xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision a06c488d)
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 "user.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 enum {
80 	MLX4_IB_IBOE_ETHERTYPE		= 0x8915
81 };
82 
83 struct mlx4_ib_sqp {
84 	struct mlx4_ib_qp	qp;
85 	int			pkey_index;
86 	u32			qkey;
87 	u32			send_psn;
88 	struct ib_ud_header	ud_header;
89 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
90 	struct ib_qp		*roce_v2_gsi;
91 };
92 
93 enum {
94 	MLX4_IB_MIN_SQ_STRIDE	= 6,
95 	MLX4_IB_CACHE_LINE_SIZE	= 64,
96 };
97 
98 enum {
99 	MLX4_RAW_QP_MTU		= 7,
100 	MLX4_RAW_QP_MSGMAX	= 31,
101 };
102 
103 #ifndef ETH_ALEN
104 #define ETH_ALEN        6
105 #endif
106 
107 static const __be32 mlx4_ib_opcode[] = {
108 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
109 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
110 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
111 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
112 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
113 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
114 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
115 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
116 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
117 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
118 	[IB_WR_REG_MR]				= cpu_to_be32(MLX4_OPCODE_FMR),
119 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
120 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
121 };
122 
123 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
124 {
125 	return container_of(mqp, struct mlx4_ib_sqp, qp);
126 }
127 
128 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
129 {
130 	if (!mlx4_is_master(dev->dev))
131 		return 0;
132 
133 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
134 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
135 		8 * MLX4_MFUNC_MAX;
136 }
137 
138 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
139 {
140 	int proxy_sqp = 0;
141 	int real_sqp = 0;
142 	int i;
143 	/* PPF or Native -- real SQP */
144 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
145 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
146 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
147 	if (real_sqp)
148 		return 1;
149 	/* VF or PF -- proxy SQP */
150 	if (mlx4_is_mfunc(dev->dev)) {
151 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
152 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] ||
153 			    qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) {
154 				proxy_sqp = 1;
155 				break;
156 			}
157 		}
158 	}
159 	if (proxy_sqp)
160 		return 1;
161 
162 	return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
163 }
164 
165 /* used for INIT/CLOSE port logic */
166 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
167 {
168 	int proxy_qp0 = 0;
169 	int real_qp0 = 0;
170 	int i;
171 	/* PPF or Native -- real QP0 */
172 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
173 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
174 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
175 	if (real_qp0)
176 		return 1;
177 	/* VF or PF -- proxy QP0 */
178 	if (mlx4_is_mfunc(dev->dev)) {
179 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
180 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) {
181 				proxy_qp0 = 1;
182 				break;
183 			}
184 		}
185 	}
186 	return proxy_qp0;
187 }
188 
189 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
190 {
191 	return mlx4_buf_offset(&qp->buf, offset);
192 }
193 
194 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
195 {
196 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
197 }
198 
199 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
200 {
201 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
202 }
203 
204 /*
205  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
206  * first four bytes of every 64 byte chunk with
207  *     0x7FFFFFF | (invalid_ownership_value << 31).
208  *
209  * When the max work request size is less than or equal to the WQE
210  * basic block size, as an optimization, we can stamp all WQEs with
211  * 0xffffffff, and skip the very first chunk of each WQE.
212  */
213 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
214 {
215 	__be32 *wqe;
216 	int i;
217 	int s;
218 	int ind;
219 	void *buf;
220 	__be32 stamp;
221 	struct mlx4_wqe_ctrl_seg *ctrl;
222 
223 	if (qp->sq_max_wqes_per_wr > 1) {
224 		s = roundup(size, 1U << qp->sq.wqe_shift);
225 		for (i = 0; i < s; i += 64) {
226 			ind = (i >> qp->sq.wqe_shift) + n;
227 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
228 						       cpu_to_be32(0xffffffff);
229 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
230 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
231 			*wqe = stamp;
232 		}
233 	} else {
234 		ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
235 		s = (ctrl->fence_size & 0x3f) << 4;
236 		for (i = 64; i < s; i += 64) {
237 			wqe = buf + i;
238 			*wqe = cpu_to_be32(0xffffffff);
239 		}
240 	}
241 }
242 
243 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
244 {
245 	struct mlx4_wqe_ctrl_seg *ctrl;
246 	struct mlx4_wqe_inline_seg *inl;
247 	void *wqe;
248 	int s;
249 
250 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
251 	s = sizeof(struct mlx4_wqe_ctrl_seg);
252 
253 	if (qp->ibqp.qp_type == IB_QPT_UD) {
254 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
255 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
256 		memset(dgram, 0, sizeof *dgram);
257 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
258 		s += sizeof(struct mlx4_wqe_datagram_seg);
259 	}
260 
261 	/* Pad the remainder of the WQE with an inline data segment. */
262 	if (size > s) {
263 		inl = wqe + s;
264 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
265 	}
266 	ctrl->srcrb_flags = 0;
267 	ctrl->fence_size = size / 16;
268 	/*
269 	 * Make sure descriptor is fully written before setting ownership bit
270 	 * (because HW can start executing as soon as we do).
271 	 */
272 	wmb();
273 
274 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
275 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
276 
277 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
278 }
279 
280 /* Post NOP WQE to prevent wrap-around in the middle of WR */
281 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
282 {
283 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
284 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
285 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
286 		ind += s;
287 	}
288 	return ind;
289 }
290 
291 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
292 {
293 	struct ib_event event;
294 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
295 
296 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
297 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
298 
299 	if (ibqp->event_handler) {
300 		event.device     = ibqp->device;
301 		event.element.qp = ibqp;
302 		switch (type) {
303 		case MLX4_EVENT_TYPE_PATH_MIG:
304 			event.event = IB_EVENT_PATH_MIG;
305 			break;
306 		case MLX4_EVENT_TYPE_COMM_EST:
307 			event.event = IB_EVENT_COMM_EST;
308 			break;
309 		case MLX4_EVENT_TYPE_SQ_DRAINED:
310 			event.event = IB_EVENT_SQ_DRAINED;
311 			break;
312 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
313 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
314 			break;
315 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
316 			event.event = IB_EVENT_QP_FATAL;
317 			break;
318 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
319 			event.event = IB_EVENT_PATH_MIG_ERR;
320 			break;
321 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
322 			event.event = IB_EVENT_QP_REQ_ERR;
323 			break;
324 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
325 			event.event = IB_EVENT_QP_ACCESS_ERR;
326 			break;
327 		default:
328 			pr_warn("Unexpected event type %d "
329 			       "on QP %06x\n", type, qp->qpn);
330 			return;
331 		}
332 
333 		ibqp->event_handler(&event, ibqp->qp_context);
334 	}
335 }
336 
337 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
338 {
339 	/*
340 	 * UD WQEs must have a datagram segment.
341 	 * RC and UC WQEs might have a remote address segment.
342 	 * MLX WQEs need two extra inline data segments (for the UD
343 	 * header and space for the ICRC).
344 	 */
345 	switch (type) {
346 	case MLX4_IB_QPT_UD:
347 		return sizeof (struct mlx4_wqe_ctrl_seg) +
348 			sizeof (struct mlx4_wqe_datagram_seg) +
349 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
350 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
351 	case MLX4_IB_QPT_PROXY_SMI:
352 	case MLX4_IB_QPT_PROXY_GSI:
353 		return sizeof (struct mlx4_wqe_ctrl_seg) +
354 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
355 	case MLX4_IB_QPT_TUN_SMI_OWNER:
356 	case MLX4_IB_QPT_TUN_GSI:
357 		return sizeof (struct mlx4_wqe_ctrl_seg) +
358 			sizeof (struct mlx4_wqe_datagram_seg);
359 
360 	case MLX4_IB_QPT_UC:
361 		return sizeof (struct mlx4_wqe_ctrl_seg) +
362 			sizeof (struct mlx4_wqe_raddr_seg);
363 	case MLX4_IB_QPT_RC:
364 		return sizeof (struct mlx4_wqe_ctrl_seg) +
365 			sizeof (struct mlx4_wqe_atomic_seg) +
366 			sizeof (struct mlx4_wqe_raddr_seg);
367 	case MLX4_IB_QPT_SMI:
368 	case MLX4_IB_QPT_GSI:
369 		return sizeof (struct mlx4_wqe_ctrl_seg) +
370 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
371 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
372 					   MLX4_INLINE_ALIGN) *
373 			      sizeof (struct mlx4_wqe_inline_seg),
374 			      sizeof (struct mlx4_wqe_data_seg)) +
375 			ALIGN(4 +
376 			      sizeof (struct mlx4_wqe_inline_seg),
377 			      sizeof (struct mlx4_wqe_data_seg));
378 	default:
379 		return sizeof (struct mlx4_wqe_ctrl_seg);
380 	}
381 }
382 
383 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
384 		       int is_user, int has_rq, struct mlx4_ib_qp *qp)
385 {
386 	/* Sanity check RQ size before proceeding */
387 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
388 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
389 		return -EINVAL;
390 
391 	if (!has_rq) {
392 		if (cap->max_recv_wr)
393 			return -EINVAL;
394 
395 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
396 	} else {
397 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
398 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
399 			return -EINVAL;
400 
401 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
402 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
403 		qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
404 	}
405 
406 	/* leave userspace return values as they were, so as not to break ABI */
407 	if (is_user) {
408 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
409 		cap->max_recv_sge = qp->rq.max_gs;
410 	} else {
411 		cap->max_recv_wr  = qp->rq.max_post =
412 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
413 		cap->max_recv_sge = min(qp->rq.max_gs,
414 					min(dev->dev->caps.max_sq_sg,
415 					    dev->dev->caps.max_rq_sg));
416 	}
417 
418 	return 0;
419 }
420 
421 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
422 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
423 {
424 	int s;
425 
426 	/* Sanity check SQ size before proceeding */
427 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
428 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
429 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
430 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
431 		return -EINVAL;
432 
433 	/*
434 	 * For MLX transport we need 2 extra S/G entries:
435 	 * one for the header and one for the checksum at the end
436 	 */
437 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
438 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
439 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
440 		return -EINVAL;
441 
442 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
443 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
444 		send_wqe_overhead(type, qp->flags);
445 
446 	if (s > dev->dev->caps.max_sq_desc_sz)
447 		return -EINVAL;
448 
449 	/*
450 	 * Hermon supports shrinking WQEs, such that a single work
451 	 * request can include multiple units of 1 << wqe_shift.  This
452 	 * way, work requests can differ in size, and do not have to
453 	 * be a power of 2 in size, saving memory and speeding up send
454 	 * WR posting.  Unfortunately, if we do this then the
455 	 * wqe_index field in CQEs can't be used to look up the WR ID
456 	 * anymore, so we do this only if selective signaling is off.
457 	 *
458 	 * Further, on 32-bit platforms, we can't use vmap() to make
459 	 * the QP buffer virtually contiguous.  Thus we have to use
460 	 * constant-sized WRs to make sure a WR is always fully within
461 	 * a single page-sized chunk.
462 	 *
463 	 * Finally, we use NOP work requests to pad the end of the
464 	 * work queue, to avoid wrap-around in the middle of WR.  We
465 	 * set NEC bit to avoid getting completions with error for
466 	 * these NOP WRs, but since NEC is only supported starting
467 	 * with firmware 2.2.232, we use constant-sized WRs for older
468 	 * firmware.
469 	 *
470 	 * And, since MLX QPs only support SEND, we use constant-sized
471 	 * WRs in this case.
472 	 *
473 	 * We look for the smallest value of wqe_shift such that the
474 	 * resulting number of wqes does not exceed device
475 	 * capabilities.
476 	 *
477 	 * We set WQE size to at least 64 bytes, this way stamping
478 	 * invalidates each WQE.
479 	 */
480 	if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
481 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
482 	    type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
483 	    !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
484 		      MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
485 		qp->sq.wqe_shift = ilog2(64);
486 	else
487 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
488 
489 	for (;;) {
490 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
491 
492 		/*
493 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
494 		 * allow HW to prefetch.
