xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision 4ed91d48259d9ddd378424d008f2e6559f7e78f8)
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/etherdevice.h>
36 #include <net/ip.h>
37 #include <linux/slab.h>
38 #include <linux/netdevice.h>
39 #include <linux/vmalloc.h>
40 
41 #include <rdma/ib_cache.h>
42 #include <rdma/ib_pack.h>
43 #include <rdma/ib_addr.h>
44 #include <rdma/ib_mad.h>
45 
46 #include <linux/mlx4/driver.h>
47 #include <linux/mlx4/qp.h>
48 
49 #include "mlx4_ib.h"
50 #include <rdma/mlx4-abi.h>
51 
52 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
53 			     struct mlx4_ib_cq *recv_cq);
54 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
55 			       struct mlx4_ib_cq *recv_cq);
56 
57 enum {
58 	MLX4_IB_ACK_REQ_FREQ	= 8,
59 };
60 
61 enum {
62 	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
63 	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
64 	MLX4_IB_LINK_TYPE_IB		= 0,
65 	MLX4_IB_LINK_TYPE_ETH		= 1
66 };
67 
68 enum {
69 	/*
70 	 * Largest possible UD header: send with GRH and immediate
71 	 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
72 	 * tag.  (LRH would only use 8 bytes, so Ethernet is the
73 	 * biggest case)
74 	 */
75 	MLX4_IB_UD_HEADER_SIZE		= 82,
76 	MLX4_IB_LSO_HEADER_SPARE	= 128,
77 };
78 
79 struct mlx4_ib_sqp {
80 	struct mlx4_ib_qp	qp;
81 	int			pkey_index;
82 	u32			qkey;
83 	u32			send_psn;
84 	struct ib_ud_header	ud_header;
85 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
86 	struct ib_qp		*roce_v2_gsi;
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 };
118 
119 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
120 {
121 	return container_of(mqp, struct mlx4_ib_sqp, qp);
122 }
123 
124 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
125 {
126 	if (!mlx4_is_master(dev->dev))
127 		return 0;
128 
129 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
130 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
131 		8 * MLX4_MFUNC_MAX;
132 }
133 
134 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
135 {
136 	int proxy_sqp = 0;
137 	int real_sqp = 0;
138 	int i;
139 	/* PPF or Native -- real SQP */
140 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
141 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
142 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
143 	if (real_sqp)
144 		return 1;
145 	/* VF or PF -- proxy SQP */
146 	if (mlx4_is_mfunc(dev->dev)) {
147 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
148 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] ||
149 			    qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) {
150 				proxy_sqp = 1;
151 				break;
152 			}
153 		}
154 	}
155 	if (proxy_sqp)
156 		return 1;
157 
158 	return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
159 }
160 
161 /* used for INIT/CLOSE port logic */
162 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
163 {
164 	int proxy_qp0 = 0;
165 	int real_qp0 = 0;
166 	int i;
167 	/* PPF or Native -- real QP0 */
168 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
169 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
170 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
171 	if (real_qp0)
172 		return 1;
173 	/* VF or PF -- proxy QP0 */
174 	if (mlx4_is_mfunc(dev->dev)) {
175 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
176 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) {
177 				proxy_qp0 = 1;
178 				break;
179 			}
180 		}
181 	}
182 	return proxy_qp0;
183 }
184 
185 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
186 {
187 	return mlx4_buf_offset(&qp->buf, offset);
188 }
189 
190 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
191 {
192 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
193 }
194 
195 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
196 {
197 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
198 }
199 
200 /*
201  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
202  * first four bytes of every 64 byte chunk with
203  *     0x7FFFFFF | (invalid_ownership_value << 31).
204  *
205  * When the max work request size is less than or equal to the WQE
206  * basic block size, as an optimization, we can stamp all WQEs with
207  * 0xffffffff, and skip the very first chunk of each WQE.
208  */
209 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
210 {
211 	__be32 *wqe;
212 	int i;
213 	int s;
214 	int ind;
215 	void *buf;
216 	__be32 stamp;
217 	struct mlx4_wqe_ctrl_seg *ctrl;
218 
219 	if (qp->sq_max_wqes_per_wr > 1) {
220 		s = roundup(size, 1U << qp->sq.wqe_shift);
221 		for (i = 0; i < s; i += 64) {
222 			ind = (i >> qp->sq.wqe_shift) + n;
223 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
224 						       cpu_to_be32(0xffffffff);
225 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
226 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
227 			*wqe = stamp;
228 		}
229 	} else {
230 		ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
231 		s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4;
232 		for (i = 64; i < s; i += 64) {
233 			wqe = buf + i;
234 			*wqe = cpu_to_be32(0xffffffff);
235 		}
236 	}
237 }
238 
239 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
240 {
241 	struct mlx4_wqe_ctrl_seg *ctrl;
242 	struct mlx4_wqe_inline_seg *inl;
243 	void *wqe;
244 	int s;
245 
246 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
247 	s = sizeof(struct mlx4_wqe_ctrl_seg);
248 
249 	if (qp->ibqp.qp_type == IB_QPT_UD) {
250 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
251 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
252 		memset(dgram, 0, sizeof *dgram);
253 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
254 		s += sizeof(struct mlx4_wqe_datagram_seg);
255 	}
256 
257 	/* Pad the remainder of the WQE with an inline data segment. */
258 	if (size > s) {
259 		inl = wqe + s;
260 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
261 	}
262 	ctrl->srcrb_flags = 0;
263 	ctrl->qpn_vlan.fence_size = size / 16;
264 	/*
265 	 * Make sure descriptor is fully written before setting ownership bit
266 	 * (because HW can start executing as soon as we do).
267 	 */
268 	wmb();
269 
270 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
271 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
272 
273 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
274 }
275 
276 /* Post NOP WQE to prevent wrap-around in the middle of WR */
277 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
278 {
279 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
280 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
281 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
282 		ind += s;
283 	}
284 	return ind;
285 }
286 
287 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
288 {
289 	struct ib_event event;
290 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
291 
292 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
293 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
294 
295 	if (ibqp->event_handler) {
296 		event.device     = ibqp->device;
297 		event.element.qp = ibqp;
298 		switch (type) {
299 		case MLX4_EVENT_TYPE_PATH_MIG:
300 			event.event = IB_EVENT_PATH_MIG;
301 			break;
302 		case MLX4_EVENT_TYPE_COMM_EST:
303 			event.event = IB_EVENT_COMM_EST;
304 			break;
305 		case MLX4_EVENT_TYPE_SQ_DRAINED:
306 			event.event = IB_EVENT_SQ_DRAINED;
307 			break;
308 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
309 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
310 			break;
311 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
312 			event.event = IB_EVENT_QP_FATAL;
313 			break;
314 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
315 			event.event = IB_EVENT_PATH_MIG_ERR;
316 			break;
317 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
318 			event.event = IB_EVENT_QP_REQ_ERR;
319 			break;
320 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
321 			event.event = IB_EVENT_QP_ACCESS_ERR;
322 			break;
323 		default:
324 			pr_warn("Unexpected event type %d "
325 			       "on QP %06x\n", type, qp->qpn);
326 			return;
327 		}
328 
329 		ibqp->event_handler(&event, ibqp->qp_context);
330 	}
331 }
332 
333 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
334 {
335 	/*
336 	 * UD WQEs must have a datagram segment.
337 	 * RC and UC WQEs might have a remote address segment.
338 	 * MLX WQEs need two extra inline data segments (for the UD
339 	 * header and space for the ICRC).
340 	 */
341 	switch (type) {
342 	case MLX4_IB_QPT_UD:
343 		return sizeof (struct mlx4_wqe_ctrl_seg) +
344 			sizeof (struct mlx4_wqe_datagram_seg) +
345 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
346 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
347 	case MLX4_IB_QPT_PROXY_SMI:
348 	case MLX4_IB_QPT_PROXY_GSI:
349 		return sizeof (struct mlx4_wqe_ctrl_seg) +
350 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
351 	case MLX4_IB_QPT_TUN_SMI_OWNER:
352 	case MLX4_IB_QPT_TUN_GSI:
353 		return sizeof (struct mlx4_wqe_ctrl_seg) +
354 			sizeof (struct mlx4_wqe_datagram_seg);
355 
356 	case MLX4_IB_QPT_UC:
357 		return sizeof (struct mlx4_wqe_ctrl_seg) +
358 			sizeof (struct mlx4_wqe_raddr_seg);
359 	case MLX4_IB_QPT_RC:
360 		return sizeof (struct mlx4_wqe_ctrl_seg) +
361 			sizeof (struct mlx4_wqe_masked_atomic_seg) +
362 			sizeof (struct mlx4_wqe_raddr_seg);
363 	case MLX4_IB_QPT_SMI:
364 	case MLX4_IB_QPT_GSI:
365 		return sizeof (struct mlx4_wqe_ctrl_seg) +
366 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
367 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
368 					   MLX4_INLINE_ALIGN) *
369 			      sizeof (struct mlx4_wqe_inline_seg),
370 			      sizeof (struct mlx4_wqe_data_seg)) +
371 			ALIGN(4 +
372 			      sizeof (struct mlx4_wqe_inline_seg),
373 			      sizeof (struct mlx4_wqe_data_seg));
374 	default:
375 		return sizeof (struct mlx4_wqe_ctrl_seg);
376 	}
377 }
378 
379 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
380 		       int is_user, int has_rq, struct mlx4_ib_qp *qp)
381 {
382 	/* Sanity check RQ size before proceeding */
383 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
384 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
385 		return -EINVAL;
386 
387 	if (!has_rq) {
388 		if (cap->max_recv_wr)
389 			return -EINVAL;
390 
391 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
392 	} else {
393 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
394 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
395 			return -EINVAL;
396 
397 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
398 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
399 		qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
400 	}
401 
402 	/* leave userspace return values as they were, so as not to break ABI */
403 	if (is_user) {
404 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
405 		cap->max_recv_sge = qp->rq.max_gs;
406 	} else {
407 		cap->max_recv_wr  = qp->rq.max_post =
408 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
409 		cap->max_recv_sge = min(qp->rq.max_gs,
410 					min(dev->dev->caps.max_sq_sg,
411 					    dev->dev->caps.max_rq_sg));
412 	}
413 
414 	return 0;
415 }
416 
417 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
418 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp,
419 			      bool shrink_wqe)
420 {
421 	int s;
422 
423 	/* Sanity check SQ size before proceeding */
424 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
425 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
426 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
427 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
428 		return -EINVAL;
429 
430 	/*
431 	 * For MLX transport we need 2 extra S/G entries:
432 	 * one for the header and one for the checksum at the end
433 	 */
434 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
435 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
436 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
437 		return -EINVAL;
438 
439 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
440 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
441 		send_wqe_overhead(type, qp->flags);
442 
443 	if (s > dev->dev->caps.max_sq_desc_sz)
444 		return -EINVAL;
445 
446 	/*
447 	 * Hermon supports shrinking WQEs, such that a single work
448 	 * request can include multiple units of 1 << wqe_shift.  This
449 	 * way, work requests can differ in size, and do not have to
450 	 * be a power of 2 in size, saving memory and speeding up send
451 	 * WR posting.  Unfortunately, if we do this then the
452 	 * wqe_index field in CQEs can't be used to look up the WR ID
453 	 * anymore, so we do this only if selective signaling is off.
454 	 *
455 	 * Further, on 32-bit platforms, we can't use vmap() to make
456 	 * the QP buffer virtually contiguous.  Thus we have to use
457 	 * constant-sized WRs to make sure a WR is always fully within
458 	 * a single page-sized chunk.
459 	 *
460 	 * Finally, we use NOP work requests to pad the end of the
461 	 * work queue, to avoid wrap-around in the middle of WR.  We
462 	 * set NEC bit to avoid getting completions with error for
463 	 * these NOP WRs, but since NEC is only supported starting
464 	 * with firmware 2.2.232, we use constant-sized WRs for older
465 	 * firmware.
466 	 *
467 	 * And, since MLX QPs only support SEND, we use constant-sized
468 	 * WRs in this case.
469 	 *
470 	 * We look for the smallest value of wqe_shift such that the
471 	 * resulting number of wqes does not exceed device
472 	 * capabilities.
473 	 *
474 	 * We set WQE size to at least 64 bytes, this way stamping
475 	 * invalidates each WQE.
476 	 */
477 	if (shrink_wqe && dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
478 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
479 	    type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
480 	    !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
481 		      MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
482 		qp->sq.wqe_shift = ilog2(64);
483 	else
484 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
485 
486 	for (;;) {
487 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
488 
489 		/*
490 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
491 		 * allow HW to prefetch.
