xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision 9cfc5c90)
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_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 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
570 			kfree(qp->sqp_proxy_rcv[i].addr);
571 			goto err;
572 		}
573 	}
574 	return 0;
575 
576 err:
577 	while (i > 0) {
578 		--i;
579 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
580 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
581 				    DMA_FROM_DEVICE);
582 		kfree(qp->sqp_proxy_rcv[i].addr);
583 	}
584 	kfree(qp->sqp_proxy_rcv);
585 	qp->sqp_proxy_rcv = NULL;
586 	return -ENOMEM;
587 }
588 
589 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
590 {
591 	int i;
592 
593 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
594 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
595 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
596 				    DMA_FROM_DEVICE);
597 		kfree(qp->sqp_proxy_rcv[i].addr);
598 	}
599 	kfree(qp->sqp_proxy_rcv);
600 }
601 
602 static int qp_has_rq(struct ib_qp_init_attr *attr)
603 {
604 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
605 		return 0;
606 
607 	return !attr->srq;
608 }
609 
610 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
611 {
612 	int i;
613 	for (i = 0; i < dev->caps.num_ports; i++) {
614 		if (qpn == dev->caps.qp0_proxy[i])
615 			return !!dev->caps.qp0_qkey[i];
616 	}
617 	return 0;
618 }
619 
620 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
621 				    struct mlx4_ib_qp *qp)
622 {
623 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
624 	mlx4_counter_free(dev->dev, qp->counter_index->index);
625 	list_del(&qp->counter_index->list);
626 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
627 
628 	kfree(qp->counter_index);
629 	qp->counter_index = NULL;
630 }
631 
632 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
633 			    struct ib_qp_init_attr *init_attr,
634 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp,
635 			    gfp_t gfp)
636 {
637 	int qpn;
638 	int err;
639 	struct mlx4_ib_sqp *sqp;
640 	struct mlx4_ib_qp *qp;
641 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
642 	struct mlx4_ib_cq *mcq;
643 	unsigned long flags;
644 
645 	/* When tunneling special qps, we use a plain UD qp */
646 	if (sqpn) {
647 		if (mlx4_is_mfunc(dev->dev) &&
648 		    (!mlx4_is_master(dev->dev) ||
649 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
650 			if (init_attr->qp_type == IB_QPT_GSI)
651 				qp_type = MLX4_IB_QPT_PROXY_GSI;
652 			else {
653 				if (mlx4_is_master(dev->dev) ||
654 				    qp0_enabled_vf(dev->dev, sqpn))
655 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
656 				else
657 					qp_type = MLX4_IB_QPT_PROXY_SMI;
658 			}
659 		}
660 		qpn = sqpn;
661 		/* add extra sg entry for tunneling */
662 		init_attr->cap.max_recv_sge++;
663 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
664 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
665 			container_of(init_attr,
666 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
667 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
668 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
669 		    !mlx4_is_master(dev->dev))
670 			return -EINVAL;
671 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
672 			qp_type = MLX4_IB_QPT_TUN_GSI;
673 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
674 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
675 					     tnl_init->port))
676 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
677 		else
678 			qp_type = MLX4_IB_QPT_TUN_SMI;
679 		/* we are definitely in the PPF here, since we are creating
680 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
681 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
682 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
683 		sqpn = qpn;
684 	}
685 
686 	if (!*caller_qp) {
687 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
688 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
689 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
690 			sqp = kzalloc(sizeof (struct mlx4_ib_sqp), gfp);
691 			if (!sqp)
692 				return -ENOMEM;
693 			qp = &sqp->qp;
694 			qp->pri.vid = 0xFFFF;
695 			qp->alt.vid = 0xFFFF;
696 		} else {
697 			qp = kzalloc(sizeof (struct mlx4_ib_qp), gfp);
698 			if (!qp)
699 				return -ENOMEM;
700 			qp->pri.vid = 0xFFFF;
701 			qp->alt.vid = 0xFFFF;
702 		}
703 	} else
704 		qp = *caller_qp;
705 
706 	qp->mlx4_ib_qp_type = qp_type;
707 
708 	mutex_init(&qp->mutex);
709 	spin_lock_init(&qp->sq.lock);
710 	spin_lock_init(&qp->rq.lock);
711 	INIT_LIST_HEAD(&qp->gid_list);
712 	INIT_LIST_HEAD(&qp->steering_rules);
713 
714 	qp->state	 = IB_QPS_RESET;
715 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
716 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
717 
718 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
719 	if (err)
720 		goto err;
721 
722 	if (pd->uobject) {
723 		struct mlx4_ib_create_qp ucmd;
724 
725 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
726 			err = -EFAULT;
727 			goto err;
728 		}
729 
730 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
731 
732 		err = set_user_sq_size(dev, qp, &ucmd);
733 		if (err)
734 			goto err;
735 
736 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
737 				       qp->buf_size, 0, 0);
738 		if (IS_ERR(qp->umem)) {
739 			err = PTR_ERR(qp->umem);
740 			goto err;
741 		}
742 
743 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
744 				    ilog2(qp->umem->page_size), &qp->mtt);
745 		if (err)
746 			goto err_buf;
747 
748 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
749 		if (err)
750 			goto err_mtt;
751 
752 		if (qp_has_rq(init_attr)) {
753 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
754 						  ucmd.db_addr, &qp->db);
755 			if (err)
756 				goto err_mtt;
757 		}
758 	} else {
759 		qp->sq_no_prefetch = 0;
760 
761 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
762 			qp->flags |= MLX4_IB_QP_LSO;
763 
764 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
765 			if (dev->steering_support ==
766 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
767 				qp->flags |= MLX4_IB_QP_NETIF;
768 			else
769 				goto err;
770 		}
771 
772 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
773 		if (err)
774 			goto err;
775 
776 		if (qp_has_rq(init_attr)) {
777 			err = mlx4_db_alloc(dev->dev, &qp->db, 0, gfp);
778 			if (err)
779 				goto err;
780 
781 			*qp->db.db = 0;
782 		}
783 
784 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf, gfp)) {
785 			err = -ENOMEM;
786 			goto err_db;
787 		}
788 
789 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
790 				    &qp->mtt);
791 		if (err)
792 			goto err_buf;
793 
794 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf, gfp);
795 		if (err)
796 			goto err_mtt;
797 
798 		qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), gfp);
799 		qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), gfp);
800 		if (!qp->sq.wrid || !qp->rq.wrid) {
801 			err = -ENOMEM;
802 			goto err_wrid;
803 		}
804 	}
805 
806 	if (sqpn) {
807 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
808 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
809 			if (alloc_proxy_bufs(pd->device, qp)) {
810 				err = -ENOMEM;
811 				goto err_wrid;
812 			}
813 		}
814 	} else {
815 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
816 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
817 		 * VLAN insertion. */
818 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
819 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
820 						    (init_attr->cap.max_send_wr ?
821 						     MLX4_RESERVE_ETH_BF_QP : 0) |
822 						    (init_attr->cap.max_recv_wr ?
823 						     MLX4_RESERVE_A0_QP : 0));
824 		else
825 			if (qp->flags & MLX4_IB_QP_NETIF)
826 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
827 			else
828 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
829 							    &qpn, 0);
830 		if (err)
831 			goto err_proxy;
832 	}
833 
834 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
835 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
836 
837 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp, gfp);
838 	if (err)
839 		goto err_qpn;
840 
841 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
842 		qp->mqp.qpn |= (1 << 23);
843 
844 	/*
845 	 * Hardware wants QPN written in big-endian order (after
846 	 * shifting) for send doorbell.  Precompute this value to save
847 	 * a little bit when posting sends.