495 		 */
496 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
497 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
498 						    qp->sq_max_wqes_per_wr +
499 						    qp->sq_spare_wqes);
500 
501 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
502 			break;
503 
504 		if (qp->sq_max_wqes_per_wr <= 1)
505 			return -EINVAL;
506 
507 		++qp->sq.wqe_shift;
508 	}
509 
510 	qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
511 			     (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
512 			 send_wqe_overhead(type, qp->flags)) /
513 		sizeof (struct mlx4_wqe_data_seg);
514 
515 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
516 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
517 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
518 		qp->rq.offset = 0;
519 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
520 	} else {
521 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
522 		qp->sq.offset = 0;
523 	}
524 
525 	cap->max_send_wr  = qp->sq.max_post =
526 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
527 	cap->max_send_sge = min(qp->sq.max_gs,
528 				min(dev->dev->caps.max_sq_sg,
529 				    dev->dev->caps.max_rq_sg));
530 	/* We don't support inline sends for kernel QPs (yet) */
531 	cap->max_inline_data = 0;
532 
533 	return 0;
534 }
535 
536 static int set_user_sq_size(struct mlx4_ib_dev *dev,
537 			    struct mlx4_ib_qp *qp,
538 			    struct mlx4_ib_create_qp *ucmd)
539 {
540 	/* Sanity check SQ size before proceeding */
541 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
542 	    ucmd->log_sq_stride >
543 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
544 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
545 		return -EINVAL;
546 
547 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
548 	qp->sq.wqe_shift = ucmd->log_sq_stride;
549 
550 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
551 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
552 
553 	return 0;
554 }
555 
556 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
557 {
558 	int i;
559 
560 	qp->sqp_proxy_rcv =
561 		kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
562 			GFP_KERNEL);
563 	if (!qp->sqp_proxy_rcv)
564 		return -ENOMEM;
565 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
566 		qp->sqp_proxy_rcv[i].addr =
567 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
568 				GFP_KERNEL);
569 		if (!qp->sqp_proxy_rcv[i].addr)
570 			goto err;
571 		qp->sqp_proxy_rcv[i].map =
572 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
573 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
574 					  DMA_FROM_DEVICE);
575 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
576 			kfree(qp->sqp_proxy_rcv[i].addr);
577 			goto err;
578 		}
579 	}
580 	return 0;
581 
582 err:
583 	while (i > 0) {
584 		--i;
585 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
586 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
587 				    DMA_FROM_DEVICE);
588 		kfree(qp->sqp_proxy_rcv[i].addr);
589 	}
590 	kfree(qp->sqp_proxy_rcv);
591 	qp->sqp_proxy_rcv = NULL;
592 	return -ENOMEM;
593 }
594 
595 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
596 {
597 	int i;
598 
599 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
600 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
601 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
602 				    DMA_FROM_DEVICE);
603 		kfree(qp->sqp_proxy_rcv[i].addr);
604 	}
605 	kfree(qp->sqp_proxy_rcv);
606 }
607 
608 static int qp_has_rq(struct ib_qp_init_attr *attr)
609 {
610 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
611 		return 0;
612 
613 	return !attr->srq;
614 }
615 
616 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
617 {
618 	int i;
619 	for (i = 0; i < dev->caps.num_ports; i++) {
620 		if (qpn == dev->caps.qp0_proxy[i])
621 			return !!dev->caps.qp0_qkey[i];
622 	}
623 	return 0;
624 }
625 
626 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
627 				    struct mlx4_ib_qp *qp)
628 {
629 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
630 	mlx4_counter_free(dev->dev, qp->counter_index->index);
631 	list_del(&qp->counter_index->list);
632 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
633 
634 	kfree(qp->counter_index);
635 	qp->counter_index = NULL;
636 }
637 
638 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
639 			    struct ib_qp_init_attr *init_attr,
640 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp,
641 			    gfp_t gfp)
642 {
643 	int qpn;
644 	int err;
645 	struct mlx4_ib_sqp *sqp;
646 	struct mlx4_ib_qp *qp;
647 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
648 	struct mlx4_ib_cq *mcq;
649 	unsigned long flags;
650 
651 	/* When tunneling special qps, we use a plain UD qp */
652 	if (sqpn) {
653 		if (mlx4_is_mfunc(dev->dev) &&
654 		    (!mlx4_is_master(dev->dev) ||
655 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
656 			if (init_attr->qp_type == IB_QPT_GSI)
657 				qp_type = MLX4_IB_QPT_PROXY_GSI;
658 			else {
659 				if (mlx4_is_master(dev->dev) ||
660 				    qp0_enabled_vf(dev->dev, sqpn))
661 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
662 				else
663 					qp_type = MLX4_IB_QPT_PROXY_SMI;
664 			}
665 		}
666 		qpn = sqpn;
667 		/* add extra sg entry for tunneling */
668 		init_attr->cap.max_recv_sge++;
669 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
670 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
671 			container_of(init_attr,
672 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
673 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
674 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
675 		    !mlx4_is_master(dev->dev))
676 			return -EINVAL;
677 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
678 			qp_type = MLX4_IB_QPT_TUN_GSI;
679 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
680 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
681 					     tnl_init->port))
682 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
683 		else
684 			qp_type = MLX4_IB_QPT_TUN_SMI;
685 		/* we are definitely in the PPF here, since we are creating
686 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
687 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
688 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
689 		sqpn = qpn;
690 	}
691 
692 	if (!*caller_qp) {
693 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
694 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
695 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
696 			sqp = kzalloc(sizeof (struct mlx4_ib_sqp), gfp);
697 			if (!sqp)
698 				return -ENOMEM;
699 			qp = &sqp->qp;
700 			qp->pri.vid = 0xFFFF;
701 			qp->alt.vid = 0xFFFF;
702 		} else {
703 			qp = kzalloc(sizeof (struct mlx4_ib_qp), gfp);
704 			if (!qp)
705 				return -ENOMEM;
706 			qp->pri.vid = 0xFFFF;
707 			qp->alt.vid = 0xFFFF;
708 		}
709 	} else
710 		qp = *caller_qp;
711 
712 	qp->mlx4_ib_qp_type = qp_type;
713 
714 	mutex_init(&qp->mutex);
715 	spin_lock_init(&qp->sq.lock);
716 	spin_lock_init(&qp->rq.lock);
717 	INIT_LIST_HEAD(&qp->gid_list);
718 	INIT_LIST_HEAD(&qp->steering_rules);
719 
720 	qp->state	 = IB_QPS_RESET;
721 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
722 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
723 
724 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
725 	if (err)
726 		goto err;
727 
728 	if (pd->uobject) {
729 		struct mlx4_ib_create_qp ucmd;
730 
731 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
732 			err = -EFAULT;
733 			goto err;
734 		}
735 
736 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
737 
738 		err = set_user_sq_size(dev, qp, &ucmd);
739 		if (err)
740 			goto err;
741 
742 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
743 				       qp->buf_size, 0, 0);
744 		if (IS_ERR(qp->umem)) {
745 			err = PTR_ERR(qp->umem);
746 			goto err;
747 		}
748 
749 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
750 				    ilog2(qp->umem->page_size), &qp->mtt);
751 		if (err)
752 			goto err_buf;
753 
754 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
755 		if (err)
756 			goto err_mtt;
757 
758 		if (qp_has_rq(init_attr)) {
759 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
760 						  ucmd.db_addr, &qp->db);
761 			if (err)
762 				goto err_mtt;
763 		}
764 	} else {
765 		qp->sq_no_prefetch = 0;
766 
767 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
768 			qp->flags |= MLX4_IB_QP_LSO;
769 
770 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
771 			if (dev->steering_support ==
772 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
773 				qp->flags |= MLX4_IB_QP_NETIF;
774 			else
775 				goto err;
776 		}
777 
778 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
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, PAGE_SIZE * 2, &qp->buf, gfp)) {
791 			err = -ENOMEM;
792 			goto err_db;
793 		}
794 
795 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
796 				    &qp->mtt);
797 		if (err)
798 			goto err_buf;
799 
800 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf, gfp);
801 		if (err)
802 			goto err_mtt;
803 
804 		qp->sq.wrid = kmalloc_array(qp->sq.wqe_cnt, sizeof(u64),
805 					gfp | __GFP_NOWARN);
806 		if (!qp->sq.wrid)
807 			qp->sq.wrid = __vmalloc(qp->sq.wqe_cnt * sizeof(u64),
808 						gfp, PAGE_KERNEL);
809 		qp->rq.wrid = kmalloc_array(qp->rq.wqe_cnt, sizeof(u64),
810 					gfp | __GFP_NOWARN);
811 		if (!qp->rq.wrid)
812 			qp->rq.wrid = __vmalloc(qp->rq.wqe_cnt * sizeof(u64),
813 						gfp, PAGE_KERNEL);
814 		if (!qp->sq.wrid || !qp->rq.wrid) {
815 			err = -ENOMEM;
816 			goto err_wrid;
817 		}
818 	}
819 
820 	if (sqpn) {
821 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
822 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
823 			if (alloc_proxy_bufs(pd->device, qp)) {
824 				err = -ENOMEM;
825 				goto err_wrid;
826 			}
827 		}
828 	} else {
829 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
830 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
831 		 * VLAN insertion. */
832 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
833 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
834 						    (init_attr->cap.max_send_wr ?
835 						     MLX4_RESERVE_ETH_BF_QP : 0) |
836 						    (init_attr->cap.max_recv_wr ?
837 						     MLX4_RESERVE_A0_QP : 0));
838 		else
839 			if (qp->flags & MLX4_IB_QP_NETIF)
840 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
841 			else
842 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
843 							    &qpn, 0);
844 		if (err)
845 			goto err_proxy;
846 	}
847 
848 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
849 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
850 
851 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp, gfp);
852 	if (err)
853 		goto err_qpn;
854 
855 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
856 		qp->mqp.qpn |= (1 << 23);
857 
858 	/*
859 	 * Hardware wants QPN written in big-endian order (after
860 	 * shifting) for send doorbell.  Precompute this value to save
861 	 * a little bit when posting sends.
