xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision 9c6d26df1fae6ad4718d51c48e6517913304ed27)
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 					  MLX4_IB_RX_HASH_DST_IPV4	|
671 					  MLX4_IB_RX_HASH_SRC_IPV6	|
672 					  MLX4_IB_RX_HASH_DST_IPV6	|
673 					  MLX4_IB_RX_HASH_SRC_PORT_TCP	|
674 					  MLX4_IB_RX_HASH_DST_PORT_TCP	|
675 					  MLX4_IB_RX_HASH_SRC_PORT_UDP	|
676 					  MLX4_IB_RX_HASH_DST_PORT_UDP)) {
677 		pr_debug("RX Hash fields_mask has unsupported mask (0x%llx)\n",
678 			 ucmd->rx_hash_fields_mask);
679 		return (-EOPNOTSUPP);
680 	}
681 
682 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) &&
683 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
684 		rss_ctx->flags = MLX4_RSS_IPV4;
685 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) ||
686 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
687 		pr_debug("RX Hash fields_mask is not supported - both IPv4 SRC and DST must be set\n");
688 		return (-EOPNOTSUPP);
689 	}
690 
691 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) &&
692 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
693 		rss_ctx->flags |= MLX4_RSS_IPV6;
694 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) ||
695 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
696 		pr_debug("RX Hash fields_mask is not supported - both IPv6 SRC and DST must be set\n");
697 		return (-EOPNOTSUPP);
698 	}
699 
700 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) &&
701 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
702 		if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS)) {
703 			pr_debug("RX Hash fields_mask for UDP is not supported\n");
704 			return (-EOPNOTSUPP);
705 		}
706 
707 		if (rss_ctx->flags & MLX4_RSS_IPV4)
708 			rss_ctx->flags |= MLX4_RSS_UDP_IPV4;
709 		if (rss_ctx->flags & MLX4_RSS_IPV6)
710 			rss_ctx->flags |= MLX4_RSS_UDP_IPV6;
711 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
712 			pr_debug("RX Hash fields_mask is not supported - UDP must be set with IPv4 or IPv6\n");
713 			return (-EOPNOTSUPP);
714 		}
715 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) ||
716 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
717 		pr_debug("RX Hash fields_mask is not supported - both UDP SRC and DST must be set\n");
718 		return (-EOPNOTSUPP);
719 	}
720 
721 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) &&
722 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
723 		if (rss_ctx->flags & MLX4_RSS_IPV4)
724 			rss_ctx->flags |= MLX4_RSS_TCP_IPV4;
725 		if (rss_ctx->flags & MLX4_RSS_IPV6)
726 			rss_ctx->flags |= MLX4_RSS_TCP_IPV6;
727 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
728 			pr_debug("RX Hash fields_mask is not supported - TCP must be set with IPv4 or IPv6\n");
729 			return (-EOPNOTSUPP);
730 		}
731 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) ||
732 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
733 		pr_debug("RX Hash fields_mask is not supported - both TCP SRC and DST must be set\n");
734 		return (-EOPNOTSUPP);
735 	}
736 
737 	if (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_INNER) {
738 		if (dev->dev->caps.tunnel_offload_mode ==
739 		    MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
740 			/*
741 			 * Hash according to inner headers if exist, otherwise
742 			 * according to outer headers.