848 	 */
849 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
850 
851 	qp->mqp.event = mlx4_ib_qp_event;
852 	if (!*caller_qp)
853 		*caller_qp = qp;
854 
855 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
856 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
857 			 to_mcq(init_attr->recv_cq));
858 	/* Maintain device to QPs access, needed for further handling
859 	 * via reset flow
860 	 */
861 	list_add_tail(&qp->qps_list, &dev->qp_list);
862 	/* Maintain CQ to QPs access, needed for further handling
863 	 * via reset flow
864 	 */
865 	mcq = to_mcq(init_attr->send_cq);
866 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
867 	mcq = to_mcq(init_attr->recv_cq);
868 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
869 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
870 			   to_mcq(init_attr->recv_cq));
871 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
872 	return 0;
873 
874 err_qpn:
875 	if (!sqpn) {
876 		if (qp->flags & MLX4_IB_QP_NETIF)
877 			mlx4_ib_steer_qp_free(dev, qpn, 1);
878 		else
879 			mlx4_qp_release_range(dev->dev, qpn, 1);
880 	}
881 err_proxy:
882 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
883 		free_proxy_bufs(pd->device, qp);
884 err_wrid:
885 	if (pd->uobject) {
886 		if (qp_has_rq(init_attr))
887 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
888 	} else {
889 		kfree(qp->sq.wrid);
890 		kfree(qp->rq.wrid);
891 	}
892 
893 err_mtt:
894 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
895 
896 err_buf:
897 	if (pd->uobject)
898 		ib_umem_release(qp->umem);
899 	else
900 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
901 
902 err_db:
903 	if (!pd->uobject && qp_has_rq(init_attr))
904 		mlx4_db_free(dev->dev, &qp->db);
905 
906 err:
907 	if (!*caller_qp)
908 		kfree(qp);
909 	return err;
910 }
911 
912 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
913 {
914 	switch (state) {
915 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
916 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
917 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
918 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
919 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
920 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
921 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
922 	default:		return -1;
923 	}
924 }
925 
926 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
927 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
928 {
929 	if (send_cq == recv_cq) {
930 		spin_lock(&send_cq->lock);
931 		__acquire(&recv_cq->lock);
932 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
933 		spin_lock(&send_cq->lock);
934 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
935 	} else {
936 		spin_lock(&recv_cq->lock);
937 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
938 	}
939 }
940 
941 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
942 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
943 {
944 	if (send_cq == recv_cq) {
945 		__release(&recv_cq->lock);
946 		spin_unlock(&send_cq->lock);
947 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
948 		spin_unlock(&recv_cq->lock);
949 		spin_unlock(&send_cq->lock);
950 	} else {
951 		spin_unlock(&send_cq->lock);
952 		spin_unlock(&recv_cq->lock);
953 	}
954 }
955 
956 static void del_gid_entries(struct mlx4_ib_qp *qp)
957 {
958 	struct mlx4_ib_gid_entry *ge, *tmp;
959 
960 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
961 		list_del(&ge->list);
962 		kfree(ge);
963 	}
964 }
965 
966 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
967 {
968 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
969 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
970 	else
971 		return to_mpd(qp->ibqp.pd);
972 }
973 
974 static void get_cqs(struct mlx4_ib_qp *qp,
975 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
976 {
977 	switch (qp->ibqp.qp_type) {
978 	case IB_QPT_XRC_TGT:
979 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
980 		*recv_cq = *send_cq;
981 		break;
982 	case IB_QPT_XRC_INI:
983 		*send_cq = to_mcq(qp->ibqp.send_cq);
984 		*recv_cq = *send_cq;
985 		break;
986 	default:
987 		*send_cq = to_mcq(qp->ibqp.send_cq);
988 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
989 		break;
990 	}
991 }
992 
993 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
994 			      int is_user)
995 {
996 	struct mlx4_ib_cq *send_cq, *recv_cq;
997 	unsigned long flags;
998 
999 	if (qp->state != IB_QPS_RESET) {
1000 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1001 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1002 			pr_warn("modify QP %06x to RESET failed.\n",
1003 			       qp->mqp.qpn);
1004 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1005 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1006 			qp->pri.smac = 0;
1007 			qp->pri.smac_port = 0;
1008 		}
1009 		if (qp->alt.smac) {
1010 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1011 			qp->alt.smac = 0;
1012 		}
1013 		if (qp->pri.vid < 0x1000) {
1014 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1015 			qp->pri.vid = 0xFFFF;
1016 			qp->pri.candidate_vid = 0xFFFF;
1017 			qp->pri.update_vid = 0;
1018 		}
1019 		if (qp->alt.vid < 0x1000) {
1020 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1021 			qp->alt.vid = 0xFFFF;
1022 			qp->alt.candidate_vid = 0xFFFF;
1023 			qp->alt.update_vid = 0;
1024 		}
1025 	}
1026 
1027 	get_cqs(qp, &send_cq, &recv_cq);
1028 
1029 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1030 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1031 
1032 	/* del from lists under both locks above to protect reset flow paths */
1033 	list_del(&qp->qps_list);
1034 	list_del(&qp->cq_send_list);
1035 	list_del(&qp->cq_recv_list);
1036 	if (!is_user) {
1037 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1038 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1039 		if (send_cq != recv_cq)
1040 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1041 	}
1042 
1043 	mlx4_qp_remove(dev->dev, &qp->mqp);
1044 
1045 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1046 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1047 
1048 	mlx4_qp_free(dev->dev, &qp->mqp);
1049 
1050 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1051 		if (qp->flags & MLX4_IB_QP_NETIF)
1052 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1053 		else
1054 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1055 	}
1056 
1057 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1058 
1059 	if (is_user) {
1060 		if (qp->rq.wqe_cnt)
1061 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
1062 					      &qp->db);
1063 		ib_umem_release(qp->umem);
1064 	} else {
1065 		kfree(qp->sq.wrid);
1066 		kfree(qp->rq.wrid);
1067 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1068 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1069 			free_proxy_bufs(&dev->ib_dev, qp);
1070 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1071 		if (qp->rq.wqe_cnt)
1072 			mlx4_db_free(dev->dev, &qp->db);
1073 	}
1074 
1075 	del_gid_entries(qp);
1076 }
1077 
1078 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1079 {
1080 	/* Native or PPF */
1081 	if (!mlx4_is_mfunc(dev->dev) ||
1082 	    (mlx4_is_master(dev->dev) &&
1083 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1084 		return  dev->dev->phys_caps.base_sqpn +
1085 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1086 			attr->port_num - 1;
1087 	}
1088 	/* PF or VF -- creating proxies */
1089 	if (attr->qp_type == IB_QPT_SMI)
1090 		return dev->dev->caps.qp0_proxy[attr->port_num - 1];
1091 	else
1092 		return dev->dev->caps.qp1_proxy[attr->port_num - 1];
1093 }
1094 
1095 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1096 				struct ib_qp_init_attr *init_attr,
1097 				struct ib_udata *udata)
1098 {
1099 	struct mlx4_ib_qp *qp = NULL;
1100 	int err;
1101 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1102 	u16 xrcdn = 0;
1103 	gfp_t gfp;
1104 
1105 	gfp = (init_attr->create_flags & MLX4_IB_QP_CREATE_USE_GFP_NOIO) ?
1106 		GFP_NOIO : GFP_KERNEL;
1107 	/*
1108 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1109 	 * and only for kernel UD QPs.
1110 	 */
1111 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1112 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1113 					MLX4_IB_SRIOV_TUNNEL_QP |
1114 					MLX4_IB_SRIOV_SQP |
1115 					MLX4_IB_QP_NETIF |
1116 					MLX4_IB_QP_CREATE_USE_GFP_NOIO))
1117 		return ERR_PTR(-EINVAL);
1118 
1119 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1120 		if (init_attr->qp_type != IB_QPT_UD)
1121 			return ERR_PTR(-EINVAL);
1122 	}
1123 
1124 	if (init_attr->create_flags &&
1125 	    ((udata && init_attr->create_flags & ~(sup_u_create_flags)) ||
1126 	     ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1127 					   MLX4_IB_QP_CREATE_USE_GFP_NOIO |
1128 					   MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)) &&
1129 	      init_attr->qp_type != IB_QPT_UD) ||
1130 	     ((init_attr->create_flags & MLX4_IB_SRIOV_SQP) &&
1131 	      init_attr->qp_type > IB_QPT_GSI)))
1132 		return ERR_PTR(-EINVAL);
1133 
1134 	switch (init_attr->qp_type) {
1135 	case IB_QPT_XRC_TGT:
1136 		pd = to_mxrcd(init_attr->xrcd)->pd;
1137 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1138 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1139 		/* fall through */
1140 	case IB_QPT_XRC_INI:
1141 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1142 			return ERR_PTR(-ENOSYS);
1143 		init_attr->recv_cq = init_attr->send_cq;
1144 		/* fall through */
1145 	case IB_QPT_RC:
1146 	case IB_QPT_UC:
1147 	case IB_QPT_RAW_PACKET:
1148 		qp = kzalloc(sizeof *qp, gfp);
1149 		if (!qp)
1150 			return ERR_PTR(-ENOMEM);
1151 		qp->pri.vid = 0xFFFF;
1152 		qp->alt.vid = 0xFFFF;
1153 		/* fall through */
1154 	case IB_QPT_UD:
1155 	{
1156 		err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1157 				       udata, 0, &qp, gfp);
1158 		if (err)
1159 			return ERR_PTR(err);
1160 
1161 		qp->ibqp.qp_num = qp->mqp.qpn;
1162 		qp->xrcdn = xrcdn;
1163 
1164 		break;
1165 	}
1166 	case IB_QPT_SMI:
1167 	case IB_QPT_GSI:
1168 	{
1169 		/* Userspace is not allowed to create special QPs: */
1170 		if (udata)
1171 			return ERR_PTR(-EINVAL);
1172 
1173 		err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1174 				       get_sqp_num(to_mdev(pd->device), init_attr),
1175 				       &qp, gfp);
1176 		if (err)
1177 			return ERR_PTR(err);
1178 
1179 		qp->port	= init_attr->port_num;
1180 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
1181 
1182 		break;
1183 	}
1184 	default:
1185 		/* Don't support raw QPs */
1186 		return ERR_PTR(-EINVAL);
1187 	}
1188 
1189 	return &qp->ibqp;
1190 }
1191 
1192 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1193 {
1194 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1195 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1196 	struct mlx4_ib_pd *pd;
1197 
1198 	if (is_qp0(dev, mqp))
1199 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1200 
1201 	if (dev->qp1_proxy[mqp->port - 1] == mqp) {
1202 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1203 		dev->qp1_proxy[mqp->port - 1] = NULL;
1204 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1205 	}
1206 
1207 	if (mqp->counter_index)
1208 		mlx4_ib_free_qp_counter(dev, mqp);
1209 
1210 	pd = get_pd(mqp);
1211 	destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1212 
1213 	if (is_sqp(dev, mqp))
1214 		kfree(to_msqp(mqp));
1215 	else
1216 		kfree(mqp);
1217 
1218 	return 0;
1219 }
1220 
1221 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1222 {
1223 	switch (type) {
1224 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1225 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1226 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1227 	case MLX4_IB_QPT_XRC_INI:
