq);
53 
54 enum {
55 	MLX4_IB_ACK_REQ_FREQ	= 8,
56 };
57 
58 enum {
59 	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
60 	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
61 	MLX4_IB_LINK_TYPE_IB		= 0,
62 	MLX4_IB_LINK_TYPE_ETH		= 1
63 };
64 
65 enum {
66 	/*
67 	 * Largest possible UD header: send with GRH and immediate
68 	 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
69 	 * tag.  (LRH would only use 8 bytes, so Ethernet is the
70 	 * biggest case)
71 	 */
72 	MLX4_IB_UD_HEADER_SIZE		= 82,
73 	MLX4_IB_LSO_HEADER_SPARE	= 128,
74 };
75 
76 enum {
77 	MLX4_IB_IBOE_ETHERTYPE		= 0x8915
78 };
79 
80 struct mlx4_ib_sqp {
81 	struct mlx4_ib_qp	qp;
82 	int			pkey_index;
83 	u32			qkey;
84 	u32			send_psn;
85 	struct ib_ud_header	ud_header;
86 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
87 };
88 
89 enum {
90 	MLX4_IB_MIN_SQ_STRIDE	= 6,
91 	MLX4_IB_CACHE_LINE_SIZE	= 64,
92 };
93 
94 enum {
95 	MLX4_RAW_QP_MTU		= 7,
96 	MLX4_RAW_QP_MSGMAX	= 31,
97 };
98 
99 #ifndef ETH_ALEN
100 #define ETH_ALEN        6
101 #endif
102 
103 static const __be32 mlx4_ib_opcode[] = {
104 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
105 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
106 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
107 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
108 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
109 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
110 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
111 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
112 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
113 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
114 	[IB_WR_FAST_REG_MR]			= cpu_to_be32(MLX4_OPCODE_FMR),
115 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
116 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
117 	[IB_WR_BIND_MW]				= cpu_to_be32(MLX4_OPCODE_BIND_MW),
118 };
119 
120 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
121 {
122 	return container_of(mqp, struct mlx4_ib_sqp, qp);
123 }
124 
125 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
126 {
127 	if (!mlx4_is_master(dev->dev))
128 		return 0;
129 
130 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
131 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
132 		8 * MLX4_MFUNC_MAX;
133 }
134 
135 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
136 {
137 	int proxy_sqp = 0;
138 	int real_sqp = 0;
139 	int i;
140 	/* PPF or Native -- real SQP */
141 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
142 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
143 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
144 	if (real_sqp)
145 		return 1;
146 	/* VF or PF -- proxy SQP */
147 	if (mlx4_is_mfunc(dev->dev)) {
148 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
149 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] ||
150 			    qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) {
151 				proxy_sqp = 1;
152 				break;
153 			}
154 		}
155 	}
156 	return proxy_sqp;
157 }
158 
159 /* used for INIT/CLOSE port logic */
160 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
161 {
162 	int proxy_qp0 = 0;
163 	int real_qp0 = 0;
164 	int i;
165 	/* PPF or Native -- real QP0 */
166 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
167 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
168 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
169 	if (real_qp0)
170 		return 1;
171 	/* VF or PF -- proxy QP0 */
172 	if (mlx4_is_mfunc(dev->dev)) {
173 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
174 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) {
175 				proxy_qp0 = 1;
176 				break;
177 			}
178 		}
179 	}
180 	return proxy_qp0;
181 }
182 
183 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
184 {
185 	return mlx4_buf_offset(&qp->buf, offset);
186 }
187 
188 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
189 {
190 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
191 }
192 
193 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
194 {
195 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
196 }
197 
198 /*
199  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
200  * first four bytes of every 64 byte chunk with
201  *     0x7FFFFFF | (invalid_ownership_value << 31).
202  *
203  * When the max work request size is less than or equal to the WQE
204  * basic block size, as an optimization, we can stamp all WQEs with
205  * 0xffffffff, and skip the very first chunk of each WQE.
206  */
207 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
208 {
209 	__be32 *wqe;
210 	int i;
211 	int s;
212 	int ind;
213 	void *buf;
214 	__be32 stamp;
215 	struct mlx4_wqe_ctrl_seg *ctrl;
216 
217 	if (qp->sq_max_wqes_per_wr > 1) {
218 		s = roundup(size, 1U << qp->sq.wqe_shift);
219 		for (i = 0; i < s; i += 64) {
220 			ind = (i >> qp->sq.wqe_shift) + n;
221 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
222 						       cpu_to_be32(0xffffffff);
223 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
224 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
225 			*wqe = stamp;
226 		}
227 	} else {
228 		ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
229 		s = (ctrl->fence_size & 0x3f) << 4;
230 		for (i = 64; i < s; i += 64) {
231 			wqe = buf + i;
232 			*wqe = cpu_to_be32(0xffffffff);
233 		}
234 	}
235 }
236 
237 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
238 {
239 	struct mlx4_wqe_ctrl_seg *ctrl;
240 	struct mlx4_wqe_inline_seg *inl;
241 	void *wqe;
242 	int s;
243 
244 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
245 	s = sizeof(struct mlx4_wqe_ctrl_seg);
246 
247 	if (qp->ibqp.qp_type == IB_QPT_UD) {
248 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
249 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
250 		memset(dgram, 0, sizeof *dgram);
251 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
252 		s += sizeof(struct mlx4_wqe_datagram_seg);
253 	}
254 
255 	/* Pad the remainder of the WQE with an inline data segment. */
256 	if (size > s) {
257 		inl = wqe + s;
258 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
259 	}
260 	ctrl->srcrb_flags = 0;
261 	ctrl->fence_size = size / 16;
262 	/*
263 	 * Make sure descriptor is fully written before setting ownership bit
264 	 * (because HW can start executing as soon as we do).
265 	 */
266 	wmb();
267 
268 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
269 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
270 
271 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
272 }
273 
274 /* Post NOP WQE to prevent wrap-around in the middle of WR */
275 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
276 {
277 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
278 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
279 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
280 		ind += s;
281 	}
282 	return ind;
283 }
284 
285 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
286 {
287 	struct ib_event event;
288 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
289 
290 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
291 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
292 
293 	if (ibqp->event_handler) {
294 		event.device     = ibqp->device;
295 		event.element.qp = ibqp;
296 		switch (type) {
297 		case MLX4_EVENT_TYPE_PATH_MIG:
298 			event.event = IB_EVENT_PATH_MIG;
299 			break;
300 		case MLX4_EVENT_TYPE_COMM_EST:
301 			event.event = IB_EVENT_COMM_EST;
302 			break;
303 		case MLX4_EVENT_TYPE_SQ_DRAINED:
304 			event.event = IB_EVENT_SQ_DRAINED;
305 			break;
306 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
307 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
308 			break;
309 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
310 			event.event = IB_EVENT_QP_FATAL;
311 			break;
312 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
313 			event.event = IB_EVENT_PATH_MIG_ERR;
314 			break;
315 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
316 			event.event = IB_EVENT_QP_REQ_ERR;
317 			break;
318 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
319 			event.event = IB_EVENT_QP_ACCESS_ERR;
320 			break;
321 		default:
322 			pr_warn("Unexpected event type %d "
323 			       "on QP %06x\n", type, qp->qpn);
324 			return;
325 		}
326 
327 		ibqp->event_handler(&event, ibqp->qp_context);
328 	}
329 }
330 
331 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
332 {
333 	/*
334 	 * UD WQEs must have a datagram segment.
335 	 * RC and UC WQEs might have a remote address segment.
336 	 * MLX WQEs need two extra inline data segments (for the UD
337 	 * header and space for the ICRC).
338 	 */
339 	switch (type) {
340 	case MLX4_IB_QPT_UD:
341 		return sizeof (struct mlx4_wqe_ctrl_seg) +
342 			sizeof (struct mlx4_wqe_datagram_seg) +
343 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
344 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
345 	case MLX4_IB_QPT_PROXY_SMI:
346 	case MLX4_IB_QPT_PROXY_GSI:
347 		return sizeof (struct mlx4_wqe_ctrl_seg) +
348 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
349 	case MLX4_IB_QPT_TUN_SMI_OWNER:
350 	case MLX4_IB_QPT_TUN_GSI:
351 		return sizeof (struct mlx4_wqe_ctrl_seg) +
352 			sizeof (struct mlx4_wqe_datagram_seg);
353 
354 	case MLX4_IB_QPT_UC:
355 		return sizeof (struct mlx4_wqe_ctrl_seg) +
356 			sizeof (struct mlx4_wqe_raddr_seg);
357 	case MLX4_IB_QPT_RC:
358 		return sizeof (struct mlx4_wqe_ctrl_seg) +
359 			sizeof (struct mlx4_wqe_atomic_seg) +
360 			sizeof (struct mlx4_wqe_raddr_seg);
361 	case MLX4_IB_QPT_SMI:
362 	case MLX4_IB_QPT_GSI:
363 		return sizeof (struct mlx4_wqe_ctrl_seg) +
364 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
365 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
366 					   MLX4_INLINE_ALIGN) *
367 			      sizeof (struct mlx4_wqe_inline_seg),
368 			      sizeof (struct mlx4_wqe_data_seg)) +
369 			ALIGN(4 +
370 			      sizeof (struct mlx4_wqe_inline_seg),
371 			      sizeof (struct mlx4_wqe_data_seg));
372 	default:
373 		return sizeof (struct mlx4_wqe_ctrl_seg);
374 	}
375 }
376 
377 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
378 		       int is_user, int has_rq, struct mlx4_ib_qp *qp)
379 {
380 	/* Sanity check RQ size before proceeding */
381 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
382 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
383 		return -EINVAL;
384 
385 	if (!has_rq) {
386 		if (cap->max_recv_wr)
387 			return -EINVAL;
388 
389 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
390 	} else {
391 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
392 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
393 			return -EINVAL;
394 
395 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
396 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
397 		qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
398 	}
399 
400 	/* leave userspace return values as they were, so as not to break ABI */
401 	if (is_user) {
402 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
403 		cap->max_recv_sge = qp->rq.max_gs;
404 	} else {
405 		cap->max_recv_wr  = qp->rq.max_post =
406 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
407 		cap->max_recv_sge = min(qp->rq.max_gs,
408 					min(dev->dev->caps.max_sq_sg,
409 					    dev->dev->caps.max_rq_sg));
410 	}
411 
412 	return 0;
413 }
414 
415 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
416 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
417 {
418 	int s;
419 
420 	/* Sanity check SQ size before proceeding */
421 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
422 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
423 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
424 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
425 		return -EINVAL;
426 
427 	/*
428 	 * For MLX transport we need 2 extra S/G entries:
429 	 * one for the header and one for the checksum at the end
430 	 */
431 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
432 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
433 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
434 		return -EINVAL;
435 
436 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
437 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
438 		send_wqe_overhead(type, qp->flags);
439 
440 	if (s > dev->dev->caps.max_sq_desc_sz)
441 		return -EINVAL;
442 
443 	/*
444 	 * Hermon supports shrinking WQEs, such that a single work
445 	 * request can include multiple units of 1 << wqe_shift.  This
446 	 * way, work requests can differ in size, and do not have to
447 	 * be a power of 2 in size, saving memory and speeding up send
448 	 * WR posting.  Unfortunately, if we do this then the
449 	 * wqe_index field in CQEs can't be used to look up the WR ID
450 	 * anymore, so we do this only if selective signaling is off.
451 	 *
452 	 * Further, on 32-bit platforms, we can't use vmap() to make
453 	 * the QP buffer virtually contiguous.  Thus we have to use
454 	 * constant-sized WRs to make sure a WR is always fully within
455 	 * a single page-sized chunk.
