xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision 2f828fb2)
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 
40 #include <rdma/ib_cache.h>
41 #include <rdma/ib_pack.h>
42 #include <rdma/ib_addr.h>
43 #include <rdma/ib_mad.h>
44 
45 #include <linux/mlx4/driver.h>
46 #include <linux/mlx4/qp.h>
47 
48 #include "mlx4_ib.h"
49 #include <rdma/mlx4-abi.h>
50 
51 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
52 			     struct mlx4_ib_cq *recv_cq);
53 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
54 			       struct mlx4_ib_cq *recv_cq);
55 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state);
56 
57 enum {
58 	MLX4_IB_ACK_REQ_FREQ	= 8,
59 };
60 
61 enum {
62 	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
63 	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
64 	MLX4_IB_LINK_TYPE_IB		= 0,
65 	MLX4_IB_LINK_TYPE_ETH		= 1
66 };
67 
68 enum {
69 	/*
70 	 * Largest possible UD header: send with GRH and immediate
71 	 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
72 	 * tag.  (LRH would only use 8 bytes, so Ethernet is the
73 	 * biggest case)
74 	 */
75 	MLX4_IB_UD_HEADER_SIZE		= 82,
76 	MLX4_IB_LSO_HEADER_SPARE	= 128,
77 };
78 
79 struct mlx4_ib_sqp {
80 	struct mlx4_ib_qp	qp;
81 	int			pkey_index;
82 	u32			qkey;
83 	u32			send_psn;
84 	struct ib_ud_header	ud_header;
85 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
86 	struct ib_qp		*roce_v2_gsi;
87 };
88 
89 enum {
90 	MLX4_IB_MIN_SQ_STRIDE	= 6,
91 	MLX4_IB_CACHE_LINE_SIZE	= 64,
92 };
93 
94 enum {
95 	MLX4_RAW_QP_MTU		= 7,
96 	MLX4_RAW_QP_MSGMAX	= 31,
97 };
98 
99 #ifndef ETH_ALEN
100 #define ETH_ALEN        6
101 #endif
102 
103 static const __be32 mlx4_ib_opcode[] = {
104 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
105 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
106 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
107 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
108 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
109 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
110 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
111 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
112 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
113 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
114 	[IB_WR_REG_MR]				= cpu_to_be32(MLX4_OPCODE_FMR),
115 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
116 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
117 };
118 
119 enum mlx4_ib_source_type {
120 	MLX4_IB_QP_SRC	= 0,
121 	MLX4_IB_RWQ_SRC	= 1,
122 };
123 
124 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
125 {
126 	return container_of(mqp, struct mlx4_ib_sqp, qp);
127 }
128 
129 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
130 {
131 	if (!mlx4_is_master(dev->dev))
132 		return 0;
133 
134 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
135 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
136 		8 * MLX4_MFUNC_MAX;
137 }
138 
139 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
140 {
141 	int proxy_sqp = 0;
142 	int real_sqp = 0;
143 	int i;
144 	/* PPF or Native -- real SQP */
145 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
146 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
147 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
148 	if (real_sqp)
149 		return 1;
150 	/* VF or PF -- proxy SQP */
151 	if (mlx4_is_mfunc(dev->dev)) {
152 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
153 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy ||
154 			    qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp1_proxy) {
155 				proxy_sqp = 1;
156 				break;
157 			}
158 		}
159 	}
160 	if (proxy_sqp)
161 		return 1;
162 
163 	return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
164 }
165 
166 /* used for INIT/CLOSE port logic */
167 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
168 {
169 	int proxy_qp0 = 0;
170 	int real_qp0 = 0;
171 	int i;
172 	/* PPF or Native -- real QP0 */
173 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
174 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
175 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
176 	if (real_qp0)
177 		return 1;
178 	/* VF or PF -- proxy QP0 */
179 	if (mlx4_is_mfunc(dev->dev)) {
180 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
181 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy) {
182 				proxy_qp0 = 1;
183 				break;
184 			}
185 		}
186 	}
187 	return proxy_qp0;
188 }
189 
190 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
191 {
192 	return mlx4_buf_offset(&qp->buf, offset);
193 }
194 
195 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
196 {
197 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
198 }
199 
200 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
201 {
202 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
203 }
204 
205 /*
206  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
207  * first four bytes of every 64 byte chunk with
208  *     0x7FFFFFF | (invalid_ownership_value << 31).
209  *
210  * When the max work request size is less than or equal to the WQE
211  * basic block size, as an optimization, we can stamp all WQEs with
212  * 0xffffffff, and skip the very first chunk of each WQE.
213  */
214 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
215 {
216 	__be32 *wqe;
217 	int i;
218 	int s;
219 	int ind;
220 	void *buf;
221 	__be32 stamp;
222 	struct mlx4_wqe_ctrl_seg *ctrl;
223 
224 	if (qp->sq_max_wqes_per_wr > 1) {
225 		s = roundup(size, 1U << qp->sq.wqe_shift);
226 		for (i = 0; i < s; i += 64) {
227 			ind = (i >> qp->sq.wqe_shift) + n;
228 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
229 						       cpu_to_be32(0xffffffff);
230 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
231 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
232 			*wqe = stamp;
233 		}
234 	} else {
235 		ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
236 		s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4;
237 		for (i = 64; i < s; i += 64) {
238 			wqe = buf + i;
239 			*wqe = cpu_to_be32(0xffffffff);
240 		}
241 	}
242 }
243 
244 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
245 {
246 	struct mlx4_wqe_ctrl_seg *ctrl;
247 	struct mlx4_wqe_inline_seg *inl;
248 	void *wqe;
249 	int s;
250 
251 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
252 	s = sizeof(struct mlx4_wqe_ctrl_seg);
253 
254 	if (qp->ibqp.qp_type == IB_QPT_UD) {
255 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
256 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
257 		memset(dgram, 0, sizeof *dgram);
258 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
259 		s += sizeof(struct mlx4_wqe_datagram_seg);
260 	}
261 
262 	/* Pad the remainder of the WQE with an inline data segment. */
263 	if (size > s) {
264 		inl = wqe + s;
265 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
266 	}
267 	ctrl->srcrb_flags = 0;
268 	ctrl->qpn_vlan.fence_size = size / 16;
269 	/*
270 	 * Make sure descriptor is fully written before setting ownership bit
271 	 * (because HW can start executing as soon as we do).
272 	 */
273 	wmb();
274 
275 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
276 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
277 
278 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
279 }
280 
281 /* Post NOP WQE to prevent wrap-around in the middle of WR */
282 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
283 {
284 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
285 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
286 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
287 		ind += s;
288 	}
289 	return ind;
290 }
291 
292 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
293 {
294 	struct ib_event event;
295 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
296 
297 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
298 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
299 
300 	if (ibqp->event_handler) {
301 		event.device     = ibqp->device;
302 		event.element.qp = ibqp;
303 		switch (type) {
304 		case MLX4_EVENT_TYPE_PATH_MIG:
305 			event.event = IB_EVENT_PATH_MIG;
306 			break;
307 		case MLX4_EVENT_TYPE_COMM_EST:
308 			event.event = IB_EVENT_COMM_EST;
309 			break;
310 		case MLX4_EVENT_TYPE_SQ_DRAINED:
311 			event.event = IB_EVENT_SQ_DRAINED;
312 			break;
313 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
314 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
315 			break;
316 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
317 			event.event = IB_EVENT_QP_FATAL;
318 			break;
319 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
320 			event.event = IB_EVENT_PATH_MIG_ERR;
321 			break;
322 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
323 			event.event = IB_EVENT_QP_REQ_ERR;
324 			break;
325 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
326 			event.event = IB_EVENT_QP_ACCESS_ERR;
327 			break;
328 		default:
329 			pr_warn("Unexpected event type %d "
330 			       "on QP %06x\n", type, qp->qpn);
331 			return;
332 		}
333 
334 		ibqp->event_handler(&event, ibqp->qp_context);
335 	}
336 }
337 
338 static void mlx4_ib_wq_event(struct mlx4_qp *qp, enum mlx4_event type)
339 {
340 	pr_warn_ratelimited("Unexpected event type %d on WQ 0x%06x. Events are not supported for WQs\n",
341 			    type, qp->qpn);
342 }
343 
344 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
345 {
346 	/*
347 	 * UD WQEs must have a datagram segment.
348 	 * RC and UC WQEs might have a remote address segment.
349 	 * MLX WQEs need two extra inline data segments (for the UD
350 	 * header and space for the ICRC).
351 	 */
352 	switch (type) {
353 	case MLX4_IB_QPT_UD:
354 		return sizeof (struct mlx4_wqe_ctrl_seg) +
355 			sizeof (struct mlx4_wqe_datagram_seg) +
356 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
357 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
358 	case MLX4_IB_QPT_PROXY_SMI:
359 	case MLX4_IB_QPT_PROXY_GSI:
360 		return sizeof (struct mlx4_wqe_ctrl_seg) +
361 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
362 	case MLX4_IB_QPT_TUN_SMI_OWNER:
363 	case MLX4_IB_QPT_TUN_GSI:
364 		return sizeof (struct mlx4_wqe_ctrl_seg) +
365 			sizeof (struct mlx4_wqe_datagram_seg);
366 
367 	case MLX4_IB_QPT_UC:
368 		return sizeof (struct mlx4_wqe_ctrl_seg) +
369 			sizeof (struct mlx4_wqe_raddr_seg);
370 	case MLX4_IB_QPT_RC:
371 		return sizeof (struct mlx4_wqe_ctrl_seg) +
372 			sizeof (struct mlx4_wqe_masked_atomic_seg) +
373 			sizeof (struct mlx4_wqe_raddr_seg);
374 	case MLX4_IB_QPT_SMI:
375 	case MLX4_IB_QPT_GSI:
376 		return sizeof (struct mlx4_wqe_ctrl_seg) +
377 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
378 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
379 					   MLX4_INLINE_ALIGN) *
380 			      sizeof (struct mlx4_wqe_inline_seg),
381 			      sizeof (struct mlx4_wqe_data_seg)) +
382 			ALIGN(4 +
383 			      sizeof (struct mlx4_wqe_inline_seg),
384 			      sizeof (struct mlx4_wqe_data_seg));
385 	default:
386 		return sizeof (struct mlx4_wqe_ctrl_seg);
387 	}
388 }
389 
390 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
391 		       int is_user, int has_rq, struct mlx4_ib_qp *qp,
392 		       u32 inl_recv_sz)
393 {
394 	/* Sanity check RQ size before proceeding */
395 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
396 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
397 		return -EINVAL;
398 
399 	if (!has_rq) {
400 		if (cap->max_recv_wr || inl_recv_sz)
401 			return -EINVAL;
402 
403 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
404 	} else {
405 		u32 max_inl_recv_sz = dev->dev->caps.max_rq_sg *
406 			sizeof(struct mlx4_wqe_data_seg);
407 		u32 wqe_size;
408 
409 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
410 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge ||
411 				inl_recv_sz > max_inl_recv_sz))
412 			return -EINVAL;
413 
414 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
415 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
416 		wqe_size = qp->rq.max_gs * sizeof(struct mlx4_wqe_data_seg);
417 		qp->rq.wqe_shift = ilog2(max_t(u32, wqe_size, inl_recv_sz));
418 	}
419 
420 	/* leave userspace return values as they were, so as not to break ABI */
421 	if (is_user) {
422 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
423 		cap->max_recv_sge = qp->rq.max_gs;
424 	} else {
425 		cap->max_recv_wr  = qp->rq.max_post =
426 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
427 		cap->max_recv_sge = min(qp->rq.max_gs,
428 					min(dev->dev->caps.max_sq_sg,
429 					    dev->dev->caps.max_rq_sg));
430 	}
431 
432 	return 0;
433 }
434 
435 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
436 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp,
437 			      bool shrink_wqe)
438 {
439 	int s;
440 
441 	/* Sanity check SQ size before proceeding */
442 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
443 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
444 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
445 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
446 		return -EINVAL;
447 
448 	/*
449 	 * For MLX transport we need 2 extra S/G entries:
450 	 * one for the header and one for the checksum at the end
451 	 */
452 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
453 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
454 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
455 		return -EINVAL;
456 
457 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
458 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
459 		send_wqe_overhead(type, qp->flags);
460 
461 	if (s > dev->dev->caps.max_sq_desc_sz)
462 		return -EINVAL;
463 
464 	/*
465 	 * Hermon supports shrinking WQEs, such that a single work
466 	 * request can include multiple units of 1 << wqe_shift.  This
467 	 * way, work requests can differ in size, and do not have to
468 	 * be a power of 2 in size, saving memory and speeding up send
469 	 * WR posting.  Unfortunately, if we do this then the
470 	 * wqe_index field in CQEs can't be used to look up the WR ID
471 	 * anymore, so we do this only if selective signaling is off.
472 	 *
473 	 * Further, on 32-bit platforms, we can't use vmap() to make
474 	 * the QP buffer virtually contiguous.  Thus we have to use
475 	 * constant-sized WRs to make sure a WR is always fully within
476 	 * a single page-sized chunk.
477 	 *
478 	 * Finally, we use NOP work requests to pad the end of the
479 	 * work queue, to avoid wrap-around in the middle of WR.  We
480 	 * set NEC bit to avoid getting completions with error for
481 	 * these NOP WRs, but since NEC is only supported starting
482 	 * with firmware 2.2.232, we use constant-sized WRs for older
483 	 * firmware.
484 	 *
485 	 * And, since MLX QPs only support SEND, we use constant-sized
486 	 * WRs in this case.
487 	 *
488 	 * We look for the smallest value of wqe_shift such that the
489 	 * resulting number of wqes does not exceed device
490 	 * capabilities.
491 	 *
492 	 * We set WQE size to at least 64 bytes, this way stamping
493 	 * invalidates each WQE.
494 	 */
495 	if (shrink_wqe && dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
496 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
497 	    type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
498 	    !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
499 		      MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
500 		qp->sq.wqe_shift = ilog2(64);
501 	else
502 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
503 
504 	for (;;) {
505 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
506 
507 		/*
508 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
509 		 * allow HW to prefetch.