862 	 */
863 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
864 
865 	qp->mqp.event = mlx4_ib_qp_event;
866 	if (!*caller_qp)
867 		*caller_qp = qp;
868 
869 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
870 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
871 			 to_mcq(init_attr->recv_cq));
872 	/* Maintain device to QPs access, needed for further handling
873 	 * via reset flow
874 	 */
875 	list_add_tail(&qp->qps_list, &dev->qp_list);
876 	/* Maintain CQ to QPs access, needed for further handling
877 	 * via reset flow
878 	 */
879 	mcq = to_mcq(init_attr->send_cq);
880 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
881 	mcq = to_mcq(init_attr->recv_cq);
882 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
883 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
884 			   to_mcq(init_attr->recv_cq));
885 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
886 	return 0;
887 
888 err_qpn:
889 	if (!sqpn) {
890 		if (qp->flags & MLX4_IB_QP_NETIF)
891 			mlx4_ib_steer_qp_free(dev, qpn, 1);
892 		else
893 			mlx4_qp_release_range(dev->dev, qpn, 1);
894 	}
895 err_proxy:
896 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
897 		free_proxy_bufs(pd->device, qp);
898 err_wrid:
899 	if (pd->uobject) {
900 		if (qp_has_rq(init_attr))
901 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
902 	} else {
903 		kvfree(qp->sq.wrid);
904 		kvfree(qp->rq.wrid);
905 	}
906 
907 err_mtt:
908 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
909 
910 err_buf:
911 	if (pd->uobject)
912 		ib_umem_release(qp->umem);
913 	else
914 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
915 
916 err_db:
917 	if (!pd->uobject && qp_has_rq(init_attr))
918 		mlx4_db_free(dev->dev, &qp->db);
919 
920 err:
921 	if (!*caller_qp)
922 		kfree(qp);
923 	return err;
924 }
925 
926 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
927 {
928 	switch (state) {
929 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
930 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
931 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
932 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
933 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
934 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
935 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
936 	default:		return -1;
937 	}
938 }
939 
940 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
941 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
942 {
943 	if (send_cq == recv_cq) {
944 		spin_lock(&send_cq->lock);
945 		__acquire(&recv_cq->lock);
946 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
947 		spin_lock(&send_cq->lock);
948 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
949 	} else {
950 		spin_lock(&recv_cq->lock);
951 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
952 	}
953 }
954 
955 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
956 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
957 {
958 	if (send_cq == recv_cq) {
959 		__release(&recv_cq->lock);
960 		spin_unlock(&send_cq->lock);
961 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
962 		spin_unlock(&recv_cq->lock);
963 		spin_unlock(&send_cq->lock);
964 	} else {
965 		spin_unlock(&send_cq->lock);
966 		spin_unlock(&recv_cq->lock);
967 	}
968 }
969 
970 static void del_gid_entries(struct mlx4_ib_qp *qp)
971 {
972 	struct mlx4_ib_gid_entry *ge, *tmp;
973 
974 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
975 		list_del(&ge->list);
976 		kfree(ge);
977 	}
978 }
979 
980 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
981 {
982 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
983 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
984 	else
985 		return to_mpd(qp->ibqp.pd);
986 }
987 
988 static void get_cqs(struct mlx4_ib_qp *qp,
989 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
990 {
991 	switch (qp->ibqp.qp_type) {
992 	case IB_QPT_XRC_TGT:
993 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
994 		*recv_cq = *send_cq;
995 		break;
996 	case IB_QPT_XRC_INI:
997 		*send_cq = to_mcq(qp->ibqp.send_cq);
998 		*recv_cq = *send_cq;
999 		break;
1000 	default:
1001 		*send_cq = to_mcq(qp->ibqp.send_cq);
1002 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
1003 		break;
1004 	}
1005 }
1006 
1007 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1008 			      int is_user)
1009 {
1010 	struct mlx4_ib_cq *send_cq, *recv_cq;
1011 	unsigned long flags;
1012 
1013 	if (qp->state != IB_QPS_RESET) {
1014 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1015 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1016 			pr_warn("modify QP %06x to RESET failed.\n",
1017 			       qp->mqp.qpn);
1018 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1019 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1020 			qp->pri.smac = 0;
1021 			qp->pri.smac_port = 0;
1022 		}
1023 		if (qp->alt.smac) {
1024 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1025 			qp->alt.smac = 0;
1026 		}
1027 		if (qp->pri.vid < 0x1000) {
1028 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1029 			qp->pri.vid = 0xFFFF;
1030 			qp->pri.candidate_vid = 0xFFFF;
1031 			qp->pri.update_vid = 0;
1032 		}
1033 		if (qp->alt.vid < 0x1000) {
1034 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1035 			qp->alt.vid = 0xFFFF;
1036 			qp->alt.candidate_vid = 0xFFFF;
1037 			qp->alt.update_vid = 0;
1038 		}
1039 	}
1040 
1041 	get_cqs(qp, &send_cq, &recv_cq);
1042 
1043 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1044 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1045 
1046 	/* del from lists under both locks above to protect reset flow paths */
1047 	list_del(&qp->qps_list);
1048 	list_del(&qp->cq_send_list);
1049 	list_del(&qp->cq_recv_list);
1050 	if (!is_user) {
1051 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1052 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1053 		if (send_cq != recv_cq)
1054 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1055 	}
1056 
1057 	mlx4_qp_remove(dev->dev, &qp->mqp);
1058 
1059 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1060 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1061 
1062 	mlx4_qp_free(dev->dev, &qp->mqp);
1063 
1064 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1065 		if (qp->flags & MLX4_IB_QP_NETIF)
1066 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1067 		else
1068 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1069 	}
1070 
1071 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1072 
1073 	if (is_user) {
1074 		if (qp->rq.wqe_cnt)
1075 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
1076 					      &qp->db);
1077 		ib_umem_release(qp->umem);
1078 	} else {
1079 		kvfree(qp->sq.wrid);
1080 		kvfree(qp->rq.wrid);
1081 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1082 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1083 			free_proxy_bufs(&dev->ib_dev, qp);
1084 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1085 		if (qp->rq.wqe_cnt)
1086 			mlx4_db_free(dev->dev, &qp->db);
1087 	}
1088 
1089 	del_gid_entries(qp);
1090 }
1091 
1092 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1093 {
1094 	/* Native or PPF */
1095 	if (!mlx4_is_mfunc(dev->dev) ||
1096 	    (mlx4_is_master(dev->dev) &&
1097 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1098 		return  dev->dev->phys_caps.base_sqpn +
1099 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1100 			attr->port_num - 1;
1101 	}
1102 	/* PF or VF -- creating proxies */
1103 	if (attr->qp_type == IB_QPT_SMI)
1104 		return dev->dev->caps.qp0_proxy[attr->port_num - 1];
1105 	else
1106 		return dev->dev->caps.qp1_proxy[attr->port_num - 1];
1107 }
1108 
1109 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1110 					struct ib_qp_init_attr *init_attr,
1111 					struct ib_udata *udata)
1112 {
1113 	struct mlx4_ib_qp *qp = NULL;
1114 	int err;
1115 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1116 	u16 xrcdn = 0;
1117 	gfp_t gfp;
1118 
1119 	gfp = (init_attr->create_flags & MLX4_IB_QP_CREATE_USE_GFP_NOIO) ?
1120 		GFP_NOIO : GFP_KERNEL;
1121 	/*
1122 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1123 	 * and only for kernel UD QPs.
1124 	 */
1125 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1126 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1127 					MLX4_IB_SRIOV_TUNNEL_QP |
1128 					MLX4_IB_SRIOV_SQP |
1129 					MLX4_IB_QP_NETIF |
1130 					MLX4_IB_QP_CREATE_ROCE_V2_GSI |
1131 					MLX4_IB_QP_CREATE_USE_GFP_NOIO))
1132 		return ERR_PTR(-EINVAL);
1133 
1134 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1135 		if (init_attr->qp_type != IB_QPT_UD)
1136 			return ERR_PTR(-EINVAL);
1137 	}
1138 
1139 	if (init_attr->create_flags) {
1140 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1141 			return ERR_PTR(-EINVAL);
1142 
1143 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1144 						 MLX4_IB_QP_CREATE_USE_GFP_NOIO |
1145 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1146 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1147 		     init_attr->qp_type != IB_QPT_UD) ||
1148 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1149 		     init_attr->qp_type > IB_QPT_GSI) ||
1150 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1151 		     init_attr->qp_type != IB_QPT_GSI))
1152 			return ERR_PTR(-EINVAL);
1153 	}
1154 
1155 	switch (init_attr->qp_type) {
1156 	case IB_QPT_XRC_TGT:
1157 		pd = to_mxrcd(init_attr->xrcd)->pd;
1158 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1159 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1160 		/* fall through */
1161 	case IB_QPT_XRC_INI:
1162 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1163 			return ERR_PTR(-ENOSYS);
1164 		init_attr->recv_cq = init_attr->send_cq;
1165 		/* fall through */
1166 	case IB_QPT_RC:
1167 	case IB_QPT_UC:
1168 	case IB_QPT_RAW_PACKET:
1169 		qp = kzalloc(sizeof *qp, gfp);
1170 		if (!qp)
1171 			return ERR_PTR(-ENOMEM);
1172 		qp->pri.vid = 0xFFFF;
1173 		qp->alt.vid = 0xFFFF;
1174 		/* fall through */
1175 	case IB_QPT_UD:
1176 	{
1177 		err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1178 				       udata, 0, &qp, gfp);
1179 		if (err)
1180 			return ERR_PTR(err);
1181 
1182 		qp->ibqp.qp_num = qp->mqp.qpn;
1183 		qp->xrcdn = xrcdn;
1184 
1185 		break;
1186 	}
1187 	case IB_QPT_SMI:
1188 	case IB_QPT_GSI:
1189 	{
1190 		int sqpn;
1191 
1192 		/* Userspace is not allowed to create special QPs: */
1193 		if (udata)
1194 			return ERR_PTR(-EINVAL);
1195 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1196 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev, 1, 1, &sqpn, 0);
1197 
1198 			if (res)
1199 				return ERR_PTR(res);
1200 		} else {
1201 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1202 		}
1203 
1204 		err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1205 				       sqpn,
1206 				       &qp, gfp);
1207 		if (err)
1208 			return ERR_PTR(err);
1209 
1210 		qp->port	= init_attr->port_num;
1211 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1212 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1213 		break;
1214 	}
1215 	default:
1216 		/* Don't support raw QPs */
1217 		return ERR_PTR(-EINVAL);
1218 	}
1219 
1220 	return &qp->ibqp;
1221 }
1222 
1223 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1224 				struct ib_qp_init_attr *init_attr,
1225 				struct ib_udata *udata) {
1226 	struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1227 	struct ib_qp *ibqp;
1228 	struct mlx4_ib_dev *dev = to_mdev(device);
1229 
1230 	ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1231 
1232 	if (!IS_ERR(ibqp) &&
1233 	    (init_attr->qp_type == IB_QPT_GSI) &&
1234 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1235 		struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1236 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1237 
1238 		if (is_eth &&
1239 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1240 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1241 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1242 
1243 			if (IS_ERR(sqp->roce_v2_gsi)) {
1244 				pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1245 				sqp->roce_v2_gsi = NULL;
1246 			} else {
1247 				sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1248 				sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1249 			}
1250 
1251 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1252 		}
1253 	}
1254 	return ibqp;
1255 }
1256 
1257 static int _mlx4_ib_destroy_qp(struct ib_qp *qp)
1258 {
1259 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1260 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1261 	struct mlx4_ib_pd *pd;
1262 
1263 	if (is_qp0(dev, mqp))
1264 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1265 
1266 	if (dev->qp1_proxy[mqp->port - 1] == mqp) {
1267 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1268 		dev->qp1_proxy[mqp->port - 1] = NULL;
1269 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1270 	}
1271 
1272 	if (mqp->counter_index)
1273 		mlx4_ib_free_qp_counter(dev, mqp);
1274 
1275 	pd = get_pd(mqp);
1276 	destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1277 
1278 	if (is_sqp(dev, mqp))
1279 		kfree(to_msqp(mqp));
1280 	else
1281 		kfree(mqp);
1282 
1283 	return 0;
1284 }
1285 
1286 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1287 {
1288 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1289 
1290 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1291 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1292 
1293 		if (sqp->roce_v2_gsi)
1294 			ib_destroy_qp(sqp->roce_v2_gsi);
1295 	}
1296 
1297 	return _mlx4_ib_destroy_qp(qp);
1298 }
1299 
1300 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1301 {
1302 	switch (type) {
1303 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1304 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1305 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1306 	case MLX4_IB_QPT_XRC_INI:
1307 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1308 	case MLX4_IB_QPT_SMI:
1309 	case MLX4_IB_QPT_GSI:
1310 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1311 
1312 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1313 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1314 						MLX4_QP_ST_MLX : -1);
1315 	case MLX4_IB_QPT_PROXY_SMI:
1316 	case MLX4_IB_QPT_TUN_SMI:
1317 	case MLX4_IB_QPT_PROXY_GSI:
1318 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1319 						MLX4_QP_ST_UD : -1);
1320 	default:			return -1;
1321 	}
1322 }
1323 
1324 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1325 				   int attr_mask)
1326 {
1327 	u8 dest_rd_atomic;
1328 	u32 access_flags;
1329 	u32 hw_access_flags = 0;
1330 
1331 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1332 		dest_rd_atomic = attr->max_dest_rd_atomic;
1333 	else
1334 		dest_rd_atomic = qp->resp_depth;
1335 
1336 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1337 		access_flags = attr->qp_access_flags;
1338 	else
1339 		access_flags = qp->atomic_rd_en;
1340 
1341 	if (!dest_rd_atomic)
1342 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1343 
1344 	if (access_flags & IB_ACCESS_REMOTE_READ)
1345 		hw_access_flags |= MLX4_QP_BIT_RRE;
1346 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1347 		hw_access_flags |= MLX4_QP_BIT_RAE;
1348 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1349 		hw_access_flags |= MLX4_QP_BIT_RWE;
1350 
1351 	return cpu_to_be32(hw_access_flags);
1352 }
1353 
1354 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1355 			    int attr_mask)
1356 {
1357 	if (attr_mask & IB_QP_PKEY_INDEX)
1358 		sqp->pkey_index = attr->pkey_index;
1359 	if (attr_mask & IB_QP_QKEY)
1360 		sqp->qkey = attr->qkey;
1361 	if (attr_mask & IB_QP_SQ_PSN)
1362 		sqp->send_psn = attr->sq_psn;
1363 }
1364 
1365 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1366 {
1367 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1368 }
1369 
1370 static int _mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
1371 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1372 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1373 {
1374 	int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
1375 		IB_LINK_LAYER_ETHERNET;
1376 	int vidx;
1377 	int smac_index;
1378 	int err;
1379 
1380 
1381 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
1382 	path->rlid	    = cpu_to_be16(ah->dlid);
1383 	if (ah->static_rate) {
1384 		path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
1385 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1386 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1387 			--path->static_rate;
1388 	} else
1389 		path->static_rate = 0;
1390 
1391 	if (ah->ah_flags & IB_AH_GRH) {
1392 		int real_sgid_index = mlx4_ib_gid_index_to_real_index(dev,
1393 								      port,
1394 								      ah->grh.sgid_index);
1395 
1396 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1397 			pr_err("sgid_index (%u) too large. max is %d\n",
1398 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1399 			return -1;
1400 		}
1401 
1402 		path->grh_mylmc |= 1 << 7;
1403 		path->mgid_index = real_sgid_index;
1404 		path->hop_limit  = ah->grh.hop_limit;
1405 		path->tclass_flowlabel =
1406 			cpu_to_be32((ah->grh.traffic_class << 20) |
1407 				    (ah->grh.flow_label));
1408 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1409 	}
1410 
1411 	if (is_eth) {
1412 		if (!(ah->ah_flags & IB_AH_GRH))
1413 			return -1;
1414 
1415 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1416 			((port - 1) << 6) | ((ah->sl & 7) << 3);
1417 
1418 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1419 		if (vlan_tag < 0x1000) {
1420 			if (smac_info->vid < 0x1000) {
1421 				/* both valid vlan ids */
1422 				if (smac_info->vid != vlan_tag) {
1423 					/* different VIDs.  unreg old and reg new */
1424 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1425 					if (err)
1426 						return err;
1427 					smac_info->candidate_vid = vlan_tag;
1428 					smac_info->candidate_vlan_index = vidx;
1429 					smac_info->candidate_vlan_port = port;
1430 					smac_info->update_vid = 1;
1431 					path->vlan_index = vidx;
1432 				} else {
1433 					path->vlan_index = smac_info->vlan_index;
1434 				}
1435 			} else {
1436 				/* no current vlan tag in qp */
1437 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1438 				if (err)
1439 					return err;
1440 				smac_info->candidate_vid = vlan_tag;
1441 				smac_info->candidate_vlan_index = vidx;
1442 				smac_info->candidate_vlan_port = port;
1443 				smac_info->update_vid = 1;
1444 				path->vlan_index = vidx;
1445 			}
1446 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1447 			path->fl = 1 << 6;
1448 		} else {
1449 			/* have current vlan tag. unregister it at modify-qp success */
1450 			if (smac_info->vid < 0x1000) {
1451 				smac_info->candidate_vid = 0xFFFF;
1452 				smac_info->update_vid = 1;
1453 			}
1454 		}
1455 
1456 		/* get smac_index for RoCE use.