456 	 *
457 	 * Finally, we use NOP work requests to pad the end of the
458 	 * work queue, to avoid wrap-around in the middle of WR.  We
459 	 * set NEC bit to avoid getting completions with error for
460 	 * these NOP WRs, but since NEC is only supported starting
461 	 * with firmware 2.2.232, we use constant-sized WRs for older
462 	 * firmware.
463 	 *
464 	 * And, since MLX QPs only support SEND, we use constant-sized
465 	 * WRs in this case.
466 	 *
467 	 * We look for the smallest value of wqe_shift such that the
468 	 * resulting number of wqes does not exceed device
469 	 * capabilities.
470 	 *
471 	 * We set WQE size to at least 64 bytes, this way stamping
472 	 * invalidates each WQE.
473 	 */
474 	if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
475 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
476 	    type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
477 	    !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
478 		      MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
479 		qp->sq.wqe_shift = ilog2(64);
480 	else
481 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
482 
483 	for (;;) {
484 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
485 
486 		/*
487 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
488 		 * allow HW to prefetch.
489 		 */
490 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
491 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
492 						    qp->sq_max_wqes_per_wr +
493 						    qp->sq_spare_wqes);
494 
495 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
496 			break;
497 
498 		if (qp->sq_max_wqes_per_wr <= 1)
499 			return -EINVAL;
500 
501 		++qp->sq.wqe_shift;
502 	}
503 
504 	qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
505 			     (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
506 			 send_wqe_overhead(type, qp->flags)) /
507 		sizeof (struct mlx4_wqe_data_seg);
508 
509 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
510 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
511 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
512 		qp->rq.offset = 0;
513 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
514 	} else {
515 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
516 		qp->sq.offset = 0;
517 	}
518 
519 	cap->max_send_wr  = qp->sq.max_post =
520 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
521 	cap->max_send_sge = min(qp->sq.max_gs,
522 				min(dev->dev->caps.max_sq_sg,
523 				    dev->dev->caps.max_rq_sg));
524 	/* We don't support inline sends for kernel QPs (yet) */
525 	cap->max_inline_data = 0;
526 
527 	return 0;
528 }
529 
530 static int set_user_sq_size(struct mlx4_ib_dev *dev,
531 			    struct mlx4_ib_qp *qp,
532 			    struct mlx4_ib_create_qp *ucmd)
533 {
534 	/* Sanity check SQ size before proceeding */
535 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
536 	    ucmd->log_sq_stride >
537 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
538 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
539 		return -EINVAL;
540 
541 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
542 	qp->sq.wqe_shift = ucmd->log_sq_stride;
543 
544 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
545 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
546 
547 	return 0;
548 }
549 
550 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
551 {
552 	int i;
553 
554 	qp->sqp_proxy_rcv =
555 		kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
556 			GFP_KERNEL);
557 	if (!qp->sqp_proxy_rcv)
558 		return -ENOMEM;
559 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
560 		qp->sqp_proxy_rcv[i].addr =
561 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
562 				GFP_KERNEL);
563 		if (!qp->sqp_proxy_rcv[i].addr)
564 			goto err;
565 		qp->sqp_proxy_rcv[i].map =
566 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
567 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
568 					  DMA_FROM_DEVICE);
569 	}
570 	return 0;
571 
572 err:
573 	while (i > 0) {
574 		--i;
575 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
576 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
577 				    DMA_FROM_DEVICE);
578 		kfree(qp->sqp_proxy_rcv[i].addr);
579 	}
580 	kfree(qp->sqp_proxy_rcv);
581 	qp->sqp_proxy_rcv = NULL;
582 	return -ENOMEM;
583 }
584 
585 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
586 {
587 	int i;
588 
589 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
590 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
591 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
592 				    DMA_FROM_DEVICE);
593 		kfree(qp->sqp_proxy_rcv[i].addr);
594 	}
595 	kfree(qp->sqp_proxy_rcv);
596 }
597 
598 static int qp_has_rq(struct ib_qp_init_attr *attr)
599 {
600 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
601 		return 0;
602 
603 	return !attr->srq;
604 }
605 
606 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
607 {
608 	int i;
609 	for (i = 0; i < dev->caps.num_ports; i++) {
610 		if (qpn == dev->caps.qp0_proxy[i])
611 			return !!dev->caps.qp0_qkey[i];
612 	}
613 	return 0;
614 }
615 
616 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
617 			    struct ib_qp_init_attr *init_attr,
618 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp,
619 			    gfp_t gfp)
620 {
621 	int qpn;
622 	int err;
623 	struct mlx4_ib_sqp *sqp;
624 	struct mlx4_ib_qp *qp;
625 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
626 	struct mlx4_ib_cq *mcq;
627 	unsigned long flags;
628 
629 	/* When tunneling special qps, we use a plain UD qp */
630 	if (sqpn) {
631 		if (mlx4_is_mfunc(dev->dev) &&
632 		    (!mlx4_is_master(dev->dev) ||
633 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
634 			if (init_attr->qp_type == IB_QPT_GSI)
635 				qp_type = MLX4_IB_QPT_PROXY_GSI;
636 			else {
637 				if (mlx4_is_master(dev->dev) ||
638 				    qp0_enabled_vf(dev->dev, sqpn))
639 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
640 				else
641 					qp_type = MLX4_IB_QPT_PROXY_SMI;
642 			}
643 		}
644 		qpn = sqpn;
645 		/* add extra sg entry for tunneling */
646 		init_attr->cap.max_recv_sge++;
647 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
648 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
649 			container_of(init_attr,
650 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
651 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
652 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
653 		    !mlx4_is_master(dev->dev))
654 			return -EINVAL;
655 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
656 			qp_type = MLX4_IB_QPT_TUN_GSI;
657 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
658 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
659 					     tnl_init->port))
660 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
661 		else
662 			qp_type = MLX4_IB_QPT_TUN_SMI;
663 		/* we are definitely in the PPF here, since we are creating
664 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
665 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
666 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
667 		sqpn = qpn;
668 	}
669 
670 	if (!*caller_qp) {
671 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
672 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
673 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
674 			sqp = kzalloc(sizeof (struct mlx4_ib_sqp), gfp);
675 			if (!sqp)
676 				return -ENOMEM;
677 			qp = &sqp->qp;
678 			qp->pri.vid = 0xFFFF;
679 			qp->alt.vid = 0xFFFF;
680 		} else {
681 			qp = kzalloc(sizeof (struct mlx4_ib_qp), gfp);
682 			if (!qp)
683 				return -ENOMEM;
684 			qp->pri.vid = 0xFFFF;
685 			qp->alt.vid = 0xFFFF;
686 		}
687 	} else
688 		qp = *caller_qp;
689 
690 	qp->mlx4_ib_qp_type = qp_type;
691 
692 	mutex_init(&qp->mutex);
693 	spin_lock_init(&qp->sq.lock);
694 	spin_lock_init(&qp->rq.lock);
695 	INIT_LIST_HEAD(&qp->gid_list);
696 	INIT_LIST_HEAD(&qp->steering_rules);
697 
698 	qp->state	 = IB_QPS_RESET;
699 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
700 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
701 
702 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
703 	if (err)
704 		goto err;
705 
706 	if (pd->uobject) {
707 		struct mlx4_ib_create_qp ucmd;
708 
709 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
710 			err = -EFAULT;
711 			goto err;
712 		}
713 
714 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
715 
716 		err = set_user_sq_size(dev, qp, &ucmd);
717 		if (err)
718 			goto err;
719 
720 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
721 				       qp->buf_size, 0, 0);
722 		if (IS_ERR(qp->umem)) {
723 			err = PTR_ERR(qp->umem);
724 			goto err;
725 		}
726 
727 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
728 				    ilog2(qp->umem->page_size), &qp->mtt);
729 		if (err)
730 			goto err_buf;
731 
732 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
733 		if (err)
734 			goto err_mtt;
735 
736 		if (qp_has_rq(init_attr)) {
737 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
738 						  ucmd.db_addr, &qp->db);
739 			if (err)
740 				goto err_mtt;
741 		}
742 	} else {
743 		qp->sq_no_prefetch = 0;
744 
745 		if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
746 			qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
747 
748 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
749 			qp->flags |= MLX4_IB_QP_LSO;
750 
751 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
752 			if (dev->steering_support ==
753 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
754 				qp->flags |= MLX4_IB_QP_NETIF;
755 			else
756 				goto err;
757 		}
758 
759 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
760 		if (err)
761 			goto err;
762 
763 		if (qp_has_rq(init_attr)) {
764 			err = mlx4_db_alloc(dev->dev, &qp->db, 0, gfp);
765 			if (err)
766 				goto err;
767 
768 			*qp->db.db = 0;
769 		}
770 
771 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf, gfp)) {
772 			err = -ENOMEM;
773 			goto err_db;
774 		}
775 
776 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
777 				    &qp->mtt);
778 		if (err)
779 			goto err_buf;
780 
781 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf, gfp);
782 		if (err)
783 			goto err_mtt;
784 
785 		qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), gfp);
786 		qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), gfp);
787 		if (!qp->sq.wrid || !qp->rq.wrid) {
788 			err = -ENOMEM;
789 			goto err_wrid;
790 		}
791 	}
792 
793 	if (sqpn) {
794 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
795 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
796 			if (alloc_proxy_bufs(pd->device, qp)) {
797 				err = -ENOMEM;
798 				goto err_wrid;
799 			}
800 		}
801 	} else {
802 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
803 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
804 		 * VLAN insertion. */
805 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
806 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
807 						    (init_attr->cap.max_send_wr ?
808 						     MLX4_RESERVE_ETH_BF_QP : 0) |
809 						    (init_attr->cap.max_recv_wr ?
810 						     MLX4_RESERVE_A0_QP : 0));
811 		else
812 			if (qp->flags & MLX4_IB_QP_NETIF)
813 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
814 			else
815 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
816 							    &qpn, 0);
817 		if (err)
818 			goto err_proxy;
819 	}
820 
821 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp, gfp);
822 	if (err)
823 		goto err_qpn;
824 
825 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
826 		qp->mqp.qpn |= (1 << 23);
827 
828 	/*
829 	 * Hardware wants QPN written in big-endian order (after
830 	 * shifting) for send doorbell.  Precompute this value to save
831 	 * a little bit when posting sends.