492 		 */
493 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
494 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
495 						    qp->sq_max_wqes_per_wr +
496 						    qp->sq_spare_wqes);
497 
498 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
499 			break;
500 
501 		if (qp->sq_max_wqes_per_wr <= 1)
502 			return -EINVAL;
503 
504 		++qp->sq.wqe_shift;
505 	}
506 
507 	qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
508 			     (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
509 			 send_wqe_overhead(type, qp->flags)) /
510 		sizeof (struct mlx4_wqe_data_seg);
511 
512 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
513 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
514 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
515 		qp->rq.offset = 0;
516 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
517 	} else {
518 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
519 		qp->sq.offset = 0;
520 	}
521 
522 	cap->max_send_wr  = qp->sq.max_post =
523 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
524 	cap->max_send_sge = min(qp->sq.max_gs,
525 				min(dev->dev->caps.max_sq_sg,
526 				    dev->dev->caps.max_rq_sg));
527 	/* We don't support inline sends for kernel QPs (yet) */
528 	cap->max_inline_data = 0;
529 
530 	return 0;
531 }
532 
533 static int set_user_sq_size(struct mlx4_ib_dev *dev,
534 			    struct mlx4_ib_qp *qp,
535 			    struct mlx4_ib_create_qp *ucmd)
536 {
537 	/* Sanity check SQ size before proceeding */
538 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
539 	    ucmd->log_sq_stride >
540 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
541 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
542 		return -EINVAL;
543 
544 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
545 	qp->sq.wqe_shift = ucmd->log_sq_stride;
546 
547 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
548 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
549 
550 	return 0;
551 }
552 
553 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
554 {
555 	int i;
556 
557 	qp->sqp_proxy_rcv =
558 		kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
559 			GFP_KERNEL);
560 	if (!qp->sqp_proxy_rcv)
561 		return -ENOMEM;
562 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
563 		qp->sqp_proxy_rcv[i].addr =
564 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
565 				GFP_KERNEL);
566 		if (!qp->sqp_proxy_rcv[i].addr)
567 			goto err;
568 		qp->sqp_proxy_rcv[i].map =
569 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
570 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
571 					  DMA_FROM_DEVICE);
572 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
573 			kfree(qp->sqp_proxy_rcv[i].addr);
574 			goto err;
575 		}
576 	}
577 	return 0;
578 
579 err:
580 	while (i > 0) {
581 		--i;
582 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
583 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
584 				    DMA_FROM_DEVICE);
585 		kfree(qp->sqp_proxy_rcv[i].addr);
586 	}
587 	kfree(qp->sqp_proxy_rcv);
588 	qp->sqp_proxy_rcv = NULL;
589 	return -ENOMEM;
590 }
591 
592 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
593 {
594 	int i;
595 
596 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
597 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
598 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
599 				    DMA_FROM_DEVICE);
600 		kfree(qp->sqp_proxy_rcv[i].addr);
601 	}
602 	kfree(qp->sqp_proxy_rcv);
603 }
604 
605 static int qp_has_rq(struct ib_qp_init_attr *attr)
606 {
607 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
608 		return 0;
609 
610 	return !attr->srq;
611 }
612 
613 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
614 {
615 	int i;
616 	for (i = 0; i < dev->caps.num_ports; i++) {
617 		if (qpn == dev->caps.qp0_proxy[i])
618 			return !!dev->caps.qp0_qkey[i];
619 	}
620 	return 0;
621 }
622 
623 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
624 				    struct mlx4_ib_qp *qp)
625 {
626 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
627 	mlx4_counter_free(dev->dev, qp->counter_index->index);
628 	list_del(&qp->counter_index->list);
629 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
630 
631 	kfree(qp->counter_index);
632 	qp->counter_index = NULL;
633 }
634 
635 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
636 			    struct ib_qp_init_attr *init_attr,
637 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp,
638 			    gfp_t gfp)
639 {
640 	int qpn;
641 	int err;
642 	struct ib_qp_cap backup_cap;
643 	struct mlx4_ib_sqp *sqp = NULL;
644 	struct mlx4_ib_qp *qp;
645 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
646 	struct mlx4_ib_cq *mcq;
647 	unsigned long flags;
648 
649 	/* When tunneling special qps, we use a plain UD qp */
650 	if (sqpn) {
651 		if (mlx4_is_mfunc(dev->dev) &&
652 		    (!mlx4_is_master(dev->dev) ||
653 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
654 			if (init_attr->qp_type == IB_QPT_GSI)
655 				qp_type = MLX4_IB_QPT_PROXY_GSI;
656 			else {
657 				if (mlx4_is_master(dev->dev) ||
658 				    qp0_enabled_vf(dev->dev, sqpn))
659 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
660 				else
661 					qp_type = MLX4_IB_QPT_PROXY_SMI;
662 			}
663 		}
664 		qpn = sqpn;
665 		/* add extra sg entry for tunneling */
666 		init_attr->cap.max_recv_sge++;
667 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
668 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
669 			container_of(init_attr,
670 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
671 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
672 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
673 		    !mlx4_is_master(dev->dev))
674 			return -EINVAL;
675 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
676 			qp_type = MLX4_IB_QPT_TUN_GSI;
677 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
678 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
679 					     tnl_init->port))
680 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
681 		else
682 			qp_type = MLX4_IB_QPT_TUN_SMI;
683 		/* we are definitely in the PPF here, since we are creating
684 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
685 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
686 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
687 		sqpn = qpn;
688 	}
689 
690 	if (!*caller_qp) {
691 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
692 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
693 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
694 			sqp = kzalloc(sizeof (struct mlx4_ib_sqp), gfp);
695 			if (!sqp)
696 				return -ENOMEM;
697 			qp = &sqp->qp;
698 			qp->pri.vid = 0xFFFF;
699 			qp->alt.vid = 0xFFFF;
700 		} else {
701 			qp = kzalloc(sizeof (struct mlx4_ib_qp), gfp);
702 			if (!qp)
703 				return -ENOMEM;
704 			qp->pri.vid = 0xFFFF;
705 			qp->alt.vid = 0xFFFF;
706 		}
707 	} else
708 		qp = *caller_qp;
709 
710 	qp->mlx4_ib_qp_type = qp_type;
711 
712 	mutex_init(&qp->mutex);
713 	spin_lock_init(&qp->sq.lock);
714 	spin_lock_init(&qp->rq.lock);
715 	INIT_LIST_HEAD(&qp->gid_list);
716 	INIT_LIST_HEAD(&qp->steering_rules);
717 
718 	qp->state	 = IB_QPS_RESET;
719 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
720 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
721 
722 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
723 	if (err)
724 		goto err;
725 
726 	if (pd->uobject) {
727 		struct mlx4_ib_create_qp ucmd;
728 
729 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
730 			err = -EFAULT;
731 			goto err;
732 		}
733 
734 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
735 
736 		err = set_user_sq_size(dev, qp, &ucmd);
737 		if (err)
738 			goto err;
739 
740 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
741 				       qp->buf_size, 0, 0);
742 		if (IS_ERR(qp->umem)) {
743 			err = PTR_ERR(qp->umem);
744 			goto err;
745 		}
746 
747 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
748 				    ilog2(qp->umem->page_size), &qp->mtt);
749 		if (err)
750 			goto err_buf;
751 
752 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
753 		if (err)
754 			goto err_mtt;
755 
756 		if (qp_has_rq(init_attr)) {
757 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
758 						  ucmd.db_addr, &qp->db);
759 			if (err)
760 				goto err_mtt;
761 		}
762 	} else {
763 		qp->sq_no_prefetch = 0;
764 
765 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
766 			qp->flags |= MLX4_IB_QP_LSO;
767 
768 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
769 			if (dev->steering_support ==
770 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
771 				qp->flags |= MLX4_IB_QP_NETIF;
772 			else
773 				goto err;
774 		}
775 
776 		memcpy(&backup_cap, &init_attr->cap, sizeof(backup_cap));
777 		err = set_kernel_sq_size(dev, &init_attr->cap,
778 					 qp_type, qp, true);
779 		if (err)
780 			goto err;
781 
782 		if (qp_has_rq(init_attr)) {
783 			err = mlx4_db_alloc(dev->dev, &qp->db, 0, gfp);
784 			if (err)
785 				goto err;
786 
787 			*qp->db.db = 0;
788 		}
789 
790 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, qp->buf_size,
791 				   &qp->buf, gfp)) {
792 			memcpy(&init_attr->cap, &backup_cap,
793 			       sizeof(backup_cap));
794 			err = set_kernel_sq_size(dev, &init_attr->cap, qp_type,
795 						 qp, false);
796 			if (err)
797 				goto err_db;
798 
799 			if (mlx4_buf_alloc(dev->dev, qp->buf_size,
800 					   PAGE_SIZE * 2, &qp->buf, gfp)) {
801 				err = -ENOMEM;
802 				goto err_db;
803 			}
804 		}
805 
806 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
807 				    &qp->mtt);
808 		if (err)
809 			goto err_buf;
810 
811 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf, gfp);
812 		if (err)
813 			goto err_mtt;
814 
815 		qp->sq.wrid = kmalloc_array(qp->sq.wqe_cnt, sizeof(u64),
816 					gfp | __GFP_NOWARN);
817 		if (!qp->sq.wrid)
818 			qp->sq.wrid = __vmalloc(qp->sq.wqe_cnt * sizeof(u64),
819 						gfp, PAGE_KERNEL);
820 		qp->rq.wrid = kmalloc_array(qp->rq.wqe_cnt, sizeof(u64),
821 					gfp | __GFP_NOWARN);
822 		if (!qp->rq.wrid)
823 			qp->rq.wrid = __vmalloc(qp->rq.wqe_cnt * sizeof(u64),
824 						gfp, PAGE_KERNEL);
825 		if (!qp->sq.wrid || !qp->rq.wrid) {
826 			err = -ENOMEM;
827 			goto err_wrid;
828 		}
829 	}
830 
831 	if (sqpn) {
832 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
833 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
834 			if (alloc_proxy_bufs(pd->device, qp)) {
835 				err = -ENOMEM;
836 				goto err_wrid;
837 			}
838 		}
839 	} else {
840 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
841 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
842 		 * VLAN insertion. */
843 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
844 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
845 						    (init_attr->cap.max_send_wr ?
846 						     MLX4_RESERVE_ETH_BF_QP : 0) |
847 						    (init_attr->cap.max_recv_wr ?
848 						     MLX4_RESERVE_A0_QP : 0));
849 		else
850 			if (qp->flags & MLX4_IB_QP_NETIF)
851 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
852 			else
853 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
854 							    &qpn, 0);
855 		if (err)
856 			goto err_proxy;
857 	}
858 
859 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
860 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
861 
862 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp, gfp);
863 	if (err)
864 		goto err_qpn;
865 
866 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
867 		qp->mqp.qpn |= (1 << 23);
868 
869 	/*
870 	 * Hardware wants QPN written in big-endian order (after
871 	 * shifting) for send doorbell.  Precompute this value to save
872 	 * a little bit when posting sends.