1457 		 * If no smac was yet assigned, register one.
1458 		 * If one was already assigned, but the new mac differs,
1459 		 * unregister the old one and register the new one.
1460 		*/
1461 		if ((!smac_info->smac && !smac_info->smac_port) ||
1462 		    smac_info->smac != smac) {
1463 			/* register candidate now, unreg if needed, after success */
1464 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1465 			if (smac_index >= 0) {
1466 				smac_info->candidate_smac_index = smac_index;
1467 				smac_info->candidate_smac = smac;
1468 				smac_info->candidate_smac_port = port;
1469 			} else {
1470 				return -EINVAL;
1471 			}
1472 		} else {
1473 			smac_index = smac_info->smac_index;
1474 		}
1475 
1476 		memcpy(path->dmac, ah->dmac, 6);
1477 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1478 		/* put MAC table smac index for IBoE */
1479 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1480 	} else {
1481 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1482 			((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1483 	}
1484 
1485 	return 0;
1486 }
1487 
1488 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1489 			 enum ib_qp_attr_mask qp_attr_mask,
1490 			 struct mlx4_ib_qp *mqp,
1491 			 struct mlx4_qp_path *path, u8 port,
1492 			 u16 vlan_id, u8 *smac)
1493 {
1494 	return _mlx4_set_path(dev, &qp->ah_attr,
1495 			      mlx4_mac_to_u64(smac),
1496 			      vlan_id,
1497 			      path, &mqp->pri, port);
1498 }
1499 
1500 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1501 			     const struct ib_qp_attr *qp,
1502 			     enum ib_qp_attr_mask qp_attr_mask,
1503 			     struct mlx4_ib_qp *mqp,
1504 			     struct mlx4_qp_path *path, u8 port)
1505 {
1506 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1507 			      0,
1508 			      0xffff,
1509 			      path, &mqp->alt, port);
1510 }
1511 
1512 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1513 {
1514 	struct mlx4_ib_gid_entry *ge, *tmp;
1515 
1516 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1517 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1518 			ge->added = 1;
1519 			ge->port = qp->port;
1520 		}
1521 	}
1522 }
1523 
1524 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1525 				    struct mlx4_ib_qp *qp,
1526 				    struct mlx4_qp_context *context)
1527 {
1528 	u64 u64_mac;
1529 	int smac_index;
1530 
1531 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1532 
1533 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1534 	if (!qp->pri.smac && !qp->pri.smac_port) {
1535 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1536 		if (smac_index >= 0) {
1537 			qp->pri.candidate_smac_index = smac_index;
1538 			qp->pri.candidate_smac = u64_mac;
1539 			qp->pri.candidate_smac_port = qp->port;
1540 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1541 		} else {
1542 			return -ENOENT;
1543 		}
1544 	}
1545 	return 0;
1546 }
1547 
1548 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1549 {
1550 	struct counter_index *new_counter_index;
1551 	int err;
1552 	u32 tmp_idx;
1553 
1554 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1555 	    IB_LINK_LAYER_ETHERNET ||
1556 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1557 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1558 		return 0;
1559 
1560 	err = mlx4_counter_alloc(dev->dev, &tmp_idx);
1561 	if (err)
1562 		return err;
1563 
1564 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1565 	if (!new_counter_index) {
1566 		mlx4_counter_free(dev->dev, tmp_idx);
1567 		return -ENOMEM;
1568 	}
1569 
1570 	new_counter_index->index = tmp_idx;
1571 	new_counter_index->allocated = 1;
1572 	qp->counter_index = new_counter_index;
1573 
1574 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
1575 	list_add_tail(&new_counter_index->list,
1576 		      &dev->counters_table[qp->port - 1].counters_list);
1577 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
1578 
1579 	return 0;
1580 }
1581 
1582 enum {
1583 	MLX4_QPC_ROCE_MODE_1 = 0,
1584 	MLX4_QPC_ROCE_MODE_2 = 2,
1585 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
1586 };
1587 
1588 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
1589 {
1590 	switch (gid_type) {
1591 	case IB_GID_TYPE_ROCE:
1592 		return MLX4_QPC_ROCE_MODE_1;
1593 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
1594 		return MLX4_QPC_ROCE_MODE_2;
1595 	default:
1596 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
1597 	}
1598 }
1599 
1600 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1601 			       const struct ib_qp_attr *attr, int attr_mask,
1602 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1603 {
1604 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1605 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1606 	struct mlx4_ib_pd *pd;
1607 	struct mlx4_ib_cq *send_cq, *recv_cq;
1608 	struct mlx4_qp_context *context;
1609 	enum mlx4_qp_optpar optpar = 0;
1610 	int sqd_event;
1611 	int steer_qp = 0;
1612 	int err = -EINVAL;
1613 	int counter_index;
1614 
1615 	/* APM is not supported under RoCE */
1616 	if (attr_mask & IB_QP_ALT_PATH &&
1617 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1618 	    IB_LINK_LAYER_ETHERNET)
1619 		return -ENOTSUPP;
1620 
1621 	context = kzalloc(sizeof *context, GFP_KERNEL);
1622 	if (!context)
1623 		return -ENOMEM;
1624 
1625 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1626 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1627 
1628 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1629 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1630 	else {
1631 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
1632 		switch (attr->path_mig_state) {
1633 		case IB_MIG_MIGRATED:
1634 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1635 			break;
1636 		case IB_MIG_REARM:
1637 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1638 			break;
1639 		case IB_MIG_ARMED:
1640 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1641 			break;
1642 		}
1643 	}
1644 
1645 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1646 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1647 	else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1648 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1649 	else if (ibqp->qp_type == IB_QPT_UD) {
1650 		if (qp->flags & MLX4_IB_QP_LSO)
1651 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
1652 					      ilog2(dev->dev->caps.max_gso_sz);
1653 		else
1654 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1655 	} else if (attr_mask & IB_QP_PATH_MTU) {
1656 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1657 			pr_err("path MTU (%u) is invalid\n",
1658 			       attr->path_mtu);
1659 			goto out;
1660 		}
1661 		context->mtu_msgmax = (attr->path_mtu << 5) |
1662 			ilog2(dev->dev->caps.max_msg_sz);
1663 	}
1664 
1665 	if (qp->rq.wqe_cnt)
1666 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1667 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
1668 
1669 	if (qp->sq.wqe_cnt)
1670 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1671 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
1672 
1673 	if (new_state == IB_QPS_RESET && qp->counter_index)
1674 		mlx4_ib_free_qp_counter(dev, qp);
1675 
1676 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1677 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1678 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1679 		if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1680 			context->param3 |= cpu_to_be32(1 << 30);
1681 	}
1682 
1683 	if (qp->ibqp.uobject)
1684 		context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
1685 	else
1686 		context->usr_page = cpu_to_be32(dev->priv_uar.index);
1687 
1688 	if (attr_mask & IB_QP_DEST_QPN)
1689 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1690 
1691 	if (attr_mask & IB_QP_PORT) {
1692 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1693 		    !(attr_mask & IB_QP_AV)) {
1694 			mlx4_set_sched(&context->pri_path, attr->port_num);
1695 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1696 		}
1697 	}
1698 
1699 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1700 		err = create_qp_lb_counter(dev, qp);
1701 		if (err)
1702 			goto out;
1703 
1704 		counter_index =
1705 			dev->counters_table[qp->port - 1].default_counter;
1706 		if (qp->counter_index)
1707 			counter_index = qp->counter_index->index;
1708 
1709 		if (counter_index != -1) {
1710 			context->pri_path.counter_index = counter_index;
1711 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1712 			if (qp->counter_index) {
1713 				context->pri_path.fl |=
1714 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
1715 				context->pri_path.vlan_control |=
1716 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
1717 			}
1718 		} else
1719 			context->pri_path.counter_index =
1720 				MLX4_SINK_COUNTER_INDEX(dev->dev);
1721 
1722 		if (qp->flags & MLX4_IB_QP_NETIF) {
1723 			mlx4_ib_steer_qp_reg(dev, qp, 1);
1724 			steer_qp = 1;
1725 		}
1726 
1727 		if (ibqp->qp_type == IB_QPT_GSI) {
1728 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
1729 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
1730 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
1731 
1732 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
1733 		}
1734 	}
1735 
1736 	if (attr_mask & IB_QP_PKEY_INDEX) {
1737 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1738 			context->pri_path.disable_pkey_check = 0x40;
1739 		context->pri_path.pkey_index = attr->pkey_index;
1740 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1741 	}
1742 
1743 	if (attr_mask & IB_QP_AV) {
1744 		u8 port_num = mlx4_is_bonded(to_mdev(ibqp->device)->dev) ? 1 :
1745 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1746 		union ib_gid gid;
1747 		struct ib_gid_attr gid_attr;
1748 		u16 vlan = 0xffff;
1749 		u8 smac[ETH_ALEN];
1750 		int status = 0;
1751 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
1752 			attr->ah_attr.ah_flags & IB_AH_GRH;
1753 
1754 		if (is_eth) {
1755 			int index = attr->ah_attr.grh.sgid_index;
1756 
1757 			status = ib_get_cached_gid(ibqp->device, port_num,
1758 						   index, &gid, &gid_attr);
1759 			if (!status && !memcmp(&gid, &zgid, sizeof(gid)))
1760 				status = -ENOENT;
1761 			if (!status && gid_attr.ndev) {
1762 				vlan = rdma_vlan_dev_vlan_id(gid_attr.ndev);
1763 				memcpy(smac, gid_attr.ndev->dev_addr, ETH_ALEN);
1764 				dev_put(gid_attr.ndev);
1765 			}
1766 		}
1767 		if (status)
1768 			goto out;
1769 
1770 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
1771 				  port_num, vlan, smac))
1772 			goto out;
1773 
1774 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1775 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
1776 
1777 		if (is_eth &&
1778 		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
1779 			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr.gid_type);
1780 
1781 			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
1782 				err = -EINVAL;
1783 				goto out;
1784 			}
1785 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
1786 		}
1787 
1788 	}
1789 
1790 	if (attr_mask & IB_QP_TIMEOUT) {
1791 		context->pri_path.ackto |= attr->timeout << 3;
1792 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1793 	}
1794 
1795 	if (attr_mask & IB_QP_ALT_PATH) {
1796 		if (attr->alt_port_num == 0 ||
1797 		    attr->alt_port_num > dev->dev->caps.num_ports)
1798 			goto out;
1799 
1800 		if (attr->alt_pkey_index >=
1801 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
1802 			goto out;
1803 
1804 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
1805 				      &context->alt_path,
1806 				      attr->alt_port_num))
1807 			goto out;
1808 
1809 		context->alt_path.pkey_index = attr->alt_pkey_index;
1810 		context->alt_path.ackto = attr->alt_timeout << 3;
1811 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1812 	}
1813 
1814 	pd = get_pd(qp);
1815 	get_cqs(qp, &send_cq, &recv_cq);
1816 	context->pd       = cpu_to_be32(pd->pdn);
1817 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1818 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1819 	context->params1  = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1820 
1821 	/* Set "fast registration enabled" for all kernel QPs */
1822 	if (!qp->ibqp.uobject)
1823 		context->params1 |= cpu_to_be32(1 << 11);
1824 
1825 	if (attr_mask & IB_QP_RNR_RETRY) {
1826 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1827 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1828 	}
1829 
1830 	if (attr_mask & IB_QP_RETRY_CNT) {
1831 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1832 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1833 	}
1834 
1835 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1836 		if (attr->max_rd_atomic)
1837 			context->params1 |=
1838 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1839 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1840 	}
1841 
1842 	if (attr_mask & IB_QP_SQ_PSN)
1843 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1844 
1845 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1846 		if (attr->max_dest_rd_atomic)
1847 			context->params2 |=
1848 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1849 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1850 	}
1851 