832 	 */
833 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
834 
835 	qp->mqp.event = mlx4_ib_qp_event;
836 	if (!*caller_qp)
837 		*caller_qp = qp;
838 
839 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
840 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
841 			 to_mcq(init_attr->recv_cq));
842 	/* Maintain device to QPs access, needed for further handling
843 	 * via reset flow
844 	 */
845 	list_add_tail(&qp->qps_list, &dev->qp_list);
846 	/* Maintain CQ to QPs access, needed for further handling
847 	 * via reset flow
848 	 */
849 	mcq = to_mcq(init_attr->send_cq);
850 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
851 	mcq = to_mcq(init_attr->recv_cq);
852 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
853 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
854 			   to_mcq(init_attr->recv_cq));
855 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
856 	return 0;
857 
858 err_qpn:
859 	if (!sqpn) {
860 		if (qp->flags & MLX4_IB_QP_NETIF)
861 			mlx4_ib_steer_qp_free(dev, qpn, 1);
862 		else
863 			mlx4_qp_release_range(dev->dev, qpn, 1);
864 	}
865 err_proxy:
866 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
867 		free_proxy_bufs(pd->device, qp);
868 err_wrid:
869 	if (pd->uobject) {
870 		if (qp_has_rq(init_attr))
871 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
872 	} else {
873 		kfree(qp->sq.wrid);
874 		kfree(qp->rq.wrid);
875 	}
876 
877 err_mtt:
878 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
879 
880 err_buf:
881 	if (pd->uobject)
882 		ib_umem_release(qp->umem);
883 	else
884 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
885 
886 err_db:
887 	if (!pd->uobject && qp_has_rq(init_attr))
888 		mlx4_db_free(dev->dev, &qp->db);
889 
890 err:
891 	if (!*caller_qp)
892 		kfree(qp);
893 	return err;
894 }
895 
896 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
897 {
898 	switch (state) {
899 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
900 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
901 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
902 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
903 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
904 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
905 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
906 	default:		return -1;
907 	}
908 }
909 
910 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
911 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
912 {
913 	if (send_cq == recv_cq) {
914 		spin_lock(&send_cq->lock);
915 		__acquire(&recv_cq->lock);
916 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
917 		spin_lock(&send_cq->lock);
918 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
919 	} else {
920 		spin_lock(&recv_cq->lock);
921 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
922 	}
923 }
924 
925 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
926 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
927 {
928 	if (send_cq == recv_cq) {
929 		__release(&recv_cq->lock);
930 		spin_unlock(&send_cq->lock);
931 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
932 		spin_unlock(&recv_cq->lock);
933 		spin_unlock(&send_cq->lock);
934 	} else {
935 		spin_unlock(&send_cq->lock);
936 		spin_unlock(&recv_cq->lock);
937 	}
938 }
939 
940 static void del_gid_entries(struct mlx4_ib_qp *qp)
941 {
942 	struct mlx4_ib_gid_entry *ge, *tmp;
943 
944 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
945 		list_del(&ge->list);
946 		kfree(ge);
947 	}
948 }
949 
950 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
951 {
952 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
953 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
954 	else
955 		return to_mpd(qp->ibqp.pd);
956 }
957 
958 static void get_cqs(struct mlx4_ib_qp *qp,
959 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
960 {
961 	switch (qp->ibqp.qp_type) {
962 	case IB_QPT_XRC_TGT:
963 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
964 		*recv_cq = *send_cq;
965 		break;
966 	case IB_QPT_XRC_INI:
967 		*send_cq = to_mcq(qp->ibqp.send_cq);
968 		*recv_cq = *send_cq;
969 		break;
970 	default:
971 		*send_cq = to_mcq(qp->ibqp.send_cq);
972 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
973 		break;
974 	}
975 }
976 
977 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
978 			      int is_user)
979 {
980 	struct mlx4_ib_cq *send_cq, *recv_cq;
981 	unsigned long flags;
982 
983 	if (qp->state != IB_QPS_RESET) {
984 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
985 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
986 			pr_warn("modify QP %06x to RESET failed.\n",
987 			       qp->mqp.qpn);
988 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
989 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
990 			qp->pri.smac = 0;
991 			qp->pri.smac_port = 0;
992 		}
993 		if (qp->alt.smac) {
994 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
995 			qp->alt.smac = 0;
996 		}
997 		if (qp->pri.vid < 0x1000) {
998 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
999 			qp->pri.vid = 0xFFFF;
1000 			qp->pri.candidate_vid = 0xFFFF;
1001 			qp->pri.update_vid = 0;
1002 		}
1003 		if (qp->alt.vid < 0x1000) {
1004 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1005 			qp->alt.vid = 0xFFFF;
1006 			qp->alt.candidate_vid = 0xFFFF;
1007 			qp->alt.update_vid = 0;
1008 		}
1009 	}
1010 
1011 	get_cqs(qp, &send_cq, &recv_cq);
1012 
1013 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1014 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1015 
1016 	/* del from lists under both locks above to protect reset flow paths */
1017 	list_del(&qp->qps_list);
1018 	list_del(&qp->cq_send_list);
1019 	list_del(&qp->cq_recv_list);
1020 	if (!is_user) {
1021 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1022 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1023 		if (send_cq != recv_cq)
1024 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1025 	}
1026 
1027 	mlx4_qp_remove(dev->dev, &qp->mqp);
1028 
1029 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1030 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1031 
1032 	mlx4_qp_free(dev->dev, &qp->mqp);
1033 
1034 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1035 		if (qp->flags & MLX4_IB_QP_NETIF)
1036 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1037 		else
1038 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1039 	}
1040 
1041 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1042 
1043 	if (is_user) {
1044 		if (qp->rq.wqe_cnt)
1045 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
1046 					      &qp->db);
1047 		ib_umem_release(qp->umem);
1048 	} else {
1049 		kfree(qp->sq.wrid);
1050 		kfree(qp->rq.wrid);
1051 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1052 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1053 			free_proxy_bufs(&dev->ib_dev, qp);
1054 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1055 		if (qp->rq.wqe_cnt)
1056 			mlx4_db_free(dev->dev, &qp->db);
1057 	}
1058 
1059 	del_gid_entries(qp);
1060 }
1061 
1062 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1063 {
1064 	/* Native or PPF */
1065 	if (!mlx4_is_mfunc(dev->dev) ||
1066 	    (mlx4_is_master(dev->dev) &&
1067 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1068 		return  dev->dev->phys_caps.base_sqpn +
1069 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1070 			attr->port_num - 1;
1071 	}
1072 	/* PF or VF -- creating proxies */
1073 	if (attr->qp_type == IB_QPT_SMI)
1074 		return dev->dev->caps.qp0_proxy[attr->port_num - 1];
1075 	else
1076 		return dev->dev->caps.qp1_proxy[attr->port_num - 1];
1077 }
1078 
1079 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1080 				struct ib_qp_init_attr *init_attr,
1081 				struct ib_udata *udata)
1082 {
1083 	struct mlx4_ib_qp *qp = NULL;
1084 	int err;
1085 	u16 xrcdn = 0;
1086 	gfp_t gfp;
1087 
1088 	gfp = (init_attr->create_flags & MLX4_IB_QP_CREATE_USE_GFP_NOIO) ?
1089 		GFP_NOIO : GFP_KERNEL;
1090 	/*
1091 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1092 	 * and only for kernel UD QPs.
1093 	 */
1094 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1095 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1096 					MLX4_IB_SRIOV_TUNNEL_QP |
1097 					MLX4_IB_SRIOV_SQP |
1098 					MLX4_IB_QP_NETIF |
1099 					MLX4_IB_QP_CREATE_USE_GFP_NOIO))
1100 		return ERR_PTR(-EINVAL);
1101 
1102 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1103 		if (init_attr->qp_type != IB_QPT_UD)
1104 			return ERR_PTR(-EINVAL);
1105 	}
1106 
1107 	if (init_attr->create_flags &&
1108 	    (udata ||
1109 	     ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP | MLX4_IB_QP_CREATE_USE_GFP_NOIO)) &&
1110 	      init_attr->qp_type != IB_QPT_UD) ||
1111 	     ((init_attr->create_flags & MLX4_IB_SRIOV_SQP) &&
1112 	      init_attr->qp_type > IB_QPT_GSI)))
1113 		return ERR_PTR(-EINVAL);
1114 
1115 	switch (init_attr->qp_type) {
1116 	case IB_QPT_XRC_TGT:
1117 		pd = to_mxrcd(init_attr->xrcd)->pd;
1118 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1119 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1120 		/* fall through */
1121 	case IB_QPT_XRC_INI:
1122 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1123 			return ERR_PTR(-ENOSYS);
1124 		init_attr->recv_cq = init_attr->send_cq;
1125 		/* fall through */
1126 	case IB_QPT_RC:
1127 	case IB_QPT_UC:
1128 	case IB_QPT_RAW_PACKET:
1129 		qp = kzalloc(sizeof *qp, gfp);
1130 		if (!qp)
1131 			return ERR_PTR(-ENOMEM);
1132 		qp->pri.vid = 0xFFFF;
1133 		qp->alt.vid = 0xFFFF;
1134 		/* fall through */
1135 	case IB_QPT_UD:
1136 	{
1137 		err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1138 				       udata, 0, &qp, gfp);
1139 		if (err)
1140 			return ERR_PTR(err);
1141 
1142 		qp->ibqp.qp_num = qp->mqp.qpn;
1143 		qp->xrcdn = xrcdn;
1144 
1145 		break;
1146 	}
1147 	case IB_QPT_SMI:
1148 	case IB_QPT_GSI:
1149 	{
1150 		/* Userspace is not allowed to create special QPs: */
1151 		if (udata)
1152 			return ERR_PTR(-EINVAL);
1153 
1154 		err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1155 				       get_sqp_num(to_mdev(pd->device), init_attr),
1156 				       &qp, gfp);
1157 		if (err)
1158 			return ERR_PTR(err);
1159 
1160 		qp->port	= init_attr->port_num;
1161 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
1162 
1163 		break;
1164 	}
1165 	default:
1166 		/* Don't support raw QPs */
1167 		return ERR_PTR(-EINVAL);
1168 	}
1169 
1170 	return &qp->ibqp;
1171 }
1172 
1173 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1174 {
1175 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1176 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1177 	struct mlx4_ib_pd *pd;
1178 
1179 	if (is_qp0(dev, mqp))
1180 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1181 
1182 	if (dev->qp1_proxy[mqp->port - 1] == mqp) {
1183 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1184 		dev->qp1_proxy[mqp->port - 1] = NULL;
1185 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1186 	}
1187 
1188 	pd = get_pd(mqp);
1189 	destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1190 
1191 	if (is_sqp(dev, mqp))
1192 		kfree(to_msqp(mqp));
1193 	else
1194 		kfree(mqp);
1195 
1196 	return 0;
1197 }
1198 
1199 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1200 {
1201 	switch (type) {
1202 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1203 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1204 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1205 	case MLX4_IB_QPT_XRC_INI:
1206 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1207 	case MLX4_IB_QPT_SMI:
1208 	case MLX4_IB_QPT_GSI:
1209 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1210 
1211 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1212 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1213 						MLX4_QP_ST_MLX : -1);
1214 	case MLX4_IB_QPT_PROXY_SMI:
1215 	case MLX4_IB_QPT_TUN_SMI:
1216 	case MLX4_IB_QPT_PROXY_GSI:
1217 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1218 						MLX4_QP_ST_UD : -1);
1219 	default:			return -1;
1220 	}
1221 }
1222 
1223 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1224 				   int attr_mask)
1225 {
1226 	u8 dest_rd_atomic;
1227 	u32 access_flags;
1228 	u32 hw_access_flags = 0;
1229 
1230 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1231 		dest_rd_atomic = attr->max_dest_rd_atomic;
1232 	else
1233 		dest_rd_atomic = qp->resp_depth;
1234 
1235 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1236 		access_flags = attr->qp_access_flags;
1237 	else
1238 		access_flags = qp->atomic_rd_en;
1239 
1240 	if (!dest_rd_atomic)
1241 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1242 
1243 	if (access_flags & IB_ACCESS_REMOTE_READ)
1244 		hw_access_flags |= MLX4_QP_BIT_RRE;
1245 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1246 		hw_access_flags |= MLX4_QP_BIT_RAE;
1247 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1248 		hw_access_flags |= MLX4_QP_BIT_RWE;
1249 
1250 	return cpu_to_be32(hw_access_flags);
1251 }
1252 
1253 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1254 			    int attr_mask)
1255 {
1256 	if (attr_mask & IB_QP_PKEY_INDEX)
1257 		sqp->pkey_index = attr->pkey_index;
1258 	if (attr_mask & IB_QP_QKEY)
1259 		sqp->qkey = attr->qkey;
1260 	if (attr_mask & IB_QP_SQ_PSN)
1261 		sqp->send_psn = attr->sq_psn;
1262 }
1263 
1264 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1265 {
1266 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1267 }
1268 
1269 static int _mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
1270 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1271 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1272 {
1273 	int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
1274 		IB_LINK_LAYER_ETHERNET;
1275 	int vidx;
1276 	int smac_index;
1277 	int err;
1278 
1279 
1280 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
1281 	path->rlid	    = cpu_to_be16(ah->dlid);
1282 	if (ah->static_rate) {
1283 		path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
1284 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1285 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1286 			--path->static_rate;
1287 	} else
1288 		path->static_rate = 0;
1289 
1290 	if (ah->ah_flags & IB_AH_GRH) {
1291 		if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
1292 			pr_err("sgid_index (%u) too large. max is %d\n",
1293 			       ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1294 			return -1;
1295 		}
1296 
1297 		path->grh_mylmc |= 1 << 7;
1298 		path->mgid_index = ah->grh.sgid_index;
1299 		path->hop_limit  = ah->grh.hop_limit;
1300 		path->tclass_flowlabel =
1301 			cpu_to_be32((ah->grh.traffic_class << 20) |
1302 				    (ah->grh.flow_label));
1303 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1304 	}
1305 
1306 	if (is_eth) {
1307 		if (!(ah->ah_flags & IB_AH_GRH))
1308 			return -1;
1309 
1310 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1311 			((port - 1) << 6) | ((ah->sl & 7) << 3);
1312 
1313 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1314 		if (vlan_tag < 0x1000) {
1315 			if (smac_info->vid < 0x1000) {
1316 				/* both valid vlan ids */
1317 				if (smac_info->vid != vlan_tag) {
1318 					/* different VIDs.  unreg old and reg new */
1319 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1320 					if (err)
1321 						return err;
1322 					smac_info->candidate_vid = vlan_tag;
1323 					smac_info->candidate_vlan_index = vidx;
1324 					smac_info->candidate_vlan_port = port;
1325 					smac_info->update_vid = 1;
1326 					path->vlan_index = vidx;
1327 				} else {
1328 					path->vlan_index = smac_info->vlan_index;
1329 				}
1330 			} else {
1331 				/* no current vlan tag in qp */
1332 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1333 				if (err)
1334 					return err;
1335 				smac_info->candidate_vid = vlan_tag;
1336 				smac_info->candidate_vlan_index = vidx;
1337 				smac_info->candidate_vlan_port = port;
1338 				smac_info->update_vid = 1;
1339 				path->vlan_index = vidx;
1340 			}
1341 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1342 			path->fl = 1 << 6;
1343 		} else {
1344 			/* have current vlan tag. unregister it at modify-qp success */
1345 			if (smac_info->vid < 0x1000) {
1346 				smac_info->candidate_vid = 0xFFFF;
1347 				smac_info->update_vid = 1;
1348 			}
1349 		}
1350 
1351 		/* get smac_index for RoCE use.