510 		 */
511 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
512 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
513 						    qp->sq_max_wqes_per_wr +
514 						    qp->sq_spare_wqes);
515 
516 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
517 			break;
518 
519 		if (qp->sq_max_wqes_per_wr <= 1)
520 			return -EINVAL;
521 
522 		++qp->sq.wqe_shift;
523 	}
524 
525 	qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
526 			     (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
527 			 send_wqe_overhead(type, qp->flags)) /
528 		sizeof (struct mlx4_wqe_data_seg);
529 
530 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
531 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
532 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
533 		qp->rq.offset = 0;
534 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
535 	} else {
536 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
537 		qp->sq.offset = 0;
538 	}
539 
540 	cap->max_send_wr  = qp->sq.max_post =
541 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
542 	cap->max_send_sge = min(qp->sq.max_gs,
543 				min(dev->dev->caps.max_sq_sg,
544 				    dev->dev->caps.max_rq_sg));
545 	/* We don't support inline sends for kernel QPs (yet) */
546 	cap->max_inline_data = 0;
547 
548 	return 0;
549 }
550 
551 static int set_user_sq_size(struct mlx4_ib_dev *dev,
552 			    struct mlx4_ib_qp *qp,
553 			    struct mlx4_ib_create_qp *ucmd)
554 {
555 	/* Sanity check SQ size before proceeding */
556 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
557 	    ucmd->log_sq_stride >
558 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
559 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
560 		return -EINVAL;
561 
562 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
563 	qp->sq.wqe_shift = ucmd->log_sq_stride;
564 
565 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
566 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
567 
568 	return 0;
569 }
570 
571 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
572 {
573 	int i;
574 
575 	qp->sqp_proxy_rcv =
576 		kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
577 			GFP_KERNEL);
578 	if (!qp->sqp_proxy_rcv)
579 		return -ENOMEM;
580 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
581 		qp->sqp_proxy_rcv[i].addr =
582 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
583 				GFP_KERNEL);
584 		if (!qp->sqp_proxy_rcv[i].addr)
585 			goto err;
586 		qp->sqp_proxy_rcv[i].map =
587 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
588 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
589 					  DMA_FROM_DEVICE);
590 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
591 			kfree(qp->sqp_proxy_rcv[i].addr);
592 			goto err;
593 		}
594 	}
595 	return 0;
596 
597 err:
598 	while (i > 0) {
599 		--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 	qp->sqp_proxy_rcv = NULL;
607 	return -ENOMEM;
608 }
609 
610 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
611 {
612 	int i;
613 
614 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
615 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
616 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
617 				    DMA_FROM_DEVICE);
618 		kfree(qp->sqp_proxy_rcv[i].addr);
619 	}
620 	kfree(qp->sqp_proxy_rcv);
621 }
622 
623 static int qp_has_rq(struct ib_qp_init_attr *attr)
624 {
625 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
626 		return 0;
627 
628 	return !attr->srq;
629 }
630 
631 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
632 {
633 	int i;
634 	for (i = 0; i < dev->caps.num_ports; i++) {
635 		if (qpn == dev->caps.spec_qps[i].qp0_proxy)
636 			return !!dev->caps.spec_qps[i].qp0_qkey;
637 	}
638 	return 0;
639 }
640 
641 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
642 				    struct mlx4_ib_qp *qp)
643 {
644 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
645 	mlx4_counter_free(dev->dev, qp->counter_index->index);
646 	list_del(&qp->counter_index->list);
647 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
648 
649 	kfree(qp->counter_index);
650 	qp->counter_index = NULL;
651 }
652 
653 static int set_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_rss *rss_ctx,
654 		      struct ib_qp_init_attr *init_attr,
655 		      struct mlx4_ib_create_qp_rss *ucmd)
656 {
657 	rss_ctx->base_qpn_tbl_sz = init_attr->rwq_ind_tbl->ind_tbl[0]->wq_num |
658 		(init_attr->rwq_ind_tbl->log_ind_tbl_size << 24);
659 
660 	if ((ucmd->rx_hash_function == MLX4_IB_RX_HASH_FUNC_TOEPLITZ) &&
661 	    (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS_TOP)) {
662 		memcpy(rss_ctx->rss_key, ucmd->rx_hash_key,
663 		       MLX4_EN_RSS_KEY_SIZE);
664 	} else {
665 		pr_debug("RX Hash function is not supported\n");
666 		return (-EOPNOTSUPP);
667 	}
668 
669 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) &&
670 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
671 		rss_ctx->flags = MLX4_RSS_IPV4;
672 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) ||
673 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
674 		pr_debug("RX Hash fields_mask is not supported - both IPv4 SRC and DST must be set\n");
675 		return (-EOPNOTSUPP);
676 	}
677 
678 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) &&
679 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
680 		rss_ctx->flags |= MLX4_RSS_IPV6;
681 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) ||
682 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
683 		pr_debug("RX Hash fields_mask is not supported - both IPv6 SRC and DST must be set\n");
684 		return (-EOPNOTSUPP);
685 	}
686 
687 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) &&
688 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
689 		if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS)) {
690 			pr_debug("RX Hash fields_mask for UDP is not supported\n");
691 			return (-EOPNOTSUPP);
692 		}
693 
694 		if (rss_ctx->flags & MLX4_RSS_IPV4) {
695 			rss_ctx->flags |= MLX4_RSS_UDP_IPV4;
696 		} else if (rss_ctx->flags & MLX4_RSS_IPV6) {
697 			rss_ctx->flags |= MLX4_RSS_UDP_IPV6;
698 		} else {
699 			pr_debug("RX Hash fields_mask is not supported - UDP must be set with IPv4 or IPv6\n");
700 			return (-EOPNOTSUPP);
701 		}
702 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) ||
703 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
704 		pr_debug("RX Hash fields_mask is not supported - both UDP SRC and DST must be set\n");
705 		return (-EOPNOTSUPP);
706 	}
707 
708 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) &&
709 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
710 		if (rss_ctx->flags & MLX4_RSS_IPV4) {
711 			rss_ctx->flags |= MLX4_RSS_TCP_IPV4;
712 		} else if (rss_ctx->flags & MLX4_RSS_IPV6) {
713 			rss_ctx->flags |= MLX4_RSS_TCP_IPV6;
714 		} else {
715 			pr_debug("RX Hash fields_mask is not supported - TCP must be set with IPv4 or IPv6\n");
716 			return (-EOPNOTSUPP);
717 		}
718 
719 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) ||
720 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
721 		pr_debug("RX Hash fields_mask is not supported - both TCP SRC and DST must be set\n");
722 		return (-EOPNOTSUPP);
723 	}
724 
725 	return 0;
726 }
727 
728 static int create_qp_rss(struct mlx4_ib_dev *dev, struct ib_pd *ibpd,
729 			 struct ib_qp_init_attr *init_attr,
730 			 struct mlx4_ib_create_qp_rss *ucmd,
731 			 struct mlx4_ib_qp *qp)
732 {
733 	int qpn;
734 	int err;
735 
736 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
737 
738 	err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 0, qp->mqp.usage);
739 	if (err)
740 		return err;
741 
742 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
743 	if (err)
744 		goto err_qpn;
745 
746 	mutex_init(&qp->mutex);
747 
748 	INIT_LIST_HEAD(&qp->gid_list);
749 	INIT_LIST_HEAD(&qp->steering_rules);
750 
751 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
752 	qp->state = IB_QPS_RESET;
753 
754 	/* Set dummy send resources to be compatible with HV and PRM */
755 	qp->sq_no_prefetch = 1;
756 	qp->sq.wqe_cnt = 1;
757 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
758 	qp->buf_size = qp->sq.wqe_cnt << MLX4_IB_MIN_SQ_STRIDE;
759 	qp->mtt = (to_mqp(
760 		   (struct ib_qp *)init_attr->rwq_ind_tbl->ind_tbl[0]))->mtt;
761 
762 	qp->rss_ctx = kzalloc(sizeof(*qp->rss_ctx), GFP_KERNEL);
763 	if (!qp->rss_ctx) {
764 		err = -ENOMEM;
765 		goto err_qp_alloc;
766 	}
767 
768 	err = set_qp_rss(dev, qp->rss_ctx, init_attr, ucmd);
769 	if (err)
770 		goto err;
771 
772 	return 0;
773 
774 err:
775 	kfree(qp->rss_ctx);
776 
777 err_qp_alloc:
778 	mlx4_qp_remove(dev->dev, &qp->mqp);
779 	mlx4_qp_free(dev->dev, &qp->mqp);
780 
781 err_qpn:
782 	mlx4_qp_release_range(dev->dev, qpn, 1);
783 	return err;
784 }
785 
786 static struct ib_qp *_mlx4_ib_create_qp_rss(struct ib_pd *pd,
787 					    struct ib_qp_init_attr *init_attr,
788 					    struct ib_udata *udata)
789 {
790 	struct mlx4_ib_qp *qp;
791 	struct mlx4_ib_create_qp_rss ucmd = {};
792 	size_t required_cmd_sz;
793 	int err;
794 
795 	if (!udata) {
796 		pr_debug("RSS QP with NULL udata\n");
797 		return ERR_PTR(-EINVAL);
798 	}
799 
800 	if (udata->outlen)
801 		return ERR_PTR(-EOPNOTSUPP);
802 
803 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) +
804 					sizeof(ucmd.reserved1);
805 	if (udata->inlen < required_cmd_sz) {
806 		pr_debug("invalid inlen\n");
807 		return ERR_PTR(-EINVAL);
808 	}
809 
810 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
811 		pr_debug("copy failed\n");
812 		return ERR_PTR(-EFAULT);
813 	}
814 
815 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)))
816 		return ERR_PTR(-EOPNOTSUPP);
817 
818 	if (ucmd.comp_mask || ucmd.reserved1)
819 		return ERR_PTR(-EOPNOTSUPP);
820 
821 	if (udata->inlen > sizeof(ucmd) &&
822 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
823 				 udata->inlen - sizeof(ucmd))) {
824 		pr_debug("inlen is not supported\n");
825 		return ERR_PTR(-EOPNOTSUPP);
826 	}
827 
828 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
829 		pr_debug("RSS QP with unsupported QP type %d\n",
830 			 init_attr->qp_type);
831 		return ERR_PTR(-EOPNOTSUPP);
832 	}
833 
834 	if (init_attr->create_flags) {
835 		pr_debug("RSS QP doesn't support create flags\n");
836 		return ERR_PTR(-EOPNOTSUPP);
837 	}
838 
839 	if (init_attr->send_cq || init_attr->cap.max_send_wr) {
840 		pr_debug("RSS QP with unsupported send attributes\n");
841 		return ERR_PTR(-EOPNOTSUPP);
842 	}
843 
844 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
845 	if (!qp)
846 		return ERR_PTR(-ENOMEM);
847 
848 	qp->pri.vid = 0xFFFF;
849 	qp->alt.vid = 0xFFFF;
850 
851 	err = create_qp_rss(to_mdev(pd->device), pd, init_attr, &ucmd, qp);
852 	if (err) {
853 		kfree(qp);
854 		return ERR_PTR(err);
855 	}
856 
857 	qp->ibqp.qp_num = qp->mqp.qpn;
858 
859 	return &qp->ibqp;
860 }
861 
862 /*
863  * This function allocates a WQN from a range which is consecutive and aligned
864  * to its size. In case the range is full, then it creates a new range and
865  * allocates WQN from it. The new range will be used for following allocations.
866  */
867 static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context,
868 			     struct mlx4_ib_qp *qp, int range_size, int *wqn)
869 {
870 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
871 	struct mlx4_wqn_range *range;
872 	int err = 0;
873 
874 	mutex_lock(&context->wqn_ranges_mutex);
875 
876 	range = list_first_entry_or_null(&context->wqn_ranges_list,
877 					 struct mlx4_wqn_range, list);
878 
879 	if (!range || (range->refcount == range->size) || range->dirty) {
880 		range = kzalloc(sizeof(*range), GFP_KERNEL);
881 		if (!range) {
882 			err = -ENOMEM;
883 			goto out;
884 		}
885 
886 		err = mlx4_qp_reserve_range(dev->dev, range_size,
887 					    range_size, &range->base_wqn, 0,
888 					    qp->mqp.usage);
889 		if (err) {
890 			kfree(range);
891 			goto out;
892 		}
893 
894 		range->size = range_size;
895 		list_add(&range->list, &context->wqn_ranges_list);
896 	} else if (range_size != 1) {
897 		/*
898 		 * Requesting a new range (>1) when last range is still open, is
899 		 * not valid.
900 		 */
901 		err = -EINVAL;
902 		goto out;
903 	}
904 
905 	qp->wqn_range = range;
906 
907 	*wqn = range->base_wqn + range->refcount;
908 
909 	range->refcount++;
910 
911 out:
912 	mutex_unlock(&context->wqn_ranges_mutex);
913 
914 	return err;
915 }
916 
917 static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context,
918 				struct mlx4_ib_qp *qp, bool dirty_release)
919 {
920 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
921 	struct mlx4_wqn_range *range;
922 
923 	mutex_lock(&context->wqn_ranges_mutex);
924 
925 	range = qp->wqn_range;
926 
927 	range->refcount--;
928 	if (!range->refcount) {
929 		mlx4_qp_release_range(dev->dev, range->base_wqn,
930 				      range->size);
931 		list_del(&range->list);
932 		kfree(range);
933 	} else if (dirty_release) {
934 	/*
935 	 * A range which one of its WQNs is destroyed, won't be able to be
936 	 * reused for further WQN allocations.
937 	 * The next created WQ will allocate a new range.
938 	 */
939 		range->dirty = 1;
940 	}
941 
942 	mutex_unlock(&context->wqn_ranges_mutex);
943 }
944 
945 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
946 			    enum mlx4_ib_source_type src,
947 			    struct ib_qp_init_attr *init_attr,
948 			    struct ib_udata *udata, int sqpn,
949 			    struct mlx4_ib_qp **caller_qp)
950 {
951 	int qpn;
952 	int err;
953 	struct ib_qp_cap backup_cap;
954 	struct mlx4_ib_sqp *sqp = NULL;
955 	struct mlx4_ib_qp *qp;
956 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
957 	struct mlx4_ib_cq *mcq;
958 	unsigned long flags;
959 	int range_size = 0;
960 
961 	/* When tunneling special qps, we use a plain UD qp */
962 	if (sqpn) {
963 		if (mlx4_is_mfunc(dev->dev) &&
964 		    (!mlx4_is_master(dev->dev) ||
965 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
966 			if (init_attr->qp_type == IB_QPT_GSI)
967 				qp_type = MLX4_IB_QPT_PROXY_GSI;
968 			else {
969 				if (mlx4_is_master(dev->dev) ||
970 				    qp0_enabled_vf(dev->dev, sqpn))
971 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
972 				else
973 					qp_type = MLX4_IB_QPT_PROXY_SMI;
974 			}
975 		}
976 		qpn = sqpn;
977 		/* add extra sg entry for tunneling */
978 		init_attr->cap.max_recv_sge++;
979 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
980 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
981 			container_of(init_attr,
982 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
983 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
984 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
985 		    !mlx4_is_master(dev->dev))
986 			return -EINVAL;
987 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
988 			qp_type = MLX4_IB_QPT_TUN_GSI;
989 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
990 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
991 					     tnl_init->port))
992 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
993 		else
994 			qp_type = MLX4_IB_QPT_TUN_SMI;
995 		/* we are definitely in the PPF here, since we are creating
996 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
997 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
998 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
999 		sqpn = qpn;
1000 	}
1001 
1002 	if (!*caller_qp) {
1003 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
1004 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
1005 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
1006 			sqp = kzalloc(sizeof(struct mlx4_ib_sqp), GFP_KERNEL);
1007 			if (!sqp)
1008 				return -ENOMEM;
1009 			qp = &sqp->qp;
1010 			qp->pri.vid = 0xFFFF;
1011 			qp->alt.vid = 0xFFFF;
1012 		} else {
1013 			qp = kzalloc(sizeof(struct mlx4_ib_qp), GFP_KERNEL);
1014 			if (!qp)
1015 				return -ENOMEM;
1016 			qp->pri.vid = 0xFFFF;
1017 			qp->alt.vid = 0xFFFF;
1018 		}
1019 	} else
1020 		qp = *caller_qp;
1021 
1022 	qp->mlx4_ib_qp_type = qp_type;
1023 
1024 	mutex_init(&qp->mutex);
1025 	spin_lock_init(&qp->sq.lock);
1026 	spin_lock_init(&qp->rq.lock);
1027 	INIT_LIST_HEAD(&qp->gid_list);
1028 	INIT_LIST_HEAD(&qp->steering_rules);
1029 
1030 	qp->state	 = IB_QPS_RESET;
1031 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1032 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1033 
1034 
1035 	if (pd->uobject) {
1036 		union {
1037 			struct mlx4_ib_create_qp qp;
1038 			struct mlx4_ib_create_wq wq;
1039 		} ucmd;
1040 		size_t copy_len;
1041 		int shift;
1042 		int n;
1043 
1044 		copy_len = (src == MLX4_IB_QP_SRC) ?
1045 			   sizeof(struct mlx4_ib_create_qp) :
1046 			   min(sizeof(struct mlx4_ib_create_wq), udata->inlen);
1047 
1048 		if (ib_copy_from_udata(&ucmd, udata, copy_len)) {
1049 			err = -EFAULT;
1050 			goto err;
1051 		}
1052 
1053 		if (src == MLX4_IB_RWQ_SRC) {
1054 			if (ucmd.wq.comp_mask || ucmd.wq.reserved[0] ||
1055 			    ucmd.wq.reserved[1] || ucmd.wq.reserved[2]) {
1056 				pr_debug("user command isn't supported\n");
1057 				err = -EOPNOTSUPP;
1058 				goto err;
1059 			}
1060 
1061 			if (ucmd.wq.log_range_size >
1062 			    ilog2(dev->dev->caps.max_rss_tbl_sz)) {
1063 				pr_debug("WQN range size must be equal or smaller than %d\n",
1064 					 dev->dev->caps.max_rss_tbl_sz);
1065 				err = -EOPNOTSUPP;
1066 				goto err;
1067 			}
1068 			range_size = 1 << ucmd.wq.log_range_size;
1069 		} else {
1070 			qp->inl_recv_sz = ucmd.qp.inl_recv_sz;
1071 		}
1072 
1073 		err = set_rq_size(dev, &init_attr->cap, !!pd->uobject,
1074 				  qp_has_rq(init_attr), qp, qp->inl_recv_sz);
1075 		if (err)
1076 			goto err;
1077 
1078 		if (src == MLX4_IB_QP_SRC) {
1079 			qp->sq_no_prefetch = ucmd.qp.sq_no_prefetch;
1080 
1081 			err = set_user_sq_size(dev, qp,
1082 					       (struct mlx4_ib_create_qp *)
1083 					       &ucmd);
1084 			if (err)
1085 				goto err;
1086 		} else {
1087 			qp->sq_no_prefetch = 1;
1088 			qp->sq.wqe_cnt = 1;
1089 			qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
1090 			/* Allocated buffer expects to have at least that SQ
1091 			 * size.