1228 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1229 	case MLX4_IB_QPT_SMI:
1230 	case MLX4_IB_QPT_GSI:
1231 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1232 
1233 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1234 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1235 						MLX4_QP_ST_MLX : -1);
1236 	case MLX4_IB_QPT_PROXY_SMI:
1237 	case MLX4_IB_QPT_TUN_SMI:
1238 	case MLX4_IB_QPT_PROXY_GSI:
1239 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1240 						MLX4_QP_ST_UD : -1);
1241 	default:			return -1;
1242 	}
1243 }
1244 
1245 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1246 				   int attr_mask)
1247 {
1248 	u8 dest_rd_atomic;
1249 	u32 access_flags;
1250 	u32 hw_access_flags = 0;
1251 
1252 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1253 		dest_rd_atomic = attr->max_dest_rd_atomic;
1254 	else
1255 		dest_rd_atomic = qp->resp_depth;
1256 
1257 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1258 		access_flags = attr->qp_access_flags;
1259 	else
1260 		access_flags = qp->atomic_rd_en;
1261 
1262 	if (!dest_rd_atomic)
1263 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1264 
1265 	if (access_flags & IB_ACCESS_REMOTE_READ)
1266 		hw_access_flags |= MLX4_QP_BIT_RRE;
1267 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1268 		hw_access_flags |= MLX4_QP_BIT_RAE;
1269 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1270 		hw_access_flags |= MLX4_QP_BIT_RWE;
1271 
1272 	return cpu_to_be32(hw_access_flags);
1273 }
1274 
1275 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1276 			    int attr_mask)
1277 {
1278 	if (attr_mask & IB_QP_PKEY_INDEX)
1279 		sqp->pkey_index = attr->pkey_index;
1280 	if (attr_mask & IB_QP_QKEY)
1281 		sqp->qkey = attr->qkey;
1282 	if (attr_mask & IB_QP_SQ_PSN)
1283 		sqp->send_psn = attr->sq_psn;
1284 }
1285 
1286 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1287 {
1288 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1289 }
1290 
1291 static int _mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
1292 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1293 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1294 {
1295 	int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
1296 		IB_LINK_LAYER_ETHERNET;
1297 	int vidx;
1298 	int smac_index;
1299 	int err;
1300 
1301 
1302 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
1303 	path->rlid	    = cpu_to_be16(ah->dlid);
1304 	if (ah->static_rate) {
1305 		path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
1306 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1307 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1308 			--path->static_rate;
1309 	} else
1310 		path->static_rate = 0;
1311 
1312 	if (ah->ah_flags & IB_AH_GRH) {
1313 		int real_sgid_index = mlx4_ib_gid_index_to_real_index(dev,
1314 								      port,
1315 								      ah->grh.sgid_index);
1316 
1317 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1318 			pr_err("sgid_index (%u) too large. max is %d\n",
1319 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1320 			return -1;
1321 		}
1322 
1323 		path->grh_mylmc |= 1 << 7;
1324 		path->mgid_index = real_sgid_index;
1325 		path->hop_limit  = ah->grh.hop_limit;
1326 		path->tclass_flowlabel =
1327 			cpu_to_be32((ah->grh.traffic_class << 20) |
1328 				    (ah->grh.flow_label));
1329 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1330 	}
1331 
1332 	if (is_eth) {
1333 		if (!(ah->ah_flags & IB_AH_GRH))
1334 			return -1;
1335 
1336 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1337 			((port - 1) << 6) | ((ah->sl & 7) << 3);
1338 
1339 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1340 		if (vlan_tag < 0x1000) {
1341 			if (smac_info->vid < 0x1000) {
1342 				/* both valid vlan ids */
1343 				if (smac_info->vid != vlan_tag) {
1344 					/* different VIDs.  unreg old and reg new */
1345 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1346 					if (err)
1347 						return err;
1348 					smac_info->candidate_vid = vlan_tag;
1349 					smac_info->candidate_vlan_index = vidx;
1350 					smac_info->candidate_vlan_port = port;
1351 					smac_info->update_vid = 1;
1352 					path->vlan_index = vidx;
1353 				} else {
1354 					path->vlan_index = smac_info->vlan_index;
1355 				}
1356 			} else {
1357 				/* no current vlan tag in qp */
1358 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1359 				if (err)
1360 					return err;
1361 				smac_info->candidate_vid = vlan_tag;
1362 				smac_info->candidate_vlan_index = vidx;
1363 				smac_info->candidate_vlan_port = port;
1364 				smac_info->update_vid = 1;
1365 				path->vlan_index = vidx;
1366 			}
1367 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1368 			path->fl = 1 << 6;
1369 		} else {
1370 			/* have current vlan tag. unregister it at modify-qp success */
1371 			if (smac_info->vid < 0x1000) {
1372 				smac_info->candidate_vid = 0xFFFF;
1373 				smac_info->update_vid = 1;
1374 			}
1375 		}
1376 
1377 		/* get smac_index for RoCE use.
1378 		 * If no smac was yet assigned, register one.
1379 		 * If one was already assigned, but the new mac differs,
1380 		 * unregister the old one and register the new one.
1381 		*/
1382 		if ((!smac_info->smac && !smac_info->smac_port) ||
1383 		    smac_info->smac != smac) {
1384 			/* register candidate now, unreg if needed, after success */
1385 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1386 			if (smac_index >= 0) {
1387 				smac_info->candidate_smac_index = smac_index;
1388 				smac_info->candidate_smac = smac;
1389 				smac_info->candidate_smac_port = port;
1390 			} else {
1391 				return -EINVAL;
1392 			}
1393 		} else {
1394 			smac_index = smac_info->smac_index;
1395 		}
1396 
1397 		memcpy(path->dmac, ah->dmac, 6);
1398 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1399 		/* put MAC table smac index for IBoE */
1400 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1401 	} else {
1402 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1403 			((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1404 	}
1405 
1406 	return 0;
1407 }
1408 
1409 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1410 			 enum ib_qp_attr_mask qp_attr_mask,
1411 			 struct mlx4_ib_qp *mqp,
1412 			 struct mlx4_qp_path *path, u8 port,
1413 			 u16 vlan_id, u8 *smac)
1414 {
1415 	return _mlx4_set_path(dev, &qp->ah_attr,
1416 			      mlx4_mac_to_u64(smac),
1417 			      vlan_id,
1418 			      path, &mqp->pri, port);
1419 }
1420 
1421 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1422 			     const struct ib_qp_attr *qp,
1423 			     enum ib_qp_attr_mask qp_attr_mask,
1424 			     struct mlx4_ib_qp *mqp,
1425 			     struct mlx4_qp_path *path, u8 port)
1426 {
1427 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1428 			      0,
1429 			      0xffff,
1430 			      path, &mqp->alt, port);
1431 }
1432 
1433 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1434 {
1435 	struct mlx4_ib_gid_entry *ge, *tmp;
1436 
1437 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1438 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1439 			ge->added = 1;
1440 			ge->port = qp->port;
1441 		}
1442 	}
1443 }
1444 
1445 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1446 				    struct mlx4_ib_qp *qp,
1447 				    struct mlx4_qp_context *context)
1448 {
1449 	u64 u64_mac;
1450 	int smac_index;
1451 
1452 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1453 
1454 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1455 	if (!qp->pri.smac && !qp->pri.smac_port) {
1456 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1457 		if (smac_index >= 0) {
1458 			qp->pri.candidate_smac_index = smac_index;
1459 			qp->pri.candidate_smac = u64_mac;
1460 			qp->pri.candidate_smac_port = qp->port;
1461 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1462 		} else {
1463 			return -ENOENT;
1464 		}
1465 	}
1466 	return 0;
1467 }
1468 
1469 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1470 {
1471 	struct counter_index *new_counter_index;
1472 	int err;
1473 	u32 tmp_idx;
1474 
1475 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1476 	    IB_LINK_LAYER_ETHERNET ||
1477 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1478 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1479 		return 0;
1480 
1481 	err = mlx4_counter_alloc(dev->dev, &tmp_idx);
1482 	if (err)
1483 		return err;
1484 
1485 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1486 	if (!new_counter_index) {
1487 		mlx4_counter_free(dev->dev, tmp_idx);
1488 		return -ENOMEM;
1489 	}
1490 
1491 	new_counter_index->index = tmp_idx;
1492 	new_counter_index->allocated = 1;
1493 	qp->counter_index = new_counter_index;
1494 
1495 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
1496 	list_add_tail(&new_counter_index->list,
1497 		      &dev->counters_table[qp->port - 1].counters_list);
1498 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
1499 
1500 	return 0;
1501 }
1502 
1503 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1504 			       const struct ib_qp_attr *attr, int attr_mask,
1505 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1506 {
1507 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1508 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1509 	struct mlx4_ib_pd *pd;
1510 	struct mlx4_ib_cq *send_cq, *recv_cq;
1511 	struct mlx4_qp_context *context;
1512 	enum mlx4_qp_optpar optpar = 0;
1513 	int sqd_event;
1514 	int steer_qp = 0;
1515 	int err = -EINVAL;
1516 	int counter_index;
1517 
1518 	/* APM is not supported under RoCE */
1519 	if (attr_mask & IB_QP_ALT_PATH &&
1520 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1521 	    IB_LINK_LAYER_ETHERNET)
1522 		return -ENOTSUPP;
1523 
1524 	context = kzalloc(sizeof *context, GFP_KERNEL);
1525 	if (!context)
1526 		return -ENOMEM;
1527 
1528 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1529 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1530 
1531 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1532 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1533 	else {
1534 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
1535 		switch (attr->path_mig_state) {
1536 		case IB_MIG_MIGRATED:
1537 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1538 			break;
1539 		case IB_MIG_REARM:
1540 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1541 			break;
1542 		case IB_MIG_ARMED:
1543 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1544 			break;
1545 		}
1546 	}
1547 
1548 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1549 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1550 	else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1551 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1552 	else if (ibqp->qp_type == IB_QPT_UD) {
1553 		if (qp->flags & MLX4_IB_QP_LSO)
1554 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
1555 					      ilog2(dev->dev->caps.max_gso_sz);
1556 		else
1557 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1558 	} else if (attr_mask & IB_QP_PATH_MTU) {
1559 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1560 			pr_err("path MTU (%u) is invalid\n",
1561 			       attr->path_mtu);
1562 			goto out;
1563 		}
1564 		context->mtu_msgmax = (attr->path_mtu << 5) |
1565 			ilog2(dev->dev->caps.max_msg_sz);
1566 	}
1567 
1568 	if (qp->rq.wqe_cnt)
1569 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1570 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
1571 
1572 	if (qp->sq.wqe_cnt)
1573 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1574 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
1575 