873 	 */
874 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
875 
876 	qp->mqp.event = mlx4_ib_qp_event;
877 	if (!*caller_qp)
878 		*caller_qp = qp;
879 
880 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
881 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
882 			 to_mcq(init_attr->recv_cq));
883 	/* Maintain device to QPs access, needed for further handling
884 	 * via reset flow
885 	 */
886 	list_add_tail(&qp->qps_list, &dev->qp_list);
887 	/* Maintain CQ to QPs access, needed for further handling
888 	 * via reset flow
889 	 */
890 	mcq = to_mcq(init_attr->send_cq);
891 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
892 	mcq = to_mcq(init_attr->recv_cq);
893 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
894 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
895 			   to_mcq(init_attr->recv_cq));
896 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
897 	return 0;
898 
899 err_qpn:
900 	if (!sqpn) {
901 		if (qp->flags & MLX4_IB_QP_NETIF)
902 			mlx4_ib_steer_qp_free(dev, qpn, 1);
903 		else
904 			mlx4_qp_release_range(dev->dev, qpn, 1);
905 	}
906 err_proxy:
907 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
908 		free_proxy_bufs(pd->device, qp);
909 err_wrid:
910 	if (pd->uobject) {
911 		if (qp_has_rq(init_attr))
912 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
913 	} else {
914 		kvfree(qp->sq.wrid);
915 		kvfree(qp->rq.wrid);
916 	}
917 
918 err_mtt:
919 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
920 
921 err_buf:
922 	if (pd->uobject)
923 		ib_umem_release(qp->umem);
924 	else
925 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
926 
927 err_db:
928 	if (!pd->uobject && qp_has_rq(init_attr))
929 		mlx4_db_free(dev->dev, &qp->db);
930 
931 err:
932 	if (sqp)
933 		kfree(sqp);
934 	else if (!*caller_qp)
935 		kfree(qp);
936 	return err;
937 }
938 
939 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
940 {
941 	switch (state) {
942 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
943 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
944 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
945 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
946 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
947 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
948 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
949 	default:		return -1;
950 	}
951 }
952 
953 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
954 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
955 {
956 	if (send_cq == recv_cq) {
957 		spin_lock(&send_cq->lock);
958 		__acquire(&recv_cq->lock);
959 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
960 		spin_lock(&send_cq->lock);
961 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
962 	} else {
963 		spin_lock(&recv_cq->lock);
964 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
965 	}
966 }
967 
968 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
969 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
970 {
971 	if (send_cq == recv_cq) {
972 		__release(&recv_cq->lock);
973 		spin_unlock(&send_cq->lock);
974 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
975 		spin_unlock(&recv_cq->lock);
976 		spin_unlock(&send_cq->lock);
977 	} else {
978 		spin_unlock(&send_cq->lock);
979 		spin_unlock(&recv_cq->lock);
980 	}
981 }
982 
983 static void del_gid_entries(struct mlx4_ib_qp *qp)
984 {
985 	struct mlx4_ib_gid_entry *ge, *tmp;
986 
987 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
988 		list_del(&ge->list);
989 		kfree(ge);
990 	}
991 }
992 
993 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
994 {
995 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
996 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
997 	else
998 		return to_mpd(qp->ibqp.pd);
999 }
1000 
1001 static void get_cqs(struct mlx4_ib_qp *qp,
1002 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1003 {
1004 	switch (qp->ibqp.qp_type) {
1005 	case IB_QPT_XRC_TGT:
1006 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1007 		*recv_cq = *send_cq;
1008 		break;
1009 	case IB_QPT_XRC_INI:
1010 		*send_cq = to_mcq(qp->ibqp.send_cq);
1011 		*recv_cq = *send_cq;
1012 		break;
1013 	default:
1014 		*send_cq = to_mcq(qp->ibqp.send_cq);
1015 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
1016 		break;
1017 	}
1018 }
1019 
1020 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1021 			      int is_user)
1022 {
1023 	struct mlx4_ib_cq *send_cq, *recv_cq;
1024 	unsigned long flags;
1025 
1026 	if (qp->state != IB_QPS_RESET) {
1027 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1028 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1029 			pr_warn("modify QP %06x to RESET failed.\n",
1030 			       qp->mqp.qpn);
1031 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1032 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1033 			qp->pri.smac = 0;
1034 			qp->pri.smac_port = 0;
1035 		}
1036 		if (qp->alt.smac) {
1037 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1038 			qp->alt.smac = 0;
1039 		}
1040 		if (qp->pri.vid < 0x1000) {
1041 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1042 			qp->pri.vid = 0xFFFF;
1043 			qp->pri.candidate_vid = 0xFFFF;
1044 			qp->pri.update_vid = 0;
1045 		}
1046 		if (qp->alt.vid < 0x1000) {
1047 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1048 			qp->alt.vid = 0xFFFF;
1049 			qp->alt.candidate_vid = 0xFFFF;
1050 			qp->alt.update_vid = 0;
1051 		}
1052 	}
1053 
1054 	get_cqs(qp, &send_cq, &recv_cq);
1055 
1056 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1057 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1058 
1059 	/* del from lists under both locks above to protect reset flow paths */
1060 	list_del(&qp->qps_list);
1061 	list_del(&qp->cq_send_list);
1062 	list_del(&qp->cq_recv_list);
1063 	if (!is_user) {
1064 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1065 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1066 		if (send_cq != recv_cq)
1067 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1068 	}
1069 
1070 	mlx4_qp_remove(dev->dev, &qp->mqp);
1071 
1072 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1073 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1074 
1075 	mlx4_qp_free(dev->dev, &qp->mqp);
1076 
1077 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1078 		if (qp->flags & MLX4_IB_QP_NETIF)
1079 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1080 		else
1081 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1082 	}
1083 
1084 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1085 
1086 	if (is_user) {
1087 		if (qp->rq.wqe_cnt)
1088 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
1089 					      &qp->db);
1090 		ib_umem_release(qp->umem);
1091 	} else {
1092 		kvfree(qp->sq.wrid);
1093 		kvfree(qp->rq.wrid);
1094 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1095 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1096 			free_proxy_bufs(&dev->ib_dev, qp);
1097 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1098 		if (qp->rq.wqe_cnt)
1099 			mlx4_db_free(dev->dev, &qp->db);
1100 	}
1101 
1102 	del_gid_entries(qp);
1103 }
1104 
1105 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1106 {
1107 	/* Native or PPF */
1108 	if (!mlx4_is_mfunc(dev->dev) ||
1109 	    (mlx4_is_master(dev->dev) &&
1110 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1111 		return  dev->dev->phys_caps.base_sqpn +
1112 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1113 			attr->port_num - 1;
1114 	}
1115 	/* PF or VF -- creating proxies */
1116 	if (attr->qp_type == IB_QPT_SMI)
1117 		return dev->dev->caps.qp0_proxy[attr->port_num - 1];
1118 	else
1119 		return dev->dev->caps.qp1_proxy[attr->port_num - 1];
1120 }
1121 
1122 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1123 					struct ib_qp_init_attr *init_attr,
1124 					struct ib_udata *udata)
1125 {
1126 	struct mlx4_ib_qp *qp = NULL;
1127 	int err;
1128 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1129 	u16 xrcdn = 0;
1130 	gfp_t gfp;
1131 
1132 	gfp = (init_attr->create_flags & MLX4_IB_QP_CREATE_USE_GFP_NOIO) ?
1133 		GFP_NOIO : GFP_KERNEL;
1134 	/*
1135 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1136 	 * and only for kernel UD QPs.
1137 	 */
1138 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1139 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1140 					MLX4_IB_SRIOV_TUNNEL_QP |
1141 					MLX4_IB_SRIOV_SQP |
1142 					MLX4_IB_QP_NETIF |
1143 					MLX4_IB_QP_CREATE_ROCE_V2_GSI |
1144 					MLX4_IB_QP_CREATE_USE_GFP_NOIO))
1145 		return ERR_PTR(-EINVAL);
1146 
1147 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1148 		if (init_attr->qp_type != IB_QPT_UD)
1149 			return ERR_PTR(-EINVAL);
1150 	}
1151 
1152 	if (init_attr->create_flags) {
1153 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1154 			return ERR_PTR(-EINVAL);
1155 
1156 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1157 						 MLX4_IB_QP_CREATE_USE_GFP_NOIO |
1158 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1159 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1160 		     init_attr->qp_type != IB_QPT_UD) ||
1161 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1162 		     init_attr->qp_type > IB_QPT_GSI) ||
1163 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1164 		     init_attr->qp_type != IB_QPT_GSI))
1165 			return ERR_PTR(-EINVAL);
1166 	}
1167 
1168 	switch (init_attr->qp_type) {
1169 	case IB_QPT_XRC_TGT:
1170 		pd = to_mxrcd(init_attr->xrcd)->pd;
1171 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1172 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1173 		/* fall through */
1174 	case IB_QPT_XRC_INI:
1175 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1176 			return ERR_PTR(-ENOSYS);
1177 		init_attr->recv_cq = init_attr->send_cq;
1178 		/* fall through */
1179 	case IB_QPT_RC:
1180 	case IB_QPT_UC:
1181 	case IB_QPT_RAW_PACKET:
1182 		qp = kzalloc(sizeof *qp, gfp);
1183 		if (!qp)
1184 			return ERR_PTR(-ENOMEM);
1185 		qp->pri.vid = 0xFFFF;
1186 		qp->alt.vid = 0xFFFF;
1187 		/* fall through */
1188 	case IB_QPT_UD:
1189 	{
1190 		err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1191 				       udata, 0, &qp, gfp);
1192 		if (err) {
1193 			kfree(qp);
1194 			return ERR_PTR(err);
1195 		}
1196 
1197 		qp->ibqp.qp_num = qp->mqp.qpn;
1198 		qp->xrcdn = xrcdn;
1199 
1200 		break;
1201 	}
1202 	case IB_QPT_SMI:
1203 	case IB_QPT_GSI:
1204 	{
1205 		int sqpn;
1206 
1207 		/* Userspace is not allowed to create special QPs: */
1208 		if (udata)
1209 			return ERR_PTR(-EINVAL);
1210 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1211 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev, 1, 1, &sqpn, 0);
1212 
1213 			if (res)
1214 				return ERR_PTR(res);
1215 		} else {
1216 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1217 		}
1218 
1219 		err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1220 				       sqpn,
1221 				       &qp, gfp);
1222 		if (err)
1223 			return ERR_PTR(err);
1224 
1225 		qp->port	= init_attr->port_num;
1226 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1227 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1228 		break;
1229 	}
1230 	default:
1231 		/* Don't support raw QPs */
1232 		return ERR_PTR(-EINVAL);
1233 	}
1234 
1235 	return &qp->ibqp;
1236 }
1237 
1238 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1239 				struct ib_qp_init_attr *init_attr,
1240 				struct ib_udata *udata) {
1241 	struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1242 	struct ib_qp *ibqp;
1243 	struct mlx4_ib_dev *dev = to_mdev(device);
1244 
1245 	ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1246 
1247 	if (!IS_ERR(ibqp) &&
1248 	    (init_attr->qp_type == IB_QPT_GSI) &&
1249 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1250 		struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1251 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1252 
1253 		if (is_eth &&
1254 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1255 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1256 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1257 
1258 			if (IS_ERR(sqp->roce_v2_gsi)) {
1259 				pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1260 				sqp->roce_v2_gsi = NULL;
1261 			} else {
1262 				sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1263 				sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1264 			}
1265 
1266 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1267 		}
1268 	}
1269 	return ibqp;
1270 }
1271 
1272 static int _mlx4_ib_destroy_qp(struct ib_qp *qp)
1273 {
1274 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1275 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1276 	struct mlx4_ib_pd *pd;
1277 
1278 	if (is_qp0(dev, mqp))
1279 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1280 
1281 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1282 	    dev->qp1_proxy[mqp->port - 1] == mqp) {
1283 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1284 		dev->qp1_proxy[mqp->port - 1] = NULL;
1285 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1286 	}
1287 
1288 	if (mqp->counter_index)
1289 		mlx4_ib_free_qp_counter(dev, mqp);
1290 
1291 	pd = get_pd(mqp);
1292 	destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1293 
1294 	if (is_sqp(dev, mqp))
1295 		kfree(to_msqp(mqp));
1296 	else
1297 		kfree(mqp);
1298 
1299 	return 0;
1300 }
1301 
1302 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1303 {
1304 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1305 
1306 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1307 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1308 
1309 		if (sqp->roce_v2_gsi)
1310 			ib_destroy_qp(sqp->roce_v2_gsi);
1311 	}
1312 
1313 	return _mlx4_ib_destroy_qp(qp);
1314 }
1315 
1316 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1317 {
1318 	switch (type) {
1319 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1320 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1321 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1322 	case MLX4_IB_QPT_XRC_INI:
1323 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1324 	case MLX4_IB_QPT_SMI:
1325 	case MLX4_IB_QPT_GSI:
1326 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1327 
1328 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1329 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1330 						MLX4_QP_ST_MLX : -1);
1331 	case MLX4_IB_QPT_PROXY_SMI:
1332 	case MLX4_IB_QPT_TUN_SMI:
1333 	case MLX4_IB_QPT_PROXY_GSI:
1334 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1335 						MLX4_QP_ST_UD : -1);
1336 	default:			return -1;
1337 	}
1338 }
1339 
1340 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1341 				   int attr_mask)
1342 {
1343 	u8 dest_rd_atomic;
1344 	u32 access_flags;
1345 	u32 hw_access_flags = 0;
1346 
1347 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1348 		dest_rd_atomic = attr->max_dest_rd_atomic;
1349 	else
1350 		dest_rd_atomic = qp->resp_depth;
1351 
1352 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1353 		access_flags = attr->qp_access_flags;
1354 	else
1355 		access_flags = qp->atomic_rd_en;
1356 
1357 	if (!dest_rd_atomic)
1358 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1359 
1360 	if (access_flags & IB_ACCESS_REMOTE_READ)
1361 		hw_access_flags |= MLX4_QP_BIT_RRE;
1362 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1363 		hw_access_flags |= MLX4_QP_BIT_RAE;
1364 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1365 		hw_access_flags |= MLX4_QP_BIT_RWE;
1366 
1367 	return cpu_to_be32(hw_access_flags);
1368 }
1369 
1370 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1371 			    int attr_mask)
1372 {
1373 	if (attr_mask & IB_QP_PKEY_INDEX)
1374 		sqp->pkey_index = attr->pkey_index;
1375 	if (attr_mask & IB_QP_QKEY)
1376 		sqp->qkey = attr->qkey;
1377 	if (attr_mask & IB_QP_SQ_PSN)
1378 		sqp->send_psn = attr->sq_psn;
1379 }
1380 
1381 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1382 {
1383 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1384 }
1385 
1386 static int _mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
1387 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1388 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1389 {
1390 	int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
1391 		IB_LINK_LAYER_ETHERNET;
1392 	int vidx;
1393 	int smac_index;
1394 	int err;
1395 
1396 
1397 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
1398 	path->rlid	    = cpu_to_be16(ah->dlid);
1399 	if (ah->static_rate) {
1400 		path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
1401 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1402 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1403 			--path->static_rate;
1404 	} else
1405 		path->static_rate = 0;
1406 
1407 	if (ah->ah_flags & IB_AH_GRH) {
1408 		int real_sgid_index = mlx4_ib_gid_index_to_real_index(dev,
1409 								      port,
1410 								      ah->grh.sgid_index);
1411 
1412 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1413 			pr_err("sgid_index (%u) too large. max is %d\n",
1414 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1415 			return -1;
1416 		}
1417 
1418 		path->grh_mylmc |= 1 << 7;
1419 		path->mgid_index = real_sgid_index;
1420 		path->hop_limit  = ah->grh.hop_limit;
1421 		path->tclass_flowlabel =
1422 			cpu_to_be32((ah->grh.traffic_class << 20) |
1423 				    (ah->grh.flow_label));
1424 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1425 	}
1426 
1427 	if (is_eth) {
1428 		if (!(ah->ah_flags & IB_AH_GRH))
1429 			return -1;
1430 
1431 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1432 			((port - 1) << 6) | ((ah->sl & 7) << 3);
1433 
1434 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1435 		if (vlan_tag < 0x1000) {
1436 			if (smac_info->vid < 0x1000) {
1437 				/* both valid vlan ids */
1438 				if (smac_info->vid != vlan_tag) {
1439 					/* different VIDs.  unreg old and reg new */
1440 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1441 					if (err)
1442 						return err;
1443 					smac_info->candidate_vid = vlan_tag;
1444 					smac_info->candidate_vlan_index = vidx;
1445 					smac_info->candidate_vlan_port = port;
1446 					smac_info->update_vid = 1;
1447 					path->vlan_index = vidx;
1448 				} else {
1449 					path->vlan_index = smac_info->vlan_index;
1450 				}
1451 			} else {
1452 				/* no current vlan tag in qp */
1453 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1454 				if (err)
1455 					return err;
1456 				smac_info->candidate_vid = vlan_tag;
1457 				smac_info->candidate_vlan_index = vidx;
1458 				smac_info->candidate_vlan_port = port;
1459 				smac_info->update_vid = 1;
1460 				path->vlan_index = vidx;
1461 			}
1462 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1463 			path->fl = 1 << 6;
1464 		} else {
1465 			/* have current vlan tag. unregister it at modify-qp success */
1466 			if (smac_info->vid < 0x1000) {
1467 				smac_info->candidate_vid = 0xFFFF;
1468 				smac_info->update_vid = 1;
1469 			}
1470 		}
1471 
1472 		/* get smac_index for RoCE use.