743 			 */
744 			rss_ctx->flags |= MLX4_RSS_BY_INNER_HEADERS_IPONLY;
745 		} else {
746 			pr_debug("RSS Hash for inner headers isn't supported\n");
747 			return (-EOPNOTSUPP);
748 		}
749 	}
750 
751 	return 0;
752 }
753 
754 static int create_qp_rss(struct mlx4_ib_dev *dev,
755 			 struct ib_qp_init_attr *init_attr,
756 			 struct mlx4_ib_create_qp_rss *ucmd,
757 			 struct mlx4_ib_qp *qp)
758 {
759 	int qpn;
760 	int err;
761 
762 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
763 
764 	err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 0, qp->mqp.usage);
765 	if (err)
766 		return err;
767 
768 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
769 	if (err)
770 		goto err_qpn;
771 
772 	mutex_init(&qp->mutex);
773 
774 	INIT_LIST_HEAD(&qp->gid_list);
775 	INIT_LIST_HEAD(&qp->steering_rules);
776 
777 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
778 	qp->state = IB_QPS_RESET;
779 
780 	/* Set dummy send resources to be compatible with HV and PRM */
781 	qp->sq_no_prefetch = 1;
782 	qp->sq.wqe_cnt = 1;
783 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
784 	qp->buf_size = qp->sq.wqe_cnt << MLX4_IB_MIN_SQ_STRIDE;
785 	qp->mtt = (to_mqp(
786 		   (struct ib_qp *)init_attr->rwq_ind_tbl->ind_tbl[0]))->mtt;
787 
788 	qp->rss_ctx = kzalloc(sizeof(*qp->rss_ctx), GFP_KERNEL);
789 	if (!qp->rss_ctx) {
790 		err = -ENOMEM;
791 		goto err_qp_alloc;
792 	}
793 
794 	err = set_qp_rss(dev, qp->rss_ctx, init_attr, ucmd);
795 	if (err)
796 		goto err;
797 
798 	return 0;
799 
800 err:
801 	kfree(qp->rss_ctx);
802 
803 err_qp_alloc:
804 	mlx4_qp_remove(dev->dev, &qp->mqp);
805 	mlx4_qp_free(dev->dev, &qp->mqp);
806 
807 err_qpn:
808 	mlx4_qp_release_range(dev->dev, qpn, 1);
809 	return err;
810 }
811 
812 static struct ib_qp *_mlx4_ib_create_qp_rss(struct ib_pd *pd,
813 					    struct ib_qp_init_attr *init_attr,
814 					    struct ib_udata *udata)
815 {
816 	struct mlx4_ib_qp *qp;
817 	struct mlx4_ib_create_qp_rss ucmd = {};
818 	size_t required_cmd_sz;
819 	int err;
820 
821 	if (!udata) {
822 		pr_debug("RSS QP with NULL udata\n");
823 		return ERR_PTR(-EINVAL);
824 	}
825 
826 	if (udata->outlen)
827 		return ERR_PTR(-EOPNOTSUPP);
828 
829 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) +
830 					sizeof(ucmd.reserved1);
831 	if (udata->inlen < required_cmd_sz) {
832 		pr_debug("invalid inlen\n");
833 		return ERR_PTR(-EINVAL);
834 	}
835 
836 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
837 		pr_debug("copy failed\n");
838 		return ERR_PTR(-EFAULT);
839 	}
840 
841 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)))
842 		return ERR_PTR(-EOPNOTSUPP);
843 
844 	if (ucmd.comp_mask || ucmd.reserved1)
845 		return ERR_PTR(-EOPNOTSUPP);
846 
847 	if (udata->inlen > sizeof(ucmd) &&
848 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
849 				 udata->inlen - sizeof(ucmd))) {
850 		pr_debug("inlen is not supported\n");
851 		return ERR_PTR(-EOPNOTSUPP);
852 	}
853 
854 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
855 		pr_debug("RSS QP with unsupported QP type %d\n",
856 			 init_attr->qp_type);
857 		return ERR_PTR(-EOPNOTSUPP);
858 	}
859 
860 	if (init_attr->create_flags) {
861 		pr_debug("RSS QP doesn't support create flags\n");
862 		return ERR_PTR(-EOPNOTSUPP);
863 	}
864 
865 	if (init_attr->send_cq || init_attr->cap.max_send_wr) {
866 		pr_debug("RSS QP with unsupported send attributes\n");
867 		return ERR_PTR(-EOPNOTSUPP);
868 	}
869 
870 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
871 	if (!qp)
872 		return ERR_PTR(-ENOMEM);
873 
874 	qp->pri.vid = 0xFFFF;
875 	qp->alt.vid = 0xFFFF;
876 
877 	err = create_qp_rss(to_mdev(pd->device), init_attr, &ucmd, qp);
878 	if (err) {
879 		kfree(qp);
880 		return ERR_PTR(err);
881 	}
882 
883 	qp->ibqp.qp_num = qp->mqp.qpn;
884 
885 	return &qp->ibqp;
886 }
887 
888 /*
889  * This function allocates a WQN from a range which is consecutive and aligned
890  * to its size. In case the range is full, then it creates a new range and
891  * allocates WQN from it. The new range will be used for following allocations.