1852 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1853 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1854 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1855 	}
1856 
1857 	if (ibqp->srq)
1858 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1859 
1860 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1861 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1862 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1863 	}
1864 	if (attr_mask & IB_QP_RQ_PSN)
1865 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1866 
1867 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1868 	if (attr_mask & IB_QP_QKEY) {
1869 		if (qp->mlx4_ib_qp_type &
1870 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1871 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1872 		else {
1873 			if (mlx4_is_mfunc(dev->dev) &&
1874 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1875 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1876 			    MLX4_RESERVED_QKEY_BASE) {
1877 				pr_err("Cannot use reserved QKEY"
1878 				       " 0x%x (range 0xffff0000..0xffffffff"
1879 				       " is reserved)\n", attr->qkey);
1880 				err = -EINVAL;
1881 				goto out;
1882 			}
1883 			context->qkey = cpu_to_be32(attr->qkey);
1884 		}
1885 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
1886 	}
1887 
1888 	if (ibqp->srq)
1889 		context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1890 
1891 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1892 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1893 
1894 	if (cur_state == IB_QPS_INIT &&
1895 	    new_state == IB_QPS_RTR  &&
1896 	    (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1897 	     ibqp->qp_type == IB_QPT_UD ||
1898 	     ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1899 		context->pri_path.sched_queue = (qp->port - 1) << 6;
1900 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1901 		    qp->mlx4_ib_qp_type &
1902 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1903 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1904 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1905 				context->pri_path.fl = 0x80;
1906 		} else {
1907 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1908 				context->pri_path.fl = 0x80;
1909 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1910 		}
1911 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1912 		    IB_LINK_LAYER_ETHERNET) {
1913 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
1914 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
1915 				context->pri_path.feup = 1 << 7; /* don't fsm */
1916 			/* handle smac_index */
1917 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
1918 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
1919 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
1920 				err = handle_eth_ud_smac_index(dev, qp, context);
1921 				if (err) {
1922 					err = -EINVAL;
1923 					goto out;
1924 				}
1925 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1926 					dev->qp1_proxy[qp->port - 1] = qp;
1927 			}
1928 		}
1929 	}
1930 
1931 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1932 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
1933 					MLX4_IB_LINK_TYPE_ETH;
1934 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1935 			/* set QP to receive both tunneled & non-tunneled packets */
1936 			if (!(context->flags & cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET)))
1937 				context->srqn = cpu_to_be32(7 << 28);
1938 		}
1939 	}
1940 
1941 	if (ibqp->qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
1942 		int is_eth = rdma_port_get_link_layer(
1943 				&dev->ib_dev, qp->port) ==
1944 				IB_LINK_LAYER_ETHERNET;
1945 		if (is_eth) {
1946 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
1947 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
1948 		}
1949 	}
1950 
1951 
1952 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1953 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1954 		sqd_event = 1;
1955 	else
1956 		sqd_event = 0;
1957 
1958 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1959 		context->rlkey_roce_mode |= (1 << 4);
1960 
1961 	/*
1962 	 * Before passing a kernel QP to the HW, make sure that the
1963 	 * ownership bits of the send queue are set and the SQ
1964 	 * headroom is stamped so that the hardware doesn't start
1965 	 * processing stale work requests.
1966 	 */
1967 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1968 		struct mlx4_wqe_ctrl_seg *ctrl;
1969 		int i;
1970 
1971 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1972 			ctrl = get_send_wqe(qp, i);
1973 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
1974 			if (qp->sq_max_wqes_per_wr == 1)
1975 				ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1976 
1977 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1978 		}
1979 	}
1980 
1981 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1982 			     to_mlx4_state(new_state), context, optpar,
1983 			     sqd_event, &qp->mqp);
1984 	if (err)
1985 		goto out;
1986 
1987 	qp->state = new_state;
1988 
1989 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1990 		qp->atomic_rd_en = attr->qp_access_flags;
1991 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1992 		qp->resp_depth = attr->max_dest_rd_atomic;
1993 	if (attr_mask & IB_QP_PORT) {
1994 		qp->port = attr->port_num;
1995 		update_mcg_macs(dev, qp);
1996 	}
1997 	if (attr_mask & IB_QP_ALT_PATH)
1998 		qp->alt_port = attr->alt_port_num;
1999 
2000 	if (is_sqp(dev, qp))
2001 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2002 
2003 	/*
2004 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
2005 	 * QP0 to RESET or ERROR, bring the link back down.
2006 	 */
2007 	if (is_qp0(dev, qp)) {
2008 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2009 			if (mlx4_INIT_PORT(dev->dev, qp->port))
2010 				pr_warn("INIT_PORT failed for port %d\n",
2011 				       qp->port);
2012 
2013 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2014 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2015 			mlx4_CLOSE_PORT(dev->dev, qp->port);
2016 	}
2017 
2018 	/*
2019 	 * If we moved a kernel QP to RESET, clean up all old CQ
2020 	 * entries and reinitialize the QP.
2021 	 */
2022 	if (new_state == IB_QPS_RESET) {
2023 		if (!ibqp->uobject) {
2024 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2025 					 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
2026 			if (send_cq != recv_cq)
2027 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2028 
2029 			qp->rq.head = 0;
2030 			qp->rq.tail = 0;
2031 			qp->sq.head = 0;
2032 			qp->sq.tail = 0;
2033 			qp->sq_next_wqe = 0;
2034 			if (qp->rq.wqe_cnt)
2035 				*qp->db.db  = 0;
2036 
2037 			if (qp->flags & MLX4_IB_QP_NETIF)
2038 				mlx4_ib_steer_qp_reg(dev, qp, 0);
2039 		}
2040 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2041 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2042 			qp->pri.smac = 0;
2043 			qp->pri.smac_port = 0;
2044 		}
2045 		if (qp->alt.smac) {
2046 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2047 			qp->alt.smac = 0;
2048 		}
2049 		if (qp->pri.vid < 0x1000) {
2050 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2051 			qp->pri.vid = 0xFFFF;
2052 			qp->pri.candidate_vid = 0xFFFF;
2053 			qp->pri.update_vid = 0;
2054 		}
2055 
2056 		if (qp->alt.vid < 0x1000) {
2057 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2058 			qp->alt.vid = 0xFFFF;
2059 			qp->alt.candidate_vid = 0xFFFF;
2060 			qp->alt.update_vid = 0;
2061 		}
2062 	}
2063 out:
2064 	if (err && qp->counter_index)
2065 		mlx4_ib_free_qp_counter(dev, qp);
2066 	if (err && steer_qp)
2067 		mlx4_ib_steer_qp_reg(dev, qp, 0);
2068 	kfree(context);
2069 	if (qp->pri.candidate_smac ||
2070 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2071 		if (err) {
2072 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2073 		} else {
2074 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2075 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2076 			qp->pri.smac = qp->pri.candidate_smac;
2077 			qp->pri.smac_index = qp->pri.candidate_smac_index;
2078 			qp->pri.smac_port = qp->pri.candidate_smac_port;
2079 		}
2080 		qp->pri.candidate_smac = 0;
2081 		qp->pri.candidate_smac_index = 0;
2082 		qp->pri.candidate_smac_port = 0;
2083 	}
2084 	if (qp->alt.candidate_smac) {
2085 		if (err) {
2086 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2087 		} else {
2088 			if (qp->alt.smac)
2089 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2090 			qp->alt.smac = qp->alt.candidate_smac;
2091 			qp->alt.smac_index = qp->alt.candidate_smac_index;
2092 			qp->alt.smac_port = qp->alt.candidate_smac_port;
2093 		}
2094 		qp->alt.candidate_smac = 0;
2095 		qp->alt.candidate_smac_index = 0;
2096 		qp->alt.candidate_smac_port = 0;
2097 	}
2098 
2099 	if (qp->pri.update_vid) {
2100 		if (err) {
2101 			if (qp->pri.candidate_vid < 0x1000)
2102 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2103 						     qp->pri.candidate_vid);
2104 		} else {
2105 			if (qp->pri.vid < 0x1000)
2106 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2107 						     qp->pri.vid);
2108 			qp->pri.vid = qp->pri.candidate_vid;
2109 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2110 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
2111 		}
2112 		qp->pri.candidate_vid = 0xFFFF;
2113 		qp->pri.update_vid = 0;
2114 	}
2115 
2116 	if (qp->alt.update_vid) {
2117 		if (err) {
2118 			if (qp->alt.candidate_vid < 0x1000)
2119 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2120 						     qp->alt.candidate_vid);
2121 		} else {
2122 			if (qp->alt.vid < 0x1000)
2123 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2124 						     qp->alt.vid);
2125 			qp->alt.vid = qp->alt.candidate_vid;
2126 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2127 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2128 		}
2129 		qp->alt.candidate_vid = 0xFFFF;
2130 		qp->alt.update_vid = 0;
2131 	}
2132 
2133 	return err;
2134 }
2135 
2136 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2137 			      int attr_mask, struct ib_udata *udata)
2138 {
2139 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2140 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2141 	enum ib_qp_state cur_state, new_state;
2142 	int err = -EINVAL;
2143 	int ll;
2144 	mutex_lock(&qp->mutex);
2145 
2146 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2147 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2148 
2149 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2150 		ll = IB_LINK_LAYER_UNSPECIFIED;
2151 	} else {
2152 		int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2153 		ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2154 	}
2155 
2156 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2157 				attr_mask, ll)) {
2158 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2159 			 "for transition %d to %d. qp_type %d,"
2160 			 " attr_mask 0x%x\n",
2161 			 ibqp->qp_num, cur_state, new_state,
2162 			 ibqp->qp_type, attr_mask);
2163 		goto out;
2164 	}
2165 
2166 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2167 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2168 			if ((ibqp->qp_type == IB_QPT_RC) ||
2169 			    (ibqp->qp_type == IB_QPT_UD) ||
2170 			    (ibqp->qp_type == IB_QPT_UC) ||
2171 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2172 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2173 				attr->port_num = mlx4_ib_bond_next_port(dev);
2174 			}
2175 		} else {
2176 			/* no sense in changing port_num
2177 			 * when ports are bonded */
2178 			attr_mask &= ~IB_QP_PORT;
2179 		}
2180 	}
2181 
2182 	if ((attr_mask & IB_QP_PORT) &&
2183 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2184 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2185 			 "for transition %d to %d. qp_type %d\n",
2186 			 ibqp->qp_num, attr->port_num, cur_state,
2187 			 new_state, ibqp->qp_type);
2188 		goto out;
2189 	}
2190 
2191 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2192 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2193 	     IB_LINK_LAYER_ETHERNET))
2194 		goto out;
2195 
2196 	if (attr_mask & IB_QP_PKEY_INDEX) {
2197 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2198 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2199 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2200 				 "for transition %d to %d. qp_type %d\n",
2201 				 ibqp->qp_num, attr->pkey_index, cur_state,
2202 				 new_state, ibqp->qp_type);
2203 			goto out;
2204 		}
2205 	}
2206 
2207 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2208 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2209 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2210 			 "Transition %d to %d. qp_type %d\n",
2211 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2212 			 new_state, ibqp->qp_type);
2213 		goto out;
2214 	}
2215 
2216 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2217 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2218 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2219 			 "Transition %d to %d. qp_type %d\n",
2220 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2221 			 new_state, ibqp->qp_type);
2222 		goto out;
2223 	}
2224 
2225 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2226 		err = 0;