1092 			 */
1093 			qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
1094 				(qp->sq.wqe_cnt << qp->sq.wqe_shift);
1095 		}
1096 
1097 		qp->umem = ib_umem_get(pd->uobject->context,
1098 				(src == MLX4_IB_QP_SRC) ? ucmd.qp.buf_addr :
1099 				ucmd.wq.buf_addr, qp->buf_size, 0, 0);
1100 		if (IS_ERR(qp->umem)) {
1101 			err = PTR_ERR(qp->umem);
1102 			goto err;
1103 		}
1104 
1105 		n = ib_umem_page_count(qp->umem);
1106 		shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
1107 		err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
1108 
1109 		if (err)
1110 			goto err_buf;
1111 
1112 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
1113 		if (err)
1114 			goto err_mtt;
1115 
1116 		if (qp_has_rq(init_attr)) {
1117 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
1118 				(src == MLX4_IB_QP_SRC) ? ucmd.qp.db_addr :
1119 				ucmd.wq.db_addr, &qp->db);
1120 			if (err)
1121 				goto err_mtt;
1122 		}
1123 		qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
1124 	} else {
1125 		err = set_rq_size(dev, &init_attr->cap, !!pd->uobject,
1126 				  qp_has_rq(init_attr), qp, 0);
1127 		if (err)
1128 			goto err;
1129 
1130 		qp->sq_no_prefetch = 0;
1131 
1132 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
1133 			qp->flags |= MLX4_IB_QP_LSO;
1134 
1135 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1136 			if (dev->steering_support ==
1137 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
1138 				qp->flags |= MLX4_IB_QP_NETIF;
1139 			else
1140 				goto err;
1141 		}
1142 
1143 		memcpy(&backup_cap, &init_attr->cap, sizeof(backup_cap));
1144 		err = set_kernel_sq_size(dev, &init_attr->cap,
1145 					 qp_type, qp, true);
1146 		if (err)
1147 			goto err;
1148 
1149 		if (qp_has_rq(init_attr)) {
1150 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
1151 			if (err)
1152 				goto err;
1153 
1154 			*qp->db.db = 0;
1155 		}
1156 
1157 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, qp->buf_size,
1158 				   &qp->buf)) {
1159 			memcpy(&init_attr->cap, &backup_cap,
1160 			       sizeof(backup_cap));
1161 			err = set_kernel_sq_size(dev, &init_attr->cap, qp_type,
1162 						 qp, false);
1163 			if (err)
1164 				goto err_db;
1165 
1166 			if (mlx4_buf_alloc(dev->dev, qp->buf_size,
1167 					   PAGE_SIZE * 2, &qp->buf)) {
1168 				err = -ENOMEM;
1169 				goto err_db;
1170 			}
1171 		}
1172 
1173 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
1174 				    &qp->mtt);
1175 		if (err)
1176 			goto err_buf;
1177 
1178 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
1179 		if (err)
1180 			goto err_mtt;
1181 
1182 		qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1183 					     sizeof(u64), GFP_KERNEL);
1184 		qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1185 					     sizeof(u64), GFP_KERNEL);
1186 		if (!qp->sq.wrid || !qp->rq.wrid) {
1187 			err = -ENOMEM;
1188 			goto err_wrid;
1189 		}
1190 		qp->mqp.usage = MLX4_RES_USAGE_DRIVER;
1191 	}
1192 
1193 	if (sqpn) {
1194 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1195 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
1196 			if (alloc_proxy_bufs(pd->device, qp)) {
1197 				err = -ENOMEM;
1198 				goto err_wrid;
1199 			}
1200 		}
1201 	} else if (src == MLX4_IB_RWQ_SRC) {
1202 		err = mlx4_ib_alloc_wqn(to_mucontext(pd->uobject->context), qp,
1203 					range_size, &qpn);
1204 		if (err)
1205 			goto err_wrid;
1206 	} else {
1207 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
1208 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
1209 		 * VLAN insertion. */
1210 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
1211 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
1212 						    (init_attr->cap.max_send_wr ?
1213 						     MLX4_RESERVE_ETH_BF_QP : 0) |
1214 						    (init_attr->cap.max_recv_wr ?
1215 						     MLX4_RESERVE_A0_QP : 0),
1216 						    qp->mqp.usage);
1217 		else
1218 			if (qp->flags & MLX4_IB_QP_NETIF)
1219 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
1220 			else
1221 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
1222 							    &qpn, 0, qp->mqp.usage);
1223 		if (err)
1224 			goto err_proxy;
1225 	}
1226 
1227 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1228 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1229 
1230 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
1231 	if (err)
1232 		goto err_qpn;
1233 
1234 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1235 		qp->mqp.qpn |= (1 << 23);
1236 
1237 	/*
1238 	 * Hardware wants QPN written in big-endian order (after
1239 	 * shifting) for send doorbell.  Precompute this value to save
1240 	 * a little bit when posting sends.
1241 	 */
1242 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1243 
1244 	qp->mqp.event = (src == MLX4_IB_QP_SRC) ? mlx4_ib_qp_event :
1245 						  mlx4_ib_wq_event;
1246 
1247 	if (!*caller_qp)
1248 		*caller_qp = qp;
1249 
1250 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1251 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
1252 			 to_mcq(init_attr->recv_cq));
1253 	/* Maintain device to QPs access, needed for further handling
1254 	 * via reset flow
1255 	 */
1256 	list_add_tail(&qp->qps_list, &dev->qp_list);
1257 	/* Maintain CQ to QPs access, needed for further handling
1258 	 * via reset flow
1259 	 */
1260 	mcq = to_mcq(init_attr->send_cq);
1261 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
1262 	mcq = to_mcq(init_attr->recv_cq);
1263 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
1264 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
1265 			   to_mcq(init_attr->recv_cq));
1266 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1267 	return 0;
1268 
1269 err_qpn:
1270 	if (!sqpn) {
1271 		if (qp->flags & MLX4_IB_QP_NETIF)
1272 			mlx4_ib_steer_qp_free(dev, qpn, 1);
1273 		else if (src == MLX4_IB_RWQ_SRC)
1274 			mlx4_ib_release_wqn(to_mucontext(pd->uobject->context),
1275 					    qp, 0);
1276 		else
1277 			mlx4_qp_release_range(dev->dev, qpn, 1);
1278 	}
1279 err_proxy:
1280 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1281 		free_proxy_bufs(pd->device, qp);
1282 err_wrid:
1283 	if (pd->uobject) {
1284 		if (qp_has_rq(init_attr))
1285 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
1286 	} else {
1287 		kvfree(qp->sq.wrid);
1288 		kvfree(qp->rq.wrid);
1289 	}
1290 
1291 err_mtt:
1292 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1293 
1294 err_buf:
1295 	if (pd->uobject)
1296 		ib_umem_release(qp->umem);
1297 	else
1298 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1299 
1300 err_db:
1301 	if (!pd->uobject && qp_has_rq(init_attr))
1302 		mlx4_db_free(dev->dev, &qp->db);
1303 
1304 err:
1305 	if (sqp)
1306 		kfree(sqp);
1307 	else if (!*caller_qp)
1308 		kfree(qp);
1309 	return err;
1310 }
1311 
1312 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
1313 {
1314 	switch (state) {
1315 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
1316 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
1317 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
1318 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
1319 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
1320 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
1321 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
1322 	default:		return -1;
1323 	}
1324 }
1325 
1326 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1327 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1328 {
1329 	if (send_cq == recv_cq) {
1330 		spin_lock(&send_cq->lock);
1331 		__acquire(&recv_cq->lock);
1332 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1333 		spin_lock(&send_cq->lock);
1334 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1335 	} else {
1336 		spin_lock(&recv_cq->lock);
1337 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1338 	}
1339 }
1340 
1341 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1342 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1343 {
1344 	if (send_cq == recv_cq) {
1345 		__release(&recv_cq->lock);
1346 		spin_unlock(&send_cq->lock);
1347 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1348 		spin_unlock(&recv_cq->lock);
1349 		spin_unlock(&send_cq->lock);
1350 	} else {
1351 		spin_unlock(&send_cq->lock);
1352 		spin_unlock(&recv_cq->lock);
1353 	}
1354 }
1355 
1356 static void del_gid_entries(struct mlx4_ib_qp *qp)
1357 {
1358 	struct mlx4_ib_gid_entry *ge, *tmp;
1359 
1360 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1361 		list_del(&ge->list);
1362 		kfree(ge);
1363 	}
1364 }
1365 
1366 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
1367 {
1368 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
1369 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
1370 	else
1371 		return to_mpd(qp->ibqp.pd);
1372 }
1373 
1374 static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src,
1375 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1376 {
1377 	switch (qp->ibqp.qp_type) {
1378 	case IB_QPT_XRC_TGT:
1379 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1380 		*recv_cq = *send_cq;
1381 		break;
1382 	case IB_QPT_XRC_INI:
1383 		*send_cq = to_mcq(qp->ibqp.send_cq);
1384 		*recv_cq = *send_cq;
1385 		break;
1386 	default:
1387 		*recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) :
1388 						     to_mcq(qp->ibwq.cq);
1389 		*send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) :
1390 						     *recv_cq;
1391 		break;
1392 	}
1393 }
1394 
1395 static void destroy_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1396 {
1397 	if (qp->state != IB_QPS_RESET) {
1398 		int i;
1399 
1400 		for (i = 0; i < (1 << qp->ibqp.rwq_ind_tbl->log_ind_tbl_size);
1401 		     i++) {
1402 			struct ib_wq *ibwq = qp->ibqp.rwq_ind_tbl->ind_tbl[i];
1403 			struct mlx4_ib_qp *wq =	to_mqp((struct ib_qp *)ibwq);
1404 
1405 			mutex_lock(&wq->mutex);
1406 
1407 			wq->rss_usecnt--;
1408 
1409 			mutex_unlock(&wq->mutex);
1410 		}
1411 
1412 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1413 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1414 			pr_warn("modify QP %06x to RESET failed.\n",
1415 				qp->mqp.qpn);
1416 	}
1417 
1418 	mlx4_qp_remove(dev->dev, &qp->mqp);
1419 	mlx4_qp_free(dev->dev, &qp->mqp);
1420 	mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1421 	del_gid_entries(qp);
1422 	kfree(qp->rss_ctx);
1423 }
1424 
1425 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1426 			      enum mlx4_ib_source_type src, int is_user)
1427 {
1428 	struct mlx4_ib_cq *send_cq, *recv_cq;
1429 	unsigned long flags;
1430 
1431 	if (qp->state != IB_QPS_RESET) {
1432 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1433 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1434 			pr_warn("modify QP %06x to RESET failed.\n",
1435 			       qp->mqp.qpn);
1436 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1437 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1438 			qp->pri.smac = 0;
1439 			qp->pri.smac_port = 0;
1440 		}
1441 		if (qp->alt.smac) {
1442 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1443 			qp->alt.smac = 0;
1444 		}
1445 		if (qp->pri.vid < 0x1000) {
1446 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1447 			qp->pri.vid = 0xFFFF;
1448 			qp->pri.candidate_vid = 0xFFFF;
1449 			qp->pri.update_vid = 0;
1450 		}
1451 		if (qp->alt.vid < 0x1000) {
1452 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1453 			qp->alt.vid = 0xFFFF;
1454 			qp->alt.candidate_vid = 0xFFFF;
1455 			qp->alt.update_vid = 0;
1456 		}
1457 	}
1458 
1459 	get_cqs(qp, src, &send_cq, &recv_cq);
1460 
1461 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1462 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1463 
1464 	/* del from lists under both locks above to protect reset flow paths */
1465 	list_del(&qp->qps_list);
1466 	list_del(&qp->cq_send_list);
1467 	list_del(&qp->cq_recv_list);
1468 	if (!is_user) {
1469 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1470 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1471 		if (send_cq != recv_cq)
1472 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1473 	}
1474 
1475 	mlx4_qp_remove(dev->dev, &qp->mqp);
1476 
1477 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1478 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1479 
1480 	mlx4_qp_free(dev->dev, &qp->mqp);
1481 
1482 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1483 		if (qp->flags & MLX4_IB_QP_NETIF)
1484 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1485 		else if (src == MLX4_IB_RWQ_SRC)
1486 			mlx4_ib_release_wqn(to_mucontext(
1487 					    qp->ibwq.uobject->context), qp, 1);
1488 		else
1489 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1490 	}
1491 
1492 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1493 
1494 	if (is_user) {
1495 		if (qp->rq.wqe_cnt) {
1496 			struct mlx4_ib_ucontext *mcontext = !src ?
1497 				to_mucontext(qp->ibqp.uobject->context) :
1498 				to_mucontext(qp->ibwq.uobject->context);
1499 			mlx4_ib_db_unmap_user(mcontext, &qp->db);
1500 		}
1501 		ib_umem_release(qp->umem);
1502 	} else {
1503 		kvfree(qp->sq.wrid);
1504 		kvfree(qp->rq.wrid);
1505 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1506 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1507 			free_proxy_bufs(&dev->ib_dev, qp);
1508 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1509 		if (qp->rq.wqe_cnt)
1510 			mlx4_db_free(dev->dev, &qp->db);
1511 	}
1512 
1513 	del_gid_entries(qp);
1514 }
1515 
1516 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1517 {
1518 	/* Native or PPF */
1519 	if (!mlx4_is_mfunc(dev->dev) ||
1520 	    (mlx4_is_master(dev->dev) &&
1521 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1522 		return  dev->dev->phys_caps.base_sqpn +
1523 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1524 			attr->port_num - 1;
1525 	}
1526 	/* PF or VF -- creating proxies */
1527 	if (attr->qp_type == IB_QPT_SMI)
1528 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp0_proxy;
1529 	else
1530 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp1_proxy;
1531 }
1532 
1533 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1534 					struct ib_qp_init_attr *init_attr,
1535 					struct ib_udata *udata)
1536 {
1537 	struct mlx4_ib_qp *qp = NULL;
1538 	int err;
1539 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1540 	u16 xrcdn = 0;
1541 
1542 	if (init_attr->rwq_ind_tbl)
1543 		return _mlx4_ib_create_qp_rss(pd, init_attr, udata);
1544 
1545 	/*
1546 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1547 	 * and only for kernel UD QPs.