892  */
893 static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context,
894 			     struct mlx4_ib_qp *qp, int range_size, int *wqn)
895 {
896 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
897 	struct mlx4_wqn_range *range;
898 	int err = 0;
899 
900 	mutex_lock(&context->wqn_ranges_mutex);
901 
902 	range = list_first_entry_or_null(&context->wqn_ranges_list,
903 					 struct mlx4_wqn_range, list);
904 
905 	if (!range || (range->refcount == range->size) || range->dirty) {
906 		range = kzalloc(sizeof(*range), GFP_KERNEL);
907 		if (!range) {
908 			err = -ENOMEM;
909 			goto out;
910 		}
911 
912 		err = mlx4_qp_reserve_range(dev->dev, range_size,
913 					    range_size, &range->base_wqn, 0,
914 					    qp->mqp.usage);
915 		if (err) {
916 			kfree(range);
917 			goto out;
918 		}
919 
920 		range->size = range_size;
921 		list_add(&range->list, &context->wqn_ranges_list);
922 	} else if (range_size != 1) {
923 		/*
924 		 * Requesting a new range (>1) when last range is still open, is
925 		 * not valid.
926 		 */
927 		err = -EINVAL;
928 		goto out;
929 	}
930 
931 	qp->wqn_range = range;
932 
933 	*wqn = range->base_wqn + range->refcount;
934 
935 	range->refcount++;
936 
937 out:
938 	mutex_unlock(&context->wqn_ranges_mutex);
939 
940 	return err;
941 }
942 
943 static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context,
944 				struct mlx4_ib_qp *qp, bool dirty_release)
945 {
946 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
947 	struct mlx4_wqn_range *range;
948 
949 	mutex_lock(&context->wqn_ranges_mutex);
950 
951 	range = qp->wqn_range;
952 
953 	range->refcount--;
954 	if (!range->refcount) {
955 		mlx4_qp_release_range(dev->dev, range->base_wqn,
956 				      range->size);
957 		list_del(&range->list);
958 		kfree(range);
959 	} else if (dirty_release) {
960 	/*
961 	 * A range which one of its WQNs is destroyed, won't be able to be
962 	 * reused for further WQN allocations.
963 	 * The next created WQ will allocate a new range.
964 	 */
965 		range->dirty = 1;
966 	}
967 
968 	mutex_unlock(&context->wqn_ranges_mutex);
969 }
970 
971 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
972 			    enum mlx4_ib_source_type src,
973 			    struct ib_qp_init_attr *init_attr,
974 			    struct ib_udata *udata, int sqpn,
975 			    struct mlx4_ib_qp **caller_qp)
976 {
977 	int qpn;
978 	int err;
979 	struct ib_qp_cap backup_cap;
980 	struct mlx4_ib_sqp *sqp = NULL;
981 	struct mlx4_ib_qp *qp;
982 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
983 	struct mlx4_ib_cq *mcq;
984 	unsigned long flags;
985 	int range_size = 0;
986 
987 	/* When tunneling special qps, we use a plain UD qp */
988 	if (sqpn) {
989 		if (mlx4_is_mfunc(dev->dev) &&
990 		    (!mlx4_is_master(dev->dev) ||
991 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
992 			if (init_attr->qp_type == IB_QPT_GSI)
993 				qp_type = MLX4_IB_QPT_PROXY_GSI;
994 			else {
995 				if (mlx4_is_master(dev->dev) ||
996 				    qp0_enabled_vf(dev->dev, sqpn))
997 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
998 				else
999 					qp_type = MLX4_IB_QPT_PROXY_SMI;
1000 			}
1001 		}
1002 		qpn = sqpn;
1003 		/* add extra sg entry for tunneling */
1004 		init_attr->cap.max_recv_sge++;
1005 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
1006 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
1007 			container_of(init_attr,
1008 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
1009 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
1010 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
1011 		    !mlx4_is_master(dev->dev))
1012 			return -EINVAL;
1013 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