1352 		 * If no smac was yet assigned, register one.
1353 		 * If one was already assigned, but the new mac differs,
1354 		 * unregister the old one and register the new one.
1355 		*/
1356 		if ((!smac_info->smac && !smac_info->smac_port) ||
1357 		    smac_info->smac != smac) {
1358 			/* register candidate now, unreg if needed, after success */
1359 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1360 			if (smac_index >= 0) {
1361 				smac_info->candidate_smac_index = smac_index;
1362 				smac_info->candidate_smac = smac;
1363 				smac_info->candidate_smac_port = port;
1364 			} else {
1365 				return -EINVAL;
1366 			}
1367 		} else {
1368 			smac_index = smac_info->smac_index;
1369 		}
1370 
1371 		memcpy(path->dmac, ah->dmac, 6);
1372 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1373 		/* put MAC table smac index for IBoE */
1374 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1375 	} else {
1376 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1377 			((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1378 	}
1379 
1380 	return 0;
1381 }
1382 
1383 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1384 			 enum ib_qp_attr_mask qp_attr_mask,
1385 			 struct mlx4_ib_qp *mqp,
1386 			 struct mlx4_qp_path *path, u8 port)
1387 {
1388 	return _mlx4_set_path(dev, &qp->ah_attr,
1389 			      mlx4_mac_to_u64((u8 *)qp->smac),
1390 			      (qp_attr_mask & IB_QP_VID) ? qp->vlan_id : 0xffff,
1391 			      path, &mqp->pri, port);
1392 }
1393 
1394 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1395 			     const struct ib_qp_attr *qp,
1396 			     enum ib_qp_attr_mask qp_attr_mask,
1397 			     struct mlx4_ib_qp *mqp,
1398 			     struct mlx4_qp_path *path, u8 port)
1399 {
1400 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1401 			      mlx4_mac_to_u64((u8 *)qp->alt_smac),
1402 			      (qp_attr_mask & IB_QP_ALT_VID) ?
1403 			      qp->alt_vlan_id : 0xffff,
1404 			      path, &mqp->alt, port);
1405 }
1406 
1407 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1408 {
1409 	struct mlx4_ib_gid_entry *ge, *tmp;
1410 
1411 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1412 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1413 			ge->added = 1;
1414 			ge->port = qp->port;
1415 		}
1416 	}
1417 }
1418 
1419 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp, u8 *smac,
1420 				    struct mlx4_qp_context *context)
1421 {
1422 	u64 u64_mac;
1423 	int smac_index;
1424 
1425 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1426 
1427 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1428 	if (!qp->pri.smac && !qp->pri.smac_port) {
1429 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1430 		if (smac_index >= 0) {
1431 			qp->pri.candidate_smac_index = smac_index;
1432 			qp->pri.candidate_smac = u64_mac;
1433 			qp->pri.candidate_smac_port = qp->port;
1434 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1435 		} else {
1436 			return -ENOENT;
1437 		}
1438 	}
1439 	return 0;
1440 }
1441 
1442 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1443 			       const struct ib_qp_attr *attr, int attr_mask,
1444 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1445 {
1446 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1447 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1448 	struct mlx4_ib_pd *pd;
1449 	struct mlx4_ib_cq *send_cq, *recv_cq;
1450 	struct mlx4_qp_context *context;
1451 	enum mlx4_qp_optpar optpar = 0;
1452 	int sqd_event;
1453 	int steer_qp = 0;
1454 	int err = -EINVAL;
1455 
1456 	/* APM is not supported under RoCE */
1457 	if (attr_mask & IB_QP_ALT_PATH &&
1458 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1459 	    IB_LINK_LAYER_ETHERNET)
1460 		return -ENOTSUPP;
1461 
1462 	context = kzalloc(sizeof *context, GFP_KERNEL);
1463 	if (!context)
1464 		return -ENOMEM;
1465 
1466 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1467 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1468 
1469 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1470 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1471 	else {
1472 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
1473 		switch (attr->path_mig_state) {
1474 		case IB_MIG_MIGRATED:
1475 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1476 			break;
1477 		case IB_MIG_REARM:
1478 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1479 			break;
1480 		case IB_MIG_ARMED:
1481 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1482 			break;
1483 		}
1484 	}
1485 
1486 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1487 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1488 	else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1489 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1490 	else if (ibqp->qp_type == IB_QPT_UD) {
1491 		if (qp->flags & MLX4_IB_QP_LSO)
1492 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
1493 					      ilog2(dev->dev->caps.max_gso_sz);
1494 		else
1495 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1496 	} else if (attr_mask & IB_QP_PATH_MTU) {
1497 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1498 			pr_err("path MTU (%u) is invalid\n",
1499 			       attr->path_mtu);
1500 			goto out;
1501 		}
1502 		context->mtu_msgmax = (attr->path_mtu << 5) |
1503 			ilog2(dev->dev->caps.max_msg_sz);
1504 	}
1505 
1506 	if (qp->rq.wqe_cnt)
1507 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1508 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
1509 
1510 	if (qp->sq.wqe_cnt)
1511 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1512 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
1513 
1514 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1515 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1516 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1517 		if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1518 			context->param3 |= cpu_to_be32(1 << 30);
1519 	}
1520 
1521 	if (qp->ibqp.uobject)
1522 		context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
1523 	else
1524 		context->usr_page = cpu_to_be32(dev->priv_uar.index);
1525 
1526 	if (attr_mask & IB_QP_DEST_QPN)
1527 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1528 
1529 	if (attr_mask & IB_QP_PORT) {
1530 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1531 		    !(attr_mask & IB_QP_AV)) {
1532 			mlx4_set_sched(&context->pri_path, attr->port_num);
1533 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1534 		}
1535 	}
1536 
1537 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1538 		if (dev->counters[qp->port - 1] != -1) {
1539 			context->pri_path.counter_index =
1540 						dev->counters[qp->port - 1];
1541 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1542 		} else
1543 			context->pri_path.counter_index = 0xff;
1544 
1545 		if (qp->flags & MLX4_IB_QP_NETIF) {
1546 			mlx4_ib_steer_qp_reg(dev, qp, 1);
1547 			steer_qp = 1;
1548 		}
1549 	}
1550 
1551 	if (attr_mask & IB_QP_PKEY_INDEX) {
1552 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1553 			context->pri_path.disable_pkey_check = 0x40;
1554 		context->pri_path.pkey_index = attr->pkey_index;
1555 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1556 	}
1557 
1558 	if (attr_mask & IB_QP_AV) {
1559 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
1560 				  attr_mask & IB_QP_PORT ?