1548 	 */
1549 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1550 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1551 					MLX4_IB_SRIOV_TUNNEL_QP |
1552 					MLX4_IB_SRIOV_SQP |
1553 					MLX4_IB_QP_NETIF |
1554 					MLX4_IB_QP_CREATE_ROCE_V2_GSI))
1555 		return ERR_PTR(-EINVAL);
1556 
1557 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1558 		if (init_attr->qp_type != IB_QPT_UD)
1559 			return ERR_PTR(-EINVAL);
1560 	}
1561 
1562 	if (init_attr->create_flags) {
1563 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1564 			return ERR_PTR(-EINVAL);
1565 
1566 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1567 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1568 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1569 		     init_attr->qp_type != IB_QPT_UD) ||
1570 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1571 		     init_attr->qp_type > IB_QPT_GSI) ||
1572 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1573 		     init_attr->qp_type != IB_QPT_GSI))
1574 			return ERR_PTR(-EINVAL);
1575 	}
1576 
1577 	switch (init_attr->qp_type) {
1578 	case IB_QPT_XRC_TGT:
1579 		pd = to_mxrcd(init_attr->xrcd)->pd;
1580 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1581 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1582 		/* fall through */
1583 	case IB_QPT_XRC_INI:
1584 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1585 			return ERR_PTR(-ENOSYS);
1586 		init_attr->recv_cq = init_attr->send_cq;
1587 		/* fall through */
1588 	case IB_QPT_RC:
1589 	case IB_QPT_UC:
1590 	case IB_QPT_RAW_PACKET:
1591 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1592 		if (!qp)
1593 			return ERR_PTR(-ENOMEM);
1594 		qp->pri.vid = 0xFFFF;
1595 		qp->alt.vid = 0xFFFF;
1596 		/* fall through */
1597 	case IB_QPT_UD:
1598 	{
1599 		err = create_qp_common(to_mdev(pd->device), pd,	MLX4_IB_QP_SRC,
1600 				       init_attr, udata, 0, &qp);
1601 		if (err) {
1602 			kfree(qp);
1603 			return ERR_PTR(err);
1604 		}
1605 
1606 		qp->ibqp.qp_num = qp->mqp.qpn;
1607 		qp->xrcdn = xrcdn;
1608 
1609 		break;
1610 	}
1611 	case IB_QPT_SMI:
1612 	case IB_QPT_GSI:
1613 	{
1614 		int sqpn;
1615 
1616 		/* Userspace is not allowed to create special QPs: */
1617 		if (udata)
1618 			return ERR_PTR(-EINVAL);
1619 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1620 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev,
1621 							1, 1, &sqpn, 0,
1622 							MLX4_RES_USAGE_DRIVER);
1623 
1624 			if (res)
1625 				return ERR_PTR(res);
1626 		} else {
1627 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1628 		}
1629 
1630 		err = create_qp_common(to_mdev(pd->device), pd, MLX4_IB_QP_SRC,
1631 				       init_attr, udata, sqpn, &qp);
1632 		if (err)
1633 			return ERR_PTR(err);
1634 
1635 		qp->port	= init_attr->port_num;
1636 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1637 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1638 		break;
1639 	}
1640 	default:
1641 		/* Don't support raw QPs */
1642 		return ERR_PTR(-EINVAL);
1643 	}
1644 
1645 	return &qp->ibqp;
1646 }
1647 
1648 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1649 				struct ib_qp_init_attr *init_attr,
1650 				struct ib_udata *udata) {
1651 	struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1652 	struct ib_qp *ibqp;
1653 	struct mlx4_ib_dev *dev = to_mdev(device);
1654 
1655 	ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1656 
1657 	if (!IS_ERR(ibqp) &&
1658 	    (init_attr->qp_type == IB_QPT_GSI) &&
1659 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1660 		struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1661 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1662 
1663 		if (is_eth &&
1664 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1665 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1666 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1667 
1668 			if (IS_ERR(sqp->roce_v2_gsi)) {
1669 				pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1670 				sqp->roce_v2_gsi = NULL;
1671 			} else {
1672 				sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1673 				sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1674 			}
1675 
1676 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1677 		}
1678 	}
1679 	return ibqp;
1680 }
1681 
1682 static int _mlx4_ib_destroy_qp(struct ib_qp *qp)
1683 {
1684 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1685 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1686 
1687 	if (is_qp0(dev, mqp))
1688 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1689 
1690 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1691 	    dev->qp1_proxy[mqp->port - 1] == mqp) {
1692 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1693 		dev->qp1_proxy[mqp->port - 1] = NULL;
1694 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1695 	}
1696 
1697 	if (mqp->counter_index)
1698 		mlx4_ib_free_qp_counter(dev, mqp);
1699 
1700 	if (qp->rwq_ind_tbl) {
1701 		destroy_qp_rss(dev, mqp);
1702 	} else {
1703 		struct mlx4_ib_pd *pd;
1704 
1705 		pd = get_pd(mqp);
1706 		destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, !!pd->ibpd.uobject);
1707 	}
1708 
1709 	if (is_sqp(dev, mqp))
1710 		kfree(to_msqp(mqp));
1711 	else
1712 		kfree(mqp);
1713 
1714 	return 0;
1715 }
1716 
1717 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1718 {
1719 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1720 
1721 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1722 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1723 
1724 		if (sqp->roce_v2_gsi)
1725 			ib_destroy_qp(sqp->roce_v2_gsi);
1726 	}
1727 
1728 	return _mlx4_ib_destroy_qp(qp);
1729 }
1730 
1731 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1732 {
1733 	switch (type) {
1734 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1735 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1736 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1737 	case MLX4_IB_QPT_XRC_INI:
1738 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1739 	case MLX4_IB_QPT_SMI:
1740 	case MLX4_IB_QPT_GSI:
1741 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1742 
1743 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1744 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1745 						MLX4_QP_ST_MLX : -1);
1746 	case MLX4_IB_QPT_PROXY_SMI:
1747 	case MLX4_IB_QPT_TUN_SMI:
1748 	case MLX4_IB_QPT_PROXY_GSI:
1749 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1750 						MLX4_QP_ST_UD : -1);
1751 	default:			return -1;
1752 	}
1753 }
1754 
1755 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1756 				   int attr_mask)
1757 {
1758 	u8 dest_rd_atomic;
1759 	u32 access_flags;
1760 	u32 hw_access_flags = 0;
1761 
1762 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1763 		dest_rd_atomic = attr->max_dest_rd_atomic;
1764 	else
1765 		dest_rd_atomic = qp->resp_depth;
1766 
1767 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1768 		access_flags = attr->qp_access_flags;
1769 	else
1770 		access_flags = qp->atomic_rd_en;
1771 
1772 	if (!dest_rd_atomic)
1773 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1774 
1775 	if (access_flags & IB_ACCESS_REMOTE_READ)
1776 		hw_access_flags |= MLX4_QP_BIT_RRE;
1777 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1778 		hw_access_flags |= MLX4_QP_BIT_RAE;
1779 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1780 		hw_access_flags |= MLX4_QP_BIT_RWE;
1781 
1782 	return cpu_to_be32(hw_access_flags);
1783 }
1784 
1785 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1786 			    int attr_mask)
1787 {
1788 	if (attr_mask & IB_QP_PKEY_INDEX)
1789 		sqp->pkey_index = attr->pkey_index;
1790 	if (attr_mask & IB_QP_QKEY)
1791 		sqp->qkey = attr->qkey;
1792 	if (attr_mask & IB_QP_SQ_PSN)
1793 		sqp->send_psn = attr->sq_psn;
1794 }
1795 
1796 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1797 {
1798 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1799 }
1800 
1801 static int _mlx4_set_path(struct mlx4_ib_dev *dev,
1802 			  const struct rdma_ah_attr *ah,
1803 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1804 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1805 {
1806 	int vidx;
1807 	int smac_index;
1808 	int err;
1809 
1810 	path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
1811 	path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
1812 	if (rdma_ah_get_static_rate(ah)) {
1813 		path->static_rate = rdma_ah_get_static_rate(ah) +
1814 				    MLX4_STAT_RATE_OFFSET;
1815 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1816 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1817 			--path->static_rate;
1818 	} else
1819 		path->static_rate = 0;
1820 
1821 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
1822 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
1823 		int real_sgid_index =
1824 			mlx4_ib_gid_index_to_real_index(dev, port,
1825 							grh->sgid_index);
1826 
1827 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1828 			pr_err("sgid_index (%u) too large. max is %d\n",
1829 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1830 			return -1;
1831 		}
1832 
1833 		path->grh_mylmc |= 1 << 7;
1834 		path->mgid_index = real_sgid_index;
1835 		path->hop_limit  = grh->hop_limit;
1836 		path->tclass_flowlabel =
1837 			cpu_to_be32((grh->traffic_class << 20) |
1838 				    (grh->flow_label));
1839 		memcpy(path->rgid, grh->dgid.raw, 16);
1840 	}
1841 
1842 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
1843 		if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
1844 			return -1;
1845 
1846 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1847 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
1848 
1849 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1850 		if (vlan_tag < 0x1000) {
1851 			if (smac_info->vid < 0x1000) {
1852 				/* both valid vlan ids */
1853 				if (smac_info->vid != vlan_tag) {
1854 					/* different VIDs.  unreg old and reg new */
1855 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1856 					if (err)
1857 						return err;
1858 					smac_info->candidate_vid = vlan_tag;
1859 					smac_info->candidate_vlan_index = vidx;
1860 					smac_info->candidate_vlan_port = port;
1861 					smac_info->update_vid = 1;
1862 					path->vlan_index = vidx;
1863 				} else {
1864 					path->vlan_index = smac_info->vlan_index;
1865 				}
1866 			} else {
1867 				/* no current vlan tag in qp */
1868 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1869 				if (err)
1870 					return err;
1871 				smac_info->candidate_vid = vlan_tag;
1872 				smac_info->candidate_vlan_index = vidx;
1873 				smac_info->candidate_vlan_port = port;
1874 				smac_info->update_vid = 1;
1875 				path->vlan_index = vidx;
1876 			}
1877 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1878 			path->fl = 1 << 6;
1879 		} else {
1880 			/* have current vlan tag. unregister it at modify-qp success */
1881 			if (smac_info->vid < 0x1000) {
1882 				smac_info->candidate_vid = 0xFFFF;
1883 				smac_info->update_vid = 1;
1884 			}
1885 		}
1886 
1887 		/* get smac_index for RoCE use.
1888 		 * If no smac was yet assigned, register one.
1889 		 * If one was already assigned, but the new mac differs,
1890 		 * unregister the old one and register the new one.
1891 		*/
1892 		if ((!smac_info->smac && !smac_info->smac_port) ||
1893 		    smac_info->smac != smac) {
1894 			/* register candidate now, unreg if needed, after success */
1895 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1896 			if (smac_index >= 0) {
1897 				smac_info->candidate_smac_index = smac_index;
1898 				smac_info->candidate_smac = smac;
1899 				smac_info->candidate_smac_port = port;
1900 			} else {
1901 				return -EINVAL;
1902 			}
1903 		} else {
1904 			smac_index = smac_info->smac_index;
1905 		}
1906 		memcpy(path->dmac, ah->roce.dmac, 6);
1907 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1908 		/* put MAC table smac index for IBoE */
1909 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1910 	} else {
1911 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1912 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
1913 	}
1914 
1915 	return 0;
1916 }
1917 
1918 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1919 			 enum ib_qp_attr_mask qp_attr_mask,
1920 			 struct mlx4_ib_qp *mqp,
1921 			 struct mlx4_qp_path *path, u8 port,
1922 			 u16 vlan_id, u8 *smac)
1923 {
1924 	return _mlx4_set_path(dev, &qp->ah_attr,
1925 			      mlx4_mac_to_u64(smac),
1926 			      vlan_id,
1927 			      path, &mqp->pri, port);
1928 }
1929 
1930 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1931 			     const struct ib_qp_attr *qp,
1932 			     enum ib_qp_attr_mask qp_attr_mask,
1933 			     struct mlx4_ib_qp *mqp,
1934 			     struct mlx4_qp_path *path, u8 port)
1935 {
1936 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1937 			      0,
1938 			      0xffff,
1939 			      path, &mqp->alt, port);
1940 }
1941 
1942 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1943 {
1944 	struct mlx4_ib_gid_entry *ge, *tmp;
1945 
1946 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1947 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1948 			ge->added = 1;
1949 			ge->port = qp->port;
1950 		}
1951 	}
1952 }
1953 
1954 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1955 				    struct mlx4_ib_qp *qp,
1956 				    struct mlx4_qp_context *context)
1957 {
1958 	u64 u64_mac;
1959 	int smac_index;
1960 
1961 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1962 
1963 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1964 	if (!qp->pri.smac && !qp->pri.smac_port) {
1965 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1966 		if (smac_index >= 0) {
1967 			qp->pri.candidate_smac_index = smac_index;
1968 			qp->pri.candidate_smac = u64_mac;
1969 			qp->pri.candidate_smac_port = qp->port;
1970 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1971 		} else {
1972 			return -ENOENT;
1973 		}
1974 	}
1975 	return 0;
1976 }
1977 
1978 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1979 {
1980 	struct counter_index *new_counter_index;
1981 	int err;
1982 	u32 tmp_idx;
1983 
1984 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1985 	    IB_LINK_LAYER_ETHERNET ||
1986 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1987 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1988 		return 0;
1989 
1990 	err = mlx4_counter_alloc(dev->dev, &tmp_idx, MLX4_RES_USAGE_DRIVER);
1991 	if (err)
1992 		return err;
1993 
1994 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1995 	if (!new_counter_index) {
1996 		mlx4_counter_free(dev->dev, tmp_idx);
1997 		return -ENOMEM;
1998 	}
1999 
2000 	new_counter_index->index = tmp_idx;
2001 	new_counter_index->allocated = 1;
2002 	qp->counter_index = new_counter_index;
2003 
2004 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
2005 	list_add_tail(&new_counter_index->list,
2006 		      &dev->counters_table[qp->port - 1].counters_list);
2007 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
2008 
2009 	return 0;
2010 }
2011 
2012 enum {
2013 	MLX4_QPC_ROCE_MODE_1 = 0,
2014 	MLX4_QPC_ROCE_MODE_2 = 2,
2015 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
2016 };
2017 
2018 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
2019 {
2020 	switch (gid_type) {
2021 	case IB_GID_TYPE_ROCE:
2022 		return MLX4_QPC_ROCE_MODE_1;
2023 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
2024 		return MLX4_QPC_ROCE_MODE_2;
2025 	default:
2026 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
2027 	}
2028 }
2029 
2030 /*
2031  * Go over all RSS QP's childes (WQs) and apply their HW state according to
2032  * their logic state if the RSS QP is the first RSS QP associated for the WQ.
2033  */
2034 static int bringup_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, u8 port_num)
2035 {
2036 	int err = 0;
2037 	int i;
2038 
2039 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2040 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2041 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2042 
2043 		mutex_lock(&wq->mutex);
2044 
2045 		/* Mlx4_ib restrictions:
2046 		 * WQ's is associated to a port according to the RSS QP it is
2047 		 * associates to.
2048 		 * In case the WQ is associated to a different port by another
2049 		 * RSS QP, return a failure.
2050 		 */
2051 		if ((wq->rss_usecnt > 0) && (wq->port != port_num)) {
2052 			err = -EINVAL;
2053 			mutex_unlock(&wq->mutex);
2054 			break;
2055 		}
2056 		wq->port = port_num;
2057 		if ((wq->rss_usecnt == 0) && (ibwq->state == IB_WQS_RDY)) {
2058 			err = _mlx4_ib_modify_wq(ibwq, IB_WQS_RDY);
2059 			if (err) {
2060 				mutex_unlock(&wq->mutex);
2061 				break;
2062 			}
2063 		}
2064 		wq->rss_usecnt++;
2065 
2066 		mutex_unlock(&wq->mutex);
2067 	}
2068 
2069 	if (i && err) {
2070 		int j;
2071 
2072 		for (j = (i - 1); j >= 0; j--) {
2073 			struct ib_wq *ibwq = ind_tbl->ind_tbl[j];
2074 			struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2075 
2076 			mutex_lock(&wq->mutex);
2077 
2078 			if ((wq->rss_usecnt == 1) &&
2079 			    (ibwq->state == IB_WQS_RDY))
2080 				if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET))
2081 					pr_warn("failed to reverse WQN=0x%06x\n",
2082 						ibwq->wq_num);
2083 			wq->rss_usecnt--;
2084 
2085 			mutex_unlock(&wq->mutex);
2086 		}
2087 	}
2088 
2089 	return err;
2090 }
2091 
2092 static void bring_down_rss_rwqs(struct ib_rwq_ind_table *ind_tbl)
2093 {
2094 	int i;
2095 
2096 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2097 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2098 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2099 
2100 		mutex_lock(&wq->mutex);
2101 
2102 		if ((wq->rss_usecnt == 1) && (ibwq->state == IB_WQS_RDY))