1014 			qp_type = MLX4_IB_QPT_TUN_GSI;
1015 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
1016 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
1017 					     tnl_init->port))
1018 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
1019 		else
1020 			qp_type = MLX4_IB_QPT_TUN_SMI;
1021 		/* we are definitely in the PPF here, since we are creating
1022 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
1023 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
1024 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
1025 		sqpn = qpn;
1026 	}
1027 
1028 	if (!*caller_qp) {
1029 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
1030 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
1031 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
1032 			sqp = kzalloc(sizeof(struct mlx4_ib_sqp), GFP_KERNEL);
1033 			if (!sqp)
1034 				return -ENOMEM;
1035 			qp = &sqp->qp;
1036 			qp->pri.vid = 0xFFFF;
1037 			qp->alt.vid = 0xFFFF;
1038 		} else {
1039 			qp = kzalloc(sizeof(struct mlx4_ib_qp), GFP_KERNEL);
1040 			if (!qp)
1041 				return -ENOMEM;
1042 			qp->pri.vid = 0xFFFF;
1043 			qp->alt.vid = 0xFFFF;
1044 		}
1045 	} else
1046 		qp = *caller_qp;
1047 
1048 	qp->mlx4_ib_qp_type = qp_type;
1049 
1050 	mutex_init(&qp->mutex);
1051 	spin_lock_init(&qp->sq.lock);
1052 	spin_lock_init(&qp->rq.lock);
1053 	INIT_LIST_HEAD(&qp->gid_list);
1054 	INIT_LIST_HEAD(&qp->steering_rules);
1055 
1056 	qp->state	 = IB_QPS_RESET;
1057 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1058 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1059 
1060 
1061 	if (pd->uobject) {
1062 		union {
1063 			struct mlx4_ib_create_qp qp;
1064 			struct mlx4_ib_create_wq wq;
1065 		} ucmd;
1066 		size_t copy_len;
1067 		int shift;
1068 		int n;
1069 
1070 		copy_len = (src == MLX4_IB_QP_SRC) ?
1071 			   sizeof(struct mlx4_ib_create_qp) :
1072 			   min(sizeof(struct mlx4_ib_create_wq), udata->inlen);
1073 
1074 		if (ib_copy_from_udata(&ucmd, udata, copy_len)) {
1075 			err = -EFAULT;
1076 			goto err;
1077 		}
1078 
1079 		if (src == MLX4_IB_RWQ_SRC) {
1080 			if (ucmd.wq.comp_mask || ucmd.wq.reserved[0] ||
1081 			    ucmd.wq.reserved[1] || ucmd.wq.reserved[2]) {
1082 				pr_debug("user command isn't supported\n");
1083 				err = -EOPNOTSUPP;
1084 				goto err;
1085 			}
1086 
1087 			if (ucmd.wq.log_range_size >
1088 			    ilog2(dev->dev->caps.max_rss_tbl_sz)) {
1089 				pr_debug("WQN range size must be equal or smaller than %d\n",
1090 					 dev->dev->caps.max_rss_tbl_sz);
1091 				err = -EOPNOTSUPP;
1092 				goto err;
1093 			}
1094 			range_size = 1 << ucmd.wq.log_range_size;
1095 		} else {
1096 			qp->inl_recv_sz = ucmd.qp.inl_recv_sz;
1097 		}
1098 
1099 		if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1100 			if (!(dev->dev->caps.flags &
1101 			      MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
1102 				pr_debug("scatter FCS is unsupported\n");
1103 				err = -EOPNOTSUPP;
1104 				goto err;
1105 			}
1106 
1107 			qp->flags |= MLX4_IB_QP_SCATTER_FCS;
1108 		}
1109 
1110 		err = set_rq_size(dev, &init_attr->cap, !!pd->uobject,
1111 				  qp_has_rq(init_attr), qp, qp->inl_recv_sz);
1112 		if (err)
1113 			goto err;
1114 
1115 		if (src == MLX4_IB_QP_SRC) {
1116 			qp->sq_no_prefetch = ucmd.qp.sq_no_prefetch;
1117 
1118 			err = set_user_sq_size(dev, qp,
1119 					       (struct mlx4_ib_create_qp *)
1120 					       &ucmd);
1121 			if (err)
1122 				goto err;
1123 		} else {
1124 			qp->sq_no_prefetch = 1;
1125 			qp->sq.wqe_cnt = 1;
1126 			qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
1127 			/* Allocated buffer expects to have at least that SQ
1128 			 * size.