1561 				  attr->port_num : qp->port))
1562 			goto out;
1563 
1564 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1565 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
1566 	}
1567 
1568 	if (attr_mask & IB_QP_TIMEOUT) {
1569 		context->pri_path.ackto |= attr->timeout << 3;
1570 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1571 	}
1572 
1573 	if (attr_mask & IB_QP_ALT_PATH) {
1574 		if (attr->alt_port_num == 0 ||
1575 		    attr->alt_port_num > dev->dev->caps.num_ports)
1576 			goto out;
1577 
1578 		if (attr->alt_pkey_index >=
1579 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
1580 			goto out;
1581 
1582 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
1583 				      &context->alt_path,
1584 				      attr->alt_port_num))
1585 			goto out;
1586 
1587 		context->alt_path.pkey_index = attr->alt_pkey_index;
1588 		context->alt_path.ackto = attr->alt_timeout << 3;
1589 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1590 	}
1591 
1592 	pd = get_pd(qp);
1593 	get_cqs(qp, &send_cq, &recv_cq);
1594 	context->pd       = cpu_to_be32(pd->pdn);
1595 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1596 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1597 	context->params1  = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1598 
1599 	/* Set "fast registration enabled" for all kernel QPs */
1600 	if (!qp->ibqp.uobject)
1601 		context->params1 |= cpu_to_be32(1 << 11);
1602 
1603 	if (attr_mask & IB_QP_RNR_RETRY) {
1604 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1605 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1606 	}
1607 
1608 	if (attr_mask & IB_QP_RETRY_CNT) {
1609 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1610 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1611 	}
1612 
1613 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1614 		if (attr->max_rd_atomic)
1615 			context->params1 |=
1616 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1617 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1618 	}
1619 
1620 	if (attr_mask & IB_QP_SQ_PSN)
1621 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1622 
1623 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1624 		if (attr->max_dest_rd_atomic)
1625 			context->params2 |=
1626 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1627 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1628 	}
1629 
1630 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1631 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1632 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1633 	}
1634 
1635 	if (ibqp->srq)
1636 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1637 
1638 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1639 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1640 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1641 	}
1642 	if (attr_mask & IB_QP_RQ_PSN)
1643 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1644 
1645 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1646 	if (attr_mask & IB_QP_QKEY) {
1647 		if (qp->mlx4_ib_qp_type &
1648 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1649 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1650 		else {
1651 			if (mlx4_is_mfunc(dev->dev) &&
1652 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1653 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1654 			    MLX4_RESERVED_QKEY_BASE) {
1655 				pr_err("Cannot use reserved QKEY"
1656 				       " 0x%x (range 0xffff0000..0xffffffff"
1657 				       " is reserved)\n", attr->qkey);
1658 				err = -EINVAL;
1659 				goto out;
1660 			}
1661 			context->qkey = cpu_to_be32(attr->qkey);
1662 		}
1663 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
1664 	}
1665 
1666 	if (ibqp->srq)
1667 		context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1668 
1669 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1670 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1671 
1672 	if (cur_state == IB_QPS_INIT &&
1673 	    new_state == IB_QPS_RTR  &&
1674 	    (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1675 	     ibqp->qp_type == IB_QPT_UD ||
1676 	     ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1677 		context->pri_path.sched_queue = (qp->port - 1) << 6;
1678 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1679 		    qp->mlx4_ib_qp_type &
1680 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1681 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1682 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1683 				context->pri_path.fl = 0x80;
1684 		} else {
1685 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1686 				context->pri_path.fl = 0x80;
1687 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1688 		}
1689 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1690 		    IB_LINK_LAYER_ETHERNET) {
1691 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
1692 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
1693 				context->pri_path.feup = 1 << 7; /* don't fsm */
1694 			/* handle smac_index */
1695 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
1696 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
1697 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
1698 				err = handle_eth_ud_smac_index(dev, qp, (u8 *)attr->smac, context);
1699 				if (err) {
1700 					err = -EINVAL;
1701 					goto out;
1702 				}
1703 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1704 					dev->qp1_proxy[qp->port - 1] = qp;
1705 			}
1706 		}
1707 	}
1708 
1709 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1710 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
1711 					MLX4_IB_LINK_TYPE_ETH;
1712 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1713 			/* set QP to receive both tunneled & non-tunneled packets */
1714 			if (!(context->flags & cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET)))
1715 				context->srqn = cpu_to_be32(7 << 28);
1716 		}
1717 	}
1718 
1719 	if (ibqp->qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
1720 		int is_eth = rdma_port_get_link_layer(
1721 				&dev->ib_dev, qp->port) ==
1722 				IB_LINK_LAYER_ETHERNET;
1723 		if (is_eth) {
1724 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
1725 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
1726 		}
1727 	}
1728 
1729 
1730 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1731 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1732 		sqd_event = 1;
1733 	else
1734 		sqd_event = 0;
1735 
1736 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1737 		context->rlkey |= (1 << 4);
1738 
1739 	/*
1740 	 * Before passing a kernel QP to the HW, make sure that the
1741 	 * ownership bits of the send queue are set and the SQ
1742 	 * headroom is stamped so that the hardware doesn't start
1743 	 * processing stale work requests.
1744 	 */
1745 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1746 		struct mlx4_wqe_ctrl_seg *ctrl;
1747 		int i;
1748 
1749 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1750 			ctrl = get_send_wqe(qp, i);
1751 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
1752 			if (qp->sq_max_wqes_per_wr == 1)
1753 				ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1754 
1755 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1756 		}
1757 	}
1758 
1759 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1760 			     to_mlx4_state(new_state), context, optpar,
1761 			     sqd_event, &qp->mqp);
1762 	if (err)
1763 		goto out;
1764 
1765 	qp->state = new_state;
1766 
1767 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1768 		qp->atomic_rd_en = attr->qp_access_flags;
1769 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1770 		qp->resp_depth = attr->max_dest_rd_atomic;
1771 	if (attr_mask & IB_QP_PORT) {
1772 		qp->port = attr->port_num;
1773 		update_mcg_macs(dev, qp);
1774 	}
1775 	if (attr_mask & IB_QP_ALT_PATH)
1776 		qp->alt_port = attr->alt_port_num;
1777 
1778 	if (is_sqp(dev, qp))
1779 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1780 
1781 	/*
1782 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
1783 	 * QP0 to RESET or ERROR, bring the link back down.
1784 	 */
1785 	if (is_qp0(dev, qp)) {
1786 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1787 			if (mlx4_INIT_PORT(dev->dev, qp->port))
1788 				pr_warn("INIT_PORT failed for port %d\n",
1789 				       qp->port);
1790 
1791 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1792 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1793 			mlx4_CLOSE_PORT(dev->dev, qp->port);
1794 	}
1795 
1796 	/*
1797 	 * If we moved a kernel QP to RESET, clean up all old CQ
1798 	 * entries and reinitialize the QP.
1799 	 */
1800 	if (new_state == IB_QPS_RESET) {
1801 		if (!ibqp->uobject) {
1802 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1803 					 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
1804 			if (send_cq != recv_cq)
1805 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1806 
1807 			qp->rq.head = 0;
1808 			qp->rq.tail = 0;
1809 			qp->sq.head = 0;
1810 			qp->sq.tail = 0;
1811 			qp->sq_next_wqe = 0;
1812 			if (qp->rq.wqe_cnt)
1813 				*qp->db.db  = 0;
1814 
1815 			if (qp->flags & MLX4_IB_QP_NETIF)
1816 				mlx4_ib_steer_qp_reg(dev, qp, 0);
1817 		}
1818 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1819 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1820 			qp->pri.smac = 0;
1821 			qp->pri.smac_port = 0;
1822 		}
1823 		if (qp->alt.smac) {
1824 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1825 			qp->alt.smac = 0;
1826 		}
1827 		if (qp->pri.vid < 0x1000) {
1828 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1829 			qp->pri.vid = 0xFFFF;
1830 			qp->pri.candidate_vid = 0xFFFF;
1831 			qp->pri.update_vid = 0;
1832 		}
1833 
1834 		if (qp->alt.vid < 0x1000) {
1835 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1836 			qp->alt.vid = 0xFFFF;
1837 			qp->alt.candidate_vid = 0xFFFF;
1838 			qp->alt.update_vid = 0;
1839 		}
1840 	}
1841 out:
1842 	if (err && steer_qp)
1843 		mlx4_ib_steer_qp_reg(dev, qp, 0);
1844 	kfree(context);
1845 	if (qp->pri.candidate_smac ||
1846 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
1847 		if (err) {
1848 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
1849 		} else {
1850 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
1851 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1852 			qp->pri.smac = qp->pri.candidate_smac;
1853 			qp->pri.smac_index = qp->pri.candidate_smac_index;
1854 			qp->pri.smac_port = qp->pri.candidate_smac_port;
1855 		}
1856 		qp->pri.candidate_smac = 0;
1857 		qp->pri.candidate_smac_index = 0;
1858 		qp->pri.candidate_smac_port = 0;
1859 	}
1860 	if (qp->alt.candidate_smac) {
1861 		if (err) {
1862 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
1863 		} else {
1864 			if (qp->alt.smac)
1865 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1866 			qp->alt.smac = qp->alt.candidate_smac;
1867 			qp->alt.smac_index = qp->alt.candidate_smac_index;
1868 			qp->alt.smac_port = qp->alt.candidate_smac_port;
1869 		}
1870 		qp->alt.candidate_smac = 0;
1871 		qp->alt.candidate_smac_index = 0;
1872 		qp->alt.candidate_smac_port = 0;
1873 	}
1874 
1875 	if (qp->pri.update_vid) {
1876 		if (err) {
1877 			if (qp->pri.candidate_vid < 0x1000)
1878 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
1879 						     qp->pri.candidate_vid);
1880 		} else {
1881 			if (qp->pri.vid < 0x1000)
1882 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
1883 						     qp->pri.vid);
1884 			qp->pri.vid = qp->pri.candidate_vid;
1885 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
1886 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
1887 		}
1888 		qp->pri.candidate_vid = 0xFFFF;
1889 		qp->pri.update_vid = 0;
1890 	}
1891 
1892 	if (qp->alt.update_vid) {
1893 		if (err) {
1894 			if (qp->alt.candidate_vid < 0x1000)
1895 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
1896 						     qp->alt.candidate_vid);
1897 		} else {
1898 			if (qp->alt.vid < 0x1000)
1899 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
1900 						     qp->alt.vid);
1901 			qp->alt.vid = qp->alt.candidate_vid;
1902 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
1903 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
1904 		}
1905 		qp->alt.candidate_vid = 0xFFFF;
1906 		qp->alt.update_vid = 0;
1907 	}
1908 
1909 	return err;
1910 }
1911 
1912 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1913 		      int attr_mask, struct ib_udata *udata)
1914 {
1915 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1916 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1917 	enum ib_qp_state cur_state, new_state;
1918 	int err = -EINVAL;
1919 	int ll;
1920 	mutex_lock(&qp->mutex);
1921 
1922 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1923 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1924 
1925 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1926 		ll = IB_LINK_LAYER_UNSPECIFIED;
1927 	} else {
1928 		int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1929 		ll = rdma_port_get_link_layer(&dev->ib_dev, port);
1930 	}
1931 
1932 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1933 				attr_mask, ll)) {
1934 		pr_debug("qpn 0x%x: invalid attribute mask specified "
1935 			 "for transition %d to %d. qp_type %d,"
1936 			 " attr_mask 0x%x\n",
1937 			 ibqp->qp_num, cur_state, new_state,
1938 			 ibqp->qp_type, attr_mask);
1939 		goto out;
1940 	}
1941 
1942 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
1943 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
1944 			if ((ibqp->qp_type == IB_QPT_RC) ||
1945 			    (ibqp->qp_type == IB_QPT_UD) ||
1946 			    (ibqp->qp_type == IB_QPT_UC) ||
1947 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
1948 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
1949 				attr->port_num = mlx4_ib_bond_next_port(dev);
1950 			}
1951 		} else {
1952 			/* no sense in changing port_num
1953 			 * when ports are bonded */
1954 			attr_mask &= ~IB_QP_PORT;
1955 		}
1956 	}
1957 
1958 	if ((attr_mask & IB_QP_PORT) &&
1959 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
1960 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
1961 			 "for transition %d to %d. qp_type %d\n",
1962 			 ibqp->qp_num, attr->port_num, cur_state,
1963 			 new_state, ibqp->qp_type);
1964 		goto out;
1965 	}
1966 
1967 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
1968 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
1969 	     IB_LINK_LAYER_ETHERNET))
1970 		goto out;
1971 
1972 	if (attr_mask & IB_QP_PKEY_INDEX) {
1973 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1974 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
1975 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
1976 				 "for transition %d to %d. qp_type %d\n",
1977 				 ibqp->qp_num, attr->pkey_index, cur_state,
1978 				 new_state, ibqp->qp_type);
1979 			goto out;
1980 		}
1981 	}
1982 
1983 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1984 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
1985 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
1986 			 "Transition %d to %d. qp_type %d\n",
1987 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
1988 			 new_state, ibqp->qp_type);
1989 		goto out;
1990 	}
1991 
1992 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1993 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
1994 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
1995 			 "Transition %d to %d. qp_type %d\n",
1996 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
1997 			 new_state, ibqp->qp_type);
1998 		goto out;
1999 	}
2000 
2001 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2002 		err = 0;
2003 		goto out;
2004 	}
2005 
2006 	err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2007 
2008 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2009 		attr->port_num = 1;
2010 
2011 out:
2012 	mutex_unlock(&qp->mutex);
2013 	return err;
2014 }
2015 
2016 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2017 {
2018 	int i;
2019 	for (i = 0; i < dev->caps.num_ports; i++) {
2020 		if (qpn == dev->caps.qp0_proxy[i] ||
2021 		    qpn == dev->caps.qp0_tunnel[i]) {
2022 			*qkey = dev->caps.qp0_qkey[i];
2023 			return 0;
2024 		}
2025 	}
2026 	return -EINVAL;
2027 }
2028 
2029 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2030 				  struct ib_send_wr *wr,
2031 				  void *wqe, unsigned *mlx_seg_len)
2032 {
2033 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2034 	struct ib_device *ib_dev = &mdev->ib_dev;
2035 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2036 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2037 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2038 	u16 pkey;
2039 	u32 qkey;
2040 	int send_size;
2041 	int header_size;
2042 	int spc;
2043 	int i;
2044 
2045 	if (wr->opcode != IB_WR_SEND)
2046 		return -EINVAL;
2047 
2048 	send_size = 0;
2049 
2050 	for (i = 0; i < wr->num_sge; ++i)
2051 		send_size += wr->sg_list[i].length;
2052 
2053 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2054 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2055 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2056 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2057 
2058 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, &sqp->ud_header);
2059 
2060 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2061 		sqp->ud_header.lrh.service_level =
2062 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2063 		sqp->ud_header.lrh.destination_lid =
2064 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2065 		sqp->ud_header.lrh.source_lid =
2066 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2067 	}
2068 
2069 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2070 
2071 	/* force loopback */
2072 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2073 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2074 
2075 	sqp->ud_header.lrh.virtual_lane    = 0;
2076 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
2077 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2078 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2079 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2080 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2081 	else
2082 		sqp->ud_header.bth.destination_qpn =
2083 			cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
2084 
2085 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2086 	if (mlx4_is_master(mdev->dev)) {
2087 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2088 			return -EINVAL;
2089 	} else {
2090 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2091 			return -EINVAL;
2092 	}
2093 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2094 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2095 
2096 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2097 	sqp->ud_header.immediate_present = 0;
2098 
2099 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2100 
2101 	/*
2102 	 * Inline data segments may not cross a 64 byte boundary.  If
2103 	 * our UD header is bigger than the space available up to the
2104 	 * next 64 byte boundary in the WQE, use two inline data
2105 	 * segments to hold the UD header.