1473 		 * If no smac was yet assigned, register one.
1474 		 * If one was already assigned, but the new mac differs,
1475 		 * unregister the old one and register the new one.
1476 		*/
1477 		if ((!smac_info->smac && !smac_info->smac_port) ||
1478 		    smac_info->smac != smac) {
1479 			/* register candidate now, unreg if needed, after success */
1480 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1481 			if (smac_index >= 0) {
1482 				smac_info->candidate_smac_index = smac_index;
1483 				smac_info->candidate_smac = smac;
1484 				smac_info->candidate_smac_port = port;
1485 			} else {
1486 				return -EINVAL;
1487 			}
1488 		} else {
1489 			smac_index = smac_info->smac_index;
1490 		}
1491 
1492 		memcpy(path->dmac, ah->dmac, 6);
1493 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1494 		/* put MAC table smac index for IBoE */
1495 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1496 	} else {
1497 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1498 			((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1499 	}
1500 
1501 	return 0;
1502 }
1503 
1504 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1505 			 enum ib_qp_attr_mask qp_attr_mask,
1506 			 struct mlx4_ib_qp *mqp,
1507 			 struct mlx4_qp_path *path, u8 port,
1508 			 u16 vlan_id, u8 *smac)
1509 {
1510 	return _mlx4_set_path(dev, &qp->ah_attr,
1511 			      mlx4_mac_to_u64(smac),
1512 			      vlan_id,
1513 			      path, &mqp->pri, port);
1514 }
1515 
1516 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1517 			     const struct ib_qp_attr *qp,
1518 			     enum ib_qp_attr_mask qp_attr_mask,
1519 			     struct mlx4_ib_qp *mqp,
1520 			     struct mlx4_qp_path *path, u8 port)
1521 {
1522 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1523 			      0,
1524 			      0xffff,
1525 			      path, &mqp->alt, port);
1526 }
1527 
1528 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1529 {
1530 	struct mlx4_ib_gid_entry *ge, *tmp;
1531 
1532 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1533 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1534 			ge->added = 1;
1535 			ge->port = qp->port;
1536 		}
1537 	}
1538 }
1539 
1540 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1541 				    struct mlx4_ib_qp *qp,
1542 				    struct mlx4_qp_context *context)
1543 {
1544 	u64 u64_mac;
1545 	int smac_index;
1546 
1547 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1548 
1549 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1550 	if (!qp->pri.smac && !qp->pri.smac_port) {
1551 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1552 		if (smac_index >= 0) {
1553 			qp->pri.candidate_smac_index = smac_index;
1554 			qp->pri.candidate_smac = u64_mac;
1555 			qp->pri.candidate_smac_port = qp->port;
1556 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1557 		} else {
1558 			return -ENOENT;
1559 		}
1560 	}
1561 	return 0;
1562 }
1563 
1564 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1565 {
1566 	struct counter_index *new_counter_index;
1567 	int err;
1568 	u32 tmp_idx;
1569 
1570 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1571 	    IB_LINK_LAYER_ETHERNET ||
1572 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1573 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1574 		return 0;
1575 
1576 	err = mlx4_counter_alloc(dev->dev, &tmp_idx);
1577 	if (err)
1578 		return err;
1579 
1580 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1581 	if (!new_counter_index) {
1582 		mlx4_counter_free(dev->dev, tmp_idx);
1583 		return -ENOMEM;
1584 	}
1585 
1586 	new_counter_index->index = tmp_idx;
1587 	new_counter_index->allocated = 1;
1588 	qp->counter_index = new_counter_index;
1589 
1590 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
1591 	list_add_tail(&new_counter_index->list,
1592 		      &dev->counters_table[qp->port - 1].counters_list);
1593 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
1594 
1595 	return 0;
1596 }
1597 
1598 enum {
1599 	MLX4_QPC_ROCE_MODE_1 = 0,
1600 	MLX4_QPC_ROCE_MODE_2 = 2,
1601 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
1602 };
1603 
1604 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
1605 {
1606 	switch (gid_type) {
1607 	case IB_GID_TYPE_ROCE:
1608 		return MLX4_QPC_ROCE_MODE_1;
1609 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
1610 		return MLX4_QPC_ROCE_MODE_2;
1611 	default:
1612 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
1613 	}
1614 }
1615 
1616 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1617 			       const struct ib_qp_attr *attr, int attr_mask,
1618 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1619 {
1620 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1621 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1622 	struct mlx4_ib_pd *pd;
1623 	struct mlx4_ib_cq *send_cq, *recv_cq;
1624 	struct mlx4_qp_context *context;
1625 	enum mlx4_qp_optpar optpar = 0;
1626 	int sqd_event;
1627 	int steer_qp = 0;
1628 	int err = -EINVAL;
1629 	int counter_index;
1630 
1631 	/* APM is not supported under RoCE */
1632 	if (attr_mask & IB_QP_ALT_PATH &&
1633 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1634 	    IB_LINK_LAYER_ETHERNET)
1635 		return -ENOTSUPP;
1636 
1637 	context = kzalloc(sizeof *context, GFP_KERNEL);
1638 	if (!context)
1639 		return -ENOMEM;
1640 
1641 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1642 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1643 
1644 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1645 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1646 	else {
1647 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
1648 		switch (attr->path_mig_state) {
1649 		case IB_MIG_MIGRATED:
1650 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1651 			break;
1652 		case IB_MIG_REARM:
1653 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1654 			break;
1655 		case IB_MIG_ARMED:
1656 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1657 			break;
1658 		}
1659 	}
1660 
1661 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1662 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1663 	else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1664 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1665 	else if (ibqp->qp_type == IB_QPT_UD) {
1666 		if (qp->flags & MLX4_IB_QP_LSO)
1667 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
1668 					      ilog2(dev->dev->caps.max_gso_sz);
1669 		else
1670 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1671 	} else if (attr_mask & IB_QP_PATH_MTU) {
1672 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1673 			pr_err("path MTU (%u) is invalid\n",
1674 			       attr->path_mtu);
1675 			goto out;
1676 		}
1677 		context->mtu_msgmax = (attr->path_mtu << 5) |
1678 			ilog2(dev->dev->caps.max_msg_sz);
1679 	}
1680 
1681 	if (qp->rq.wqe_cnt)
1682 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1683 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
1684 
1685 	if (qp->sq.wqe_cnt)
1686 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1687 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
1688 
1689 	if (new_state == IB_QPS_RESET && qp->counter_index)
1690 		mlx4_ib_free_qp_counter(dev, qp);
1691 
1692 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1693 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1694 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1695 		if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1696 			context->param3 |= cpu_to_be32(1 << 30);
1697 	}
1698 
1699 	if (qp->ibqp.uobject)
1700 		context->usr_page = cpu_to_be32(
1701 			mlx4_to_hw_uar_index(dev->dev,
1702 					     to_mucontext(ibqp->uobject->context)->uar.index));
1703 	else
1704 		context->usr_page = cpu_to_be32(
1705 			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
1706 
1707 	if (attr_mask & IB_QP_DEST_QPN)
1708 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1709 
1710 	if (attr_mask & IB_QP_PORT) {
1711 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1712 		    !(attr_mask & IB_QP_AV)) {
1713 			mlx4_set_sched(&context->pri_path, attr->port_num);
1714 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1715 		}
1716 	}
1717 
1718 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1719 		err = create_qp_lb_counter(dev, qp);
1720 		if (err)
1721 			goto out;
1722 
1723 		counter_index =
1724 			dev->counters_table[qp->port - 1].default_counter;
1725 		if (qp->counter_index)
1726 			counter_index = qp->counter_index->index;
1727 
1728 		if (counter_index != -1) {
1729 			context->pri_path.counter_index = counter_index;
1730 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1731 			if (qp->counter_index) {
1732 				context->pri_path.fl |=
1733 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
1734 				context->pri_path.vlan_control |=
1735 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
1736 			}
1737 		} else
1738 			context->pri_path.counter_index =
1739 				MLX4_SINK_COUNTER_INDEX(dev->dev);
1740 
1741 		if (qp->flags & MLX4_IB_QP_NETIF) {
1742 			mlx4_ib_steer_qp_reg(dev, qp, 1);
1743 			steer_qp = 1;
1744 		}
1745 
1746 		if (ibqp->qp_type == IB_QPT_GSI) {
1747 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
1748 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
1749 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
1750 
1751 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
1752 		}
1753 	}
1754 
1755 	if (attr_mask & IB_QP_PKEY_INDEX) {
1756 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1757 			context->pri_path.disable_pkey_check = 0x40;
1758 		context->pri_path.pkey_index = attr->pkey_index;
1759 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1760 	}
1761 
1762 	if (attr_mask & IB_QP_AV) {
1763 		u8 port_num = mlx4_is_bonded(to_mdev(ibqp->device)->dev) ? 1 :
1764 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1765 		union ib_gid gid;
1766 		struct ib_gid_attr gid_attr = {.gid_type = IB_GID_TYPE_IB};
1767 		u16 vlan = 0xffff;
1768 		u8 smac[ETH_ALEN];
1769 		int status = 0;
1770 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
1771 			attr->ah_attr.ah_flags & IB_AH_GRH;
1772 
1773 		if (is_eth && attr->ah_attr.ah_flags & IB_AH_GRH) {
1774 			int index = attr->ah_attr.grh.sgid_index;
1775 
1776 			status = ib_get_cached_gid(ibqp->device, port_num,
1777 						   index, &gid, &gid_attr);
1778 			if (!status && !memcmp(&gid, &zgid, sizeof(gid)))
1779 				status = -ENOENT;
1780 			if (!status && gid_attr.ndev) {
1781 				vlan = rdma_vlan_dev_vlan_id(gid_attr.ndev);
1782 				memcpy(smac, gid_attr.ndev->dev_addr, ETH_ALEN);
1783 				dev_put(gid_attr.ndev);
1784 			}
1785 		}
1786 		if (status)
1787 			goto out;
1788 
1789 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
1790 				  port_num, vlan, smac))
1791 			goto out;
1792 
1793 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1794 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
1795 
1796 		if (is_eth &&
1797 		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
1798 			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr.gid_type);
1799 
1800 			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
1801 				err = -EINVAL;
1802 				goto out;
1803 			}
1804 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
1805 		}
1806 
1807 	}
1808 
1809 	if (attr_mask & IB_QP_TIMEOUT) {
1810 		context->pri_path.ackto |= attr->timeout << 3;
1811 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1812 	}
1813 
1814 	if (attr_mask & IB_QP_ALT_PATH) {
1815 		if (attr->alt_port_num == 0 ||
1816 		    attr->alt_port_num > dev->dev->caps.num_ports)
1817 			goto out;
1818 
1819 		if (attr->alt_pkey_index >=
1820 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
1821 			goto out;
1822 
1823 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
1824 				      &context->alt_path,
1825 				      attr->alt_port_num))
1826 			goto out;
1827 
1828 		context->alt_path.pkey_index = attr->alt_pkey_index;
1829 		context->alt_path.ackto = attr->alt_timeout << 3;
1830 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1831 	}
1832 
1833 	pd = get_pd(qp);
1834 	get_cqs(qp, &send_cq, &recv_cq);
1835 	context->pd       = cpu_to_be32(pd->pdn);
1836 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1837 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1838 	context->params1  = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1839 
1840 	/* Set "fast registration enabled" for all kernel QPs */
1841 	if (!qp->ibqp.uobject)
1842 		context->params1 |= cpu_to_be32(1 << 11);
1843 
1844 	if (attr_mask & IB_QP_RNR_RETRY) {
1845 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1846 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1847 	}
1848 
1849 	if (attr_mask & IB_QP_RETRY_CNT) {
1850 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1851 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1852 	}
1853 
1854 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1855 		if (attr->max_rd_atomic)
1856 			context->params1 |=
1857 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1858 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1859 	}
1860 
1861 	if (attr_mask & IB_QP_SQ_PSN)
1862 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1863 
1864 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1865 		if (attr->max_dest_rd_atomic)
1866 			context->params2 |=
1867 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1868 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1869 	}
1870 
1871 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1872 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1873 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1874 	}
1875 
1876 	if (ibqp->srq)
1877 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1878 