2103 			if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET))
2104 				pr_warn("failed to reverse WQN=%x\n",
2105 					ibwq->wq_num);
2106 		wq->rss_usecnt--;
2107 
2108 		mutex_unlock(&wq->mutex);
2109 	}
2110 }
2111 
2112 static void fill_qp_rss_context(struct mlx4_qp_context *context,
2113 				struct mlx4_ib_qp *qp)
2114 {
2115 	struct mlx4_rss_context *rss_context;
2116 
2117 	rss_context = (void *)context + offsetof(struct mlx4_qp_context,
2118 			pri_path) + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
2119 
2120 	rss_context->base_qpn = cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz);
2121 	rss_context->default_qpn =
2122 		cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz & 0xffffff);
2123 	if (qp->rss_ctx->flags & (MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6))
2124 		rss_context->base_qpn_udp = rss_context->default_qpn;
2125 	rss_context->flags = qp->rss_ctx->flags;
2126 	/* Currently support just toeplitz */
2127 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
2128 
2129 	memcpy(rss_context->rss_key, qp->rss_ctx->rss_key,
2130 	       MLX4_EN_RSS_KEY_SIZE);
2131 }
2132 
2133 static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type,
2134 			       const struct ib_qp_attr *attr, int attr_mask,
2135 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
2136 {
2137 	struct ib_uobject *ibuobject;
2138 	struct ib_srq  *ibsrq;
2139 	struct ib_rwq_ind_table *rwq_ind_tbl;
2140 	enum ib_qp_type qp_type;
2141 	struct mlx4_ib_dev *dev;
2142 	struct mlx4_ib_qp *qp;
2143 	struct mlx4_ib_pd *pd;
2144 	struct mlx4_ib_cq *send_cq, *recv_cq;
2145 	struct mlx4_qp_context *context;
2146 	enum mlx4_qp_optpar optpar = 0;
2147 	int sqd_event;
2148 	int steer_qp = 0;
2149 	int err = -EINVAL;
2150 	int counter_index;
2151 
2152 	if (src_type == MLX4_IB_RWQ_SRC) {
2153 		struct ib_wq *ibwq;
2154 
2155 		ibwq	    = (struct ib_wq *)src;
2156 		ibuobject   = ibwq->uobject;
2157 		ibsrq	    = NULL;
2158 		rwq_ind_tbl = NULL;
2159 		qp_type     = IB_QPT_RAW_PACKET;
2160 		qp	    = to_mqp((struct ib_qp *)ibwq);
2161 		dev	    = to_mdev(ibwq->device);
2162 		pd	    = to_mpd(ibwq->pd);
2163 	} else {
2164 		struct ib_qp *ibqp;
2165 
2166 		ibqp	    = (struct ib_qp *)src;
2167 		ibuobject   = ibqp->uobject;
2168 		ibsrq	    = ibqp->srq;
2169 		rwq_ind_tbl = ibqp->rwq_ind_tbl;
2170 		qp_type     = ibqp->qp_type;
2171 		qp	    = to_mqp(ibqp);
2172 		dev	    = to_mdev(ibqp->device);
2173 		pd	    = get_pd(qp);
2174 	}
2175 
2176 	/* APM is not supported under RoCE */
2177 	if (attr_mask & IB_QP_ALT_PATH &&
2178 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2179 	    IB_LINK_LAYER_ETHERNET)
2180 		return -ENOTSUPP;
2181 
2182 	context = kzalloc(sizeof *context, GFP_KERNEL);
2183 	if (!context)
2184 		return -ENOMEM;
2185 
2186 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
2187 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
2188 
2189 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
2190 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2191 	else {
2192 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
2193 		switch (attr->path_mig_state) {
2194 		case IB_MIG_MIGRATED:
2195 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2196 			break;
2197 		case IB_MIG_REARM:
2198 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
2199 			break;
2200 		case IB_MIG_ARMED:
2201 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
2202 			break;
2203 		}
2204 	}
2205 
2206 	if (qp->inl_recv_sz)
2207 		context->param3 |= cpu_to_be32(1 << 25);
2208 
2209 	if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI)
2210 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
2211 	else if (qp_type == IB_QPT_RAW_PACKET)
2212 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
2213 	else if (qp_type == IB_QPT_UD) {
2214 		if (qp->flags & MLX4_IB_QP_LSO)
2215 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
2216 					      ilog2(dev->dev->caps.max_gso_sz);
2217 		else
2218 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 13;
2219 	} else if (attr_mask & IB_QP_PATH_MTU) {
2220 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
2221 			pr_err("path MTU (%u) is invalid\n",
2222 			       attr->path_mtu);
2223 			goto out;
2224 		}
2225 		context->mtu_msgmax = (attr->path_mtu << 5) |
2226 			ilog2(dev->dev->caps.max_msg_sz);
2227 	}
2228 
2229 	if (!rwq_ind_tbl) { /* PRM RSS receive side should be left zeros */
2230 		if (qp->rq.wqe_cnt)
2231 			context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
2232 		context->rq_size_stride |= qp->rq.wqe_shift - 4;
2233 	}
2234 
2235 	if (qp->sq.wqe_cnt)
2236 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
2237 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
2238 
2239 	if (new_state == IB_QPS_RESET && qp->counter_index)
2240 		mlx4_ib_free_qp_counter(dev, qp);
2241 
2242 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2243 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
2244 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
2245 		if (qp_type == IB_QPT_RAW_PACKET)
2246 			context->param3 |= cpu_to_be32(1 << 30);
2247 	}
2248 
2249 	if (ibuobject)
2250 		context->usr_page = cpu_to_be32(
2251 			mlx4_to_hw_uar_index(dev->dev,
2252 					     to_mucontext(ibuobject->context)
2253 					     ->uar.index));
2254 	else
2255 		context->usr_page = cpu_to_be32(
2256 			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
2257 
2258 	if (attr_mask & IB_QP_DEST_QPN)
2259 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
2260 
2261 	if (attr_mask & IB_QP_PORT) {
2262 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
2263 		    !(attr_mask & IB_QP_AV)) {
2264 			mlx4_set_sched(&context->pri_path, attr->port_num);
2265 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
2266 		}
2267 	}
2268 
2269 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
2270 		err = create_qp_lb_counter(dev, qp);
2271 		if (err)
2272 			goto out;
2273 
2274 		counter_index =
2275 			dev->counters_table[qp->port - 1].default_counter;
2276 		if (qp->counter_index)
2277 			counter_index = qp->counter_index->index;
2278 
2279 		if (counter_index != -1) {
2280 			context->pri_path.counter_index = counter_index;
2281 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
2282 			if (qp->counter_index) {
2283 				context->pri_path.fl |=
2284 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
2285 				context->pri_path.vlan_control |=
2286 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
2287 			}
2288 		} else
2289 			context->pri_path.counter_index =
2290 				MLX4_SINK_COUNTER_INDEX(dev->dev);
2291 
2292 		if (qp->flags & MLX4_IB_QP_NETIF) {
2293 			mlx4_ib_steer_qp_reg(dev, qp, 1);
2294 			steer_qp = 1;
2295 		}
2296 
2297 		if (qp_type == IB_QPT_GSI) {
2298 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
2299 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
2300 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
2301 
2302 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2303 		}
2304 	}
2305 
2306 	if (attr_mask & IB_QP_PKEY_INDEX) {
2307 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2308 			context->pri_path.disable_pkey_check = 0x40;
2309 		context->pri_path.pkey_index = attr->pkey_index;
2310 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
2311 	}
2312 
2313 	if (attr_mask & IB_QP_AV) {
2314 		u8 port_num = mlx4_is_bonded(dev->dev) ? 1 :
2315 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2316 		union ib_gid gid;
2317 		struct ib_gid_attr gid_attr = {.gid_type = IB_GID_TYPE_IB};
2318 		u16 vlan = 0xffff;
2319 		u8 smac[ETH_ALEN];
2320 		int status = 0;
2321 		int is_eth =
2322 			rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
2323 			rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
2324 
2325 		if (is_eth) {
2326 			int index =
2327 				rdma_ah_read_grh(&attr->ah_attr)->sgid_index;
2328 
2329 			status = ib_get_cached_gid(&dev->ib_dev, port_num,
2330 						   index, &gid, &gid_attr);
2331 			if (!status && !memcmp(&gid, &zgid, sizeof(gid)))
2332 				status = -ENOENT;
2333 			if (!status && gid_attr.ndev) {
2334 				vlan = rdma_vlan_dev_vlan_id(gid_attr.ndev);
2335 				memcpy(smac, gid_attr.ndev->dev_addr, ETH_ALEN);
2336 				dev_put(gid_attr.ndev);
2337 			}
2338 		}
2339 		if (status)
2340 			goto out;
2341 
2342 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
2343 				  port_num, vlan, smac))
2344 			goto out;
2345 
2346 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
2347 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
2348 
2349 		if (is_eth &&
2350 		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
2351 			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr.gid_type);
2352 
2353 			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
2354 				err = -EINVAL;
2355 				goto out;
2356 			}
2357 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2358 		}
2359 
2360 	}
2361 
2362 	if (attr_mask & IB_QP_TIMEOUT) {
2363 		context->pri_path.ackto |= attr->timeout << 3;
2364 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
2365 	}
2366 
2367 	if (attr_mask & IB_QP_ALT_PATH) {
2368 		if (attr->alt_port_num == 0 ||
2369 		    attr->alt_port_num > dev->dev->caps.num_ports)
2370 			goto out;
2371 
2372 		if (attr->alt_pkey_index >=
2373 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
2374 			goto out;
2375 
2376 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
2377 				      &context->alt_path,
2378 				      attr->alt_port_num))
2379 			goto out;
2380 
2381 		context->alt_path.pkey_index = attr->alt_pkey_index;
2382 		context->alt_path.ackto = attr->alt_timeout << 3;
2383 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
2384 	}
2385 
2386 	context->pd = cpu_to_be32(pd->pdn);
2387 
2388 	if (!rwq_ind_tbl) {
2389 		context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
2390 		get_cqs(qp, src_type, &send_cq, &recv_cq);
2391 	} else { /* Set dummy CQs to be compatible with HV and PRM */
2392 		send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq);
2393 		recv_cq = send_cq;
2394 	}
2395 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
2396 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
2397 
2398 	/* Set "fast registration enabled" for all kernel QPs */
2399 	if (!ibuobject)
2400 		context->params1 |= cpu_to_be32(1 << 11);
2401 
2402 	if (attr_mask & IB_QP_RNR_RETRY) {
2403 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2404 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
2405 	}
2406 
2407 	if (attr_mask & IB_QP_RETRY_CNT) {
2408 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2409 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
2410 	}
2411 
2412 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2413 		if (attr->max_rd_atomic)
2414 			context->params1 |=
2415 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2416 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
2417 	}
2418 
2419 	if (attr_mask & IB_QP_SQ_PSN)
2420 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2421 
2422 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2423 		if (attr->max_dest_rd_atomic)
2424 			context->params2 |=
2425 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2426 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
2427 	}
2428 
2429 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
2430 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
2431 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
2432 	}
2433 
2434 	if (ibsrq)
2435 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
2436 
2437 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
2438 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2439 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
2440 	}
2441 	if (attr_mask & IB_QP_RQ_PSN)
2442 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2443 
2444 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
2445 	if (attr_mask & IB_QP_QKEY) {
2446 		if (qp->mlx4_ib_qp_type &
2447 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
2448 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2449 		else {
2450 			if (mlx4_is_mfunc(dev->dev) &&
2451 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
2452 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
2453 			    MLX4_RESERVED_QKEY_BASE) {
2454 				pr_err("Cannot use reserved QKEY"
2455 				       " 0x%x (range 0xffff0000..0xffffffff"
2456 				       " is reserved)\n", attr->qkey);
2457 				err = -EINVAL;
2458 				goto out;
2459 			}
2460 			context->qkey = cpu_to_be32(attr->qkey);
2461 		}
2462 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
2463 	}
2464 
2465 	if (ibsrq)
2466 		context->srqn = cpu_to_be32(1 << 24 |
2467 					    to_msrq(ibsrq)->msrq.srqn);
2468 
2469 	if (qp->rq.wqe_cnt &&
2470 	    cur_state == IB_QPS_RESET &&
2471 	    new_state == IB_QPS_INIT)
2472 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2473 
2474 	if (cur_state == IB_QPS_INIT &&
2475 	    new_state == IB_QPS_RTR  &&
2476 	    (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI ||
2477 	     qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) {
2478 		context->pri_path.sched_queue = (qp->port - 1) << 6;
2479 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2480 		    qp->mlx4_ib_qp_type &
2481 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
2482 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
2483 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
2484 				context->pri_path.fl = 0x80;
2485 		} else {
2486 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2487 				context->pri_path.fl = 0x80;
2488 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
2489 		}
2490 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2491 		    IB_LINK_LAYER_ETHERNET) {
2492 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
2493 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
2494 				context->pri_path.feup = 1 << 7; /* don't fsm */
2495 			/* handle smac_index */
2496 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
2497 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
2498 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
2499 				err = handle_eth_ud_smac_index(dev, qp, context);
2500 				if (err) {
2501 					err = -EINVAL;
2502 					goto out;
2503 				}
2504 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
2505 					dev->qp1_proxy[qp->port - 1] = qp;
2506 			}
2507 		}
2508 	}
2509 
2510 	if (qp_type == IB_QPT_RAW_PACKET) {
2511 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
2512 					MLX4_IB_LINK_TYPE_ETH;
2513 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
2514 			/* set QP to receive both tunneled & non-tunneled packets */
2515 			if (!rwq_ind_tbl)
2516 				context->srqn = cpu_to_be32(7 << 28);
2517 		}
2518 	}
2519 
2520 	if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
2521 		int is_eth = rdma_port_get_link_layer(
2522 				&dev->ib_dev, qp->port) ==
2523 				IB_LINK_LAYER_ETHERNET;
2524 		if (is_eth) {
2525 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
2526 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
2527 		}
2528 	}
2529 
2530 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2531 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2532 		sqd_event = 1;
2533 	else
2534 		sqd_event = 0;
2535 
2536 	if (!ibuobject &&
2537 	    cur_state == IB_QPS_RESET &&
2538 	    new_state == IB_QPS_INIT)
2539 		context->rlkey_roce_mode |= (1 << 4);
2540 
2541 	/*
2542 	 * Before passing a kernel QP to the HW, make sure that the
2543 	 * ownership bits of the send queue are set and the SQ
2544 	 * headroom is stamped so that the hardware doesn't start
2545 	 * processing stale work requests.
2546 	 */
2547 	if (!ibuobject &&
2548 	    cur_state == IB_QPS_RESET &&
2549 	    new_state == IB_QPS_INIT) {
2550 		struct mlx4_wqe_ctrl_seg *ctrl;
2551 		int i;
2552 
2553 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
2554 			ctrl = get_send_wqe(qp, i);
2555 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
2556 			if (qp->sq_max_wqes_per_wr == 1)
2557 				ctrl->qpn_vlan.fence_size =
2558 						1 << (qp->sq.wqe_shift - 4);
2559 
2560 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
2561 		}
2562 	}
2563 
2564 	if (rwq_ind_tbl	&&
2565 	    cur_state == IB_QPS_RESET &&
2566 	    new_state == IB_QPS_INIT) {
2567 		fill_qp_rss_context(context, qp);
2568 		context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET);
2569 	}
2570 
2571 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2572 			     to_mlx4_state(new_state), context, optpar,
2573 			     sqd_event, &qp->mqp);
2574 	if (err)
2575 		goto out;
2576 
2577 	qp->state = new_state;
2578 
2579 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2580 		qp->atomic_rd_en = attr->qp_access_flags;
2581 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2582 		qp->resp_depth = attr->max_dest_rd_atomic;
2583 	if (attr_mask & IB_QP_PORT) {
2584 		qp->port = attr->port_num;
2585 		update_mcg_macs(dev, qp);
2586 	}
2587 	if (attr_mask & IB_QP_ALT_PATH)
2588 		qp->alt_port = attr->alt_port_num;
2589 
2590 	if (is_sqp(dev, qp))
2591 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2592 
2593 	/*
2594 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
2595 	 * QP0 to RESET or ERROR, bring the link back down.
2596 	 */
2597 	if (is_qp0(dev, qp)) {
2598 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2599 			if (mlx4_INIT_PORT(dev->dev, qp->port))
2600 				pr_warn("INIT_PORT failed for port %d\n",
2601 				       qp->port);
2602 
2603 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2604 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2605 			mlx4_CLOSE_PORT(dev->dev, qp->port);
2606 	}
2607 
2608 	/*
2609 	 * If we moved a kernel QP to RESET, clean up all old CQ
2610 	 * entries and reinitialize the QP.