1576 	if (new_state == IB_QPS_RESET && qp->counter_index)
1577 		mlx4_ib_free_qp_counter(dev, qp);
1578 
1579 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1580 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1581 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1582 		if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1583 			context->param3 |= cpu_to_be32(1 << 30);
1584 	}
1585 
1586 	if (qp->ibqp.uobject)
1587 		context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
1588 	else
1589 		context->usr_page = cpu_to_be32(dev->priv_uar.index);
1590 
1591 	if (attr_mask & IB_QP_DEST_QPN)
1592 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1593 
1594 	if (attr_mask & IB_QP_PORT) {
1595 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1596 		    !(attr_mask & IB_QP_AV)) {
1597 			mlx4_set_sched(&context->pri_path, attr->port_num);
1598 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1599 		}
1600 	}
1601 
1602 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1603 		err = create_qp_lb_counter(dev, qp);
1604 		if (err)
1605 			goto out;
1606 
1607 		counter_index =
1608 			dev->counters_table[qp->port - 1].default_counter;
1609 		if (qp->counter_index)
1610 			counter_index = qp->counter_index->index;
1611 
1612 		if (counter_index != -1) {
1613 			context->pri_path.counter_index = counter_index;
1614 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1615 			if (qp->counter_index) {
1616 				context->pri_path.fl |=
1617 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
1618 				context->pri_path.vlan_control |=
1619 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
1620 			}
1621 		} else
1622 			context->pri_path.counter_index =
1623 				MLX4_SINK_COUNTER_INDEX(dev->dev);
1624 
1625 		if (qp->flags & MLX4_IB_QP_NETIF) {
1626 			mlx4_ib_steer_qp_reg(dev, qp, 1);
1627 			steer_qp = 1;
1628 		}
1629 	}
1630 
1631 	if (attr_mask & IB_QP_PKEY_INDEX) {
1632 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1633 			context->pri_path.disable_pkey_check = 0x40;
1634 		context->pri_path.pkey_index = attr->pkey_index;
1635 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1636 	}
1637 
1638 	if (attr_mask & IB_QP_AV) {
1639 		u8 port_num = mlx4_is_bonded(to_mdev(ibqp->device)->dev) ? 1 :
1640 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1641 		union ib_gid gid;
1642 		struct ib_gid_attr gid_attr;
1643 		u16 vlan = 0xffff;
1644 		u8 smac[ETH_ALEN];
1645 		int status = 0;
1646 
1647 		if (rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
1648 		    attr->ah_attr.ah_flags & IB_AH_GRH) {
1649 			int index = attr->ah_attr.grh.sgid_index;
1650 
1651 			status = ib_get_cached_gid(ibqp->device, port_num,
1652 						   index, &gid, &gid_attr);
1653 			if (!status && !memcmp(&gid, &zgid, sizeof(gid)))
1654 				status = -ENOENT;
1655 			if (!status && gid_attr.ndev) {
1656 				vlan = rdma_vlan_dev_vlan_id(gid_attr.ndev);
1657 				memcpy(smac, gid_attr.ndev->dev_addr, ETH_ALEN);
1658 				dev_put(gid_attr.ndev);
1659 			}
1660 		}
1661 		if (status)
1662 			goto out;
1663 
1664 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
1665 				  port_num, vlan, smac))
1666 			goto out;
1667 
1668 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1669 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
1670 	}
1671 
1672 	if (attr_mask & IB_QP_TIMEOUT) {
1673 		context->pri_path.ackto |= attr->timeout << 3;
1674 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1675 	}
1676 
1677 	if (attr_mask & IB_QP_ALT_PATH) {
1678 		if (attr->alt_port_num == 0 ||
1679 		    attr->alt_port_num > dev->dev->caps.num_ports)
1680 			goto out;
1681 
1682 		if (attr->alt_pkey_index >=
1683 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
1684 			goto out;
1685 
1686 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
1687 				      &context->alt_path,
1688 				      attr->alt_port_num))
1689 			goto out;
1690 
1691 		context->alt_path.pkey_index = attr->alt_pkey_index;
1692 		context->alt_path.ackto = attr->alt_timeout << 3;
1693 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1694 	}
1695 
1696 	pd = get_pd(qp);
1697 	get_cqs(qp, &send_cq, &recv_cq);
1698 	context->pd       = cpu_to_be32(pd->pdn);
1699 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1700 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1701 	context->params1  = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1702 
1703 	/* Set "fast registration enabled" for all kernel QPs */
1704 	if (!qp->ibqp.uobject)
1705 		context->params1 |= cpu_to_be32(1 << 11);
1706 
1707 	if (attr_mask & IB_QP_RNR_RETRY) {
1708 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1709 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1710 	}
1711 
1712 	if (attr_mask & IB_QP_RETRY_CNT) {
1713 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1714 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1715 	}
1716 
1717 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1718 		if (attr->max_rd_atomic)
1719 			context->params1 |=
1720 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1721 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1722 	}
1723 
1724 	if (attr_mask & IB_QP_SQ_PSN)
1725 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1726 
1727 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1728 		if (attr->max_dest_rd_atomic)
1729 			context->params2 |=
1730 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1731 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1732 	}
1733 
1734 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1735 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1736 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1737 	}
1738 
1739 	if (ibqp->srq)
1740 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1741 
1742 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1743 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1744 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1745 	}
1746 	if (attr_mask & IB_QP_RQ_PSN)
1747 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1748 
1749 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1750 	if (attr_mask & IB_QP_QKEY) {
1751 		if (qp->mlx4_ib_qp_type &
1752 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1753 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1754 		else {
1755 			if (mlx4_is_mfunc(dev->dev) &&
1756 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1757 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1758 			    MLX4_RESERVED_QKEY_BASE) {
1759 				pr_err("Cannot use reserved QKEY"
1760 				       " 0x%x (range 0xffff0000..0xffffffff"
1761 				       " is reserved)\n", attr->qkey);
1762 				err = -EINVAL;
1763 				goto out;
1764 			}
1765 			context->qkey = cpu_to_be32(attr->qkey);
1766 		}
1767 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
1768 	}
1769 
1770 	if (ibqp->srq)
1771 		context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1772 
1773 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1774 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1775 
1776 	if (cur_state == IB_QPS_INIT &&
1777 	    new_state == IB_QPS_RTR  &&
1778 	    (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1779 	     ibqp->qp_type == IB_QPT_UD ||
1780 	     ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1781 		context->pri_path.sched_queue = (qp->port - 1) << 6;
1782 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1783 		    qp->mlx4_ib_qp_type &
1784 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1785 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1786 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1787 				context->pri_path.fl = 0x80;
1788 		} else {
1789 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1790 				context->pri_path.fl = 0x80;
1791 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1792 		}
1793 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1794 		    IB_LINK_LAYER_ETHERNET) {
1795 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
1796 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
1797 				context->pri_path.feup = 1 << 7; /* don't fsm */
1798 			/* handle smac_index */
1799 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
1800 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
1801 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
1802 				err = handle_eth_ud_smac_index(dev, qp, context);
1803 				if (err) {
1804 					err = -EINVAL;
1805 					goto out;
1806 				}
1807 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1808 					dev->qp1_proxy[qp->port - 1] = qp;
1809 			}
1810 		}
1811 	}
1812 
1813 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1814 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
1815 					MLX4_IB_LINK_TYPE_ETH;
1816 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1817 			/* set QP to receive both tunneled & non-tunneled packets */
1818 			if (!(context->flags & cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET)))
1819 				context->srqn = cpu_to_be32(7 << 28);
1820 		}
1821 	}
1822 
1823 	if (ibqp->qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
1824 		int is_eth = rdma_port_get_link_layer(
1825 				&dev->ib_dev, qp->port) ==
1826 				IB_LINK_LAYER_ETHERNET;
1827 		if (is_eth) {
1828 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
1829 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
1830 		}
1831 	}
1832 
1833 
1834 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1835 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1836 		sqd_event = 1;
1837 	else
1838 		sqd_event = 0;
1839 
1840 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1841 		context->rlkey |= (1 << 4);
1842 
1843 	/*
1844 	 * Before passing a kernel QP to the HW, make sure that the
1845 	 * ownership bits of the send queue are set and the SQ
1846 	 * headroom is stamped so that the hardware doesn't start
1847 	 * processing stale work requests.
1848 	 */
1849 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1850 		struct mlx4_wqe_ctrl_seg *ctrl;
1851 		int i;
1852 
1853 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1854 			ctrl = get_send_wqe(qp, i);
1855 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
1856 			if (qp->sq_max_wqes_per_wr == 1)
1857 				ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1858 
1859 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1860 		}
1861 	}
1862 
1863 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1864 			     to_mlx4_state(new_state), context, optpar,
1865 			     sqd_event, &qp->mqp);
1866 	if (err)
1867 		goto out;
1868 
1869 	qp->state = new_state;
1870 
1871 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1872 		qp->atomic_rd_en = attr->qp_access_flags;
1873 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1874 		qp->resp_depth = attr->max_dest_rd_atomic;
1875 	if (attr_mask & IB_QP_PORT) {
1876 		qp->port = attr->port_num;
1877 		update_mcg_macs(dev, qp);
1878 	}
1879 	if (attr_mask & IB_QP_ALT_PATH)
1880 		qp->alt_port = attr->alt_port_num;
1881 
1882 	if (is_sqp(dev, qp))
1883 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1884 
1885 	/*
1886 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
1887 	 * QP0 to RESET or ERROR, bring the link back down.
1888 	 */
1889 	if (is_qp0(dev, qp)) {
1890 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1891 			if (mlx4_INIT_PORT(dev->dev, qp->port))
1892 				pr_warn("INIT_PORT failed for port %d\n",
1893 				       qp->port);
1894 
1895 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1896 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1897 			mlx4_CLOSE_PORT(dev->dev, qp->port);
1898 	}
1899 
1900 	/*
1901 	 * If we moved a kernel QP to RESET, clean up all old CQ
1902 	 * entries and reinitialize the QP.