1879 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1880 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1881 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1882 	}
1883 	if (attr_mask & IB_QP_RQ_PSN)
1884 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1885 
1886 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1887 	if (attr_mask & IB_QP_QKEY) {
1888 		if (qp->mlx4_ib_qp_type &
1889 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1890 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1891 		else {
1892 			if (mlx4_is_mfunc(dev->dev) &&
1893 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1894 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1895 			    MLX4_RESERVED_QKEY_BASE) {
1896 				pr_err("Cannot use reserved QKEY"
1897 				       " 0x%x (range 0xffff0000..0xffffffff"
1898 				       " is reserved)\n", attr->qkey);
1899 				err = -EINVAL;
1900 				goto out;
1901 			}
1902 			context->qkey = cpu_to_be32(attr->qkey);
1903 		}
1904 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
1905 	}
1906 
1907 	if (ibqp->srq)
1908 		context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1909 
1910 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1911 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1912 
1913 	if (cur_state == IB_QPS_INIT &&
1914 	    new_state == IB_QPS_RTR  &&
1915 	    (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1916 	     ibqp->qp_type == IB_QPT_UD ||
1917 	     ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1918 		context->pri_path.sched_queue = (qp->port - 1) << 6;
1919 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1920 		    qp->mlx4_ib_qp_type &
1921 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1922 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1923 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1924 				context->pri_path.fl = 0x80;
1925 		} else {
1926 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1927 				context->pri_path.fl = 0x80;
1928 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1929 		}
1930 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1931 		    IB_LINK_LAYER_ETHERNET) {
1932 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
1933 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
1934 				context->pri_path.feup = 1 << 7; /* don't fsm */
1935 			/* handle smac_index */
1936 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
1937 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
1938 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
1939 				err = handle_eth_ud_smac_index(dev, qp, context);
1940 				if (err) {
1941 					err = -EINVAL;
1942 					goto out;
1943 				}
1944 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1945 					dev->qp1_proxy[qp->port - 1] = qp;
1946 			}
1947 		}
1948 	}
1949 
1950 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1951 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
1952 					MLX4_IB_LINK_TYPE_ETH;
1953 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1954 			/* set QP to receive both tunneled & non-tunneled packets */
1955 			if (!(context->flags & cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET)))
1956 				context->srqn = cpu_to_be32(7 << 28);
1957 		}
1958 	}
1959 
1960 	if (ibqp->qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
1961 		int is_eth = rdma_port_get_link_layer(
1962 				&dev->ib_dev, qp->port) ==
1963 				IB_LINK_LAYER_ETHERNET;
1964 		if (is_eth) {
1965 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
1966 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
1967 		}
1968 	}
1969 
1970 
1971 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1972 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1973 		sqd_event = 1;
1974 	else
1975 		sqd_event = 0;
1976 
1977 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1978 		context->rlkey_roce_mode |= (1 << 4);
1979 
1980 	/*
1981 	 * Before passing a kernel QP to the HW, make sure that the
1982 	 * ownership bits of the send queue are set and the SQ
1983 	 * headroom is stamped so that the hardware doesn't start
1984 	 * processing stale work requests.
1985 	 */
1986 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1987 		struct mlx4_wqe_ctrl_seg *ctrl;
1988 		int i;
1989 
1990 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1991 			ctrl = get_send_wqe(qp, i);
1992 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
1993 			if (qp->sq_max_wqes_per_wr == 1)
1994 				ctrl->qpn_vlan.fence_size =
1995 						1 << (qp->sq.wqe_shift - 4);
1996 
1997 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1998 		}
1999 	}
2000 
2001 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2002 			     to_mlx4_state(new_state), context, optpar,
2003 			     sqd_event, &qp->mqp);
2004 	if (err)
2005 		goto out;
2006 
2007 	qp->state = new_state;
2008 
2009 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2010 		qp->atomic_rd_en = attr->qp_access_flags;
2011 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2012 		qp->resp_depth = attr->max_dest_rd_atomic;
2013 	if (attr_mask & IB_QP_PORT) {
2014 		qp->port = attr->port_num;
2015 		update_mcg_macs(dev, qp);
2016 	}
2017 	if (attr_mask & IB_QP_ALT_PATH)
2018 		qp->alt_port = attr->alt_port_num;
2019 
2020 	if (is_sqp(dev, qp))
2021 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2022 
2023 	/*
2024 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
2025 	 * QP0 to RESET or ERROR, bring the link back down.
2026 	 */
2027 	if (is_qp0(dev, qp)) {
2028 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2029 			if (mlx4_INIT_PORT(dev->dev, qp->port))
2030 				pr_warn("INIT_PORT failed for port %d\n",
2031 				       qp->port);
2032 
2033 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2034 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2035 			mlx4_CLOSE_PORT(dev->dev, qp->port);
2036 	}
2037 
2038 	/*
2039 	 * If we moved a kernel QP to RESET, clean up all old CQ
2040 	 * entries and reinitialize the QP.
2041 	 */
2042 	if (new_state == IB_QPS_RESET) {
2043 		if (!ibqp->uobject) {
2044 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2045 					 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
2046 			if (send_cq != recv_cq)
2047 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2048 
2049 			qp->rq.head = 0;
2050 			qp->rq.tail = 0;
2051 			qp->sq.head = 0;
2052 			qp->sq.tail = 0;
2053 			qp->sq_next_wqe = 0;
2054 			if (qp->rq.wqe_cnt)
2055 				*qp->db.db  = 0;
2056 
2057 			if (qp->flags & MLX4_IB_QP_NETIF)
2058 				mlx4_ib_steer_qp_reg(dev, qp, 0);
2059 		}
2060 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2061 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2062 			qp->pri.smac = 0;
2063 			qp->pri.smac_port = 0;
2064 		}
2065 		if (qp->alt.smac) {
2066 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2067 			qp->alt.smac = 0;
2068 		}
2069 		if (qp->pri.vid < 0x1000) {
2070 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2071 			qp->pri.vid = 0xFFFF;
2072 			qp->pri.candidate_vid = 0xFFFF;
2073 			qp->pri.update_vid = 0;
2074 		}
2075 
2076 		if (qp->alt.vid < 0x1000) {
2077 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2078 			qp->alt.vid = 0xFFFF;
2079 			qp->alt.candidate_vid = 0xFFFF;
2080 			qp->alt.update_vid = 0;
2081 		}
2082 	}
2083 out:
2084 	if (err && qp->counter_index)
2085 		mlx4_ib_free_qp_counter(dev, qp);
2086 	if (err && steer_qp)
2087 		mlx4_ib_steer_qp_reg(dev, qp, 0);
2088 	kfree(context);
2089 	if (qp->pri.candidate_smac ||
2090 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2091 		if (err) {
2092 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2093 		} else {
2094 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2095 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2096 			qp->pri.smac = qp->pri.candidate_smac;
2097 			qp->pri.smac_index = qp->pri.candidate_smac_index;
2098 			qp->pri.smac_port = qp->pri.candidate_smac_port;
2099 		}
2100 		qp->pri.candidate_smac = 0;
2101 		qp->pri.candidate_smac_index = 0;
2102 		qp->pri.candidate_smac_port = 0;
2103 	}
2104 	if (qp->alt.candidate_smac) {
2105 		if (err) {
2106 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2107 		} else {
2108 			if (qp->alt.smac)
2109 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2110 			qp->alt.smac = qp->alt.candidate_smac;
2111 			qp->alt.smac_index = qp->alt.candidate_smac_index;
2112 			qp->alt.smac_port = qp->alt.candidate_smac_port;
2113 		}
2114 		qp->alt.candidate_smac = 0;
2115 		qp->alt.candidate_smac_index = 0;
2116 		qp->alt.candidate_smac_port = 0;
2117 	}
2118 
2119 	if (qp->pri.update_vid) {
2120 		if (err) {
2121 			if (qp->pri.candidate_vid < 0x1000)
2122 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2123 						     qp->pri.candidate_vid);
2124 		} else {
2125 			if (qp->pri.vid < 0x1000)
2126 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2127 						     qp->pri.vid);
2128 			qp->pri.vid = qp->pri.candidate_vid;
2129 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2130 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
2131 		}
2132 		qp->pri.candidate_vid = 0xFFFF;
2133 		qp->pri.update_vid = 0;
2134 	}
2135 
2136 	if (qp->alt.update_vid) {
2137 		if (err) {
2138 			if (qp->alt.candidate_vid < 0x1000)
2139 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2140 						     qp->alt.candidate_vid);
2141 		} else {
2142 			if (qp->alt.vid < 0x1000)
2143 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2144 						     qp->alt.vid);
2145 			qp->alt.vid = qp->alt.candidate_vid;
2146 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2147 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2148 		}
2149 		qp->alt.candidate_vid = 0xFFFF;
2150 		qp->alt.update_vid = 0;
2151 	}
2152 
2153 	return err;
2154 }
2155 
2156 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2157 			      int attr_mask, struct ib_udata *udata)
2158 {
2159 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2160 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2161 	enum ib_qp_state cur_state, new_state;
2162 	int err = -EINVAL;
2163 	int ll;
2164 	mutex_lock(&qp->mutex);
2165 
2166 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2167 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2168 
2169 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2170 		ll = IB_LINK_LAYER_UNSPECIFIED;
2171 	} else {
2172 		int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2173 		ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2174 	}
2175 
2176 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2177 				attr_mask, ll)) {
2178 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2179 			 "for transition %d to %d. qp_type %d,"
2180 			 " attr_mask 0x%x\n",
2181 			 ibqp->qp_num, cur_state, new_state,
2182 			 ibqp->qp_type, attr_mask);
2183 		goto out;
2184 	}
2185 
2186 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2187 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2188 			if ((ibqp->qp_type == IB_QPT_RC) ||
2189 			    (ibqp->qp_type == IB_QPT_UD) ||
2190 			    (ibqp->qp_type == IB_QPT_UC) ||
2191 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2192 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2193 				attr->port_num = mlx4_ib_bond_next_port(dev);
2194 			}
2195 		} else {
2196 			/* no sense in changing port_num
2197 			 * when ports are bonded */
2198 			attr_mask &= ~IB_QP_PORT;
2199 		}
2200 	}
2201 
2202 	if ((attr_mask & IB_QP_PORT) &&
2203 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2204 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2205 			 "for transition %d to %d. qp_type %d\n",
2206 			 ibqp->qp_num, attr->port_num, cur_state,
2207 			 new_state, ibqp->qp_type);
2208 		goto out;
2209 	}
2210 
2211 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2212 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2213 	     IB_LINK_LAYER_ETHERNET))
2214 		goto out;
2215 
2216 	if (attr_mask & IB_QP_PKEY_INDEX) {
2217 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2218 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2219 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2220 				 "for transition %d to %d. qp_type %d\n",
2221 				 ibqp->qp_num, attr->pkey_index, cur_state,
2222 				 new_state, ibqp->qp_type);
2223 			goto out;
2224 		}
2225 	}
2226 
2227 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2228 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2229 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2230 			 "Transition %d to %d. qp_type %d\n",
2231 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2232 			 new_state, ibqp->qp_type);
2233 		goto out;
2234 	}
2235 
2236 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2237 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2238 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2239 			 "Transition %d to %d. qp_type %d\n",
2240 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2241 			 new_state, ibqp->qp_type);
2242 		goto out;
2243 	}
2244 
2245 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2246 		err = 0;
2247 		goto out;
2248 	}
2249 
2250 	err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2251 
2252 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2253 		attr->port_num = 1;
2254 
2255 out:
2256 	mutex_unlock(&qp->mutex);
2257 	return err;
2258 }
2259 
2260 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2261 		      int attr_mask, struct ib_udata *udata)
2262 {
2263 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2264 	int ret;
2265 
2266 	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2267 
2268 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2269 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
2270 		int err = 0;
2271 