2611 	 */
2612 	if (new_state == IB_QPS_RESET) {
2613 		if (!ibuobject) {
2614 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2615 					 ibsrq ? to_msrq(ibsrq) : NULL);
2616 			if (send_cq != recv_cq)
2617 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2618 
2619 			qp->rq.head = 0;
2620 			qp->rq.tail = 0;
2621 			qp->sq.head = 0;
2622 			qp->sq.tail = 0;
2623 			qp->sq_next_wqe = 0;
2624 			if (qp->rq.wqe_cnt)
2625 				*qp->db.db  = 0;
2626 
2627 			if (qp->flags & MLX4_IB_QP_NETIF)
2628 				mlx4_ib_steer_qp_reg(dev, qp, 0);
2629 		}
2630 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2631 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2632 			qp->pri.smac = 0;
2633 			qp->pri.smac_port = 0;
2634 		}
2635 		if (qp->alt.smac) {
2636 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2637 			qp->alt.smac = 0;
2638 		}
2639 		if (qp->pri.vid < 0x1000) {
2640 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2641 			qp->pri.vid = 0xFFFF;
2642 			qp->pri.candidate_vid = 0xFFFF;
2643 			qp->pri.update_vid = 0;
2644 		}
2645 
2646 		if (qp->alt.vid < 0x1000) {
2647 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2648 			qp->alt.vid = 0xFFFF;
2649 			qp->alt.candidate_vid = 0xFFFF;
2650 			qp->alt.update_vid = 0;
2651 		}
2652 	}
2653 out:
2654 	if (err && qp->counter_index)
2655 		mlx4_ib_free_qp_counter(dev, qp);
2656 	if (err && steer_qp)
2657 		mlx4_ib_steer_qp_reg(dev, qp, 0);
2658 	kfree(context);
2659 	if (qp->pri.candidate_smac ||
2660 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2661 		if (err) {
2662 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2663 		} else {
2664 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2665 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2666 			qp->pri.smac = qp->pri.candidate_smac;
2667 			qp->pri.smac_index = qp->pri.candidate_smac_index;
2668 			qp->pri.smac_port = qp->pri.candidate_smac_port;
2669 		}
2670 		qp->pri.candidate_smac = 0;
2671 		qp->pri.candidate_smac_index = 0;
2672 		qp->pri.candidate_smac_port = 0;
2673 	}
2674 	if (qp->alt.candidate_smac) {
2675 		if (err) {
2676 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2677 		} else {
2678 			if (qp->alt.smac)
2679 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2680 			qp->alt.smac = qp->alt.candidate_smac;
2681 			qp->alt.smac_index = qp->alt.candidate_smac_index;
2682 			qp->alt.smac_port = qp->alt.candidate_smac_port;
2683 		}
2684 		qp->alt.candidate_smac = 0;
2685 		qp->alt.candidate_smac_index = 0;
2686 		qp->alt.candidate_smac_port = 0;
2687 	}
2688 
2689 	if (qp->pri.update_vid) {
2690 		if (err) {
2691 			if (qp->pri.candidate_vid < 0x1000)
2692 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2693 						     qp->pri.candidate_vid);
2694 		} else {
2695 			if (qp->pri.vid < 0x1000)
2696 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2697 						     qp->pri.vid);
2698 			qp->pri.vid = qp->pri.candidate_vid;
2699 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2700 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
2701 		}
2702 		qp->pri.candidate_vid = 0xFFFF;
2703 		qp->pri.update_vid = 0;
2704 	}
2705 
2706 	if (qp->alt.update_vid) {
2707 		if (err) {
2708 			if (qp->alt.candidate_vid < 0x1000)
2709 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2710 						     qp->alt.candidate_vid);
2711 		} else {
2712 			if (qp->alt.vid < 0x1000)
2713 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2714 						     qp->alt.vid);
2715 			qp->alt.vid = qp->alt.candidate_vid;
2716 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2717 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2718 		}
2719 		qp->alt.candidate_vid = 0xFFFF;
2720 		qp->alt.update_vid = 0;
2721 	}
2722 
2723 	return err;
2724 }
2725 
2726 enum {
2727 	MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK = (IB_QP_STATE	|
2728 					      IB_QP_PORT),
2729 };
2730 
2731 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2732 			      int attr_mask, struct ib_udata *udata)
2733 {
2734 	enum rdma_link_layer ll = IB_LINK_LAYER_UNSPECIFIED;
2735 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2736 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2737 	enum ib_qp_state cur_state, new_state;
2738 	int err = -EINVAL;
2739 	mutex_lock(&qp->mutex);
2740 
2741 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2742 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2743 
2744 	if (cur_state != new_state || cur_state != IB_QPS_RESET) {
2745 		int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2746 		ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2747 	}
2748 
2749 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2750 				attr_mask, ll)) {
2751 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2752 			 "for transition %d to %d. qp_type %d,"
2753 			 " attr_mask 0x%x\n",
2754 			 ibqp->qp_num, cur_state, new_state,
2755 			 ibqp->qp_type, attr_mask);
2756 		goto out;
2757 	}
2758 
2759 	if (ibqp->rwq_ind_tbl) {
2760 		if (!(((cur_state == IB_QPS_RESET) &&
2761 		       (new_state == IB_QPS_INIT)) ||
2762 		      ((cur_state == IB_QPS_INIT)  &&
2763 		       (new_state == IB_QPS_RTR)))) {
2764 			pr_debug("qpn 0x%x: RSS QP unsupported transition %d to %d\n",
2765 				 ibqp->qp_num, cur_state, new_state);
2766 
2767 			err = -EOPNOTSUPP;
2768 			goto out;
2769 		}
2770 
2771 		if (attr_mask & ~MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK) {
2772 			pr_debug("qpn 0x%x: RSS QP unsupported attribute mask 0x%x for transition %d to %d\n",
2773 				 ibqp->qp_num, attr_mask, cur_state, new_state);
2774 
2775 			err = -EOPNOTSUPP;
2776 			goto out;
2777 		}
2778 	}
2779 
2780 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2781 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2782 			if ((ibqp->qp_type == IB_QPT_RC) ||
2783 			    (ibqp->qp_type == IB_QPT_UD) ||
2784 			    (ibqp->qp_type == IB_QPT_UC) ||
2785 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2786 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2787 				attr->port_num = mlx4_ib_bond_next_port(dev);
2788 			}
2789 		} else {
2790 			/* no sense in changing port_num
2791 			 * when ports are bonded */
2792 			attr_mask &= ~IB_QP_PORT;
2793 		}
2794 	}
2795 
2796 	if ((attr_mask & IB_QP_PORT) &&
2797 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2798 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2799 			 "for transition %d to %d. qp_type %d\n",
2800 			 ibqp->qp_num, attr->port_num, cur_state,
2801 			 new_state, ibqp->qp_type);
2802 		goto out;
2803 	}
2804 
2805 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2806 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2807 	     IB_LINK_LAYER_ETHERNET))
2808 		goto out;
2809 
2810 	if (attr_mask & IB_QP_PKEY_INDEX) {
2811 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2812 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2813 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2814 				 "for transition %d to %d. qp_type %d\n",
2815 				 ibqp->qp_num, attr->pkey_index, cur_state,
2816 				 new_state, ibqp->qp_type);
2817 			goto out;
2818 		}
2819 	}
2820 
2821 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2822 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2823 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2824 			 "Transition %d to %d. qp_type %d\n",
2825 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2826 			 new_state, ibqp->qp_type);
2827 		goto out;
2828 	}
2829 
2830 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2831 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2832 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2833 			 "Transition %d to %d. qp_type %d\n",
2834 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2835 			 new_state, ibqp->qp_type);
2836 		goto out;
2837 	}
2838 
2839 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2840 		err = 0;
2841 		goto out;
2842 	}
2843 
2844 	if (ibqp->rwq_ind_tbl && (new_state == IB_QPS_INIT)) {
2845 		err = bringup_rss_rwqs(ibqp->rwq_ind_tbl, attr->port_num);
2846 		if (err)
2847 			goto out;
2848 	}
2849 
2850 	err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask,
2851 				  cur_state, new_state);
2852 
2853 	if (ibqp->rwq_ind_tbl && err)
2854 		bring_down_rss_rwqs(ibqp->rwq_ind_tbl);
2855 
2856 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2857 		attr->port_num = 1;
2858 
2859 out:
2860 	mutex_unlock(&qp->mutex);
2861 	return err;
2862 }
2863 
2864 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2865 		      int attr_mask, struct ib_udata *udata)
2866 {
2867 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2868 	int ret;
2869 
2870 	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2871 
2872 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2873 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
2874 		int err = 0;
2875 
2876 		if (sqp->roce_v2_gsi)
2877 			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2878 		if (err)
2879 			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2880 			       err);
2881 	}
2882 	return ret;
2883 }
2884 
2885 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2886 {
2887 	int i;
2888 	for (i = 0; i < dev->caps.num_ports; i++) {
2889 		if (qpn == dev->caps.spec_qps[i].qp0_proxy ||
2890 		    qpn == dev->caps.spec_qps[i].qp0_tunnel) {
2891 			*qkey = dev->caps.spec_qps[i].qp0_qkey;
2892 			return 0;
2893 		}
2894 	}
2895 	return -EINVAL;
2896 }
2897 
2898 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2899 				  struct ib_ud_wr *wr,
2900 				  void *wqe, unsigned *mlx_seg_len)
2901 {
2902 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2903 	struct ib_device *ib_dev = &mdev->ib_dev;
2904 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2905 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2906 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2907 	u16 pkey;
2908 	u32 qkey;
2909 	int send_size;
2910 	int header_size;
2911 	int spc;
2912 	int i;
2913 
2914 	if (wr->wr.opcode != IB_WR_SEND)
2915 		return -EINVAL;
2916 
2917 	send_size = 0;
2918 
2919 	for (i = 0; i < wr->wr.num_sge; ++i)
2920 		send_size += wr->wr.sg_list[i].length;
2921 
2922 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2923 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2924 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2925 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2926 
2927 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2928 
2929 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2930 		sqp->ud_header.lrh.service_level =
2931 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2932 		sqp->ud_header.lrh.destination_lid =
2933 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2934 		sqp->ud_header.lrh.source_lid =
2935 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2936 	}
2937 
2938 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2939 
2940 	/* force loopback */
2941 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2942 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2943 
2944 	sqp->ud_header.lrh.virtual_lane    = 0;
2945 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2946 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2947 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2948 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2949 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2950 	else
2951 		sqp->ud_header.bth.destination_qpn =
2952 			cpu_to_be32(mdev->dev->caps.spec_qps[sqp->qp.port - 1].qp0_tunnel);
2953 
2954 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2955 	if (mlx4_is_master(mdev->dev)) {
2956 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2957 			return -EINVAL;
2958 	} else {
2959 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2960 			return -EINVAL;
2961 	}
2962 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2963 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2964 
2965 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2966 	sqp->ud_header.immediate_present = 0;
2967 
2968 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2969 
2970 	/*
2971 	 * Inline data segments may not cross a 64 byte boundary.  If
2972 	 * our UD header is bigger than the space available up to the
2973 	 * next 64 byte boundary in the WQE, use two inline data
2974 	 * segments to hold the UD header.
2975 	 */
2976 	spc = MLX4_INLINE_ALIGN -
2977 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2978 	if (header_size <= spc) {
2979 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2980 		memcpy(inl + 1, sqp->header_buf, header_size);
2981 		i = 1;
2982 	} else {
2983 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2984 		memcpy(inl + 1, sqp->header_buf, spc);
2985 
2986 		inl = (void *) (inl + 1) + spc;
2987 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2988 		/*
2989 		 * Need a barrier here to make sure all the data is
2990 		 * visible before the byte_count field is set.
2991 		 * Otherwise the HCA prefetcher could grab the 64-byte
2992 		 * chunk with this inline segment and get a valid (!=
2993 		 * 0xffffffff) byte count but stale data, and end up
2994 		 * generating a packet with bad headers.
2995 		 *
2996 		 * The first inline segment's byte_count field doesn't
2997 		 * need a barrier, because it comes after a
2998 		 * control/MLX segment and therefore is at an offset
2999 		 * of 16 mod 64.
3000 		 */
3001 		wmb();
3002 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3003 		i = 2;
3004 	}
3005 
3006 	*mlx_seg_len =
3007 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3008 	return 0;
3009 }
3010 
3011 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
3012 {
3013 	union sl2vl_tbl_to_u64 tmp_vltab;
3014 	u8 vl;
3015 
3016 	if (sl > 15)
3017 		return 0xf;
3018 	tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
3019 	vl = tmp_vltab.sl8[sl >> 1];
3020 	if (sl & 1)
3021 		vl &= 0x0f;
3022 	else
3023 		vl >>= 4;
3024 	return vl;
3025 }
3026 
3027 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
3028 				int index, union ib_gid *gid,
3029 				enum ib_gid_type *gid_type)
3030 {
3031 	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
3032 	struct mlx4_port_gid_table *port_gid_table;
3033 	unsigned long flags;
3034 
3035 	port_gid_table = &iboe->gids[port_num - 1];
3036 	spin_lock_irqsave(&iboe->lock, flags);
3037 	memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
3038 	*gid_type = port_gid_table->gids[index].gid_type;
3039 	spin_unlock_irqrestore(&iboe->lock, flags);
3040 	if (!memcmp(gid, &zgid, sizeof(*gid)))
3041 		return -ENOENT;
3042 
3043 	return 0;
3044 }
3045 
3046 #define MLX4_ROCEV2_QP1_SPORT 0xC000
3047 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_ud_wr *wr,
3048 			    void *wqe, unsigned *mlx_seg_len)
3049 {
3050 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
3051 	struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
3052 	struct mlx4_wqe_mlx_seg *mlx = wqe;
3053 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
3054 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
3055 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3056 	union ib_gid sgid;
3057 	u16 pkey;
3058 	int send_size;
3059 	int header_size;
3060 	int spc;
3061 	int i;
3062 	int err = 0;
3063 	u16 vlan = 0xffff;
3064 	bool is_eth;
3065 	bool is_vlan = false;
3066 	bool is_grh;
3067 	bool is_udp = false;
3068 	int ip_version = 0;
3069 
3070 	send_size = 0;
3071 	for (i = 0; i < wr->wr.num_sge; ++i)
3072 		send_size += wr->wr.sg_list[i].length;
3073 
3074 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
3075 	is_grh = mlx4_ib_ah_grh_present(ah);
3076 	if (is_eth) {
3077 		enum ib_gid_type gid_type;
3078 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3079 			/* When multi-function is enabled, the ib_core gid
3080 			 * indexes don't necessarily match the hw ones, so
3081 			 * we must use our own cache */
3082 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
3083 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
3084 							   ah->av.ib.gid_index, &sgid.raw[0]);
3085 			if (err)
3086 				return err;
3087 		} else  {
3088 			err = fill_gid_by_hw_index(ibdev, sqp->qp.port,
3089 					    ah->av.ib.gid_index,
3090 					    &sgid, &gid_type);
3091 			if (!err) {
3092 				is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
3093 				if (is_udp) {
3094 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
3095 						ip_version = 4;
3096 					else
3097 						ip_version = 6;
3098 					is_grh = false;
3099 				}
3100 			} else {
3101 				return err;
3102 			}
3103 		}
3104 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
3105 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
3106 			is_vlan = 1;
3107 		}
3108 	}
3109 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
3110 			  ip_version, is_udp, 0, &sqp->ud_header);
3111 	if (err)
3112 		return err;
3113 
3114 	if (!is_eth) {
3115 		sqp->ud_header.lrh.service_level =
3116 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
3117 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
3118 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
3119 	}
3120 
3121 	if (is_grh || (ip_version == 6)) {
3122 		sqp->ud_header.grh.traffic_class =
3123 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3124 		sqp->ud_header.grh.flow_label    =
3125 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
3126 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
3127 		if (is_eth) {
3128 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
3129 		} else {
3130 			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3131 				/* When multi-function is enabled, the ib_core gid
3132 				 * indexes don't necessarily match the hw ones, so
3133 				 * we must use our own cache
3134 				 */
3135 				sqp->ud_header.grh.source_gid.global.subnet_prefix =
3136 					cpu_to_be64(atomic64_read(&(to_mdev(ib_dev)->sriov.
3137 								    demux[sqp->qp.port - 1].
3138 								    subnet_prefix)));
3139 				sqp->ud_header.grh.source_gid.global.interface_id =
3140 					to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
3141 						       guid_cache[ah->av.ib.gid_index];
3142 			} else {
3143 				ib_get_cached_gid(ib_dev,
3144 						  be32_to_cpu(ah->av.ib.port_pd) >> 24,
3145 						  ah->av.ib.gid_index,
3146 						  &sqp->ud_header.grh.source_gid, NULL);
3147 			}
3148 		}
3149 		memcpy(sqp->ud_header.grh.destination_gid.raw,
3150 		       ah->av.ib.dgid, 16);
3151 	}
3152 
3153 	if (ip_version == 4) {
3154 		sqp->ud_header.ip4.tos =
3155 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3156 		sqp->ud_header.ip4.id = 0;
3157 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
3158 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
3159 
3160 		memcpy(&sqp->ud_header.ip4.saddr,
3161 		       sgid.raw + 12, 4);
3162 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
3163 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
3164 	}
3165 
3166 	if (is_udp) {
3167 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
3168 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
3169 		sqp->ud_header.udp.csum = 0;
3170 	}
3171 
3172 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
3173 
3174 	if (!is_eth) {
3175 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
3176 					  (sqp->ud_header.lrh.destination_lid ==
3177 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
3178 					  (sqp->ud_header.lrh.service_level << 8));
3179 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
3180 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
3181 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
3182 	}
3183 
3184 	switch (wr->wr.opcode) {
3185 	case IB_WR_SEND:
3186 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
3187 		sqp->ud_header.immediate_present = 0;
3188 		break;
3189 	case IB_WR_SEND_WITH_IMM:
3190 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
3191 		sqp->ud_header.immediate_present = 1;
3192 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
3193 		break;
3194 	default:
3195 		return -EINVAL;
3196 	}
3197 
3198 	if (is_eth) {
3199 		struct in6_addr in6;
3200 		u16 ether_type;
3201 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
3202 
3203 		ether_type = (!is_udp) ? ETH_P_IBOE:
3204 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
3205 
3206 		mlx->sched_prio = cpu_to_be16(pcp);
3207 
3208 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
3209 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
3210 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
3211 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
3212 		memcpy(&in6, sgid.raw, sizeof(in6));
3213 
3214 
3215 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
3216 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
3217 		if (!is_vlan) {
3218 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
3219 		} else {
3220 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
3221 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
3222 		}
3223 	} else {
3224 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 :
3225 							sl_to_vl(to_mdev(ib_dev),
3226 								 sqp->ud_header.lrh.service_level,
3227 								 sqp->qp.port);
3228 		if (sqp->qp.ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
3229 			return -EINVAL;
3230 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
3231 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
3232 	}
3233 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
3234 	if (!sqp->qp.ibqp.qp_num)
3235 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
3236 	else
3237 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
3238 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
3239 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
3240 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
3241 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
3242 					       sqp->qkey : wr->remote_qkey);
3243 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
3244 
3245 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3246 
3247 	if (0) {
3248 		pr_err("built UD header of size %d:\n", header_size);
3249 		for (i = 0; i < header_size / 4; ++i) {
3250 			if (i % 8 == 0)
3251 				pr_err("  [%02x] ", i * 4);
3252 			pr_cont(" %08x",
3253 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
3254 			if ((i + 1) % 8 == 0)
3255 				pr_cont("\n");
3256 		}
3257 		pr_err("\n");
3258 	}
3259 
3260 	/*
3261 	 * Inline data segments may not cross a 64 byte boundary.  If
3262 	 * our UD header is bigger than the space available up to the
3263 	 * next 64 byte boundary in the WQE, use two inline data
3264 	 * segments to hold the UD header.
3265 	 */
3266 	spc = MLX4_INLINE_ALIGN -
3267 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3268 	if (header_size <= spc) {
3269 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3270 		memcpy(inl + 1, sqp->header_buf, header_size);
3271 		i = 1;
3272 	} else {
3273 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3274 		memcpy(inl + 1, sqp->header_buf, spc);
3275 
3276 		inl = (void *) (inl + 1) + spc;
3277 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3278 		/*
3279 		 * Need a barrier here to make sure all the data is
3280 		 * visible before the byte_count field is set.
3281 		 * Otherwise the HCA prefetcher could grab the 64-byte
3282 		 * chunk with this inline segment and get a valid (!=
3283 		 * 0xffffffff) byte count but stale data, and end up
3284 		 * generating a packet with bad headers.
3285 		 *
3286 		 * The first inline segment's byte_count field doesn't
3287 		 * need a barrier, because it comes after a
3288 		 * control/MLX segment and therefore is at an offset
3289 		 * of 16 mod 64.