1903 	 */
1904 	if (new_state == IB_QPS_RESET) {
1905 		if (!ibqp->uobject) {
1906 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1907 					 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
1908 			if (send_cq != recv_cq)
1909 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1910 
1911 			qp->rq.head = 0;
1912 			qp->rq.tail = 0;
1913 			qp->sq.head = 0;
1914 			qp->sq.tail = 0;
1915 			qp->sq_next_wqe = 0;
1916 			if (qp->rq.wqe_cnt)
1917 				*qp->db.db  = 0;
1918 
1919 			if (qp->flags & MLX4_IB_QP_NETIF)
1920 				mlx4_ib_steer_qp_reg(dev, qp, 0);
1921 		}
1922 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1923 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1924 			qp->pri.smac = 0;
1925 			qp->pri.smac_port = 0;
1926 		}
1927 		if (qp->alt.smac) {
1928 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1929 			qp->alt.smac = 0;
1930 		}
1931 		if (qp->pri.vid < 0x1000) {
1932 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1933 			qp->pri.vid = 0xFFFF;
1934 			qp->pri.candidate_vid = 0xFFFF;
1935 			qp->pri.update_vid = 0;
1936 		}
1937 
1938 		if (qp->alt.vid < 0x1000) {
1939 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1940 			qp->alt.vid = 0xFFFF;
1941 			qp->alt.candidate_vid = 0xFFFF;
1942 			qp->alt.update_vid = 0;
1943 		}
1944 	}
1945 out:
1946 	if (err && qp->counter_index)
1947 		mlx4_ib_free_qp_counter(dev, qp);
1948 	if (err && steer_qp)
1949 		mlx4_ib_steer_qp_reg(dev, qp, 0);
1950 	kfree(context);
1951 	if (qp->pri.candidate_smac ||
1952 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
1953 		if (err) {
1954 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
1955 		} else {
1956 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
1957 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1958 			qp->pri.smac = qp->pri.candidate_smac;
1959 			qp->pri.smac_index = qp->pri.candidate_smac_index;
1960 			qp->pri.smac_port = qp->pri.candidate_smac_port;
1961 		}
1962 		qp->pri.candidate_smac = 0;
1963 		qp->pri.candidate_smac_index = 0;
1964 		qp->pri.candidate_smac_port = 0;
1965 	}
1966 	if (qp->alt.candidate_smac) {
1967 		if (err) {
1968 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
1969 		} else {
1970 			if (qp->alt.smac)
1971 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1972 			qp->alt.smac = qp->alt.candidate_smac;
1973 			qp->alt.smac_index = qp->alt.candidate_smac_index;
1974 			qp->alt.smac_port = qp->alt.candidate_smac_port;
1975 		}
1976 		qp->alt.candidate_smac = 0;
1977 		qp->alt.candidate_smac_index = 0;
1978 		qp->alt.candidate_smac_port = 0;
1979 	}
1980 
1981 	if (qp->pri.update_vid) {
1982 		if (err) {
1983 			if (qp->pri.candidate_vid < 0x1000)
1984 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
1985 						     qp->pri.candidate_vid);
1986 		} else {
1987 			if (qp->pri.vid < 0x1000)
1988 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
1989 						     qp->pri.vid);
1990 			qp->pri.vid = qp->pri.candidate_vid;
1991 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
1992 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
1993 		}
1994 		qp->pri.candidate_vid = 0xFFFF;
1995 		qp->pri.update_vid = 0;
1996 	}
1997 
1998 	if (qp->alt.update_vid) {
1999 		if (err) {
2000 			if (qp->alt.candidate_vid < 0x1000)
2001 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2002 						     qp->alt.candidate_vid);
2003 		} else {
2004 			if (qp->alt.vid < 0x1000)
2005 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2006 						     qp->alt.vid);
2007 			qp->alt.vid = qp->alt.candidate_vid;
2008 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2009 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2010 		}
2011 		qp->alt.candidate_vid = 0xFFFF;
2012 		qp->alt.update_vid = 0;
2013 	}
2014 
2015 	return err;
2016 }
2017 
2018 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2019 		      int attr_mask, struct ib_udata *udata)
2020 {
2021 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2022 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2023 	enum ib_qp_state cur_state, new_state;
2024 	int err = -EINVAL;
2025 	int ll;
2026 	mutex_lock(&qp->mutex);
2027 
2028 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2029 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2030 
2031 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2032 		ll = IB_LINK_LAYER_UNSPECIFIED;
2033 	} else {
2034 		int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2035 		ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2036 	}
2037 
2038 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2039 				attr_mask, ll)) {
2040 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2041 			 "for transition %d to %d. qp_type %d,"
2042 			 " attr_mask 0x%x\n",
2043 			 ibqp->qp_num, cur_state, new_state,
2044 			 ibqp->qp_type, attr_mask);
2045 		goto out;
2046 	}
2047 
2048 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2049 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2050 			if ((ibqp->qp_type == IB_QPT_RC) ||
2051 			    (ibqp->qp_type == IB_QPT_UD) ||
2052 			    (ibqp->qp_type == IB_QPT_UC) ||
2053 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2054 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2055 				attr->port_num = mlx4_ib_bond_next_port(dev);
2056 			}
2057 		} else {
2058 			/* no sense in changing port_num
2059 			 * when ports are bonded */
2060 			attr_mask &= ~IB_QP_PORT;
2061 		}
2062 	}
2063 
2064 	if ((attr_mask & IB_QP_PORT) &&
2065 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2066 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2067 			 "for transition %d to %d. qp_type %d\n",
2068 			 ibqp->qp_num, attr->port_num, cur_state,
2069 			 new_state, ibqp->qp_type);
2070 		goto out;
2071 	}
2072 
2073 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2074 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2075 	     IB_LINK_LAYER_ETHERNET))
2076 		goto out;
2077 
2078 	if (attr_mask & IB_QP_PKEY_INDEX) {
2079 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2080 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2081 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2082 				 "for transition %d to %d. qp_type %d\n",
2083 				 ibqp->qp_num, attr->pkey_index, cur_state,
2084 				 new_state, ibqp->qp_type);
2085 			goto out;
2086 		}
2087 	}
2088 
2089 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2090 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2091 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2092 			 "Transition %d to %d. qp_type %d\n",
2093 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2094 			 new_state, ibqp->qp_type);
2095 		goto out;
2096 	}
2097 
2098 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2099 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2100 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2101 			 "Transition %d to %d. qp_type %d\n",
2102 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2103 			 new_state, ibqp->qp_type);
2104 		goto out;
2105 	}
2106 
2107 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2108 		err = 0;
2109 		goto out;
2110 	}
2111 
2112 	err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2113 
2114 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2115 		attr->port_num = 1;
2116 
2117 out:
2118 	mutex_unlock(&qp->mutex);
2119 	return err;
2120 }
2121 
2122 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2123 {
2124 	int i;
2125 	for (i = 0; i < dev->caps.num_ports; i++) {
2126 		if (qpn == dev->caps.qp0_proxy[i] ||
2127 		    qpn == dev->caps.qp0_tunnel[i]) {
2128 			*qkey = dev->caps.qp0_qkey[i];
2129 			return 0;
2130 		}
2131 	}
2132 	return -EINVAL;
2133 }
2134 
2135 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2136 				  struct ib_ud_wr *wr,
2137 				  void *wqe, unsigned *mlx_seg_len)
2138 {
2139 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2140 	struct ib_device *ib_dev = &mdev->ib_dev;
2141 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2142 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2143 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2144 	u16 pkey;
2145 	u32 qkey;
2146 	int send_size;
2147 	int header_size;
2148 	int spc;
2149 	int i;
2150 
2151 	if (wr->wr.opcode != IB_WR_SEND)
2152 		return -EINVAL;
2153 
2154 	send_size = 0;
2155 
2156 	for (i = 0; i < wr->wr.num_sge; ++i)
2157 		send_size += wr->wr.sg_list[i].length;
2158 
2159 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2160 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2161 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2162 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2163 
2164 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, &sqp->ud_header);
2165 
2166 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2167 		sqp->ud_header.lrh.service_level =
2168 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2169 		sqp->ud_header.lrh.destination_lid =
2170 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2171 		sqp->ud_header.lrh.source_lid =
2172 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2173 	}
2174 
2175 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2176 
2177 	/* force loopback */
2178 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2179 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2180 
2181 	sqp->ud_header.lrh.virtual_lane    = 0;
2182 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2183 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2184 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2185 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2186 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2187 	else
2188 		sqp->ud_header.bth.destination_qpn =
2189 			cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
2190 
2191 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2192 	if (mlx4_is_master(mdev->dev)) {
2193 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2194 			return -EINVAL;
2195 	} else {
2196 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2197 			return -EINVAL;
2198 	}
2199 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2200 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2201 
2202 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2203 	sqp->ud_header.immediate_present = 0;
2204 
2205 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2206 
2207 	/*
2208 	 * Inline data segments may not cross a 64 byte boundary.  If
2209 	 * our UD header is bigger than the space available up to the
2210 	 * next 64 byte boundary in the WQE, use two inline data
2211 	 * segments to hold the UD header.
2212 	 */
2213 	spc = MLX4_INLINE_ALIGN -
2214 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2215 	if (header_size <= spc) {
2216 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2217 		memcpy(inl + 1, sqp->header_buf, header_size);
2218 		i = 1;
2219 	} else {
2220 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2221 		memcpy(inl + 1, sqp->header_buf, spc);
2222 
2223 		inl = (void *) (inl + 1) + spc;
2224 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2225 		/*
2226 		 * Need a barrier here to make sure all the data is
2227 		 * visible before the byte_count field is set.
2228 		 * Otherwise the HCA prefetcher could grab the 64-byte
2229 		 * chunk with this inline segment and get a valid (!=
2230 		 * 0xffffffff) byte count but stale data, and end up
2231 		 * generating a packet with bad headers.
2232 		 *
2233 		 * The first inline segment's byte_count field doesn't
2234 		 * need a barrier, because it comes after a
2235 		 * control/MLX segment and therefore is at an offset
2236 		 * of 16 mod 64.