1129 			 */
1130 			qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
1131 				(qp->sq.wqe_cnt << qp->sq.wqe_shift);
1132 		}
1133 
1134 		qp->umem = ib_umem_get(pd->uobject->context,
1135 				(src == MLX4_IB_QP_SRC) ? ucmd.qp.buf_addr :
1136 				ucmd.wq.buf_addr, qp->buf_size, 0, 0);
1137 		if (IS_ERR(qp->umem)) {
1138 			err = PTR_ERR(qp->umem);
1139 			goto err;
1140 		}
1141 
1142 		n = ib_umem_page_count(qp->umem);
1143 		shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
1144 		err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
1145 
1146 		if (err)
1147 			goto err_buf;
1148 
1149 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
1150 		if (err)
1151 			goto err_mtt;
1152 
1153 		if (qp_has_rq(init_attr)) {
1154 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
1155 				(src == MLX4_IB_QP_SRC) ? ucmd.qp.db_addr :
1156 				ucmd.wq.db_addr, &qp->db);
1157 			if (err)
1158 				goto err_mtt;
1159 		}
1160 		qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
1161 	} else {
1162 		err = set_rq_size(dev, &init_attr->cap, !!pd->uobject,
1163 				  qp_has_rq(init_attr), qp, 0);
1164 		if (err)
1165 			goto err;
1166 
1167 		qp->sq_no_prefetch = 0;
1168 
1169 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
1170 			qp->flags |= MLX4_IB_QP_LSO;
1171 
1172 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1173 			if (dev->steering_support ==
1174 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
1175 				qp->flags |= MLX4_IB_QP_NETIF;
1176 			else
1177 				goto err;
1178 		}
1179 
1180 		memcpy(&backup_cap, &init_attr->cap, sizeof(backup_cap));
1181 		err = set_kernel_sq_size(dev, &init_attr->cap,
1182 					 qp_type, qp, true);
1183 		if (err)
1184 			goto err;
1185 
1186 		if (qp_has_rq(init_attr)) {
1187 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
1188 			if (err)
1189 				goto err;
1190 
1191 			*qp->db.db = 0;
1192 		}
1193 
1194 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, qp->buf_size,
1195 				   &qp->buf)) {
1196 			memcpy(&init_attr->cap, &backup_cap,
1197 			       sizeof(backup_cap));
1198 			err = set_kernel_sq_size(dev, &init_attr->cap, qp_type,
1199 						 qp, false);
1200 			if (err)
1201 				goto err_db;
1202 
1203 			if (mlx4_buf_alloc(dev->dev, qp->buf_size,
1204 					   PAGE_SIZE * 2, &qp->buf)) {
1205 				err = -ENOMEM;
1206 				goto err_db;
1207 			}
1208 		}
1209 
1210 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
1211 				    &qp->mtt);
1212 		if (err)
1213 			goto err_buf;
1214 
1215 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
1216 		if (err)
1217 			goto err_mtt;
1218 
1219 		qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1220 					     sizeof(u64), GFP_KERNEL);
1221 		qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1222 					     sizeof(u64), GFP_KERNEL);
1223 		if (!qp->sq.wrid || !qp->rq.wrid) {
1224 			err = -ENOMEM;
1225 			goto err_wrid;
1226 		}
1227 		qp->mqp.usage = MLX4_RES_USAGE_DRIVER;
1228 	}
1229 
1230 	if (sqpn) {
1231 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1232 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
1233 			if (alloc_proxy_bufs(pd->device, qp)) {
1234 				err = -ENOMEM;
1235 				goto err_wrid;
1236 			}
1237 		}
1238 	} else if (src == MLX4_IB_RWQ_SRC) {
1239 		err = mlx4_ib_alloc_wqn(to_mucontext(pd->uobject->context), qp,
1240 					range_size, &qpn);
1241 		if (err)
1242 			goto err_wrid;
1243 	} else {
1244 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
1245 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
1246 		 * VLAN insertion. */
1247 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
1248 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
1249 						    (init_attr->cap.max_send_wr ?