2272 		if (sqp->roce_v2_gsi)
2273 			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2274 		if (err)
2275 			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2276 			       err);
2277 	}
2278 	return ret;
2279 }
2280 
2281 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2282 {
2283 	int i;
2284 	for (i = 0; i < dev->caps.num_ports; i++) {
2285 		if (qpn == dev->caps.qp0_proxy[i] ||
2286 		    qpn == dev->caps.qp0_tunnel[i]) {
2287 			*qkey = dev->caps.qp0_qkey[i];
2288 			return 0;
2289 		}
2290 	}
2291 	return -EINVAL;
2292 }
2293 
2294 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2295 				  struct ib_ud_wr *wr,
2296 				  void *wqe, unsigned *mlx_seg_len)
2297 {
2298 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2299 	struct ib_device *ib_dev = &mdev->ib_dev;
2300 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2301 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2302 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2303 	u16 pkey;
2304 	u32 qkey;
2305 	int send_size;
2306 	int header_size;
2307 	int spc;
2308 	int i;
2309 
2310 	if (wr->wr.opcode != IB_WR_SEND)
2311 		return -EINVAL;
2312 
2313 	send_size = 0;
2314 
2315 	for (i = 0; i < wr->wr.num_sge; ++i)
2316 		send_size += wr->wr.sg_list[i].length;
2317 
2318 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2319 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2320 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2321 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2322 
2323 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2324 
2325 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2326 		sqp->ud_header.lrh.service_level =
2327 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2328 		sqp->ud_header.lrh.destination_lid =
2329 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2330 		sqp->ud_header.lrh.source_lid =
2331 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2332 	}
2333 
2334 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2335 
2336 	/* force loopback */
2337 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2338 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2339 
2340 	sqp->ud_header.lrh.virtual_lane    = 0;
2341 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2342 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2343 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2344 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2345 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2346 	else
2347 		sqp->ud_header.bth.destination_qpn =
2348 			cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
2349 
2350 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2351 	if (mlx4_is_master(mdev->dev)) {
2352 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2353 			return -EINVAL;
2354 	} else {
2355 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2356 			return -EINVAL;
2357 	}
2358 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2359 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2360 
2361 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2362 	sqp->ud_header.immediate_present = 0;
2363 
2364 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2365 
2366 	/*
2367 	 * Inline data segments may not cross a 64 byte boundary.  If
2368 	 * our UD header is bigger than the space available up to the
2369 	 * next 64 byte boundary in the WQE, use two inline data
2370 	 * segments to hold the UD header.
2371 	 */
2372 	spc = MLX4_INLINE_ALIGN -
2373 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2374 	if (header_size <= spc) {
2375 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2376 		memcpy(inl + 1, sqp->header_buf, header_size);
2377 		i = 1;
2378 	} else {
2379 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2380 		memcpy(inl + 1, sqp->header_buf, spc);
2381 
2382 		inl = (void *) (inl + 1) + spc;
2383 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2384 		/*
2385 		 * Need a barrier here to make sure all the data is
2386 		 * visible before the byte_count field is set.
2387 		 * Otherwise the HCA prefetcher could grab the 64-byte
2388 		 * chunk with this inline segment and get a valid (!=
2389 		 * 0xffffffff) byte count but stale data, and end up
2390 		 * generating a packet with bad headers.
2391 		 *
2392 		 * The first inline segment's byte_count field doesn't
2393 		 * need a barrier, because it comes after a
2394 		 * control/MLX segment and therefore is at an offset
2395 		 * of 16 mod 64.
2396 		 */
2397 		wmb();
2398 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2399 		i = 2;
2400 	}
2401 
2402 	*mlx_seg_len =
2403 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2404 	return 0;
2405 }
2406 
2407 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
2408 {
2409 	union sl2vl_tbl_to_u64 tmp_vltab;
2410 	u8 vl;
2411 
2412 	if (sl > 15)
2413 		return 0xf;
2414 	tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
2415 	vl = tmp_vltab.sl8[sl >> 1];
2416 	if (sl & 1)
2417 		vl &= 0x0f;
2418 	else
2419 		vl >>= 4;
2420 	return vl;
2421 }
2422 
2423 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
2424 				int index, union ib_gid *gid,
2425 				enum ib_gid_type *gid_type)
2426 {
2427 	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
2428 	struct mlx4_port_gid_table *port_gid_table;
2429 	unsigned long flags;
2430 
2431 	port_gid_table = &iboe->gids[port_num - 1];
2432 	spin_lock_irqsave(&iboe->lock, flags);
2433 	memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
2434 	*gid_type = port_gid_table->gids[index].gid_type;
2435 	spin_unlock_irqrestore(&iboe->lock, flags);
2436 	if (!memcmp(gid, &zgid, sizeof(*gid)))
2437 		return -ENOENT;
2438 
2439 	return 0;
2440 }
2441 
2442 #define MLX4_ROCEV2_QP1_SPORT 0xC000
2443 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_ud_wr *wr,
2444 			    void *wqe, unsigned *mlx_seg_len)
2445 {
2446 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
2447 	struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
2448 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2449 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2450 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2451 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2452 	union ib_gid sgid;
2453 	u16 pkey;
2454 	int send_size;
2455 	int header_size;
2456 	int spc;
2457 	int i;
2458 	int err = 0;
2459 	u16 vlan = 0xffff;
2460 	bool is_eth;
2461 	bool is_vlan = false;
2462 	bool is_grh;
2463 	bool is_udp = false;
2464 	int ip_version = 0;
2465 
2466 	send_size = 0;
2467 	for (i = 0; i < wr->wr.num_sge; ++i)
2468 		send_size += wr->wr.sg_list[i].length;
2469 
2470 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2471 	is_grh = mlx4_ib_ah_grh_present(ah);
2472 	if (is_eth) {
2473 		enum ib_gid_type gid_type;
2474 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2475 			/* When multi-function is enabled, the ib_core gid
2476 			 * indexes don't necessarily match the hw ones, so
2477 			 * we must use our own cache */
2478 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2479 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
2480 							   ah->av.ib.gid_index, &sgid.raw[0]);
2481 			if (err)
2482 				return err;
2483 		} else  {
2484 			err = fill_gid_by_hw_index(ibdev, sqp->qp.port,
2485 					    ah->av.ib.gid_index,
2486 					    &sgid, &gid_type);
2487 			if (!err) {
2488 				is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
2489 				if (is_udp) {
2490 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
2491 						ip_version = 4;
2492 					else
2493 						ip_version = 6;
2494 					is_grh = false;
2495 				}
2496 			} else {
2497 				return err;
2498 			}
2499 		}
2500 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
2501 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
2502 			is_vlan = 1;
2503 		}
2504 	}
2505 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
2506 			  ip_version, is_udp, 0, &sqp->ud_header);
2507 	if (err)
2508 		return err;
2509 
2510 	if (!is_eth) {
2511 		sqp->ud_header.lrh.service_level =
2512 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2513 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
2514 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2515 	}
2516 
2517 	if (is_grh || (ip_version == 6)) {
2518 		sqp->ud_header.grh.traffic_class =
2519 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2520 		sqp->ud_header.grh.flow_label    =
2521 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
2522 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
2523 		if (is_eth) {
2524 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
2525 		} else {
2526 			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2527 				/* When multi-function is enabled, the ib_core gid
2528 				 * indexes don't necessarily match the hw ones, so
2529 				 * we must use our own cache
2530 				 */
2531 				sqp->ud_header.grh.source_gid.global.subnet_prefix =
2532 					cpu_to_be64(atomic64_read(&(to_mdev(ib_dev)->sriov.
2533 								    demux[sqp->qp.port - 1].
2534 								    subnet_prefix)));
2535 				sqp->ud_header.grh.source_gid.global.interface_id =
2536 					to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2537 						       guid_cache[ah->av.ib.gid_index];
2538 			} else {
2539 				ib_get_cached_gid(ib_dev,
2540 						  be32_to_cpu(ah->av.ib.port_pd) >> 24,
2541 						  ah->av.ib.gid_index,
2542 						  &sqp->ud_header.grh.source_gid, NULL);
2543 			}
2544 		}
2545 		memcpy(sqp->ud_header.grh.destination_gid.raw,
2546 		       ah->av.ib.dgid, 16);
2547 	}
2548 
2549 	if (ip_version == 4) {
2550 		sqp->ud_header.ip4.tos =
2551 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2552 		sqp->ud_header.ip4.id = 0;
2553 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
2554 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
2555 
2556 		memcpy(&sqp->ud_header.ip4.saddr,
2557 		       sgid.raw + 12, 4);
2558 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
2559 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
2560 	}
2561 
2562 	if (is_udp) {
2563 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
2564 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
2565 		sqp->ud_header.udp.csum = 0;
2566 	}
2567 
2568 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2569 
2570 	if (!is_eth) {
2571 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
2572 					  (sqp->ud_header.lrh.destination_lid ==
2573 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
2574 					  (sqp->ud_header.lrh.service_level << 8));
2575 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
2576 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
2577 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
2578 	}
2579 
2580 	switch (wr->wr.opcode) {
2581 	case IB_WR_SEND:
2582 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
2583 		sqp->ud_header.immediate_present = 0;
2584 		break;
2585 	case IB_WR_SEND_WITH_IMM:
2586 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2587 		sqp->ud_header.immediate_present = 1;
2588 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
2589 		break;
2590 	default:
2591 		return -EINVAL;
2592 	}
2593 
2594 	if (is_eth) {
2595 		struct in6_addr in6;
2596 		u16 ether_type;
2597 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
2598 
2599 		ether_type = (!is_udp) ? ETH_P_IBOE:
2600 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
2601 
2602 		mlx->sched_prio = cpu_to_be16(pcp);
2603 
2604 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
2605 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
2606 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
2607 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
2608 		memcpy(&in6, sgid.raw, sizeof(in6));
2609 
2610 
2611 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
2612 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
2613 		if (!is_vlan) {
2614 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
2615 		} else {
2616 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
2617 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
2618 		}
2619 	} else {
2620 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 :
2621 							sl_to_vl(to_mdev(ib_dev),
2622 								 sqp->ud_header.lrh.service_level,
2623 								 sqp->qp.port);
2624 		if (sqp->qp.ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
2625 			return -EINVAL;
2626 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
2627 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
2628 	}
2629 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2630 	if (!sqp->qp.ibqp.qp_num)
2631 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
2632 	else
2633 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
2634 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2635 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2636 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2637 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
2638 					       sqp->qkey : wr->remote_qkey);
2639 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
2640 
2641 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2642 
2643 	if (0) {
2644 		pr_err("built UD header of size %d:\n", header_size);
2645 		for (i = 0; i < header_size / 4; ++i) {
2646 			if (i % 8 == 0)
2647 				pr_err("  [%02x] ", i * 4);
2648 			pr_cont(" %08x",
2649 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
2650 			if ((i + 1) % 8 == 0)
2651 				pr_cont("\n");
2652 		}
2653 		pr_err("\n");
2654 	}
2655 
2656 	/*
2657 	 * Inline data segments may not cross a 64 byte boundary.  If
2658 	 * our UD header is bigger than the space available up to the
2659 	 * next 64 byte boundary in the WQE, use two inline data
2660 	 * segments to hold the UD header.