2237 		 */
2238 		wmb();
2239 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2240 		i = 2;
2241 	}
2242 
2243 	*mlx_seg_len =
2244 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2245 	return 0;
2246 }
2247 
2248 static void mlx4_u64_to_smac(u8 *dst_mac, u64 src_mac)
2249 {
2250 	int i;
2251 
2252 	for (i = ETH_ALEN; i; i--) {
2253 		dst_mac[i - 1] = src_mac & 0xff;
2254 		src_mac >>= 8;
2255 	}
2256 }
2257 
2258 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_ud_wr *wr,
2259 			    void *wqe, unsigned *mlx_seg_len)
2260 {
2261 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
2262 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2263 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2264 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2265 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2266 	union ib_gid sgid;
2267 	u16 pkey;
2268 	int send_size;
2269 	int header_size;
2270 	int spc;
2271 	int i;
2272 	int err = 0;
2273 	u16 vlan = 0xffff;
2274 	bool is_eth;
2275 	bool is_vlan = false;
2276 	bool is_grh;
2277 
2278 	send_size = 0;
2279 	for (i = 0; i < wr->wr.num_sge; ++i)
2280 		send_size += wr->wr.sg_list[i].length;
2281 
2282 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2283 	is_grh = mlx4_ib_ah_grh_present(ah);
2284 	if (is_eth) {
2285 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2286 			/* When multi-function is enabled, the ib_core gid
2287 			 * indexes don't necessarily match the hw ones, so
2288 			 * we must use our own cache */
2289 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2290 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
2291 							   ah->av.ib.gid_index, &sgid.raw[0]);
2292 			if (err)
2293 				return err;
2294 		} else  {
2295 			err = ib_get_cached_gid(ib_dev,
2296 						be32_to_cpu(ah->av.ib.port_pd) >> 24,
2297 						ah->av.ib.gid_index, &sgid,
2298 						NULL);
2299 			if (!err && !memcmp(&sgid, &zgid, sizeof(sgid)))
2300 				err = -ENOENT;
2301 			if (err)
2302 				return err;
2303 		}
2304 
2305 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
2306 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
2307 			is_vlan = 1;
2308 		}
2309 	}
2310 	ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh, 0, &sqp->ud_header);
2311 
2312 	if (!is_eth) {
2313 		sqp->ud_header.lrh.service_level =
2314 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2315 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
2316 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2317 	}
2318 
2319 	if (is_grh) {
2320 		sqp->ud_header.grh.traffic_class =
2321 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2322 		sqp->ud_header.grh.flow_label    =
2323 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
2324 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
2325 		if (is_eth)
2326 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
2327 		else {
2328 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2329 			/* When multi-function is enabled, the ib_core gid
2330 			 * indexes don't necessarily match the hw ones, so
2331 			 * we must use our own cache */
2332 			sqp->ud_header.grh.source_gid.global.subnet_prefix =
2333 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2334 						       subnet_prefix;
2335 			sqp->ud_header.grh.source_gid.global.interface_id =
2336 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2337 					       guid_cache[ah->av.ib.gid_index];
2338 		} else
2339 			ib_get_cached_gid(ib_dev,
2340 					  be32_to_cpu(ah->av.ib.port_pd) >> 24,
2341 					  ah->av.ib.gid_index,
2342 					  &sqp->ud_header.grh.source_gid, NULL);
2343 		}
2344 		memcpy(sqp->ud_header.grh.destination_gid.raw,
2345 		       ah->av.ib.dgid, 16);
2346 	}
2347 
2348 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2349 
2350 	if (!is_eth) {
2351 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
2352 					  (sqp->ud_header.lrh.destination_lid ==
2353 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
2354 					  (sqp->ud_header.lrh.service_level << 8));
2355 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
2356 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
2357 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
2358 	}
2359 
2360 	switch (wr->wr.opcode) {
2361 	case IB_WR_SEND:
2362 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
2363 		sqp->ud_header.immediate_present = 0;
2364 		break;
2365 	case IB_WR_SEND_WITH_IMM:
2366 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2367 		sqp->ud_header.immediate_present = 1;
2368 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
2369 		break;
2370 	default:
2371 		return -EINVAL;
2372 	}
2373 
2374 	if (is_eth) {
2375 		struct in6_addr in6;
2376 
2377 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
2378 
2379 		mlx->sched_prio = cpu_to_be16(pcp);
2380 
2381 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
2382 		/* FIXME: cache smac value? */
2383 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
2384 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
2385 		memcpy(&in6, sgid.raw, sizeof(in6));
2386 
2387 		if (!mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2388 			u64 mac = atomic64_read(&to_mdev(ib_dev)->iboe.mac[sqp->qp.port - 1]);
2389 			u8 smac[ETH_ALEN];
2390 
2391 			mlx4_u64_to_smac(smac, mac);
2392 			memcpy(sqp->ud_header.eth.smac_h, smac, ETH_ALEN);
2393 		} else {
2394 			/* use the src mac of the tunnel */
2395 			memcpy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac, ETH_ALEN);
2396 		}
2397 
2398 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
2399 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
2400 		if (!is_vlan) {
2401 			sqp->ud_header.eth.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2402 		} else {
2403 			sqp->ud_header.vlan.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2404 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
2405 		}
2406 	} else {
2407 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
2408 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
2409 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
2410 	}
2411 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2412 	if (!sqp->qp.ibqp.qp_num)
2413 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
2414 	else
2415 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
2416 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2417 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2418 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2419 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
2420 					       sqp->qkey : wr->remote_qkey);
2421 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
2422 
2423 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2424 
2425 	if (0) {
2426 		pr_err("built UD header of size %d:\n", header_size);
2427 		for (i = 0; i < header_size / 4; ++i) {
2428 			if (i % 8 == 0)
2429 				pr_err("  [%02x] ", i * 4);
2430 			pr_cont(" %08x",
2431 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
2432 			if ((i + 1) % 8 == 0)
2433 				pr_cont("\n");
2434 		}
2435 		pr_err("\n");
2436 	}
2437 
2438 	/*
2439 	 * Inline data segments may not cross a 64 byte boundary.  If
2440 	 * our UD header is bigger than the space available up to the
2441 	 * next 64 byte boundary in the WQE, use two inline data
2442 	 * segments to hold the UD header.
2443 	 */
2444 	spc = MLX4_INLINE_ALIGN -
2445 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2446 	if (header_size <= spc) {
2447 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2448 		memcpy(inl + 1, sqp->header_buf, header_size);
2449 		i = 1;
2450 	} else {
2451 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2452 		memcpy(inl + 1, sqp->header_buf, spc);
2453 
2454 		inl = (void *) (inl + 1) + spc;
2455 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2456 		/*
2457 		 * Need a barrier here to make sure all the data is
2458 		 * visible before the byte_count field is set.
2459 		 * Otherwise the HCA prefetcher could grab the 64-byte
2460 		 * chunk with this inline segment and get a valid (!=
2461 		 * 0xffffffff) byte count but stale data, and end up
2462 		 * generating a packet with bad headers.
2463 		 *
2464 		 * The first inline segment's byte_count field doesn't
2465 		 * need a barrier, because it comes after a
2466 		 * control/MLX segment and therefore is at an offset
2467 		 * of 16 mod 64.
2468 		 */
2469 		wmb();
2470 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2471 		i = 2;
2472 	}
2473 
2474 	*mlx_seg_len =
2475 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2476 	return 0;
2477 }
2478 
2479 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2480 {
2481 	unsigned cur;
2482 	struct mlx4_ib_cq *cq;
2483 
2484 	cur = wq->head - wq->tail;
2485 	if (likely(cur + nreq < wq->max_post))
2486 		return 0;
2487 
2488 	cq = to_mcq(ib_cq);
2489 	spin_lock(&cq->lock);
2490 	cur = wq->head - wq->tail;
2491 	spin_unlock(&cq->lock);
2492 
2493 	return cur + nreq >= wq->max_post;
2494 }
2495 
2496 static __be32 convert_access(int acc)
2497 {
2498 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
2499 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
2500 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
2501 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
2502 	       (acc & IB_ACCESS_REMOTE_READ   ?
2503 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
2504 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
2505 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
2506 }
2507 
2508 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
2509 			struct ib_reg_wr *wr)
2510 {
2511 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
2512 
2513 	fseg->flags		= convert_access(wr->access);
2514 	fseg->mem_key		= cpu_to_be32(wr->key);
2515 	fseg->buf_list		= cpu_to_be64(mr->page_map);
2516 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
2517 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
2518 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
2519 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
2520 	fseg->reserved[0]	= 0;
2521 	fseg->reserved[1]	= 0;
2522 }
2523 
2524 static void set_bind_seg(struct mlx4_wqe_bind_seg *bseg,
2525 		struct ib_bind_mw_wr *wr)
2526 {
2527 	bseg->flags1 =
2528 		convert_access(wr->bind_info.mw_access_flags) &
2529 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ  |
2530 			    MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE |
2531 			    MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC);
2532 	bseg->flags2 = 0;
2533 	if (wr->mw->type == IB_MW_TYPE_2)
2534 		bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_TYPE_2);
2535 	if (wr->bind_info.mw_access_flags & IB_ZERO_BASED)
2536 		bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_ZERO_BASED);
2537 	bseg->new_rkey = cpu_to_be32(wr->rkey);
2538 	bseg->lkey = cpu_to_be32(wr->bind_info.mr->lkey);
2539 	bseg->addr = cpu_to_be64(wr->bind_info.addr);
2540 	bseg->length = cpu_to_be64(wr->bind_info.length);
2541 }
2542 
2543 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2544 {
2545 	memset(iseg, 0, sizeof(*iseg));
2546 	iseg->mem_key = cpu_to_be32(rkey);
2547 }
2548 
2549 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2550 					  u64 remote_addr, u32 rkey)
2551 {
2552 	rseg->raddr    = cpu_to_be64(remote_addr);
2553 	rseg->rkey     = cpu_to_be32(rkey);
2554 	rseg->reserved = 0;
2555 }
2556 
2557 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
2558 		struct ib_atomic_wr *wr)
2559 {
2560 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2561 		aseg->swap_add = cpu_to_be64(wr->swap);
2562 		aseg->compare  = cpu_to_be64(wr->compare_add);
2563 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2564 		aseg->swap_add = cpu_to_be64(wr->compare_add);
2565 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
2566 	} else {
2567 		aseg->swap_add = cpu_to_be64(wr->compare_add);
2568 		aseg->compare  = 0;
2569 	}
2570 
2571 }
2572 
2573 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2574 				  struct ib_atomic_wr *wr)
2575 {
2576 	aseg->swap_add		= cpu_to_be64(wr->swap);
2577 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
2578 	aseg->compare		= cpu_to_be64(wr->compare_add);
2579 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
2580 }
2581 
2582 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2583 			     struct ib_ud_wr *wr)
2584 {
2585 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
2586 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
2587 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
2588 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
2589 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
2590 }
2591 
2592 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2593 				    struct mlx4_wqe_datagram_seg *dseg,
2594 				    struct ib_ud_wr *wr,
2595 				    enum mlx4_ib_qp_type qpt)
2596 {
2597 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
2598 	struct mlx4_av sqp_av = {0};
2599 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
2600 
2601 	/* force loopback */
2602 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2603 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2604 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2605 			cpu_to_be32(0xf0000000);
2606 
2607 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2608 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
2609 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2610 	else
2611 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp0_tunnel[port - 1]);
2612 	/* Use QKEY from the QP context, which is set by master */
2613 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2614 }
2615 
2616 static void build_tunnel_header(struct ib_ud_wr *wr, void *wqe, unsigned *mlx_seg_len)
2617 {
2618 	struct mlx4_wqe_inline_seg *inl = wqe;
2619 	struct mlx4_ib_tunnel_header hdr;
2620 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2621 	int spc;
2622 	int i;
2623 
2624 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2625 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
2626 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
2627 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
2628 	memcpy(hdr.mac, ah->av.eth.mac, 6);
2629 	hdr.vlan = ah->av.eth.vlan;
2630 
2631 	spc = MLX4_INLINE_ALIGN -
2632 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2633 	if (sizeof (hdr) <= spc) {
2634 		memcpy(inl + 1, &hdr, sizeof (hdr));
2635 		wmb();
2636 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
2637 		i = 1;
2638 	} else {
2639 		memcpy(inl + 1, &hdr, spc);
2640 		wmb();
2641 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2642 
2643 		inl = (void *) (inl + 1) + spc;
2644 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2645 		wmb();
2646 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
2647 		i = 2;
2648 	}
2649 
2650 	*mlx_seg_len =
2651 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2652 }
2653 
2654 static void set_mlx_icrc_seg(void *dseg)
2655 {
2656 	u32 *t = dseg;
2657 	struct mlx4_wqe_inline_seg *iseg = dseg;
2658 
2659 	t[1] = 0;
2660 
2661 	/*
2662 	 * Need a barrier here before writing the byte_count field to
2663 	 * make sure that all the data is visible before the
2664 	 * byte_count field is set.  Otherwise, if the segment begins
2665 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2666 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2667 	 * stale data, and end up sending the wrong data.