2227 		goto out;
2228 	}
2229 
2230 	err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2231 
2232 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2233 		attr->port_num = 1;
2234 
2235 out:
2236 	mutex_unlock(&qp->mutex);
2237 	return err;
2238 }
2239 
2240 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2241 		      int attr_mask, struct ib_udata *udata)
2242 {
2243 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2244 	int ret;
2245 
2246 	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2247 
2248 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2249 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
2250 		int err = 0;
2251 
2252 		if (sqp->roce_v2_gsi)
2253 			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2254 		if (err)
2255 			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2256 			       err);
2257 	}
2258 	return ret;
2259 }
2260 
2261 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2262 {
2263 	int i;
2264 	for (i = 0; i < dev->caps.num_ports; i++) {
2265 		if (qpn == dev->caps.qp0_proxy[i] ||
2266 		    qpn == dev->caps.qp0_tunnel[i]) {
2267 			*qkey = dev->caps.qp0_qkey[i];
2268 			return 0;
2269 		}
2270 	}
2271 	return -EINVAL;
2272 }
2273 
2274 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2275 				  struct ib_ud_wr *wr,
2276 				  void *wqe, unsigned *mlx_seg_len)
2277 {
2278 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2279 	struct ib_device *ib_dev = &mdev->ib_dev;
2280 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2281 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2282 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2283 	u16 pkey;
2284 	u32 qkey;
2285 	int send_size;
2286 	int header_size;
2287 	int spc;
2288 	int i;
2289 
2290 	if (wr->wr.opcode != IB_WR_SEND)
2291 		return -EINVAL;
2292 
2293 	send_size = 0;
2294 
2295 	for (i = 0; i < wr->wr.num_sge; ++i)
2296 		send_size += wr->wr.sg_list[i].length;
2297 
2298 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2299 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2300 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2301 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2302 
2303 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2304 
2305 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2306 		sqp->ud_header.lrh.service_level =
2307 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2308 		sqp->ud_header.lrh.destination_lid =
2309 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2310 		sqp->ud_header.lrh.source_lid =
2311 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2312 	}
2313 
2314 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2315 
2316 	/* force loopback */
2317 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2318 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2319 
2320 	sqp->ud_header.lrh.virtual_lane    = 0;
2321 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2322 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2323 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2324 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2325 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2326 	else
2327 		sqp->ud_header.bth.destination_qpn =
2328 			cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
2329 
2330 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2331 	if (mlx4_is_master(mdev->dev)) {
2332 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2333 			return -EINVAL;
2334 	} else {
2335 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2336 			return -EINVAL;
2337 	}
2338 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2339 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2340 
2341 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2342 	sqp->ud_header.immediate_present = 0;
2343 
2344 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2345 
2346 	/*
2347 	 * Inline data segments may not cross a 64 byte boundary.  If
2348 	 * our UD header is bigger than the space available up to the
2349 	 * next 64 byte boundary in the WQE, use two inline data
2350 	 * segments to hold the UD header.
2351 	 */
2352 	spc = MLX4_INLINE_ALIGN -
2353 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2354 	if (header_size <= spc) {
2355 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2356 		memcpy(inl + 1, sqp->header_buf, header_size);
2357 		i = 1;
2358 	} else {
2359 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2360 		memcpy(inl + 1, sqp->header_buf, spc);
2361 
2362 		inl = (void *) (inl + 1) + spc;
2363 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2364 		/*
2365 		 * Need a barrier here to make sure all the data is
2366 		 * visible before the byte_count field is set.
2367 		 * Otherwise the HCA prefetcher could grab the 64-byte
2368 		 * chunk with this inline segment and get a valid (!=
2369 		 * 0xffffffff) byte count but stale data, and end up
2370 		 * generating a packet with bad headers.
2371 		 *
2372 		 * The first inline segment's byte_count field doesn't
2373 		 * need a barrier, because it comes after a
2374 		 * control/MLX segment and therefore is at an offset
2375 		 * of 16 mod 64.
2376 		 */
2377 		wmb();
2378 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2379 		i = 2;
2380 	}
2381 
2382 	*mlx_seg_len =
2383 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2384 	return 0;
2385 }
2386 
2387 #define MLX4_ROCEV2_QP1_SPORT 0xC000
2388 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_ud_wr *wr,
2389 			    void *wqe, unsigned *mlx_seg_len)
2390 {
2391 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
2392 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2393 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2394 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2395 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2396 	union ib_gid sgid;
2397 	u16 pkey;
2398 	int send_size;
2399 	int header_size;
2400 	int spc;
2401 	int i;
2402 	int err = 0;
2403 	u16 vlan = 0xffff;
2404 	bool is_eth;
2405 	bool is_vlan = false;
2406 	bool is_grh;
2407 	bool is_udp = false;
2408 	int ip_version = 0;
2409 
2410 	send_size = 0;
2411 	for (i = 0; i < wr->wr.num_sge; ++i)
2412 		send_size += wr->wr.sg_list[i].length;
2413 
2414 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2415 	is_grh = mlx4_ib_ah_grh_present(ah);
2416 	if (is_eth) {
2417 		struct ib_gid_attr gid_attr;
2418 
2419 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2420 			/* When multi-function is enabled, the ib_core gid
2421 			 * indexes don't necessarily match the hw ones, so
2422 			 * we must use our own cache */
2423 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2424 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
2425 							   ah->av.ib.gid_index, &sgid.raw[0]);
2426 			if (err)
2427 				return err;
2428 		} else  {
2429 			err = ib_get_cached_gid(ib_dev,
2430 						be32_to_cpu(ah->av.ib.port_pd) >> 24,
2431 						ah->av.ib.gid_index, &sgid,
2432 						&gid_attr);
2433 			if (!err) {
2434 				if (gid_attr.ndev)
2435 					dev_put(gid_attr.ndev);
2436 				if (!memcmp(&sgid, &zgid, sizeof(sgid)))
2437 					err = -ENOENT;
2438 			}
2439 			if (!err) {
2440 				is_udp = gid_attr.gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
2441 				if (is_udp) {
2442 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
2443 						ip_version = 4;
2444 					else
2445 						ip_version = 6;
2446 					is_grh = false;
2447 				}
2448 			} else {
2449 				return err;
2450 			}
2451 		}
2452 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
2453 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
2454 			is_vlan = 1;
2455 		}
2456 	}
2457 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
2458 			  ip_version, is_udp, 0, &sqp->ud_header);
2459 	if (err)
2460 		return err;
2461 
2462 	if (!is_eth) {
2463 		sqp->ud_header.lrh.service_level =
2464 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2465 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
2466 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2467 	}
2468 
2469 	if (is_grh || (ip_version == 6)) {
2470 		sqp->ud_header.grh.traffic_class =
2471 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2472 		sqp->ud_header.grh.flow_label    =
2473 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
2474 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
2475 		if (is_eth)
2476 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
2477 		else {
2478 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2479 			/* When multi-function is enabled, the ib_core gid
2480 			 * indexes don't necessarily match the hw ones, so
2481 			 * we must use our own cache */
2482 			sqp->ud_header.grh.source_gid.global.subnet_prefix =
2483 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2484 						       subnet_prefix;
2485 			sqp->ud_header.grh.source_gid.global.interface_id =
2486 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2487 					       guid_cache[ah->av.ib.gid_index];
2488 		} else
2489 			ib_get_cached_gid(ib_dev,
2490 					  be32_to_cpu(ah->av.ib.port_pd) >> 24,
2491 					  ah->av.ib.gid_index,
2492 					  &sqp->ud_header.grh.source_gid, NULL);
2493 		}
2494 		memcpy(sqp->ud_header.grh.destination_gid.raw,
2495 		       ah->av.ib.dgid, 16);
2496 	}
2497 
2498 	if (ip_version == 4) {
2499 		sqp->ud_header.ip4.tos =
2500 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2501 		sqp->ud_header.ip4.id = 0;
2502 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
2503 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
2504 
2505 		memcpy(&sqp->ud_header.ip4.saddr,
2506 		       sgid.raw + 12, 4);
2507 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
2508 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
2509 	}
2510 
2511 	if (is_udp) {
2512 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
2513 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
2514 		sqp->ud_header.udp.csum = 0;
2515 	}
2516 
2517 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2518 
2519 	if (!is_eth) {
2520 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
2521 					  (sqp->ud_header.lrh.destination_lid ==
2522 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
2523 					  (sqp->ud_header.lrh.service_level << 8));
2524 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
2525 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
2526 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
2527 	}
2528 
2529 	switch (wr->wr.opcode) {
2530 	case IB_WR_SEND:
2531 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
2532 		sqp->ud_header.immediate_present = 0;
2533 		break;
2534 	case IB_WR_SEND_WITH_IMM:
2535 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2536 		sqp->ud_header.immediate_present = 1;
2537 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
2538 		break;
2539 	default:
2540 		return -EINVAL;
2541 	}
2542 
2543 	if (is_eth) {
2544 		struct in6_addr in6;
2545 		u16 ether_type;
2546 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
2547 
2548 		ether_type = (!is_udp) ? MLX4_IB_IBOE_ETHERTYPE :
2549 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
2550 
2551 		mlx->sched_prio = cpu_to_be16(pcp);
2552 
2553 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
2554 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
2555 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
2556 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
2557 		memcpy(&in6, sgid.raw, sizeof(in6));
2558 
2559 
2560 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
2561 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
2562 		if (!is_vlan) {
2563 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
2564 		} else {
2565 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
2566 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
2567 		}
2568 	} else {
2569 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
2570 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
2571 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
2572 	}
2573 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2574 	if (!sqp->qp.ibqp.qp_num)
2575 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
2576 	else
2577 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
2578 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2579 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2580 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2581 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
2582 					       sqp->qkey : wr->remote_qkey);
2583 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
2584 
2585 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2586 
2587 	if (0) {
2588 		pr_err("built UD header of size %d:\n", header_size);
2589 		for (i = 0; i < header_size / 4; ++i) {
2590 			if (i % 8 == 0)
2591 				pr_err("  [%02x] ", i * 4);
2592 			pr_cont(" %08x",
2593 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
2594 			if ((i + 1) % 8 == 0)
2595 				pr_cont("\n");
2596 		}
2597 		pr_err("\n");
2598 	}
2599 
2600 	/*
2601 	 * Inline data segments may not cross a 64 byte boundary.  If
2602 	 * our UD header is bigger than the space available up to the
2603 	 * next 64 byte boundary in the WQE, use two inline data
2604 	 * segments to hold the UD header.