2106 	 */
2107 	spc = MLX4_INLINE_ALIGN -
2108 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2109 	if (header_size <= spc) {
2110 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2111 		memcpy(inl + 1, sqp->header_buf, header_size);
2112 		i = 1;
2113 	} else {
2114 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2115 		memcpy(inl + 1, sqp->header_buf, spc);
2116 
2117 		inl = (void *) (inl + 1) + spc;
2118 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2119 		/*
2120 		 * Need a barrier here to make sure all the data is
2121 		 * visible before the byte_count field is set.
2122 		 * Otherwise the HCA prefetcher could grab the 64-byte
2123 		 * chunk with this inline segment and get a valid (!=
2124 		 * 0xffffffff) byte count but stale data, and end up
2125 		 * generating a packet with bad headers.
2126 		 *
2127 		 * The first inline segment's byte_count field doesn't
2128 		 * need a barrier, because it comes after a
2129 		 * control/MLX segment and therefore is at an offset
2130 		 * of 16 mod 64.
2131 		 */
2132 		wmb();
2133 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2134 		i = 2;
2135 	}
2136 
2137 	*mlx_seg_len =
2138 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2139 	return 0;
2140 }
2141 
2142 static void mlx4_u64_to_smac(u8 *dst_mac, u64 src_mac)
2143 {
2144 	int i;
2145 
2146 	for (i = ETH_ALEN; i; i--) {
2147 		dst_mac[i - 1] = src_mac & 0xff;
2148 		src_mac >>= 8;
2149 	}
2150 }
2151 
2152 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
2153 			    void *wqe, unsigned *mlx_seg_len)
2154 {
2155 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
2156 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2157 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2158 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2159 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2160 	union ib_gid sgid;
2161 	u16 pkey;
2162 	int send_size;
2163 	int header_size;
2164 	int spc;
2165 	int i;
2166 	int err = 0;
2167 	u16 vlan = 0xffff;
2168 	bool is_eth;
2169 	bool is_vlan = false;
2170 	bool is_grh;
2171 
2172 	send_size = 0;
2173 	for (i = 0; i < wr->num_sge; ++i)
2174 		send_size += wr->sg_list[i].length;
2175 
2176 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2177 	is_grh = mlx4_ib_ah_grh_present(ah);
2178 	if (is_eth) {
2179 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2180 			/* When multi-function is enabled, the ib_core gid
2181 			 * indexes don't necessarily match the hw ones, so
2182 			 * we must use our own cache */
2183 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2184 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
2185 							   ah->av.ib.gid_index, &sgid.raw[0]);
2186 			if (err)
2187 				return err;
2188 		} else  {
2189 			err = ib_get_cached_gid(ib_dev,
2190 						be32_to_cpu(ah->av.ib.port_pd) >> 24,
2191 						ah->av.ib.gid_index, &sgid);
2192 			if (err)
2193 				return err;
2194 		}
2195 
2196 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
2197 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
2198 			is_vlan = 1;
2199 		}
2200 	}
2201 	ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh, 0, &sqp->ud_header);
2202 
2203 	if (!is_eth) {
2204 		sqp->ud_header.lrh.service_level =
2205 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2206 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
2207 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2208 	}
2209 
2210 	if (is_grh) {
2211 		sqp->ud_header.grh.traffic_class =
2212 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2213 		sqp->ud_header.grh.flow_label    =
2214 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
2215 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
2216 		if (is_eth)
2217 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
2218 		else {
2219 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2220 			/* When multi-function is enabled, the ib_core gid
2221 			 * indexes don't necessarily match the hw ones, so
2222 			 * we must use our own cache */
2223 			sqp->ud_header.grh.source_gid.global.subnet_prefix =
2224 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2225 						       subnet_prefix;
2226 			sqp->ud_header.grh.source_gid.global.interface_id =
2227 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2228 					       guid_cache[ah->av.ib.gid_index];
2229 		} else
2230 			ib_get_cached_gid(ib_dev,
2231 					  be32_to_cpu(ah->av.ib.port_pd) >> 24,
2232 					  ah->av.ib.gid_index,
2233 					  &sqp->ud_header.grh.source_gid);
2234 		}
2235 		memcpy(sqp->ud_header.grh.destination_gid.raw,
2236 		       ah->av.ib.dgid, 16);
2237 	}
2238 
2239 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2240 
2241 	if (!is_eth) {
2242 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
2243 					  (sqp->ud_header.lrh.destination_lid ==
2244 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
2245 					  (sqp->ud_header.lrh.service_level << 8));
2246 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
2247 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
2248 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
2249 	}
2250 
2251 	switch (wr->opcode) {
2252 	case IB_WR_SEND:
2253 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
2254 		sqp->ud_header.immediate_present = 0;
2255 		break;
2256 	case IB_WR_SEND_WITH_IMM:
2257 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2258 		sqp->ud_header.immediate_present = 1;
2259 		sqp->ud_header.immediate_data    = wr->ex.imm_data;
2260 		break;
2261 	default:
2262 		return -EINVAL;
2263 	}
2264 
2265 	if (is_eth) {
2266 		struct in6_addr in6;
2267 
2268 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
2269 
2270 		mlx->sched_prio = cpu_to_be16(pcp);
2271 
2272 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
2273 		/* FIXME: cache smac value? */
2274 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
2275 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
2276 		memcpy(&in6, sgid.raw, sizeof(in6));
2277 
2278 		if (!mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2279 			u64 mac = atomic64_read(&to_mdev(ib_dev)->iboe.mac[sqp->qp.port - 1]);
2280 			u8 smac[ETH_ALEN];
2281 
2282 			mlx4_u64_to_smac(smac, mac);
2283 			memcpy(sqp->ud_header.eth.smac_h, smac, ETH_ALEN);
2284 		} else {
2285 			/* use the src mac of the tunnel */
2286 			memcpy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac, ETH_ALEN);
2287 		}
2288 
2289 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
2290 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
2291 		if (!is_vlan) {
2292 			sqp->ud_header.eth.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2293 		} else {
2294 			sqp->ud_header.vlan.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2295 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
2296 		}
2297 	} else {
2298 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
2299 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
2300 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
2301 	}
2302 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
2303 	if (!sqp->qp.ibqp.qp_num)
2304 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
2305 	else
2306 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
2307 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2308 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2309 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2310 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
2311 					       sqp->qkey : wr->wr.ud.remote_qkey);
2312 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
2313 
2314 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2315 
2316 	if (0) {
2317 		pr_err("built UD header of size %d:\n", header_size);
2318 		for (i = 0; i < header_size / 4; ++i) {
2319 			if (i % 8 == 0)
2320 				pr_err("  [%02x] ", i * 4);
2321 			pr_cont(" %08x",
2322 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
2323 			if ((i + 1) % 8 == 0)
2324 				pr_cont("\n");
2325 		}
2326 		pr_err("\n");
2327 	}
2328 
2329 	/*
2330 	 * Inline data segments may not cross a 64 byte boundary.  If
2331 	 * our UD header is bigger than the space available up to the
2332 	 * next 64 byte boundary in the WQE, use two inline data
2333 	 * segments to hold the UD header.
2334 	 */
2335 	spc = MLX4_INLINE_ALIGN -
2336 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2337 	if (header_size <= spc) {
2338 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2339 		memcpy(inl + 1, sqp->header_buf, header_size);
2340 		i = 1;
2341 	} else {
2342 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2343 		memcpy(inl + 1, sqp->header_buf, spc);
2344 
2345 		inl = (void *) (inl + 1) + spc;
2346 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2347 		/*
2348 		 * Need a barrier here to make sure all the data is
2349 		 * visible before the byte_count field is set.
2350 		 * Otherwise the HCA prefetcher could grab the 64-byte
2351 		 * chunk with this inline segment and get a valid (!=
2352 		 * 0xffffffff) byte count but stale data, and end up
2353 		 * generating a packet with bad headers.
2354 		 *
2355 		 * The first inline segment's byte_count field doesn't
2356 		 * need a barrier, because it comes after a
2357 		 * control/MLX segment and therefore is at an offset
2358 		 * of 16 mod 64.
2359 		 */
2360 		wmb();
2361 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2362 		i = 2;
2363 	}
2364 
2365 	*mlx_seg_len =
2366 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2367 	return 0;
2368 }
2369 
2370 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2371 {
2372 	unsigned cur;
2373 	struct mlx4_ib_cq *cq;
2374 
2375 	cur = wq->head - wq->tail;
2376 	if (likely(cur + nreq < wq->max_post))
2377 		return 0;
2378 
2379 	cq = to_mcq(ib_cq);
2380 	spin_lock(&cq->lock);
2381 	cur = wq->head - wq->tail;
2382 	spin_unlock(&cq->lock);
2383 
2384 	return cur + nreq >= wq->max_post;
2385 }
2386 
2387 static __be32 convert_access(int acc)
2388 {
2389 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
2390 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
2391 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
2392 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
2393 	       (acc & IB_ACCESS_REMOTE_READ   ?