2668 	 */
2669 	wmb();
2670 
2671 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2672 }
2673 
2674 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2675 {
2676 	dseg->lkey       = cpu_to_be32(sg->lkey);
2677 	dseg->addr       = cpu_to_be64(sg->addr);
2678 
2679 	/*
2680 	 * Need a barrier here before writing the byte_count field to
2681 	 * make sure that all the data is visible before the
2682 	 * byte_count field is set.  Otherwise, if the segment begins
2683 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2684 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2685 	 * stale data, and end up sending the wrong data.
2686 	 */
2687 	wmb();
2688 
2689 	dseg->byte_count = cpu_to_be32(sg->length);
2690 }
2691 
2692 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2693 {
2694 	dseg->byte_count = cpu_to_be32(sg->length);
2695 	dseg->lkey       = cpu_to_be32(sg->lkey);
2696 	dseg->addr       = cpu_to_be64(sg->addr);
2697 }
2698 
2699 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_ud_wr *wr,
2700 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2701 			 __be32 *lso_hdr_sz, __be32 *blh)
2702 {
2703 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
2704 
2705 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2706 		*blh = cpu_to_be32(1 << 6);
2707 
2708 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2709 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
2710 		return -EINVAL;
2711 
2712 	memcpy(wqe->header, wr->header, wr->hlen);
2713 
2714 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
2715 	*lso_seg_len = halign;
2716 	return 0;
2717 }
2718 
2719 static __be32 send_ieth(struct ib_send_wr *wr)
2720 {
2721 	switch (wr->opcode) {
2722 	case IB_WR_SEND_WITH_IMM:
2723 	case IB_WR_RDMA_WRITE_WITH_IMM:
2724 		return wr->ex.imm_data;
2725 
2726 	case IB_WR_SEND_WITH_INV:
2727 		return cpu_to_be32(wr->ex.invalidate_rkey);
2728 
2729 	default:
2730 		return 0;
2731 	}
2732 }
2733 
2734 static void add_zero_len_inline(void *wqe)
2735 {
2736 	struct mlx4_wqe_inline_seg *inl = wqe;
2737 	memset(wqe, 0, 16);
2738 	inl->byte_count = cpu_to_be32(1 << 31);
2739 }
2740 
2741 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2742 		      struct ib_send_wr **bad_wr)
2743 {
2744 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2745 	void *wqe;
2746 	struct mlx4_wqe_ctrl_seg *ctrl;
2747 	struct mlx4_wqe_data_seg *dseg;
2748 	unsigned long flags;
2749 	int nreq;
2750 	int err = 0;
2751 	unsigned ind;
2752 	int uninitialized_var(stamp);
2753 	int uninitialized_var(size);
2754 	unsigned uninitialized_var(seglen);
2755 	__be32 dummy;
2756 	__be32 *lso_wqe;
2757 	__be32 uninitialized_var(lso_hdr_sz);
2758 	__be32 blh;
2759 	int i;
2760 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
2761 
2762 	spin_lock_irqsave(&qp->sq.lock, flags);
2763 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
2764 		err = -EIO;
2765 		*bad_wr = wr;
2766 		nreq = 0;
2767 		goto out;
2768 	}
2769 
2770 	ind = qp->sq_next_wqe;
2771 
2772 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2773 		lso_wqe = &dummy;
2774 		blh = 0;
2775 
2776 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2777 			err = -ENOMEM;
2778 			*bad_wr = wr;
2779 			goto out;
2780 		}
2781 
2782 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
2783 			err = -EINVAL;
2784 			*bad_wr = wr;
2785 			goto out;
2786 		}
2787 
2788 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
2789 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
2790 
2791 		ctrl->srcrb_flags =
2792 			(wr->send_flags & IB_SEND_SIGNALED ?
2793 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
2794 			(wr->send_flags & IB_SEND_SOLICITED ?
2795 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
2796 			((wr->send_flags & IB_SEND_IP_CSUM) ?
2797 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
2798 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
2799 			qp->sq_signal_bits;
2800 
2801 		ctrl->imm = send_ieth(wr);
2802 
2803 		wqe += sizeof *ctrl;
2804 		size = sizeof *ctrl / 16;
2805 
2806 		switch (qp->mlx4_ib_qp_type) {
2807 		case MLX4_IB_QPT_RC:
2808 		case MLX4_IB_QPT_UC:
2809 			switch (wr->opcode) {
2810 			case IB_WR_ATOMIC_CMP_AND_SWP:
2811 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2812 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
2813 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
2814 					      atomic_wr(wr)->rkey);
2815 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2816 
2817 				set_atomic_seg(wqe, atomic_wr(wr));
2818 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
2819 
2820 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2821 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
2822 
2823 				break;
2824 
2825 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2826 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
2827 					      atomic_wr(wr)->rkey);
2828 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2829 
2830 				set_masked_atomic_seg(wqe, atomic_wr(wr));
2831 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
2832 
2833 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2834 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
2835 
2836 				break;
2837 
2838 			case IB_WR_RDMA_READ:
2839 			case IB_WR_RDMA_WRITE:
2840 			case IB_WR_RDMA_WRITE_WITH_IMM:
2841 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
2842 					      rdma_wr(wr)->rkey);
2843 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2844 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
2845 				break;
2846 
2847 			case IB_WR_LOCAL_INV:
2848 				ctrl->srcrb_flags |=
2849 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2850 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
2851 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
2852 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
2853 				break;
2854 
2855 			case IB_WR_REG_MR:
2856 				ctrl->srcrb_flags |=
2857 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2858 				set_reg_seg(wqe, reg_wr(wr));
2859 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
2860 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
2861 				break;
2862 
2863 			case IB_WR_BIND_MW:
2864 				ctrl->srcrb_flags |=
2865 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2866 				set_bind_seg(wqe, bind_mw_wr(wr));
2867 				wqe  += sizeof(struct mlx4_wqe_bind_seg);
2868 				size += sizeof(struct mlx4_wqe_bind_seg) / 16;
2869 				break;
2870 			default:
2871 				/* No extra segments required for sends */
2872 				break;
2873 			}
2874 			break;
2875 
2876 		case MLX4_IB_QPT_TUN_SMI_OWNER:
2877 			err =  build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
2878 					ctrl, &seglen);
2879 			if (unlikely(err)) {
2880 				*bad_wr = wr;
2881 				goto out;
2882 			}
2883 			wqe  += seglen;
2884 			size += seglen / 16;
2885 			break;
2886 		case MLX4_IB_QPT_TUN_SMI:
2887 		case MLX4_IB_QPT_TUN_GSI:
2888 			/* this is a UD qp used in MAD responses to slaves. */
2889 			set_datagram_seg(wqe, ud_wr(wr));
2890 			/* set the forced-loopback bit in the data seg av */
2891 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
2892 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2893 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2894 			break;
2895 		case MLX4_IB_QPT_UD:
2896 			set_datagram_seg(wqe, ud_wr(wr));
2897 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2898 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2899 
2900 			if (wr->opcode == IB_WR_LSO) {
2901 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
2902 						&lso_hdr_sz, &blh);
2903 				if (unlikely(err)) {
2904 					*bad_wr = wr;
2905 					goto out;
2906 				}
2907 				lso_wqe = (__be32 *) wqe;
2908 				wqe  += seglen;
2909 				size += seglen / 16;
2910 			}
2911 			break;
2912 
2913 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
2914 			err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
2915 					ctrl, &seglen);
2916 			if (unlikely(err)) {
2917 				*bad_wr = wr;
2918 				goto out;
2919 			}
2920 			wqe  += seglen;
2921 			size += seglen / 16;
2922 			/* to start tunnel header on a cache-line boundary */
2923 			add_zero_len_inline(wqe);
2924 			wqe += 16;
2925 			size++;
2926 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
2927 			wqe  += seglen;
2928 			size += seglen / 16;
2929 			break;
2930 		case MLX4_IB_QPT_PROXY_SMI:
2931 		case MLX4_IB_QPT_PROXY_GSI:
2932 			/* If we are tunneling special qps, this is a UD qp.