2605 	 */
2606 	spc = MLX4_INLINE_ALIGN -
2607 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2608 	if (header_size <= spc) {
2609 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2610 		memcpy(inl + 1, sqp->header_buf, header_size);
2611 		i = 1;
2612 	} else {
2613 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2614 		memcpy(inl + 1, sqp->header_buf, spc);
2615 
2616 		inl = (void *) (inl + 1) + spc;
2617 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2618 		/*
2619 		 * Need a barrier here to make sure all the data is
2620 		 * visible before the byte_count field is set.
2621 		 * Otherwise the HCA prefetcher could grab the 64-byte
2622 		 * chunk with this inline segment and get a valid (!=
2623 		 * 0xffffffff) byte count but stale data, and end up
2624 		 * generating a packet with bad headers.
2625 		 *
2626 		 * The first inline segment's byte_count field doesn't
2627 		 * need a barrier, because it comes after a
2628 		 * control/MLX segment and therefore is at an offset
2629 		 * of 16 mod 64.
2630 		 */
2631 		wmb();
2632 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2633 		i = 2;
2634 	}
2635 
2636 	*mlx_seg_len =
2637 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2638 	return 0;
2639 }
2640 
2641 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2642 {
2643 	unsigned cur;
2644 	struct mlx4_ib_cq *cq;
2645 
2646 	cur = wq->head - wq->tail;
2647 	if (likely(cur + nreq < wq->max_post))
2648 		return 0;
2649 
2650 	cq = to_mcq(ib_cq);
2651 	spin_lock(&cq->lock);
2652 	cur = wq->head - wq->tail;
2653 	spin_unlock(&cq->lock);
2654 
2655 	return cur + nreq >= wq->max_post;
2656 }
2657 
2658 static __be32 convert_access(int acc)
2659 {
2660 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
2661 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
2662 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
2663 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
2664 	       (acc & IB_ACCESS_REMOTE_READ   ?
2665 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
2666 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
2667 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
2668 }
2669 
2670 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
2671 			struct ib_reg_wr *wr)
2672 {
2673 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
2674 
2675 	fseg->flags		= convert_access(wr->access);
2676 	fseg->mem_key		= cpu_to_be32(wr->key);
2677 	fseg->buf_list		= cpu_to_be64(mr->page_map);
2678 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
2679 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
2680 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
2681 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
2682 	fseg->reserved[0]	= 0;
2683 	fseg->reserved[1]	= 0;
2684 }
2685 
2686 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2687 {
2688 	memset(iseg, 0, sizeof(*iseg));
2689 	iseg->mem_key = cpu_to_be32(rkey);
2690 }
2691 
2692 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2693 					  u64 remote_addr, u32 rkey)
2694 {
2695 	rseg->raddr    = cpu_to_be64(remote_addr);
2696 	rseg->rkey     = cpu_to_be32(rkey);
2697 	rseg->reserved = 0;
2698 }
2699 
2700 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
2701 		struct ib_atomic_wr *wr)
2702 {
2703 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2704 		aseg->swap_add = cpu_to_be64(wr->swap);
2705 		aseg->compare  = cpu_to_be64(wr->compare_add);
2706 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2707 		aseg->swap_add = cpu_to_be64(wr->compare_add);
2708 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
2709 	} else {
2710 		aseg->swap_add = cpu_to_be64(wr->compare_add);
2711 		aseg->compare  = 0;
2712 	}
2713 
2714 }
2715 
2716 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2717 				  struct ib_atomic_wr *wr)
2718 {
2719 	aseg->swap_add		= cpu_to_be64(wr->swap);
2720 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
2721 	aseg->compare		= cpu_to_be64(wr->compare_add);
2722 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
2723 }
2724 
2725 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2726 			     struct ib_ud_wr *wr)
2727 {
2728 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
2729 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
2730 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
2731 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
2732 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
2733 }
2734 
2735 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2736 				    struct mlx4_wqe_datagram_seg *dseg,
2737 				    struct ib_ud_wr *wr,
2738 				    enum mlx4_ib_qp_type qpt)
2739 {
2740 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
2741 	struct mlx4_av sqp_av = {0};
2742 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
2743 
2744 	/* force loopback */
2745 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2746 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2747 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2748 			cpu_to_be32(0xf0000000);
2749 
2750 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2751 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
2752 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2753 	else
2754 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp0_tunnel[port - 1]);
2755 	/* Use QKEY from the QP context, which is set by master */
2756 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2757 }
2758 
2759 static void build_tunnel_header(struct ib_ud_wr *wr, void *wqe, unsigned *mlx_seg_len)
2760 {
2761 	struct mlx4_wqe_inline_seg *inl = wqe;
2762 	struct mlx4_ib_tunnel_header hdr;
2763 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2764 	int spc;
2765 	int i;
2766 
2767 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2768 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
2769 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
2770 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
2771 	memcpy(hdr.mac, ah->av.eth.mac, 6);
2772 	hdr.vlan = ah->av.eth.vlan;
2773 
2774 	spc = MLX4_INLINE_ALIGN -
2775 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2776 	if (sizeof (hdr) <= spc) {
2777 		memcpy(inl + 1, &hdr, sizeof (hdr));
2778 		wmb();
2779 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
2780 		i = 1;
2781 	} else {
2782 		memcpy(inl + 1, &hdr, spc);
2783 		wmb();
2784 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2785 
2786 		inl = (void *) (inl + 1) + spc;
2787 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2788 		wmb();
2789 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
2790 		i = 2;
2791 	}
2792 
2793 	*mlx_seg_len =
2794 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2795 }
2796 
2797 static void set_mlx_icrc_seg(void *dseg)
2798 {
2799 	u32 *t = dseg;
2800 	struct mlx4_wqe_inline_seg *iseg = dseg;
2801 
2802 	t[1] = 0;
2803 
2804 	/*
2805 	 * Need a barrier here before writing the byte_count field to
2806 	 * make sure that all the data is visible before the
2807 	 * byte_count field is set.  Otherwise, if the segment begins
2808 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2809 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2810 	 * stale data, and end up sending the wrong data.
2811 	 */
2812 	wmb();
2813 
2814 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2815 }
2816 
2817 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2818 {
2819 	dseg->lkey       = cpu_to_be32(sg->lkey);
2820 	dseg->addr       = cpu_to_be64(sg->addr);
2821 
2822 	/*
2823 	 * Need a barrier here before writing the byte_count field to
2824 	 * make sure that all the data is visible before the
2825 	 * byte_count field is set.  Otherwise, if the segment begins
2826 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2827 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2828 	 * stale data, and end up sending the wrong data.
2829 	 */
2830 	wmb();
2831 
2832 	dseg->byte_count = cpu_to_be32(sg->length);
2833 }
2834 
2835 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2836 {
2837 	dseg->byte_count = cpu_to_be32(sg->length);
2838 	dseg->lkey       = cpu_to_be32(sg->lkey);
2839 	dseg->addr       = cpu_to_be64(sg->addr);
2840 }
2841 
2842 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_ud_wr *wr,
2843 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2844 			 __be32 *lso_hdr_sz, __be32 *blh)
2845 {
2846 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
2847 
2848 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2849 		*blh = cpu_to_be32(1 << 6);
2850 
2851 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2852 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
2853 		return -EINVAL;
2854 
2855 	memcpy(wqe->header, wr->header, wr->hlen);
2856 
2857 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
2858 	*lso_seg_len = halign;
2859 	return 0;
2860 }
2861 
2862 static __be32 send_ieth(struct ib_send_wr *wr)
2863 {
2864 	switch (wr->opcode) {
2865 	case IB_WR_SEND_WITH_IMM:
2866 	case IB_WR_RDMA_WRITE_WITH_IMM:
2867 		return wr->ex.imm_data;
2868 
2869 	case IB_WR_SEND_WITH_INV:
2870 		return cpu_to_be32(wr->ex.invalidate_rkey);
2871 
2872 	default:
2873 		return 0;
2874 	}
2875 }
2876 
2877 static void add_zero_len_inline(void *wqe)
2878 {
2879 	struct mlx4_wqe_inline_seg *inl = wqe;
2880 	memset(wqe, 0, 16);
2881 	inl->byte_count = cpu_to_be32(1 << 31);
2882 }
2883 
2884 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2885 		      struct ib_send_wr **bad_wr)
2886 {
2887 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2888 	void *wqe;
2889 	struct mlx4_wqe_ctrl_seg *ctrl;
2890 	struct mlx4_wqe_data_seg *dseg;
2891 	unsigned long flags;
2892 	int nreq;
2893 	int err = 0;
2894 	unsigned ind;
2895 	int uninitialized_var(stamp);
2896 	int uninitialized_var(size);
2897 	unsigned uninitialized_var(seglen);
2898 	__be32 dummy;
2899 	__be32 *lso_wqe;
2900 	__be32 uninitialized_var(lso_hdr_sz);
2901 	__be32 blh;
2902 	int i;
2903 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
2904 
2905 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2906 		struct mlx4_ib_sqp *sqp = to_msqp(qp);
2907 
2908 		if (sqp->roce_v2_gsi) {
2909 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
2910 			struct ib_gid_attr gid_attr;
2911 			union ib_gid gid;
2912 
2913 			if (!ib_get_cached_gid(ibqp->device,
2914 					       be32_to_cpu(ah->av.ib.port_pd) >> 24,
2915 					       ah->av.ib.gid_index, &gid,
2916 					       &gid_attr)) {
2917 				if (gid_attr.ndev)
2918 					dev_put(gid_attr.ndev);
2919 				qp = (gid_attr.gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
2920 					to_mqp(sqp->roce_v2_gsi) : qp;
2921 			} else {
2922 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
2923 				       ah->av.ib.gid_index);
2924 			}
2925 		}
2926 	}
2927 
2928 	spin_lock_irqsave(&qp->sq.lock, flags);
2929 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
2930 		err = -EIO;
2931 		*bad_wr = wr;
2932 		nreq = 0;
2933 		goto out;
2934 	}
2935 
2936 	ind = qp->sq_next_wqe;
2937 
2938 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2939 		lso_wqe = &dummy;
2940 		blh = 0;
2941 
2942 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2943 			err = -ENOMEM;
2944 			*bad_wr = wr;
2945 			goto out;
2946 		}
2947 
2948 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
2949 			err = -EINVAL;
2950 			*bad_wr = wr;
2951 			goto out;
2952 		}
2953 
2954 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
2955 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
2956 
2957 		ctrl->srcrb_flags =
2958 			(wr->send_flags & IB_SEND_SIGNALED ?
2959 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
2960 			(wr->send_flags & IB_SEND_SOLICITED ?
2961 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
2962 			((wr->send_flags & IB_SEND_IP_CSUM) ?