3290 		 */
3291 		wmb();
3292 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3293 		i = 2;
3294 	}
3295 
3296 	*mlx_seg_len =
3297 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3298 	return 0;
3299 }
3300 
3301 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3302 {
3303 	unsigned cur;
3304 	struct mlx4_ib_cq *cq;
3305 
3306 	cur = wq->head - wq->tail;
3307 	if (likely(cur + nreq < wq->max_post))
3308 		return 0;
3309 
3310 	cq = to_mcq(ib_cq);
3311 	spin_lock(&cq->lock);
3312 	cur = wq->head - wq->tail;
3313 	spin_unlock(&cq->lock);
3314 
3315 	return cur + nreq >= wq->max_post;
3316 }
3317 
3318 static __be32 convert_access(int acc)
3319 {
3320 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
3321 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
3322 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
3323 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
3324 	       (acc & IB_ACCESS_REMOTE_READ   ?
3325 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
3326 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
3327 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
3328 }
3329 
3330 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
3331 			struct ib_reg_wr *wr)
3332 {
3333 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
3334 
3335 	fseg->flags		= convert_access(wr->access);
3336 	fseg->mem_key		= cpu_to_be32(wr->key);
3337 	fseg->buf_list		= cpu_to_be64(mr->page_map);
3338 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
3339 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
3340 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
3341 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
3342 	fseg->reserved[0]	= 0;
3343 	fseg->reserved[1]	= 0;
3344 }
3345 
3346 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
3347 {
3348 	memset(iseg, 0, sizeof(*iseg));
3349 	iseg->mem_key = cpu_to_be32(rkey);
3350 }
3351 
3352 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
3353 					  u64 remote_addr, u32 rkey)
3354 {
3355 	rseg->raddr    = cpu_to_be64(remote_addr);
3356 	rseg->rkey     = cpu_to_be32(rkey);
3357 	rseg->reserved = 0;
3358 }
3359 
3360 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
3361 		struct ib_atomic_wr *wr)
3362 {
3363 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
3364 		aseg->swap_add = cpu_to_be64(wr->swap);
3365 		aseg->compare  = cpu_to_be64(wr->compare_add);
3366 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
3367 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3368 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
3369 	} else {
3370 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3371 		aseg->compare  = 0;
3372 	}
3373 
3374 }
3375 
3376 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
3377 				  struct ib_atomic_wr *wr)
3378 {
3379 	aseg->swap_add		= cpu_to_be64(wr->swap);
3380 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
3381 	aseg->compare		= cpu_to_be64(wr->compare_add);
3382 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
3383 }
3384 
3385 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
3386 			     struct ib_ud_wr *wr)
3387 {
3388 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
3389 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
3390 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
3391 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
3392 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
3393 }
3394 
3395 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
3396 				    struct mlx4_wqe_datagram_seg *dseg,
3397 				    struct ib_ud_wr *wr,
3398 				    enum mlx4_ib_qp_type qpt)
3399 {
3400 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
3401 	struct mlx4_av sqp_av = {0};
3402 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
3403 
3404 	/* force loopback */
3405 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
3406 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
3407 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
3408 			cpu_to_be32(0xf0000000);
3409 
3410 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
3411 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
3412 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp1_tunnel);
3413 	else
3414 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp0_tunnel);
3415 	/* Use QKEY from the QP context, which is set by master */
3416 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
3417 }
3418 
3419 static void build_tunnel_header(struct ib_ud_wr *wr, void *wqe, unsigned *mlx_seg_len)
3420 {
3421 	struct mlx4_wqe_inline_seg *inl = wqe;
3422 	struct mlx4_ib_tunnel_header hdr;
3423 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3424 	int spc;
3425 	int i;
3426 
3427 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
3428 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
3429 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
3430 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
3431 	memcpy(hdr.mac, ah->av.eth.mac, 6);
3432 	hdr.vlan = ah->av.eth.vlan;
3433 
3434 	spc = MLX4_INLINE_ALIGN -
3435 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3436 	if (sizeof (hdr) <= spc) {
3437 		memcpy(inl + 1, &hdr, sizeof (hdr));
3438 		wmb();
3439 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
3440 		i = 1;
3441 	} else {
3442 		memcpy(inl + 1, &hdr, spc);
3443 		wmb();
3444 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3445 
3446 		inl = (void *) (inl + 1) + spc;
3447 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
3448 		wmb();
3449 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
3450 		i = 2;
3451 	}
3452 
3453 	*mlx_seg_len =
3454 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
3455 }
3456 
3457 static void set_mlx_icrc_seg(void *dseg)
3458 {
3459 	u32 *t = dseg;
3460 	struct mlx4_wqe_inline_seg *iseg = dseg;
3461 
3462 	t[1] = 0;
3463 
3464 	/*
3465 	 * Need a barrier here before writing the byte_count field to
3466 	 * make sure that all the data is visible before the
3467 	 * byte_count field is set.  Otherwise, if the segment begins
3468 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3469 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3470 	 * stale data, and end up sending the wrong data.
3471 	 */
3472 	wmb();
3473 
3474 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
3475 }
3476 
3477 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3478 {
3479 	dseg->lkey       = cpu_to_be32(sg->lkey);
3480 	dseg->addr       = cpu_to_be64(sg->addr);
3481 
3482 	/*
3483 	 * Need a barrier here before writing the byte_count field to
3484 	 * make sure that all the data is visible before the
3485 	 * byte_count field is set.  Otherwise, if the segment begins
3486 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3487 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3488 	 * stale data, and end up sending the wrong data.
3489 	 */
3490 	wmb();
3491 
3492 	dseg->byte_count = cpu_to_be32(sg->length);
3493 }
3494 
3495 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3496 {
3497 	dseg->byte_count = cpu_to_be32(sg->length);
3498 	dseg->lkey       = cpu_to_be32(sg->lkey);
3499 	dseg->addr       = cpu_to_be64(sg->addr);
3500 }
3501 
3502 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_ud_wr *wr,
3503 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
3504 			 __be32 *lso_hdr_sz, __be32 *blh)
3505 {
3506 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
3507 
3508 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
3509 		*blh = cpu_to_be32(1 << 6);
3510 
3511 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
3512 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
3513 		return -EINVAL;
3514 
3515 	memcpy(wqe->header, wr->header, wr->hlen);
3516 
3517 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
3518 	*lso_seg_len = halign;
3519 	return 0;
3520 }
3521 
3522 static __be32 send_ieth(struct ib_send_wr *wr)
3523 {
3524 	switch (wr->opcode) {
3525 	case IB_WR_SEND_WITH_IMM:
3526 	case IB_WR_RDMA_WRITE_WITH_IMM:
3527 		return wr->ex.imm_data;
3528 
3529 	case IB_WR_SEND_WITH_INV:
3530 		return cpu_to_be32(wr->ex.invalidate_rkey);
3531 
3532 	default:
3533 		return 0;
3534 	}
3535 }
3536 
3537 static void add_zero_len_inline(void *wqe)
3538 {
3539 	struct mlx4_wqe_inline_seg *inl = wqe;
3540 	memset(wqe, 0, 16);
3541 	inl->byte_count = cpu_to_be32(1 << 31);
3542 }
3543 
3544 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
3545 		      struct ib_send_wr **bad_wr)
3546 {
3547 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3548 	void *wqe;
3549 	struct mlx4_wqe_ctrl_seg *ctrl;
3550 	struct mlx4_wqe_data_seg *dseg;
3551 	unsigned long flags;
3552 	int nreq;
3553 	int err = 0;
3554 	unsigned ind;
3555 	int uninitialized_var(stamp);
3556 	int uninitialized_var(size);
3557 	unsigned uninitialized_var(seglen);
3558 	__be32 dummy;
3559 	__be32 *lso_wqe;
3560 	__be32 uninitialized_var(lso_hdr_sz);
3561 	__be32 blh;
3562 	int i;
3563 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3564 
3565 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
3566 		struct mlx4_ib_sqp *sqp = to_msqp(qp);
3567 
3568 		if (sqp->roce_v2_gsi) {
3569 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
3570 			enum ib_gid_type gid_type;
3571 			union ib_gid gid;
3572 
3573 			if (!fill_gid_by_hw_index(mdev, sqp->qp.port,
3574 					   ah->av.ib.gid_index,
3575 					   &gid, &gid_type))
3576 				qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
3577 						to_mqp(sqp->roce_v2_gsi) : qp;
3578 			else
3579 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
3580 				       ah->av.ib.gid_index);
3581 		}
3582 	}
3583 
3584 	spin_lock_irqsave(&qp->sq.lock, flags);
3585 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3586 		err = -EIO;
3587 		*bad_wr = wr;
3588 		nreq = 0;
3589 		goto out;
3590 	}
3591 
3592 	ind = qp->sq_next_wqe;
3593 
3594 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3595 		lso_wqe = &dummy;
3596 		blh = 0;
3597 
3598 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
3599 			err = -ENOMEM;
3600 			*bad_wr = wr;
3601 			goto out;
3602 		}
3603 
3604 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3605 			err = -EINVAL;
3606 			*bad_wr = wr;
3607 			goto out;
3608 		}
3609 
3610 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3611 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3612 
3613 		ctrl->srcrb_flags =
3614 			(wr->send_flags & IB_SEND_SIGNALED ?
3615 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3616 			(wr->send_flags & IB_SEND_SOLICITED ?
3617 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3618 			((wr->send_flags & IB_SEND_IP_CSUM) ?
3619 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3620 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3621 			qp->sq_signal_bits;
3622 
3623 		ctrl->imm = send_ieth(wr);
3624 
3625 		wqe += sizeof *ctrl;
3626 		size = sizeof *ctrl / 16;
3627 
3628 		switch (qp->mlx4_ib_qp_type) {
3629 		case MLX4_IB_QPT_RC:
3630 		case MLX4_IB_QPT_UC:
3631 			switch (wr->opcode) {
3632 			case IB_WR_ATOMIC_CMP_AND_SWP:
3633 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3634 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3635 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3636 					      atomic_wr(wr)->rkey);
3637 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3638 
3639 				set_atomic_seg(wqe, atomic_wr(wr));
3640 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
3641 
3642 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3643 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3644 
3645 				break;
3646 
3647 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3648 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3649 					      atomic_wr(wr)->rkey);
3650 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3651 
3652 				set_masked_atomic_seg(wqe, atomic_wr(wr));
3653 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
3654 
3655 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3656 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3657 
3658 				break;
3659 
3660 			case IB_WR_RDMA_READ:
3661 			case IB_WR_RDMA_WRITE:
3662 			case IB_WR_RDMA_WRITE_WITH_IMM:
3663 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3664 					      rdma_wr(wr)->rkey);
3665 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3666 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3667 				break;
3668 
3669 			case IB_WR_LOCAL_INV:
3670 				ctrl->srcrb_flags |=
3671 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3672 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3673 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3674 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3675 				break;
3676 
3677 			case IB_WR_REG_MR:
3678 				ctrl->srcrb_flags |=
3679 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3680 				set_reg_seg(wqe, reg_wr(wr));
3681 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3682 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3683 				break;
3684 
3685 			default:
3686 				/* No extra segments required for sends */
3687 				break;
3688 			}
3689 			break;
3690 
3691 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3692 			err =  build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3693 					ctrl, &seglen);
3694 			if (unlikely(err)) {
3695 				*bad_wr = wr;
3696 				goto out;
3697 			}
3698 			wqe  += seglen;
3699 			size += seglen / 16;
3700 			break;
3701 		case MLX4_IB_QPT_TUN_SMI:
3702 		case MLX4_IB_QPT_TUN_GSI:
3703 			/* this is a UD qp used in MAD responses to slaves. */
3704 			set_datagram_seg(wqe, ud_wr(wr));
3705 			/* set the forced-loopback bit in the data seg av */
3706 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3707 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3708 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3709 			break;
3710 		case MLX4_IB_QPT_UD:
3711 			set_datagram_seg(wqe, ud_wr(wr));
3712 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3713 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3714 
3715 			if (wr->opcode == IB_WR_LSO) {
3716 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3717 						&lso_hdr_sz, &blh);
3718 				if (unlikely(err)) {
3719 					*bad_wr = wr;
3720 					goto out;
3721 				}
3722 				lso_wqe = (__be32 *) wqe;
3723 				wqe  += seglen;
3724 				size += seglen / 16;
3725 			}
3726 			break;
3727 
3728 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3729 			err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3730 					ctrl, &seglen);
3731 			if (unlikely(err)) {
3732 				*bad_wr = wr;
3733 				goto out;
3734 			}
3735 			wqe  += seglen;
3736 			size += seglen / 16;
3737 			/* to start tunnel header on a cache-line boundary */
3738 			add_zero_len_inline(wqe);
3739 			wqe += 16;
3740 			size++;
3741 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3742 			wqe  += seglen;
3743 			size += seglen / 16;
3744 			break;
3745 		case MLX4_IB_QPT_PROXY_SMI:
3746 		case MLX4_IB_QPT_PROXY_GSI:
3747 			/* If we are tunneling special qps, this is a UD qp.
3748 			 * In this case we first add a UD segment targeting
3749 			 * the tunnel qp, and then add a header with address
3750 			 * information */
3751 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3752 						ud_wr(wr),
3753 						qp->mlx4_ib_qp_type);
3754 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3755 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3756 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3757 			wqe  += seglen;
3758 			size += seglen / 16;
3759 			break;
3760 
3761 		case MLX4_IB_QPT_SMI:
3762 		case MLX4_IB_QPT_GSI:
3763 			err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
3764 					&seglen);
3765 			if (unlikely(err)) {
3766 				*bad_wr = wr;
3767 				goto out;
3768 			}
3769 			wqe  += seglen;
3770 			size += seglen / 16;
3771 			break;
3772 
3773 		default:
3774 			break;
3775 		}
3776 
3777 		/*
3778 		 * Write data segments in reverse order, so as to
3779 		 * overwrite cacheline stamp last within each
3780 		 * cacheline.  This avoids issues with WQE
3781 		 * prefetching.
3782 		 */
3783 
3784 		dseg = wqe;
3785 		dseg += wr->num_sge - 1;
3786 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3787 
3788 		/* Add one more inline data segment for ICRC for MLX sends */
3789 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3790 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3791 			     qp->mlx4_ib_qp_type &
3792 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3793 			set_mlx_icrc_seg(dseg + 1);
3794 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3795 		}
3796 
3797 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3798 			set_data_seg(dseg, wr->sg_list + i);
3799 
3800 		/*
3801 		 * Possibly overwrite stamping in cacheline with LSO
3802 		 * segment only after making sure all data segments
3803 		 * are written.
3804 		 */
3805 		wmb();
3806 		*lso_wqe = lso_hdr_sz;
3807 
3808 		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3809 					     MLX4_WQE_CTRL_FENCE : 0) | size;
3810 
3811 		/*
3812 		 * Make sure descriptor is fully written before
3813 		 * setting ownership bit (because HW can start
3814 		 * executing as soon as we do).
3815 		 */
3816 		wmb();
3817 
3818 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3819 			*bad_wr = wr;
3820 			err = -EINVAL;
3821 			goto out;
3822 		}
3823 
3824 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3825 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3826 
3827 		stamp = ind + qp->sq_spare_wqes;
3828 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
3829 
3830 		/*
3831 		 * We can improve latency by not stamping the last
3832 		 * send queue WQE until after ringing the doorbell, so
3833 		 * only stamp here if there are still more WQEs to post.
3834 		 *
3835 		 * Same optimization applies to padding with NOP wqe
3836 		 * in case of WQE shrinking (used to prevent wrap-around
3837 		 * in the middle of WR).
3838 		 */
3839 		if (wr->next) {
3840 			stamp_send_wqe(qp, stamp, size * 16);
3841 			ind = pad_wraparound(qp, ind);
3842 		}
3843 	}
3844 
3845 out:
3846 	if (likely(nreq)) {
3847 		qp->sq.head += nreq;
3848 
3849 		/*
3850 		 * Make sure that descriptors are written before
3851 		 * doorbell record.
3852 		 */
3853 		wmb();
3854 
3855 		writel(qp->doorbell_qpn,
3856 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3857 
3858 		/*
3859 		 * Make sure doorbells don't leak out of SQ spinlock
3860 		 * and reach the HCA out of order.