2394 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
2395 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
2396 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
2397 }
2398 
2399 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr)
2400 {
2401 	struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
2402 	int i;
2403 
2404 	for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i)
2405 		mfrpl->mapped_page_list[i] =
2406 			cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] |
2407 				    MLX4_MTT_FLAG_PRESENT);
2408 
2409 	fseg->flags		= convert_access(wr->wr.fast_reg.access_flags);
2410 	fseg->mem_key		= cpu_to_be32(wr->wr.fast_reg.rkey);
2411 	fseg->buf_list		= cpu_to_be64(mfrpl->map);
2412 	fseg->start_addr	= cpu_to_be64(wr->wr.fast_reg.iova_start);
2413 	fseg->reg_len		= cpu_to_be64(wr->wr.fast_reg.length);
2414 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
2415 	fseg->page_size		= cpu_to_be32(wr->wr.fast_reg.page_shift);
2416 	fseg->reserved[0]	= 0;
2417 	fseg->reserved[1]	= 0;
2418 }
2419 
2420 static void set_bind_seg(struct mlx4_wqe_bind_seg *bseg, struct ib_send_wr *wr)
2421 {
2422 	bseg->flags1 =
2423 		convert_access(wr->wr.bind_mw.bind_info.mw_access_flags) &
2424 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ  |
2425 			    MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE |
2426 			    MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC);
2427 	bseg->flags2 = 0;
2428 	if (wr->wr.bind_mw.mw->type == IB_MW_TYPE_2)
2429 		bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_TYPE_2);
2430 	if (wr->wr.bind_mw.bind_info.mw_access_flags & IB_ZERO_BASED)
2431 		bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_ZERO_BASED);
2432 	bseg->new_rkey = cpu_to_be32(wr->wr.bind_mw.rkey);
2433 	bseg->lkey = cpu_to_be32(wr->wr.bind_mw.bind_info.mr->lkey);
2434 	bseg->addr = cpu_to_be64(wr->wr.bind_mw.bind_info.addr);
2435 	bseg->length = cpu_to_be64(wr->wr.bind_mw.bind_info.length);
2436 }
2437 
2438 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2439 {
2440 	memset(iseg, 0, sizeof(*iseg));
2441 	iseg->mem_key = cpu_to_be32(rkey);
2442 }
2443 
2444 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2445 					  u64 remote_addr, u32 rkey)
2446 {
2447 	rseg->raddr    = cpu_to_be64(remote_addr);
2448 	rseg->rkey     = cpu_to_be32(rkey);
2449 	rseg->reserved = 0;
2450 }
2451 
2452 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
2453 {
2454 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2455 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
2456 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
2457 	} else if (wr->opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2458 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2459 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add_mask);
2460 	} else {
2461 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2462 		aseg->compare  = 0;
2463 	}
2464 
2465 }
2466 
2467 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2468 				  struct ib_send_wr *wr)
2469 {
2470 	aseg->swap_add		= cpu_to_be64(wr->wr.atomic.swap);
2471 	aseg->swap_add_mask	= cpu_to_be64(wr->wr.atomic.swap_mask);
2472 	aseg->compare		= cpu_to_be64(wr->wr.atomic.compare_add);
2473 	aseg->compare_mask	= cpu_to_be64(wr->wr.atomic.compare_add_mask);
2474 }
2475 
2476 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2477 			     struct ib_send_wr *wr)
2478 {
2479 	memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
2480 	dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2481 	dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2482 	dseg->vlan = to_mah(wr->wr.ud.ah)->av.eth.vlan;
2483 	memcpy(dseg->mac, to_mah(wr->wr.ud.ah)->av.eth.mac, 6);
2484 }
2485 
2486 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2487 				    struct mlx4_wqe_datagram_seg *dseg,
2488 				    struct ib_send_wr *wr,
2489 				    enum mlx4_ib_qp_type qpt)
2490 {
2491 	union mlx4_ext_av *av = &to_mah(wr->wr.ud.ah)->av;
2492 	struct mlx4_av sqp_av = {0};
2493 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
2494 
2495 	/* force loopback */
2496 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2497 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2498 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2499 			cpu_to_be32(0xf0000000);
2500 
2501 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2502 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
2503 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2504 	else
2505 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp0_tunnel[port - 1]);
2506 	/* Use QKEY from the QP context, which is set by master */
2507 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2508 }
2509 
2510 static void build_tunnel_header(struct ib_send_wr *wr, void *wqe, unsigned *mlx_seg_len)
2511 {
2512 	struct mlx4_wqe_inline_seg *inl = wqe;
2513 	struct mlx4_ib_tunnel_header hdr;
2514 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2515 	int spc;
2516 	int i;
2517 
2518 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2519 	hdr.remote_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2520 	hdr.pkey_index = cpu_to_be16(wr->wr.ud.pkey_index);
2521 	hdr.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2522 	memcpy(hdr.mac, ah->av.eth.mac, 6);
2523 	hdr.vlan = ah->av.eth.vlan;
2524 
2525 	spc = MLX4_INLINE_ALIGN -
2526 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2527 	if (sizeof (hdr) <= spc) {
2528 		memcpy(inl + 1, &hdr, sizeof (hdr));
2529 		wmb();
2530 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
2531 		i = 1;
2532 	} else {
2533 		memcpy(inl + 1, &hdr, spc);
2534 		wmb();
2535 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2536 
2537 		inl = (void *) (inl + 1) + spc;
2538 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2539 		wmb();
2540 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
2541 		i = 2;
2542 	}
2543 
2544 	*mlx_seg_len =
2545 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2546 }
2547 
2548 static void set_mlx_icrc_seg(void *dseg)
2549 {
2550 	u32 *t = dseg;
2551 	struct mlx4_wqe_inline_seg *iseg = dseg;
2552 
2553 	t[1] = 0;
2554 
2555 	/*
2556 	 * Need a barrier here before writing the byte_count field to
2557 	 * make sure that all the data is visible before the
2558 	 * byte_count field is set.  Otherwise, if the segment begins
2559 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2560 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2561 	 * stale data, and end up sending the wrong data.
2562 	 */
2563 	wmb();
2564 
2565 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2566 }
2567 
2568 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2569 {
2570 	dseg->lkey       = cpu_to_be32(sg->lkey);
2571 	dseg->addr       = cpu_to_be64(sg->addr);
2572 
2573 	/*
2574 	 * Need a barrier here before writing the byte_count field to
2575 	 * make sure that all the data is visible before the
2576 	 * byte_count field is set.  Otherwise, if the segment begins
2577 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2578 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2579 	 * stale data, and end up sending the wrong data.
2580 	 */
2581 	wmb();
2582 
2583 	dseg->byte_count = cpu_to_be32(sg->length);
2584 }
2585 
2586 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2587 {
2588 	dseg->byte_count = cpu_to_be32(sg->length);
2589 	dseg->lkey       = cpu_to_be32(sg->lkey);
2590 	dseg->addr       = cpu_to_be64(sg->addr);
2591 }
2592 
2593 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr,
2594 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2595 			 __be32 *lso_hdr_sz, __be32 *blh)
2596 {
2597 	unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
2598 
2599 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2600 		*blh = cpu_to_be32(1 << 6);
2601 
2602 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2603 		     wr->num_sge > qp->sq.max_gs - (halign >> 4)))
2604 		return -EINVAL;
2605 
2606 	memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
2607 
2608 	*lso_hdr_sz  = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 |
2609 				   wr->wr.ud.hlen);
2610 	*lso_seg_len = halign;
2611 	return 0;
2612 }
2613 
2614 static __be32 send_ieth(struct ib_send_wr *wr)
2615 {
2616 	switch (wr->opcode) {
2617 	case IB_WR_SEND_WITH_IMM:
2618 	case IB_WR_RDMA_WRITE_WITH_IMM:
2619 		return wr->ex.imm_data;
2620 
2621 	case IB_WR_SEND_WITH_INV:
2622 		return cpu_to_be32(wr->ex.invalidate_rkey);
2623 
2624 	default:
2625 		return 0;
2626 	}
2627 }
2628 
2629 static void add_zero_len_inline(void *wqe)
2630 {
2631 	struct mlx4_wqe_inline_seg *inl = wqe;
2632 	memset(wqe, 0, 16);
2633 	inl->byte_count = cpu_to_be32(1 << 31);
2634 }
2635 
2636 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2637 		      struct ib_send_wr **bad_wr)
2638 {
2639 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2640 	void *wqe;
2641 	struct mlx4_wqe_ctrl_seg *ctrl;
2642 	struct mlx4_wqe_data_seg *dseg;
2643 	unsigned long flags;
2644 	int nreq;
2645 	int err = 0;
2646 	unsigned ind;
2647 	int uninitialized_var(stamp);
2648 	int uninitialized_var(size);
2649 	unsigned uninitialized_var(seglen);
2650 	__be32 dummy;
2651 	__be32 *lso_wqe;
2652 	__be32 uninitialized_var(lso_hdr_sz);
2653 	__be32 blh;
2654 	int i;
2655 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
2656 
2657 	spin_lock_irqsave(&qp->sq.lock, flags);
2658 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
2659 		err = -EIO;
2660 		*bad_wr = wr;
2661 		nreq = 0;
2662 		goto out;
2663 	}
2664 
2665 	ind = qp->sq_next_wqe;
2666 
2667 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2668 		lso_wqe = &dummy;
2669 		blh = 0;
2670 
2671 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2672 			err = -ENOMEM;
2673 			*bad_wr = wr;
2674 			goto out;
2675 		}
2676 
2677 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
2678 			err = -EINVAL;
2679 			*bad_wr = wr;
2680 			goto out;
2681 		}
2682 
2683 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
2684 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
2685 
2686 		ctrl->srcrb_flags =
2687 			(wr->send_flags & IB_SEND_SIGNALED ?
2688 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
2689 			(wr->send_flags & IB_SEND_SOLICITED ?
2690 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
2691 			((wr->send_flags & IB_SEND_IP_CSUM) ?
2692 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
2693 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
2694 			qp->sq_signal_bits;
2695 
2696 		ctrl->imm = send_ieth(wr);
2697 
2698 		wqe += sizeof *ctrl;
2699 		size = sizeof *ctrl / 16;
2700 
2701 		switch (qp->mlx4_ib_qp_type) {
2702 		case MLX4_IB_QPT_RC:
2703 		case MLX4_IB_QPT_UC:
2704 			switch (wr->opcode) {
2705 			case IB_WR_ATOMIC_CMP_AND_SWP:
2706 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2707 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
2708 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2709 					      wr->wr.atomic.rkey);
2710 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2711 
2712 				set_atomic_seg(wqe, wr);
2713 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
2714 
2715 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2716 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
2717 
2718 				break;
2719 
2720 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2721 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2722 					      wr->wr.atomic.rkey);
2723 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2724 
2725 				set_masked_atomic_seg(wqe, wr);
2726 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
2727 
2728 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2729 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
2730 
2731 				break;
2732 
2733 			case IB_WR_RDMA_READ:
2734 			case IB_WR_RDMA_WRITE:
2735 			case IB_WR_RDMA_WRITE_WITH_IMM:
2736 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2737 					      wr->wr.rdma.rkey);
2738 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2739 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
2740 				break;
2741 
2742 			case IB_WR_LOCAL_INV:
2743 				ctrl->srcrb_flags |=
2744 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2745 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
2746 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
2747 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
2748 				break;
2749 
2750 			case IB_WR_FAST_REG_MR:
2751 				ctrl->srcrb_flags |=
2752 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2753 				set_fmr_seg(wqe, wr);
2754 				wqe  += sizeof (struct mlx4_wqe_fmr_seg);
2755 				size += sizeof (struct mlx4_wqe_fmr_seg) / 16;
2756 				break;
2757 
2758 			case IB_WR_BIND_MW:
2759 				ctrl->srcrb_flags |=
2760 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2761 				set_bind_seg(wqe, wr);
2762 				wqe  += sizeof(struct mlx4_wqe_bind_seg);
2763 				size += sizeof(struct mlx4_wqe_bind_seg) / 16;
2764 				break;
2765 			default:
2766 				/* No extra segments required for sends */
2767 				break;
2768 			}
2769 			break;
2770 
2771 		case MLX4_IB_QPT_TUN_SMI_OWNER:
2772 			err =  build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2773 			if (unlikely(err)) {
2774 				*bad_wr = wr;
2775 				goto out;
2776 			}
2777 			wqe  += seglen;
2778 			size += seglen / 16;
2779 			break;
2780 		case MLX4_IB_QPT_TUN_SMI:
2781 		case MLX4_IB_QPT_TUN_GSI:
2782 			/* this is a UD qp used in MAD responses to slaves. */
2783 			set_datagram_seg(wqe, wr);
2784 			/* set the forced-loopback bit in the data seg av */
2785 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
2786 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2787 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2788 			break;
2789 		case MLX4_IB_QPT_UD:
2790 			set_datagram_seg(wqe, wr);
2791 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2792 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2793 
2794 			if (wr->opcode == IB_WR_LSO) {
2795 				err = build_lso_seg(wqe, wr, qp, &seglen, &lso_hdr_sz, &blh);
2796 				if (unlikely(err)) {
2797 					*bad_wr = wr;
2798 					goto out;
2799 				}
2800 				lso_wqe = (__be32 *) wqe;
2801 				wqe  += seglen;
2802 				size += seglen / 16;
2803 			}
2804 			break;
2805 
2806 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
2807 			err = build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2808 			if (unlikely(err)) {
2809 				*bad_wr = wr;
2810 				goto out;
2811 			}
2812 			wqe  += seglen;
2813 			size += seglen / 16;
2814 			/* to start tunnel header on a cache-line boundary */
2815 			add_zero_len_inline(wqe);
2816 			wqe += 16;
2817 			size++;
2818 			build_tunnel_header(wr, wqe, &seglen);
2819 			wqe  += seglen;
2820 			size += seglen / 16;
2821 			break;
2822 		case MLX4_IB_QPT_PROXY_SMI:
2823 		case MLX4_IB_QPT_PROXY_GSI:
2824 			/* If we are tunneling special qps, this is a UD qp.