2661 	 */
2662 	spc = MLX4_INLINE_ALIGN -
2663 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2664 	if (header_size <= spc) {
2665 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2666 		memcpy(inl + 1, sqp->header_buf, header_size);
2667 		i = 1;
2668 	} else {
2669 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2670 		memcpy(inl + 1, sqp->header_buf, spc);
2671 
2672 		inl = (void *) (inl + 1) + spc;
2673 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2674 		/*
2675 		 * Need a barrier here to make sure all the data is
2676 		 * visible before the byte_count field is set.
2677 		 * Otherwise the HCA prefetcher could grab the 64-byte
2678 		 * chunk with this inline segment and get a valid (!=
2679 		 * 0xffffffff) byte count but stale data, and end up
2680 		 * generating a packet with bad headers.
2681 		 *
2682 		 * The first inline segment's byte_count field doesn't
2683 		 * need a barrier, because it comes after a
2684 		 * control/MLX segment and therefore is at an offset
2685 		 * of 16 mod 64.
2686 		 */
2687 		wmb();
2688 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2689 		i = 2;
2690 	}
2691 
2692 	*mlx_seg_len =
2693 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2694 	return 0;
2695 }
2696 
2697 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2698 {
2699 	unsigned cur;
2700 	struct mlx4_ib_cq *cq;
2701 
2702 	cur = wq->head - wq->tail;
2703 	if (likely(cur + nreq < wq->max_post))
2704 		return 0;
2705 
2706 	cq = to_mcq(ib_cq);
2707 	spin_lock(&cq->lock);
2708 	cur = wq->head - wq->tail;
2709 	spin_unlock(&cq->lock);
2710 
2711 	return cur + nreq >= wq->max_post;
2712 }
2713 
2714 static __be32 convert_access(int acc)
2715 {
2716 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
2717 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
2718 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
2719 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
2720 	       (acc & IB_ACCESS_REMOTE_READ   ?
2721 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
2722 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
2723 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
2724 }
2725 
2726 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
2727 			struct ib_reg_wr *wr)
2728 {
2729 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
2730 
2731 	fseg->flags		= convert_access(wr->access);
2732 	fseg->mem_key		= cpu_to_be32(wr->key);
2733 	fseg->buf_list		= cpu_to_be64(mr->page_map);
2734 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
2735 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
2736 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
2737 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
2738 	fseg->reserved[0]	= 0;
2739 	fseg->reserved[1]	= 0;
2740 }
2741 
2742 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2743 {
2744 	memset(iseg, 0, sizeof(*iseg));
2745 	iseg->mem_key = cpu_to_be32(rkey);
2746 }
2747 
2748 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2749 					  u64 remote_addr, u32 rkey)
2750 {
2751 	rseg->raddr    = cpu_to_be64(remote_addr);
2752 	rseg->rkey     = cpu_to_be32(rkey);
2753 	rseg->reserved = 0;
2754 }
2755 
2756 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
2757 		struct ib_atomic_wr *wr)
2758 {
2759 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2760 		aseg->swap_add = cpu_to_be64(wr->swap);
2761 		aseg->compare  = cpu_to_be64(wr->compare_add);
2762 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2763 		aseg->swap_add = cpu_to_be64(wr->compare_add);
2764 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
2765 	} else {
2766 		aseg->swap_add = cpu_to_be64(wr->compare_add);
2767 		aseg->compare  = 0;
2768 	}
2769 
2770 }
2771 
2772 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2773 				  struct ib_atomic_wr *wr)
2774 {
2775 	aseg->swap_add		= cpu_to_be64(wr->swap);
2776 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
2777 	aseg->compare		= cpu_to_be64(wr->compare_add);
2778 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
2779 }
2780 
2781 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2782 			     struct ib_ud_wr *wr)
2783 {
2784 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
2785 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
2786 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
2787 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
2788 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
2789 }
2790 
2791 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2792 				    struct mlx4_wqe_datagram_seg *dseg,
2793 				    struct ib_ud_wr *wr,
2794 				    enum mlx4_ib_qp_type qpt)
2795 {
2796 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
2797 	struct mlx4_av sqp_av = {0};
2798 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
2799 
2800 	/* force loopback */
2801 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2802 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2803 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2804 			cpu_to_be32(0xf0000000);
2805 
2806 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2807 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
2808 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2809 	else
2810 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp0_tunnel[port - 1]);
2811 	/* Use QKEY from the QP context, which is set by master */
2812 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2813 }
2814 
2815 static void build_tunnel_header(struct ib_ud_wr *wr, void *wqe, unsigned *mlx_seg_len)
2816 {
2817 	struct mlx4_wqe_inline_seg *inl = wqe;
2818 	struct mlx4_ib_tunnel_header hdr;
2819 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2820 	int spc;
2821 	int i;
2822 
2823 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2824 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
2825 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
2826 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
2827 	memcpy(hdr.mac, ah->av.eth.mac, 6);
2828 	hdr.vlan = ah->av.eth.vlan;
2829 
2830 	spc = MLX4_INLINE_ALIGN -
2831 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2832 	if (sizeof (hdr) <= spc) {
2833 		memcpy(inl + 1, &hdr, sizeof (hdr));
2834 		wmb();
2835 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
2836 		i = 1;
2837 	} else {
2838 		memcpy(inl + 1, &hdr, spc);
2839 		wmb();
2840 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2841 
2842 		inl = (void *) (inl + 1) + spc;
2843 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2844 		wmb();
2845 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
2846 		i = 2;
2847 	}
2848 
2849 	*mlx_seg_len =
2850 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2851 }
2852 
2853 static void set_mlx_icrc_seg(void *dseg)
2854 {
2855 	u32 *t = dseg;
2856 	struct mlx4_wqe_inline_seg *iseg = dseg;
2857 
2858 	t[1] = 0;
2859 
2860 	/*
2861 	 * Need a barrier here before writing the byte_count field to
2862 	 * make sure that all the data is visible before the
2863 	 * byte_count field is set.  Otherwise, if the segment begins
2864 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2865 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2866 	 * stale data, and end up sending the wrong data.
2867 	 */
2868 	wmb();
2869 
2870 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2871 }
2872 
2873 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2874 {
2875 	dseg->lkey       = cpu_to_be32(sg->lkey);
2876 	dseg->addr       = cpu_to_be64(sg->addr);
2877 
2878 	/*
2879 	 * Need a barrier here before writing the byte_count field to
2880 	 * make sure that all the data is visible before the
2881 	 * byte_count field is set.  Otherwise, if the segment begins
2882 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2883 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2884 	 * stale data, and end up sending the wrong data.
2885 	 */
2886 	wmb();
2887 
2888 	dseg->byte_count = cpu_to_be32(sg->length);
2889 }
2890 
2891 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2892 {
2893 	dseg->byte_count = cpu_to_be32(sg->length);
2894 	dseg->lkey       = cpu_to_be32(sg->lkey);
2895 	dseg->addr       = cpu_to_be64(sg->addr);
2896 }
2897 
2898 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_ud_wr *wr,
2899 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2900 			 __be32 *lso_hdr_sz, __be32 *blh)
2901 {
2902 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
2903 
2904 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2905 		*blh = cpu_to_be32(1 << 6);
2906 
2907 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2908 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
2909 		return -EINVAL;
2910 
2911 	memcpy(wqe->header, wr->header, wr->hlen);
2912 
2913 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
2914 	*lso_seg_len = halign;
2915 	return 0;
2916 }
2917 
2918 static __be32 send_ieth(struct ib_send_wr *wr)
2919 {
2920 	switch (wr->opcode) {
2921 	case IB_WR_SEND_WITH_IMM:
2922 	case IB_WR_RDMA_WRITE_WITH_IMM:
2923 		return wr->ex.imm_data;
2924 
2925 	case IB_WR_SEND_WITH_INV:
2926 		return cpu_to_be32(wr->ex.invalidate_rkey);
2927 
2928 	default:
2929 		return 0;
2930 	}
2931 }
2932 
2933 static void add_zero_len_inline(void *wqe)
2934 {
2935 	struct mlx4_wqe_inline_seg *inl = wqe;
2936 	memset(wqe, 0, 16);
2937 	inl->byte_count = cpu_to_be32(1 << 31);
2938 }
2939 
2940 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2941 		      struct ib_send_wr **bad_wr)
2942 {
2943 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2944 	void *wqe;
2945 	struct mlx4_wqe_ctrl_seg *ctrl;
2946 	struct mlx4_wqe_data_seg *dseg;
2947 	unsigned long flags;
2948 	int nreq;
2949 	int err = 0;
2950 	unsigned ind;
2951 	int uninitialized_var(stamp);
2952 	int uninitialized_var(size);
2953 	unsigned uninitialized_var(seglen);
2954 	__be32 dummy;
2955 	__be32 *lso_wqe;
2956 	__be32 uninitialized_var(lso_hdr_sz);
2957 	__be32 blh;
2958 	int i;
2959 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
2960 
2961 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2962 		struct mlx4_ib_sqp *sqp = to_msqp(qp);
2963 
2964 		if (sqp->roce_v2_gsi) {
2965 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
2966 			enum ib_gid_type gid_type;
2967 			union ib_gid gid;
2968 
2969 			if (!fill_gid_by_hw_index(mdev, sqp->qp.port,
2970 					   ah->av.ib.gid_index,
2971 					   &gid, &gid_type))
2972 				qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
2973 						to_mqp(sqp->roce_v2_gsi) : qp;
2974 			else
2975 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
2976 				       ah->av.ib.gid_index);
2977 		}
2978 	}
2979 
2980 	spin_lock_irqsave(&qp->sq.lock, flags);
2981 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
2982 		err = -EIO;
2983 		*bad_wr = wr;
2984 		nreq = 0;
2985 		goto out;
2986 	}
2987 
2988 	ind = qp->sq_next_wqe;
2989 
2990 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2991 		lso_wqe = &dummy;
2992 		blh = 0;
2993 
2994 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2995 			err = -ENOMEM;
2996 			*bad_wr = wr;
2997 			goto out;
2998 		}
2999 
3000 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3001 			err = -EINVAL;
3002 			*bad_wr = wr;
3003 			goto out;
3004 		}
3005 
3006 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3007 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3008 
3009 		ctrl->srcrb_flags =
3010 			(wr->send_flags & IB_SEND_SIGNALED ?
3011 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3012 			(wr->send_flags & IB_SEND_SOLICITED ?
3013 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3014 			((wr->send_flags & IB_SEND_IP_CSUM) ?