3861 		 */
3862 		mmiowb();
3863 
3864 		stamp_send_wqe(qp, stamp, size * 16);
3865 
3866 		ind = pad_wraparound(qp, ind);
3867 		qp->sq_next_wqe = ind;
3868 	}
3869 
3870 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3871 
3872 	return err;
3873 }
3874 
3875 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3876 		      struct ib_recv_wr **bad_wr)
3877 {
3878 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3879 	struct mlx4_wqe_data_seg *scat;
3880 	unsigned long flags;
3881 	int err = 0;
3882 	int nreq;
3883 	int ind;
3884 	int max_gs;
3885 	int i;
3886 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3887 
3888 	max_gs = qp->rq.max_gs;
3889 	spin_lock_irqsave(&qp->rq.lock, flags);
3890 
3891 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3892 		err = -EIO;
3893 		*bad_wr = wr;
3894 		nreq = 0;
3895 		goto out;
3896 	}
3897 
3898 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3899 
3900 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3901 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3902 			err = -ENOMEM;
3903 			*bad_wr = wr;
3904 			goto out;
3905 		}
3906 
3907 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3908 			err = -EINVAL;
3909 			*bad_wr = wr;
3910 			goto out;
3911 		}
3912 
3913 		scat = get_recv_wqe(qp, ind);
3914 
3915 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3916 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3917 			ib_dma_sync_single_for_device(ibqp->device,
3918 						      qp->sqp_proxy_rcv[ind].map,
3919 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3920 						      DMA_FROM_DEVICE);
3921 			scat->byte_count =
3922 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3923 			/* use dma lkey from upper layer entry */
3924 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3925 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3926 			scat++;
3927 			max_gs--;
3928 		}
3929 
3930 		for (i = 0; i < wr->num_sge; ++i)
3931 			__set_data_seg(scat + i, wr->sg_list + i);
3932 
3933 		if (i < max_gs) {
3934 			scat[i].byte_count = 0;
3935 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3936 			scat[i].addr       = 0;
3937 		}
3938 
3939 		qp->rq.wrid[ind] = wr->wr_id;
3940 
3941 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3942 	}
3943 
3944 out:
3945 	if (likely(nreq)) {
3946 		qp->rq.head += nreq;
3947 
3948 		/*
3949 		 * Make sure that descriptors are written before
3950 		 * doorbell record.
3951 		 */
3952 		wmb();
3953 
3954 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3955 	}
3956 
3957 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3958 
3959 	return err;
3960 }
3961 
3962 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3963 {
3964 	switch (mlx4_state) {
3965 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3966 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3967 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3968 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3969 	case MLX4_QP_STATE_SQ_DRAINING:
3970 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3971 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3972 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3973 	default:		     return -1;
3974 	}
3975 }
3976 
3977 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3978 {
3979 	switch (mlx4_mig_state) {
3980 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3981 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3982 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3983 	default: return -1;
3984 	}
3985 }
3986 
3987 static int to_ib_qp_access_flags(int mlx4_flags)
3988 {
3989 	int ib_flags = 0;
3990 
3991 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3992 		ib_flags |= IB_ACCESS_REMOTE_READ;
3993 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3994 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3995 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3996 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3997 
3998 	return ib_flags;
3999 }
4000 
4001 static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
4002 			    struct rdma_ah_attr *ah_attr,
4003 			    struct mlx4_qp_path *path)
4004 {
4005 	struct mlx4_dev *dev = ibdev->dev;
4006 	u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
4007 
4008 	memset(ah_attr, 0, sizeof(*ah_attr));
4009 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
4010 	if (port_num == 0 || port_num > dev->caps.num_ports)
4011 		return;
4012 
4013 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
4014 		rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
4015 			       ((path->sched_queue & 4) << 1));
4016 	else
4017 		rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
4018 	rdma_ah_set_port_num(ah_attr, port_num);
4019 
4020 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
4021 	rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
4022 	rdma_ah_set_static_rate(ah_attr,
4023 				path->static_rate ? path->static_rate - 5 : 0);
4024 	if (path->grh_mylmc & (1 << 7)) {
4025 		rdma_ah_set_grh(ah_attr, NULL,
4026 				be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
4027 				path->mgid_index,
4028 				path->hop_limit,
4029 				(be32_to_cpu(path->tclass_flowlabel)
4030 				 >> 20) & 0xff);
4031 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
4032 	}
4033 }
4034 
4035 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
4036 		     struct ib_qp_init_attr *qp_init_attr)
4037 {
4038 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
4039 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
4040 	struct mlx4_qp_context context;
4041 	int mlx4_state;
4042 	int err = 0;
4043 
4044 	if (ibqp->rwq_ind_tbl)
4045 		return -EOPNOTSUPP;
4046 
4047 	mutex_lock(&qp->mutex);
4048 
4049 	if (qp->state == IB_QPS_RESET) {
4050 		qp_attr->qp_state = IB_QPS_RESET;
4051 		goto done;
4052 	}
4053 
4054 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
4055 	if (err) {
4056 		err = -EINVAL;
4057 		goto out;
4058 	}
4059 
4060 	mlx4_state = be32_to_cpu(context.flags) >> 28;
4061 
4062 	qp->state		     = to_ib_qp_state(mlx4_state);
4063 	qp_attr->qp_state	     = qp->state;
4064 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
4065 	qp_attr->path_mig_state	     =
4066 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
4067 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
4068 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
4069 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
4070 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
4071 	qp_attr->qp_access_flags     =
4072 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
4073 
4074 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4075 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
4076 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
4077 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
4078 		qp_attr->alt_port_num	=
4079 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
4080 	}
4081 
4082 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
4083 	if (qp_attr->qp_state == IB_QPS_INIT)
4084 		qp_attr->port_num = qp->port;
4085 	else
4086 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
4087 
4088 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4089 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
4090 
4091 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
4092 
4093 	qp_attr->max_dest_rd_atomic =
4094 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
4095 	qp_attr->min_rnr_timer	    =
4096 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
4097 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
4098 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
4099 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
4100 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
4101 
4102 done:
4103 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4104 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4105 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4106 
4107 	if (!ibqp->uobject) {
4108 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
4109 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4110 	} else {
4111 		qp_attr->cap.max_send_wr  = 0;
4112 		qp_attr->cap.max_send_sge = 0;
4113 	}
4114 
4115 	/*
4116 	 * We don't support inline sends for kernel QPs (yet), and we
4117 	 * don't know what userspace's value should be.
4118 	 */
4119 	qp_attr->cap.max_inline_data = 0;
4120 
4121 	qp_init_attr->cap	     = qp_attr->cap;
4122 
4123 	qp_init_attr->create_flags = 0;
4124 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4125 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4126 
4127 	if (qp->flags & MLX4_IB_QP_LSO)
4128 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
4129 
4130 	if (qp->flags & MLX4_IB_QP_NETIF)
4131 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
4132 
4133 	qp_init_attr->sq_sig_type =
4134 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
4135 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4136 
4137 out:
4138 	mutex_unlock(&qp->mutex);
4139 	return err;
4140 }
4141 
4142 struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
4143 				struct ib_wq_init_attr *init_attr,
4144 				struct ib_udata *udata)
4145 {
4146 	struct mlx4_ib_dev *dev;
4147 	struct ib_qp_init_attr ib_qp_init_attr;
4148 	struct mlx4_ib_qp *qp;
4149 	struct mlx4_ib_create_wq ucmd;
4150 	int err, required_cmd_sz;
4151 
4152 	if (!(udata && pd->uobject))
4153 		return ERR_PTR(-EINVAL);
4154 
4155 	required_cmd_sz = offsetof(typeof(ucmd), comp_mask) +
4156 			  sizeof(ucmd.comp_mask);
4157 	if (udata->inlen < required_cmd_sz) {
4158 		pr_debug("invalid inlen\n");
4159 		return ERR_PTR(-EINVAL);
4160 	}
4161 
4162 	if (udata->inlen > sizeof(ucmd) &&
4163 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4164 				 udata->inlen - sizeof(ucmd))) {
4165 		pr_debug("inlen is not supported\n");
4166 		return ERR_PTR(-EOPNOTSUPP);
4167 	}
4168 
4169 	if (udata->outlen)
4170 		return ERR_PTR(-EOPNOTSUPP);
4171 
4172 	dev = to_mdev(pd->device);
4173 
4174 	if (init_attr->wq_type != IB_WQT_RQ) {
4175 		pr_debug("unsupported wq type %d\n", init_attr->wq_type);
4176 		return ERR_PTR(-EOPNOTSUPP);
4177 	}
4178 
4179 	if (init_attr->create_flags) {
4180 		pr_debug("unsupported create_flags %u\n",
4181 			 init_attr->create_flags);
4182 		return ERR_PTR(-EOPNOTSUPP);
4183 	}
4184 
4185 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
4186 	if (!qp)
4187 		return ERR_PTR(-ENOMEM);
4188 
4189 	qp->pri.vid = 0xFFFF;
4190 	qp->alt.vid = 0xFFFF;
4191 
4192 	memset(&ib_qp_init_attr, 0, sizeof(ib_qp_init_attr));
4193 	ib_qp_init_attr.qp_context = init_attr->wq_context;
4194 	ib_qp_init_attr.qp_type = IB_QPT_RAW_PACKET;
4195 	ib_qp_init_attr.cap.max_recv_wr = init_attr->max_wr;
4196 	ib_qp_init_attr.cap.max_recv_sge = init_attr->max_sge;
4197 	ib_qp_init_attr.recv_cq = init_attr->cq;
4198 	ib_qp_init_attr.send_cq = ib_qp_init_attr.recv_cq; /* Dummy CQ */
4199 
4200 	err = create_qp_common(dev, pd, MLX4_IB_RWQ_SRC, &ib_qp_init_attr,
4201 			       udata, 0, &qp);
4202 	if (err) {
4203 		kfree(qp);
4204 		return ERR_PTR(err);
4205 	}
4206 
4207 	qp->ibwq.event_handler = init_attr->event_handler;
4208 	qp->ibwq.wq_num = qp->mqp.qpn;
4209 	qp->ibwq.state = IB_WQS_RESET;
4210 
4211 	return &qp->ibwq;
4212 }
4213 
4214 static int ib_wq2qp_state(enum ib_wq_state state)
4215 {
4216 	switch (state) {
4217 	case IB_WQS_RESET:
4218 		return IB_QPS_RESET;
4219 	case IB_WQS_RDY:
4220 		return IB_QPS_RTR;
4221 	default:
4222 		return IB_QPS_ERR;
4223 	}
4224 }
4225 
4226 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state)
4227 {
4228 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4229 	enum ib_qp_state qp_cur_state;
4230 	enum ib_qp_state qp_new_state;
4231 	int attr_mask;
4232 	int err;
4233 
4234 	/* ib_qp.state represents the WQ HW state while ib_wq.state represents
4235 	 * the WQ logic state.
4236 	 */
4237 	qp_cur_state = qp->state;
4238 	qp_new_state = ib_wq2qp_state(new_state);
4239 
4240 	if (ib_wq2qp_state(new_state) == qp_cur_state)
4241 		return 0;
4242 
4243 	if (new_state == IB_WQS_RDY) {
4244 		struct ib_qp_attr attr = {};
4245 
4246 		attr.port_num = qp->port;
4247 		attr_mask = IB_QP_PORT;
4248 
4249 		err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, &attr,
4250 					  attr_mask, IB_QPS_RESET, IB_QPS_INIT);
4251 		if (err) {
4252 			pr_debug("WQN=0x%06x failed to apply RST->INIT on the HW QP\n",
4253 				 ibwq->wq_num);
4254 			return err;
4255 		}
4256 
4257 		qp_cur_state = IB_QPS_INIT;
4258 	}
4259 
4260 	attr_mask = 0;
4261 	err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, attr_mask,
4262 				  qp_cur_state,  qp_new_state);
4263 
4264 	if (err && (qp_cur_state == IB_QPS_INIT)) {
4265 		qp_new_state = IB_QPS_RESET;
4266 		if (__mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL,
4267 					attr_mask, IB_QPS_INIT, IB_QPS_RESET)) {
4268 			pr_warn("WQN=0x%06x failed with reverting HW's resources failure\n",
4269 				ibwq->wq_num);
4270 			qp_new_state = IB_QPS_INIT;
4271 		}
4272 	}
4273 
4274 	qp->state = qp_new_state;
4275 
4276 	return err;
4277 }
4278 
4279 int mlx4_ib_modify_wq(struct ib_wq *ibwq, struct ib_wq_attr *wq_attr,
4280 		      u32 wq_attr_mask, struct ib_udata *udata)
4281 {
4282 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4283 	struct mlx4_ib_modify_wq ucmd = {};
4284 	size_t required_cmd_sz;
4285 	enum ib_wq_state cur_state, new_state;
4286 	int err = 0;
4287 
4288 	required_cmd_sz = offsetof(typeof(ucmd), reserved) +
4289 				   sizeof(ucmd.reserved);
4290 	if (udata->inlen < required_cmd_sz)
4291 		return -EINVAL;
4292 
4293 	if (udata->inlen > sizeof(ucmd) &&
4294 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4295 				 udata->inlen - sizeof(ucmd)))
4296 		return -EOPNOTSUPP;
4297 
4298 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4299 		return -EFAULT;
4300 
4301 	if (ucmd.comp_mask || ucmd.reserved)
4302 		return -EOPNOTSUPP;
4303 
4304 	if (wq_attr_mask & IB_WQ_FLAGS)
4305 		return -EOPNOTSUPP;
4306 
4307 	cur_state = wq_attr_mask & IB_WQ_CUR_STATE ? wq_attr->curr_wq_state :
4308 						     ibwq->state;
4309 	new_state = wq_attr_mask & IB_WQ_STATE ? wq_attr->wq_state : cur_state;
4310 
4311 	if (cur_state  < IB_WQS_RESET || cur_state  > IB_WQS_ERR ||
4312 	    new_state < IB_WQS_RESET || new_state > IB_WQS_ERR)
4313 		return -EINVAL;
4314 
4315 	if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR))
4316 		return -EINVAL;
4317 
4318 	if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET))
4319 		return -EINVAL;
4320 
4321 	/* Need to protect against the parent RSS which also may modify WQ
4322 	 * state.
4323 	 */
4324 	mutex_lock(&qp->mutex);
4325 
4326 	/* Can update HW state only if a RSS QP has already associated to this
4327 	 * WQ, so we can apply its port on the WQ.
4328 	 */
4329 	if (qp->rss_usecnt)
4330 		err = _mlx4_ib_modify_wq(ibwq, new_state);
4331 
4332 	if (!err)
4333 		ibwq->state = new_state;
4334 
4335 	mutex_unlock(&qp->mutex);
4336 
4337 	return err;
4338 }
4339 
4340 int mlx4_ib_destroy_wq(struct ib_wq *ibwq)
4341 {
4342 	struct mlx4_ib_dev *dev = to_mdev(ibwq->device);
4343 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4344 
4345 	if (qp->counter_index)
4346 		mlx4_ib_free_qp_counter(dev, qp);
4347 
4348 	destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, 1);
4349 
4350 	kfree(qp);
4351 
4352 	return 0;
4353 }
4354 
4355 struct ib_rwq_ind_table
4356 *mlx4_ib_create_rwq_ind_table(struct ib_device *device,
4357 			      struct ib_rwq_ind_table_init_attr *init_attr,
4358 			      struct ib_udata *udata)
4359 {
4360 	struct ib_rwq_ind_table *rwq_ind_table;
4361 	struct mlx4_ib_create_rwq_ind_tbl_resp resp = {};
4362 	unsigned int ind_tbl_size = 1 << init_attr->log_ind_tbl_size;
4363 	unsigned int base_wqn;
4364 	size_t min_resp_len;
4365 	int i;
4366 	int err;
4367 
4368 	if (udata->inlen > 0 &&
4369 	    !ib_is_udata_cleared(udata, 0,
4370 				 udata->inlen))
4371 		return ERR_PTR(-EOPNOTSUPP);
4372 
4373 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4374 	if (udata->outlen && udata->outlen < min_resp_len)
4375 		return ERR_PTR(-EINVAL);
4376 
4377 	if (ind_tbl_size >
4378 	    device->attrs.rss_caps.max_rwq_indirection_table_size) {
4379 		pr_debug("log_ind_tbl_size = %d is bigger than supported = %d\n",
4380 			 ind_tbl_size,
4381 			 device->attrs.rss_caps.max_rwq_indirection_table_size);
4382 		return ERR_PTR(-EINVAL);
4383 	}
4384 
4385 	base_wqn = init_attr->ind_tbl[0]->wq_num;
4386 
4387 	if (base_wqn % ind_tbl_size) {
4388 		pr_debug("WQN=0x%x isn't aligned with indirection table size\n",
4389 			 base_wqn);
4390 		return ERR_PTR(-EINVAL);
4391 	}
4392 
4393 	for (i = 1; i < ind_tbl_size; i++) {
4394 		if (++base_wqn != init_attr->ind_tbl[i]->wq_num) {
4395 			pr_debug("indirection table's WQNs aren't consecutive\n");
4396 			return ERR_PTR(-EINVAL);
4397 		}
4398 	}
4399 
4400 	rwq_ind_table = kzalloc(sizeof(*rwq_ind_table), GFP_KERNEL);
4401 	if (!rwq_ind_table)
4402 		return ERR_PTR(-ENOMEM);
4403 
4404 	if (udata->outlen) {
4405 		resp.response_length = offsetof(typeof(resp), response_length) +
4406 					sizeof(resp.response_length);
4407 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4408 		if (err)
4409 			goto err;
4410 	}
4411 
4412 	return rwq_ind_table;
4413 
4414 err:
4415 	kfree(rwq_ind_table);
4416 	return ERR_PTR(err);
4417 }
4418 
4419 int mlx4_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4420 {
4421 	kfree(ib_rwq_ind_tbl);
4422 	return 0;
4423 }
4424