2963 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
2964 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
2965 			qp->sq_signal_bits;
2966 
2967 		ctrl->imm = send_ieth(wr);
2968 
2969 		wqe += sizeof *ctrl;
2970 		size = sizeof *ctrl / 16;
2971 
2972 		switch (qp->mlx4_ib_qp_type) {
2973 		case MLX4_IB_QPT_RC:
2974 		case MLX4_IB_QPT_UC:
2975 			switch (wr->opcode) {
2976 			case IB_WR_ATOMIC_CMP_AND_SWP:
2977 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2978 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
2979 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
2980 					      atomic_wr(wr)->rkey);
2981 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2982 
2983 				set_atomic_seg(wqe, atomic_wr(wr));
2984 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
2985 
2986 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2987 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
2988 
2989 				break;
2990 
2991 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2992 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
2993 					      atomic_wr(wr)->rkey);
2994 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2995 
2996 				set_masked_atomic_seg(wqe, atomic_wr(wr));
2997 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
2998 
2999 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3000 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3001 
3002 				break;
3003 
3004 			case IB_WR_RDMA_READ:
3005 			case IB_WR_RDMA_WRITE:
3006 			case IB_WR_RDMA_WRITE_WITH_IMM:
3007 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3008 					      rdma_wr(wr)->rkey);
3009 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3010 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3011 				break;
3012 
3013 			case IB_WR_LOCAL_INV:
3014 				ctrl->srcrb_flags |=
3015 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3016 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3017 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3018 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3019 				break;
3020 
3021 			case IB_WR_REG_MR:
3022 				ctrl->srcrb_flags |=
3023 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3024 				set_reg_seg(wqe, reg_wr(wr));
3025 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3026 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3027 				break;
3028 
3029 			default:
3030 				/* No extra segments required for sends */
3031 				break;
3032 			}
3033 			break;
3034 
3035 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3036 			err =  build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3037 					ctrl, &seglen);
3038 			if (unlikely(err)) {
3039 				*bad_wr = wr;
3040 				goto out;
3041 			}
3042 			wqe  += seglen;
3043 			size += seglen / 16;
3044 			break;
3045 		case MLX4_IB_QPT_TUN_SMI:
3046 		case MLX4_IB_QPT_TUN_GSI:
3047 			/* this is a UD qp used in MAD responses to slaves. */
3048 			set_datagram_seg(wqe, ud_wr(wr));
3049 			/* set the forced-loopback bit in the data seg av */
3050 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3051 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3052 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3053 			break;
3054 		case MLX4_IB_QPT_UD:
3055 			set_datagram_seg(wqe, ud_wr(wr));
3056 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3057 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3058 
3059 			if (wr->opcode == IB_WR_LSO) {
3060 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3061 						&lso_hdr_sz, &blh);
3062 				if (unlikely(err)) {
3063 					*bad_wr = wr;
3064 					goto out;
3065 				}
3066 				lso_wqe = (__be32 *) wqe;
3067 				wqe  += seglen;
3068 				size += seglen / 16;
3069 			}
3070 			break;
3071 
3072 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3073 			err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3074 					ctrl, &seglen);
3075 			if (unlikely(err)) {
3076 				*bad_wr = wr;
3077 				goto out;
3078 			}
3079 			wqe  += seglen;
3080 			size += seglen / 16;
3081 			/* to start tunnel header on a cache-line boundary */
3082 			add_zero_len_inline(wqe);
3083 			wqe += 16;
3084 			size++;
3085 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3086 			wqe  += seglen;
3087 			size += seglen / 16;
3088 			break;
3089 		case MLX4_IB_QPT_PROXY_SMI:
3090 		case MLX4_IB_QPT_PROXY_GSI:
3091 			/* If we are tunneling special qps, this is a UD qp.
3092 			 * In this case we first add a UD segment targeting
3093 			 * the tunnel qp, and then add a header with address
3094 			 * information */
3095 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3096 						ud_wr(wr),
3097 						qp->mlx4_ib_qp_type);
3098 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3099 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3100 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3101 			wqe  += seglen;
3102 			size += seglen / 16;
3103 			break;
3104 
3105 		case MLX4_IB_QPT_SMI:
3106 		case MLX4_IB_QPT_GSI:
3107 			err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
3108 					&seglen);
3109 			if (unlikely(err)) {
3110 				*bad_wr = wr;
3111 				goto out;
3112 			}
3113 			wqe  += seglen;
3114 			size += seglen / 16;
3115 			break;
3116 
3117 		default:
3118 			break;
3119 		}
3120 
3121 		/*
3122 		 * Write data segments in reverse order, so as to
3123 		 * overwrite cacheline stamp last within each
3124 		 * cacheline.  This avoids issues with WQE
3125 		 * prefetching.
3126 		 */
3127 
3128 		dseg = wqe;
3129 		dseg += wr->num_sge - 1;
3130 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3131 
3132 		/* Add one more inline data segment for ICRC for MLX sends */
3133 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3134 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3135 			     qp->mlx4_ib_qp_type &
3136 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3137 			set_mlx_icrc_seg(dseg + 1);
3138 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3139 		}
3140 
3141 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3142 			set_data_seg(dseg, wr->sg_list + i);
3143 
3144 		/*
3145 		 * Possibly overwrite stamping in cacheline with LSO
3146 		 * segment only after making sure all data segments
3147 		 * are written.
3148 		 */
3149 		wmb();
3150 		*lso_wqe = lso_hdr_sz;
3151 
3152 		ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
3153 				    MLX4_WQE_CTRL_FENCE : 0) | size;
3154 
3155 		/*
3156 		 * Make sure descriptor is fully written before
3157 		 * setting ownership bit (because HW can start
3158 		 * executing as soon as we do).
3159 		 */
3160 		wmb();
3161 
3162 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3163 			*bad_wr = wr;
3164 			err = -EINVAL;
3165 			goto out;
3166 		}
3167 
3168 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3169 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3170 
3171 		stamp = ind + qp->sq_spare_wqes;
3172 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
3173 
3174 		/*
3175 		 * We can improve latency by not stamping the last
3176 		 * send queue WQE until after ringing the doorbell, so
3177 		 * only stamp here if there are still more WQEs to post.
3178 		 *
3179 		 * Same optimization applies to padding with NOP wqe
3180 		 * in case of WQE shrinking (used to prevent wrap-around
3181 		 * in the middle of WR).
3182 		 */
3183 		if (wr->next) {
3184 			stamp_send_wqe(qp, stamp, size * 16);
3185 			ind = pad_wraparound(qp, ind);
3186 		}
3187 	}
3188 
3189 out:
3190 	if (likely(nreq)) {
3191 		qp->sq.head += nreq;
3192 
3193 		/*
3194 		 * Make sure that descriptors are written before
3195 		 * doorbell record.
3196 		 */
3197 		wmb();
3198 
3199 		writel(qp->doorbell_qpn,
3200 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3201 
3202 		/*
3203 		 * Make sure doorbells don't leak out of SQ spinlock
3204 		 * and reach the HCA out of order.
3205 		 */
3206 		mmiowb();
3207 
3208 		stamp_send_wqe(qp, stamp, size * 16);
3209 
3210 		ind = pad_wraparound(qp, ind);
3211 		qp->sq_next_wqe = ind;
3212 	}
3213 
3214 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3215 
3216 	return err;
3217 }
3218 
3219 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3220 		      struct ib_recv_wr **bad_wr)
3221 {
3222 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3223 	struct mlx4_wqe_data_seg *scat;
3224 	unsigned long flags;
3225 	int err = 0;
3226 	int nreq;
3227 	int ind;
3228 	int max_gs;
3229 	int i;
3230 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3231 
3232 	max_gs = qp->rq.max_gs;
3233 	spin_lock_irqsave(&qp->rq.lock, flags);
3234 
3235 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3236 		err = -EIO;
3237 		*bad_wr = wr;
3238 		nreq = 0;
3239 		goto out;
3240 	}
3241 
3242 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3243 
3244 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3245 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3246 			err = -ENOMEM;
3247 			*bad_wr = wr;
3248 			goto out;
3249 		}
3250 
3251 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3252 			err = -EINVAL;
3253 			*bad_wr = wr;
3254 			goto out;
3255 		}
3256 
3257 		scat = get_recv_wqe(qp, ind);
3258 
3259 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3260 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3261 			ib_dma_sync_single_for_device(ibqp->device,
3262 						      qp->sqp_proxy_rcv[ind].map,
3263 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3264 						      DMA_FROM_DEVICE);
3265 			scat->byte_count =
3266 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3267 			/* use dma lkey from upper layer entry */
3268 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3269 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3270 			scat++;
3271 			max_gs--;
3272 		}
3273 
3274 		for (i = 0; i < wr->num_sge; ++i)
3275 			__set_data_seg(scat + i, wr->sg_list + i);
3276 
3277 		if (i < max_gs) {
3278 			scat[i].byte_count = 0;
3279 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3280 			scat[i].addr       = 0;
3281 		}
3282 
3283 		qp->rq.wrid[ind] = wr->wr_id;
3284 
3285 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3286 	}
3287 
3288 out:
3289 	if (likely(nreq)) {
3290 		qp->rq.head += nreq;
3291 
3292 		/*
3293 		 * Make sure that descriptors are written before
3294 		 * doorbell record.
3295 		 */
3296 		wmb();
3297 
3298 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3299 	}
3300 
3301 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3302 
3303 	return err;
3304 }
3305 
3306 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3307 {
3308 	switch (mlx4_state) {
3309 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3310 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3311 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3312 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3313 	case MLX4_QP_STATE_SQ_DRAINING:
3314 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3315 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3316 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3317 	default:		     return -1;
3318 	}
3319 }
3320 
3321 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3322 {
3323 	switch (mlx4_mig_state) {
3324 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3325 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3326 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3327 	default: return -1;
3328 	}
3329 }
3330 
3331 static int to_ib_qp_access_flags(int mlx4_flags)
3332 {
3333 	int ib_flags = 0;
3334 
3335 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3336 		ib_flags |= IB_ACCESS_REMOTE_READ;
3337 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3338 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3339 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3340 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3341 
3342 	return ib_flags;
3343 }
3344 
3345 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
3346 				struct mlx4_qp_path *path)
3347 {
3348 	struct mlx4_dev *dev = ibdev->dev;
3349 	int is_eth;
3350 
3351 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
3352 	ib_ah_attr->port_num	  = path->sched_queue & 0x40 ? 2 : 1;
3353 
3354 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
3355 		return;
3356 
3357 	is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
3358 		IB_LINK_LAYER_ETHERNET;
3359 	if (is_eth)
3360 		ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
3361 		((path->sched_queue & 4) << 1);
3362 	else
3363 		ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
3364 
3365 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
3366 	ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
3367 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
3368 	ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
3369 	if (ib_ah_attr->ah_flags) {
3370 		ib_ah_attr->grh.sgid_index = path->mgid_index;
3371 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
3372 		ib_ah_attr->grh.traffic_class =
3373 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3374 		ib_ah_attr->grh.flow_label =
3375 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3376 		memcpy(ib_ah_attr->grh.dgid.raw,
3377 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
3378 	}
3379 }
3380 
3381 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3382 		     struct ib_qp_init_attr *qp_init_attr)
3383 {
3384 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3385 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3386 	struct mlx4_qp_context context;
3387 	int mlx4_state;
3388 	int err = 0;
3389 
3390 	mutex_lock(&qp->mutex);
3391 
3392 	if (qp->state == IB_QPS_RESET) {
3393 		qp_attr->qp_state = IB_QPS_RESET;
3394 		goto done;
3395 	}
3396 
3397 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3398 	if (err) {
3399 		err = -EINVAL;
3400 		goto out;
3401 	}
3402 
3403 	mlx4_state = be32_to_cpu(context.flags) >> 28;
3404 
3405 	qp->state		     = to_ib_qp_state(mlx4_state);
3406 	qp_attr->qp_state	     = qp->state;
3407 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
3408 	qp_attr->path_mig_state	     =
3409 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3410 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
3411 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3412 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
3413 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
3414 	qp_attr->qp_access_flags     =
3415 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
3416 
3417 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3418 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3419 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3420 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3421 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
3422 	}
3423 
3424 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3425 	if (qp_attr->qp_state == IB_QPS_INIT)
3426 		qp_attr->port_num = qp->port;
3427 	else
3428 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3429 
3430 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3431 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3432 
3433 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3434 
3435 	qp_attr->max_dest_rd_atomic =
3436 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3437 	qp_attr->min_rnr_timer	    =
3438 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3439 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
3440 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
3441 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
3442 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
3443 
3444 done:
3445 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
3446 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
3447 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
3448 
3449 	if (!ibqp->uobject) {
3450 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
3451 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
3452 	} else {
3453 		qp_attr->cap.max_send_wr  = 0;
3454 		qp_attr->cap.max_send_sge = 0;
3455 	}
3456 
3457 	/*
3458 	 * We don't support inline sends for kernel QPs (yet), and we
3459 	 * don't know what userspace's value should be.
3460 	 */
3461 	qp_attr->cap.max_inline_data = 0;
3462 
3463 	qp_init_attr->cap	     = qp_attr->cap;
3464 
3465 	qp_init_attr->create_flags = 0;
3466 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3467 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3468 
3469 	if (qp->flags & MLX4_IB_QP_LSO)
3470 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
3471 
3472 	if (qp->flags & MLX4_IB_QP_NETIF)
3473 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
3474 
3475 	qp_init_attr->sq_sig_type =
3476 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
3477 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3478 
3479 out:
3480 	mutex_unlock(&qp->mutex);
3481 	return err;
3482 }
3483 
3484