3015 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3016 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3017 			qp->sq_signal_bits;
3018 
3019 		ctrl->imm = send_ieth(wr);
3020 
3021 		wqe += sizeof *ctrl;
3022 		size = sizeof *ctrl / 16;
3023 
3024 		switch (qp->mlx4_ib_qp_type) {
3025 		case MLX4_IB_QPT_RC:
3026 		case MLX4_IB_QPT_UC:
3027 			switch (wr->opcode) {
3028 			case IB_WR_ATOMIC_CMP_AND_SWP:
3029 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3030 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3031 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3032 					      atomic_wr(wr)->rkey);
3033 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3034 
3035 				set_atomic_seg(wqe, atomic_wr(wr));
3036 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
3037 
3038 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3039 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3040 
3041 				break;
3042 
3043 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3044 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3045 					      atomic_wr(wr)->rkey);
3046 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3047 
3048 				set_masked_atomic_seg(wqe, atomic_wr(wr));
3049 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
3050 
3051 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3052 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3053 
3054 				break;
3055 
3056 			case IB_WR_RDMA_READ:
3057 			case IB_WR_RDMA_WRITE:
3058 			case IB_WR_RDMA_WRITE_WITH_IMM:
3059 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3060 					      rdma_wr(wr)->rkey);
3061 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3062 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3063 				break;
3064 
3065 			case IB_WR_LOCAL_INV:
3066 				ctrl->srcrb_flags |=
3067 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3068 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3069 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3070 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3071 				break;
3072 
3073 			case IB_WR_REG_MR:
3074 				ctrl->srcrb_flags |=
3075 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3076 				set_reg_seg(wqe, reg_wr(wr));
3077 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3078 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3079 				break;
3080 
3081 			default:
3082 				/* No extra segments required for sends */
3083 				break;
3084 			}
3085 			break;
3086 
3087 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3088 			err =  build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3089 					ctrl, &seglen);
3090 			if (unlikely(err)) {
3091 				*bad_wr = wr;
3092 				goto out;
3093 			}
3094 			wqe  += seglen;
3095 			size += seglen / 16;
3096 			break;
3097 		case MLX4_IB_QPT_TUN_SMI:
3098 		case MLX4_IB_QPT_TUN_GSI:
3099 			/* this is a UD qp used in MAD responses to slaves. */
3100 			set_datagram_seg(wqe, ud_wr(wr));
3101 			/* set the forced-loopback bit in the data seg av */
3102 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3103 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3104 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3105 			break;
3106 		case MLX4_IB_QPT_UD:
3107 			set_datagram_seg(wqe, ud_wr(wr));
3108 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3109 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3110 
3111 			if (wr->opcode == IB_WR_LSO) {
3112 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3113 						&lso_hdr_sz, &blh);
3114 				if (unlikely(err)) {
3115 					*bad_wr = wr;
3116 					goto out;
3117 				}
3118 				lso_wqe = (__be32 *) wqe;
3119 				wqe  += seglen;
3120 				size += seglen / 16;
3121 			}
3122 			break;
3123 
3124 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3125 			err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3126 					ctrl, &seglen);
3127 			if (unlikely(err)) {
3128 				*bad_wr = wr;
3129 				goto out;
3130 			}
3131 			wqe  += seglen;
3132 			size += seglen / 16;
3133 			/* to start tunnel header on a cache-line boundary */
3134 			add_zero_len_inline(wqe);
3135 			wqe += 16;
3136 			size++;
3137 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3138 			wqe  += seglen;
3139 			size += seglen / 16;
3140 			break;
3141 		case MLX4_IB_QPT_PROXY_SMI:
3142 		case MLX4_IB_QPT_PROXY_GSI:
3143 			/* If we are tunneling special qps, this is a UD qp.
3144 			 * In this case we first add a UD segment targeting
3145 			 * the tunnel qp, and then add a header with address
3146 			 * information */
3147 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3148 						ud_wr(wr),
3149 						qp->mlx4_ib_qp_type);
3150 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3151 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3152 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3153 			wqe  += seglen;
3154 			size += seglen / 16;
3155 			break;
3156 
3157 		case MLX4_IB_QPT_SMI:
3158 		case MLX4_IB_QPT_GSI:
3159 			err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
3160 					&seglen);
3161 			if (unlikely(err)) {
3162 				*bad_wr = wr;
3163 				goto out;
3164 			}
3165 			wqe  += seglen;
3166 			size += seglen / 16;
3167 			break;
3168 
3169 		default:
3170 			break;
3171 		}
3172 
3173 		/*
3174 		 * Write data segments in reverse order, so as to
3175 		 * overwrite cacheline stamp last within each
3176 		 * cacheline.  This avoids issues with WQE
3177 		 * prefetching.
3178 		 */
3179 
3180 		dseg = wqe;
3181 		dseg += wr->num_sge - 1;
3182 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3183 
3184 		/* Add one more inline data segment for ICRC for MLX sends */
3185 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3186 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3187 			     qp->mlx4_ib_qp_type &
3188 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3189 			set_mlx_icrc_seg(dseg + 1);
3190 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3191 		}
3192 
3193 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3194 			set_data_seg(dseg, wr->sg_list + i);
3195 
3196 		/*
3197 		 * Possibly overwrite stamping in cacheline with LSO
3198 		 * segment only after making sure all data segments
3199 		 * are written.
3200 		 */
3201 		wmb();
3202 		*lso_wqe = lso_hdr_sz;
3203 
3204 		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3205 					     MLX4_WQE_CTRL_FENCE : 0) | size;
3206 
3207 		/*
3208 		 * Make sure descriptor is fully written before
3209 		 * setting ownership bit (because HW can start
3210 		 * executing as soon as we do).
3211 		 */
3212 		wmb();
3213 
3214 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3215 			*bad_wr = wr;
3216 			err = -EINVAL;
3217 			goto out;
3218 		}
3219 
3220 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3221 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3222 
3223 		stamp = ind + qp->sq_spare_wqes;
3224 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
3225 
3226 		/*
3227 		 * We can improve latency by not stamping the last
3228 		 * send queue WQE until after ringing the doorbell, so
3229 		 * only stamp here if there are still more WQEs to post.
3230 		 *
3231 		 * Same optimization applies to padding with NOP wqe
3232 		 * in case of WQE shrinking (used to prevent wrap-around
3233 		 * in the middle of WR).
3234 		 */
3235 		if (wr->next) {
3236 			stamp_send_wqe(qp, stamp, size * 16);
3237 			ind = pad_wraparound(qp, ind);
3238 		}
3239 	}
3240 
3241 out:
3242 	if (likely(nreq)) {
3243 		qp->sq.head += nreq;
3244 
3245 		/*
3246 		 * Make sure that descriptors are written before
3247 		 * doorbell record.
3248 		 */
3249 		wmb();
3250 
3251 		writel(qp->doorbell_qpn,
3252 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3253 
3254 		/*
3255 		 * Make sure doorbells don't leak out of SQ spinlock
3256 		 * and reach the HCA out of order.
3257 		 */
3258 		mmiowb();
3259 
3260 		stamp_send_wqe(qp, stamp, size * 16);
3261 
3262 		ind = pad_wraparound(qp, ind);
3263 		qp->sq_next_wqe = ind;
3264 	}
3265 
3266 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3267 
3268 	return err;
3269 }
3270 
3271 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3272 		      struct ib_recv_wr **bad_wr)
3273 {
3274 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3275 	struct mlx4_wqe_data_seg *scat;
3276 	unsigned long flags;
3277 	int err = 0;
3278 	int nreq;
3279 	int ind;
3280 	int max_gs;
3281 	int i;
3282 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3283 
3284 	max_gs = qp->rq.max_gs;
3285 	spin_lock_irqsave(&qp->rq.lock, flags);
3286 
3287 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3288 		err = -EIO;
3289 		*bad_wr = wr;
3290 		nreq = 0;
3291 		goto out;
3292 	}
3293 
3294 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3295 
3296 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3297 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3298 			err = -ENOMEM;
3299 			*bad_wr = wr;
3300 			goto out;
3301 		}
3302 
3303 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3304 			err = -EINVAL;
3305 			*bad_wr = wr;
3306 			goto out;
3307 		}
3308 
3309 		scat = get_recv_wqe(qp, ind);
3310 
3311 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3312 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3313 			ib_dma_sync_single_for_device(ibqp->device,
3314 						      qp->sqp_proxy_rcv[ind].map,
3315 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3316 						      DMA_FROM_DEVICE);
3317 			scat->byte_count =
3318 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3319 			/* use dma lkey from upper layer entry */
3320 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3321 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3322 			scat++;
3323 			max_gs--;
3324 		}
3325 
3326 		for (i = 0; i < wr->num_sge; ++i)
3327 			__set_data_seg(scat + i, wr->sg_list + i);
3328 
3329 		if (i < max_gs) {
3330 			scat[i].byte_count = 0;
3331 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3332 			scat[i].addr       = 0;
3333 		}
3334 
3335 		qp->rq.wrid[ind] = wr->wr_id;
3336 
3337 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3338 	}
3339 
3340 out:
3341 	if (likely(nreq)) {
3342 		qp->rq.head += nreq;
3343 
3344 		/*
3345 		 * Make sure that descriptors are written before
3346 		 * doorbell record.
3347 		 */
3348 		wmb();
3349 
3350 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3351 	}
3352 
3353 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3354 
3355 	return err;
3356 }
3357 
3358 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3359 {
3360 	switch (mlx4_state) {
3361 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3362 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3363 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3364 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3365 	case MLX4_QP_STATE_SQ_DRAINING:
3366 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3367 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3368 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3369 	default:		     return -1;
3370 	}
3371 }
3372 
3373 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3374 {
3375 	switch (mlx4_mig_state) {
3376 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3377 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3378 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3379 	default: return -1;
3380 	}
3381 }
3382 
3383 static int to_ib_qp_access_flags(int mlx4_flags)
3384 {
3385 	int ib_flags = 0;
3386 
3387 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3388 		ib_flags |= IB_ACCESS_REMOTE_READ;
3389 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3390 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3391 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3392 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3393 
3394 	return ib_flags;
3395 }
3396 
3397 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
3398 				struct mlx4_qp_path *path)
3399 {
3400 	struct mlx4_dev *dev = ibdev->dev;
3401 	int is_eth;
3402 
3403 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
3404 	ib_ah_attr->port_num	  = path->sched_queue & 0x40 ? 2 : 1;
3405 
3406 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
3407 		return;
3408 
3409 	is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
3410 		IB_LINK_LAYER_ETHERNET;
3411 	if (is_eth)
3412 		ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
3413 		((path->sched_queue & 4) << 1);
3414 	else
3415 		ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
3416 
3417 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
3418 	ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
3419 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
3420 	ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
3421 	if (ib_ah_attr->ah_flags) {
3422 		ib_ah_attr->grh.sgid_index = path->mgid_index;
3423 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
3424 		ib_ah_attr->grh.traffic_class =
3425 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3426 		ib_ah_attr->grh.flow_label =
3427 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3428 		memcpy(ib_ah_attr->grh.dgid.raw,
3429 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
3430 	}
3431 }
3432 
3433 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3434 		     struct ib_qp_init_attr *qp_init_attr)
3435 {
3436 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3437 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3438 	struct mlx4_qp_context context;
3439 	int mlx4_state;
3440 	int err = 0;
3441 
3442 	mutex_lock(&qp->mutex);
3443 
3444 	if (qp->state == IB_QPS_RESET) {
3445 		qp_attr->qp_state = IB_QPS_RESET;
3446 		goto done;
3447 	}
3448 
3449 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3450 	if (err) {
3451 		err = -EINVAL;
3452 		goto out;
3453 	}
3454 
3455 	mlx4_state = be32_to_cpu(context.flags) >> 28;
3456 
3457 	qp->state		     = to_ib_qp_state(mlx4_state);
3458 	qp_attr->qp_state	     = qp->state;
3459 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
3460 	qp_attr->path_mig_state	     =
3461 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3462 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
3463 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3464 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
3465 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
3466 	qp_attr->qp_access_flags     =
3467 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
3468 
3469 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3470 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3471 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3472 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3473 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
3474 	}
3475 
3476 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3477 	if (qp_attr->qp_state == IB_QPS_INIT)
3478 		qp_attr->port_num = qp->port;
3479 	else
3480 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3481 
3482 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3483 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3484 
3485 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3486 
3487 	qp_attr->max_dest_rd_atomic =
3488 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3489 	qp_attr->min_rnr_timer	    =
3490 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3491 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
3492 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
3493 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
3494 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
3495 
3496 done:
3497 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
3498 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
3499 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
3500 
3501 	if (!ibqp->uobject) {
3502 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
3503 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
3504 	} else {
3505 		qp_attr->cap.max_send_wr  = 0;
3506 		qp_attr->cap.max_send_sge = 0;
3507 	}
3508 
3509 	/*
3510 	 * We don't support inline sends for kernel QPs (yet), and we
3511 	 * don't know what userspace's value should be.
3512 	 */
3513 	qp_attr->cap.max_inline_data = 0;
3514 
3515 	qp_init_attr->cap	     = qp_attr->cap;
3516 
3517 	qp_init_attr->create_flags = 0;
3518 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3519 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3520 
3521 	if (qp->flags & MLX4_IB_QP_LSO)
3522 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
3523 
3524 	if (qp->flags & MLX4_IB_QP_NETIF)
3525 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
3526 
3527 	qp_init_attr->sq_sig_type =
3528 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
3529 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3530 
3531 out:
3532 	mutex_unlock(&qp->mutex);
3533 	return err;
3534 }
3535 
3536