2825 			 * In this case we first add a UD segment targeting
2826 			 * the tunnel qp, and then add a header with address
2827 			 * information */
2828 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe, wr,
2829 						qp->mlx4_ib_qp_type);
2830 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2831 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2832 			build_tunnel_header(wr, wqe, &seglen);
2833 			wqe  += seglen;
2834 			size += seglen / 16;
2835 			break;
2836 
2837 		case MLX4_IB_QPT_SMI:
2838 		case MLX4_IB_QPT_GSI:
2839 			err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
2840 			if (unlikely(err)) {
2841 				*bad_wr = wr;
2842 				goto out;
2843 			}
2844 			wqe  += seglen;
2845 			size += seglen / 16;
2846 			break;
2847 
2848 		default:
2849 			break;
2850 		}
2851 
2852 		/*
2853 		 * Write data segments in reverse order, so as to
2854 		 * overwrite cacheline stamp last within each
2855 		 * cacheline.  This avoids issues with WQE
2856 		 * prefetching.
2857 		 */
2858 
2859 		dseg = wqe;
2860 		dseg += wr->num_sge - 1;
2861 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
2862 
2863 		/* Add one more inline data segment for ICRC for MLX sends */
2864 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2865 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
2866 			     qp->mlx4_ib_qp_type &
2867 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
2868 			set_mlx_icrc_seg(dseg + 1);
2869 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
2870 		}
2871 
2872 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
2873 			set_data_seg(dseg, wr->sg_list + i);
2874 
2875 		/*
2876 		 * Possibly overwrite stamping in cacheline with LSO
2877 		 * segment only after making sure all data segments
2878 		 * are written.
2879 		 */
2880 		wmb();
2881 		*lso_wqe = lso_hdr_sz;
2882 
2883 		ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
2884 				    MLX4_WQE_CTRL_FENCE : 0) | size;
2885 
2886 		/*
2887 		 * Make sure descriptor is fully written before
2888 		 * setting ownership bit (because HW can start
2889 		 * executing as soon as we do).
2890 		 */
2891 		wmb();
2892 
2893 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
2894 			*bad_wr = wr;
2895 			err = -EINVAL;
2896 			goto out;
2897 		}
2898 
2899 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
2900 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
2901 
2902 		stamp = ind + qp->sq_spare_wqes;
2903 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
2904 
2905 		/*
2906 		 * We can improve latency by not stamping the last
2907 		 * send queue WQE until after ringing the doorbell, so
2908 		 * only stamp here if there are still more WQEs to post.
2909 		 *
2910 		 * Same optimization applies to padding with NOP wqe
2911 		 * in case of WQE shrinking (used to prevent wrap-around
2912 		 * in the middle of WR).
2913 		 */
2914 		if (wr->next) {
2915 			stamp_send_wqe(qp, stamp, size * 16);
2916 			ind = pad_wraparound(qp, ind);
2917 		}
2918 	}
2919 
2920 out:
2921 	if (likely(nreq)) {
2922 		qp->sq.head += nreq;
2923 
2924 		/*
2925 		 * Make sure that descriptors are written before
2926 		 * doorbell record.
2927 		 */
2928 		wmb();
2929 
2930 		writel(qp->doorbell_qpn,
2931 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
2932 
2933 		/*
2934 		 * Make sure doorbells don't leak out of SQ spinlock
2935 		 * and reach the HCA out of order.
2936 		 */
2937 		mmiowb();
2938 
2939 		stamp_send_wqe(qp, stamp, size * 16);
2940 
2941 		ind = pad_wraparound(qp, ind);
2942 		qp->sq_next_wqe = ind;
2943 	}
2944 
2945 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2946 
2947 	return err;
2948 }
2949 
2950 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2951 		      struct ib_recv_wr **bad_wr)
2952 {
2953 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2954 	struct mlx4_wqe_data_seg *scat;
2955 	unsigned long flags;
2956 	int err = 0;
2957 	int nreq;
2958 	int ind;
2959 	int max_gs;
2960 	int i;
2961 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
2962 
2963 	max_gs = qp->rq.max_gs;
2964 	spin_lock_irqsave(&qp->rq.lock, flags);
2965 
2966 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
2967 		err = -EIO;
2968 		*bad_wr = wr;
2969 		nreq = 0;
2970 		goto out;
2971 	}
2972 
2973 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
2974 
2975 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2976 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2977 			err = -ENOMEM;
2978 			*bad_wr = wr;
2979 			goto out;
2980 		}
2981 
2982 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2983 			err = -EINVAL;
2984 			*bad_wr = wr;
2985 			goto out;
2986 		}
2987 
2988 		scat = get_recv_wqe(qp, ind);
2989 
2990 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
2991 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
2992 			ib_dma_sync_single_for_device(ibqp->device,
2993 						      qp->sqp_proxy_rcv[ind].map,
2994 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
2995 						      DMA_FROM_DEVICE);
2996 			scat->byte_count =
2997 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
2998 			/* use dma lkey from upper layer entry */
2999 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3000 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3001 			scat++;
3002 			max_gs--;
3003 		}
3004 
3005 		for (i = 0; i < wr->num_sge; ++i)
3006 			__set_data_seg(scat + i, wr->sg_list + i);
3007 
3008 		if (i < max_gs) {
3009 			scat[i].byte_count = 0;
3010 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3011 			scat[i].addr       = 0;
3012 		}
3013 
3014 		qp->rq.wrid[ind] = wr->wr_id;
3015 
3016 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3017 	}
3018 
3019 out:
3020 	if (likely(nreq)) {
3021 		qp->rq.head += nreq;
3022 
3023 		/*
3024 		 * Make sure that descriptors are written before
3025 		 * doorbell record.
3026 		 */
3027 		wmb();
3028 
3029 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3030 	}
3031 
3032 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3033 
3034 	return err;
3035 }
3036 
3037 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3038 {
3039 	switch (mlx4_state) {
3040 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3041 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3042 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3043 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3044 	case MLX4_QP_STATE_SQ_DRAINING:
3045 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3046 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3047 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3048 	default:		     return -1;
3049 	}
3050 }
3051 
3052 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3053 {
3054 	switch (mlx4_mig_state) {
3055 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3056 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3057 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3058 	default: return -1;
3059 	}
3060 }
3061 
3062 static int to_ib_qp_access_flags(int mlx4_flags)
3063 {
3064 	int ib_flags = 0;
3065 
3066 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3067 		ib_flags |= IB_ACCESS_REMOTE_READ;
3068 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3069 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3070 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3071 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3072 
3073 	return ib_flags;
3074 }
3075 
3076 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
3077 				struct mlx4_qp_path *path)
3078 {
3079 	struct mlx4_dev *dev = ibdev->dev;
3080 	int is_eth;
3081 
3082 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
3083 	ib_ah_attr->port_num	  = path->sched_queue & 0x40 ? 2 : 1;
3084 
3085 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
3086 		return;
3087 
3088 	is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
3089 		IB_LINK_LAYER_ETHERNET;
3090 	if (is_eth)
3091 		ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
3092 		((path->sched_queue & 4) << 1);
3093 	else
3094 		ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
3095 
3096 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
3097 	ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
3098 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
3099 	ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
3100 	if (ib_ah_attr->ah_flags) {
3101 		ib_ah_attr->grh.sgid_index = path->mgid_index;
3102 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
3103 		ib_ah_attr->grh.traffic_class =
3104 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3105 		ib_ah_attr->grh.flow_label =
3106 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3107 		memcpy(ib_ah_attr->grh.dgid.raw,
3108 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
3109 	}
3110 }
3111 
3112 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3113 		     struct ib_qp_init_attr *qp_init_attr)
3114 {
3115 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3116 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3117 	struct mlx4_qp_context context;
3118 	int mlx4_state;
3119 	int err = 0;
3120 
3121 	mutex_lock(&qp->mutex);
3122 
3123 	if (qp->state == IB_QPS_RESET) {
3124 		qp_attr->qp_state = IB_QPS_RESET;
3125 		goto done;
3126 	}
3127 
3128 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3129 	if (err) {
3130 		err = -EINVAL;
3131 		goto out;
3132 	}
3133 
3134 	mlx4_state = be32_to_cpu(context.flags) >> 28;
3135 
3136 	qp->state		     = to_ib_qp_state(mlx4_state);
3137 	qp_attr->qp_state	     = qp->state;
3138 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
3139 	qp_attr->path_mig_state	     =
3140 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3141 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
3142 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3143 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
3144 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
3145 	qp_attr->qp_access_flags     =
3146 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
3147 
3148 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3149 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3150 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3151 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3152 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
3153 	}
3154 
3155 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3156 	if (qp_attr->qp_state == IB_QPS_INIT)
3157 		qp_attr->port_num = qp->port;
3158 	else
3159 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3160 
3161 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3162 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3163 
3164 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3165 
3166 	qp_attr->max_dest_rd_atomic =
3167 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3168 	qp_attr->min_rnr_timer	    =
3169 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3170 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
3171 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
3172 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
3173 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
3174 
3175 done:
3176 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
3177 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
3178 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
3179 
3180 	if (!ibqp->uobject) {
3181 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
3182 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
3183 	} else {
3184 		qp_attr->cap.max_send_wr  = 0;
3185 		qp_attr->cap.max_send_sge = 0;
3186 	}
3187 
3188 	/*
3189 	 * We don't support inline sends for kernel QPs (yet), and we
3190 	 * don't know what userspace's value should be.
3191 	 */
3192 	qp_attr->cap.max_inline_data = 0;
3193 
3194 	qp_init_attr->cap	     = qp_attr->cap;
3195 
3196 	qp_init_attr->create_flags = 0;
3197 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3198 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3199 
3200 	if (qp->flags & MLX4_IB_QP_LSO)
3201 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
3202 
3203 	if (qp->flags & MLX4_IB_QP_NETIF)
3204 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
3205 
3206 	qp_init_attr->sq_sig_type =
3207 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
3208 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3209 
3210 out:
3211 	mutex_unlock(&qp->mutex);
3212 	return err;
3213 }
3214 
3215