2933 			 * In this case we first add a UD segment targeting
2934 			 * the tunnel qp, and then add a header with address
2935 			 * information */
2936 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
2937 						ud_wr(wr),
2938 						qp->mlx4_ib_qp_type);
2939 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2940 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2941 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
2942 			wqe  += seglen;
2943 			size += seglen / 16;
2944 			break;
2945 
2946 		case MLX4_IB_QPT_SMI:
2947 		case MLX4_IB_QPT_GSI:
2948 			err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
2949 					&seglen);
2950 			if (unlikely(err)) {
2951 				*bad_wr = wr;
2952 				goto out;
2953 			}
2954 			wqe  += seglen;
2955 			size += seglen / 16;
2956 			break;
2957 
2958 		default:
2959 			break;
2960 		}
2961 
2962 		/*
2963 		 * Write data segments in reverse order, so as to
2964 		 * overwrite cacheline stamp last within each
2965 		 * cacheline.  This avoids issues with WQE
2966 		 * prefetching.
2967 		 */
2968 
2969 		dseg = wqe;
2970 		dseg += wr->num_sge - 1;
2971 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
2972 
2973 		/* Add one more inline data segment for ICRC for MLX sends */
2974 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2975 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
2976 			     qp->mlx4_ib_qp_type &
2977 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
2978 			set_mlx_icrc_seg(dseg + 1);
2979 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
2980 		}
2981 
2982 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
2983 			set_data_seg(dseg, wr->sg_list + i);
2984 
2985 		/*
2986 		 * Possibly overwrite stamping in cacheline with LSO
2987 		 * segment only after making sure all data segments
2988 		 * are written.
2989 		 */
2990 		wmb();
2991 		*lso_wqe = lso_hdr_sz;
2992 
2993 		ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
2994 				    MLX4_WQE_CTRL_FENCE : 0) | size;
2995 
2996 		/*
2997 		 * Make sure descriptor is fully written before
2998 		 * setting ownership bit (because HW can start
2999 		 * executing as soon as we do).
3000 		 */
3001 		wmb();
3002 
3003 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3004 			*bad_wr = wr;
3005 			err = -EINVAL;
3006 			goto out;
3007 		}
3008 
3009 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3010 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3011 
3012 		stamp = ind + qp->sq_spare_wqes;
3013 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
3014 
3015 		/*
3016 		 * We can improve latency by not stamping the last
3017 		 * send queue WQE until after ringing the doorbell, so
3018 		 * only stamp here if there are still more WQEs to post.
3019 		 *
3020 		 * Same optimization applies to padding with NOP wqe
3021 		 * in case of WQE shrinking (used to prevent wrap-around
3022 		 * in the middle of WR).
3023 		 */
3024 		if (wr->next) {
3025 			stamp_send_wqe(qp, stamp, size * 16);
3026 			ind = pad_wraparound(qp, ind);
3027 		}
3028 	}
3029 
3030 out:
3031 	if (likely(nreq)) {
3032 		qp->sq.head += nreq;
3033 
3034 		/*
3035 		 * Make sure that descriptors are written before
3036 		 * doorbell record.
3037 		 */
3038 		wmb();
3039 
3040 		writel(qp->doorbell_qpn,
3041 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3042 
3043 		/*
3044 		 * Make sure doorbells don't leak out of SQ spinlock
3045 		 * and reach the HCA out of order.
3046 		 */
3047 		mmiowb();
3048 
3049 		stamp_send_wqe(qp, stamp, size * 16);
3050 
3051 		ind = pad_wraparound(qp, ind);
3052 		qp->sq_next_wqe = ind;
3053 	}
3054 
3055 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3056 
3057 	return err;
3058 }
3059 
3060 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3061 		      struct ib_recv_wr **bad_wr)
3062 {
3063 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3064 	struct mlx4_wqe_data_seg *scat;
3065 	unsigned long flags;
3066 	int err = 0;
3067 	int nreq;
3068 	int ind;
3069 	int max_gs;
3070 	int i;
3071 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3072 
3073 	max_gs = qp->rq.max_gs;
3074 	spin_lock_irqsave(&qp->rq.lock, flags);
3075 
3076 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3077 		err = -EIO;
3078 		*bad_wr = wr;
3079 		nreq = 0;
3080 		goto out;
3081 	}
3082 
3083 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3084 
3085 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3086 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3087 			err = -ENOMEM;
3088 			*bad_wr = wr;
3089 			goto out;
3090 		}
3091 
3092 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3093 			err = -EINVAL;
3094 			*bad_wr = wr;
3095 			goto out;
3096 		}
3097 
3098 		scat = get_recv_wqe(qp, ind);
3099 
3100 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3101 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3102 			ib_dma_sync_single_for_device(ibqp->device,
3103 						      qp->sqp_proxy_rcv[ind].map,
3104 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3105 						      DMA_FROM_DEVICE);
3106 			scat->byte_count =
3107 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3108 			/* use dma lkey from upper layer entry */
3109 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3110 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3111 			scat++;
3112 			max_gs--;
3113 		}
3114 
3115 		for (i = 0; i < wr->num_sge; ++i)
3116 			__set_data_seg(scat + i, wr->sg_list + i);
3117 
3118 		if (i < max_gs) {
3119 			scat[i].byte_count = 0;
3120 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3121 			scat[i].addr       = 0;
3122 		}
3123 
3124 		qp->rq.wrid[ind] = wr->wr_id;
3125 
3126 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3127 	}
3128 
3129 out:
3130 	if (likely(nreq)) {
3131 		qp->rq.head += nreq;
3132 
3133 		/*
3134 		 * Make sure that descriptors are written before
3135 		 * doorbell record.
3136 		 */
3137 		wmb();
3138 
3139 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3140 	}
3141 
3142 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3143 
3144 	return err;
3145 }
3146 
3147 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3148 {
3149 	switch (mlx4_state) {
3150 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3151 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3152 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3153 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3154 	case MLX4_QP_STATE_SQ_DRAINING:
3155 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3156 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3157 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3158 	default:		     return -1;
3159 	}
3160 }
3161 
3162 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3163 {
3164 	switch (mlx4_mig_state) {
3165 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3166 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3167 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3168 	default: return -1;
3169 	}
3170 }
3171 
3172 static int to_ib_qp_access_flags(int mlx4_flags)
3173 {
3174 	int ib_flags = 0;
3175 
3176 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3177 		ib_flags |= IB_ACCESS_REMOTE_READ;
3178 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3179 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3180 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3181 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3182 
3183 	return ib_flags;
3184 }
3185 
3186 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
3187 				struct mlx4_qp_path *path)
3188 {
3189 	struct mlx4_dev *dev = ibdev->dev;
3190 	int is_eth;
3191 
3192 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
3193 	ib_ah_attr->port_num	  = path->sched_queue & 0x40 ? 2 : 1;
3194 
3195 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
3196 		return;
3197 
3198 	is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
3199 		IB_LINK_LAYER_ETHERNET;
3200 	if (is_eth)
3201 		ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
3202 		((path->sched_queue & 4) << 1);
3203 	else
3204 		ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
3205 
3206 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
3207 	ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
3208 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
3209 	ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
3210 	if (ib_ah_attr->ah_flags) {
3211 		ib_ah_attr->grh.sgid_index = path->mgid_index;
3212 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
3213 		ib_ah_attr->grh.traffic_class =
3214 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3215 		ib_ah_attr->grh.flow_label =
3216 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3217 		memcpy(ib_ah_attr->grh.dgid.raw,
3218 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
3219 	}
3220 }
3221 
3222 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3223 		     struct ib_qp_init_attr *qp_init_attr)
3224 {
3225 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3226 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3227 	struct mlx4_qp_context context;
3228 	int mlx4_state;
3229 	int err = 0;
3230 
3231 	mutex_lock(&qp->mutex);
3232 
3233 	if (qp->state == IB_QPS_RESET) {
3234 		qp_attr->qp_state = IB_QPS_RESET;
3235 		goto done;
3236 	}
3237 
3238 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3239 	if (err) {
3240 		err = -EINVAL;
3241 		goto out;
3242 	}
3243 
3244 	mlx4_state = be32_to_cpu(context.flags) >> 28;
3245 
3246 	qp->state		     = to_ib_qp_state(mlx4_state);
3247 	qp_attr->qp_state	     = qp->state;
3248 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
3249 	qp_attr->path_mig_state	     =
3250 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3251 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
3252 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3253 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
3254 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
3255 	qp_attr->qp_access_flags     =
3256 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
3257 
3258 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3259 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3260 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3261 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3262 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
3263 	}
3264 
3265 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3266 	if (qp_attr->qp_state == IB_QPS_INIT)
3267 		qp_attr->port_num = qp->port;
3268 	else
3269 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3270 
3271 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3272 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3273 
3274 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3275 
3276 	qp_attr->max_dest_rd_atomic =
3277 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3278 	qp_attr->min_rnr_timer	    =
3279 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3280 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
3281 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
3282 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
3283 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
3284 
3285 done:
3286 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
3287 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
3288 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
3289 
3290 	if (!ibqp->uobject) {
3291 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
3292 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
3293 	} else {
3294 		qp_attr->cap.max_send_wr  = 0;
3295 		qp_attr->cap.max_send_sge = 0;
3296 	}
3297 
3298 	/*
3299 	 * We don't support inline sends for kernel QPs (yet), and we
3300 	 * don't know what userspace's value should be.
3301 	 */
3302 	qp_attr->cap.max_inline_data = 0;
3303 
3304 	qp_init_attr->cap	     = qp_attr->cap;
3305 
3306 	qp_init_attr->create_flags = 0;
3307 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3308 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3309 
3310 	if (qp->flags & MLX4_IB_QP_LSO)
3311 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
3312 
3313 	if (qp->flags & MLX4_IB_QP_NETIF)
3314 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
3315 
3316 	qp_init_attr->sq_sig_type =
3317 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
3318 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3319 
3320 out:
3321 	mutex_unlock(&qp->mutex);
3322 	return err;
3323 }
3324 
3325