1250 						     MLX4_RESERVE_ETH_BF_QP : 0) |
1251 						    (init_attr->cap.max_recv_wr ?
1252 						     MLX4_RESERVE_A0_QP : 0),
1253 						    qp->mqp.usage);
1254 		else
1255 			if (qp->flags & MLX4_IB_QP_NETIF)
1256 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
1257 			else
1258 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
1259 							    &qpn, 0, qp->mqp.usage);
1260 		if (err)
1261 			goto err_proxy;
1262 	}
1263 
1264 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1265 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1266 
1267 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
1268 	if (err)
1269 		goto err_qpn;
1270 
1271 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1272 		qp->mqp.qpn |= (1 << 23);
1273 
1274 	/*
1275 	 * Hardware wants QPN written in big-endian order (after
1276 	 * shifting) for send doorbell.  Precompute this value to save
1277 	 * a little bit when posting sends.
1278 	 */
1279 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1280 
1281 	qp->mqp.event = (src == MLX4_IB_QP_SRC) ? mlx4_ib_qp_event :
1282 						  mlx4_ib_wq_event;
1283 
1284 	if (!*caller_qp)
1285 		*caller_qp = qp;
1286 
1287 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1288 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
1289 			 to_mcq(init_attr->recv_cq));
1290 	/* Maintain device to QPs access, needed for further handling
1291 	 * via reset flow
1292 	 */
1293 	list_add_tail(&qp->qps_list, &dev->qp_list);
1294 	/* Maintain CQ to QPs access, needed for further handling
1295 	 * via reset flow
1296 	 */
1297 	mcq = to_mcq(init_attr->send_cq);
1298 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
1299 	mcq = to_mcq(init_attr->recv_cq);
1300 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
1301 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
1302 			   to_mcq(init_attr->recv_cq));
1303 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1304 	return 0;
1305 
1306 err_qpn:
1307 	if (!sqpn) {
1308 		if (qp->flags & MLX4_IB_QP_NETIF)
1309 			mlx4_ib_steer_qp_free(dev, qpn, 1);
1310 		else if (src == MLX4_IB_RWQ_SRC)
1311 			mlx4_ib_release_wqn(to_mucontext(pd->uobject->context),
1312 					    qp, 0);
1313 		else
1314 			mlx4_qp_release_range(dev->dev, qpn, 1);
1315 	}
1316 err_proxy:
1317 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1318 		free_proxy_bufs(pd->device, qp);
1319 err_wrid:
1320 	if (pd->uobject) {
1321 		if (qp_has_rq(init_attr))
1322 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
1323 	} else {
1324 		kvfree(qp->sq.wrid);
1325 		kvfree(qp->rq.wrid);
1326 	}
1327 
1328 err_mtt:
1329 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1330 
1331 err_buf:
1332 	if (pd->uobject)
1333 		ib_umem_release(qp->umem);
1334 	else
1335 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1336 
1337 err_db:
1338 	if (!pd->uobject && qp_has_rq(init_attr))
1339 		mlx4_db_free(dev->dev, &qp->db);
1340 
1341 err:
1342 	if (sqp)
1343 		kfree(sqp);
1344 	else if (!*caller_qp)
1345 		kfree(qp);
1346 	return err;
1347 }
1348 
1349 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
1350 {
1351 	switch (state) {
1352 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
1353 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
1354 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
1355 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
1356 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
1357 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
1358 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
1359 	default:		return -1;
1360 	}
1361 }
1362 
1363 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1364 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1365 {
1366 	if (send_cq == recv_cq) {
1367 		spin_lock(&send_cq->lock);
1368 		__acquire(&recv_cq->lock);
1369 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1370 		spin_lock(&send_cq->lock);
1371 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1372 	} else {
1373 		spin_lock(&recv_cq->lock);
1374 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1375 	}
1376 }
1377 
1378 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1379 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1380 {
1381 	if (send_cq == recv_cq) {
1382 		__release(&recv_cq->lock);
1383 		spin_unlock(&send_cq->lock);
1384 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1385 		spin_unlock(&recv_cq->lock);
1386 		spin_unlock(&send_cq->lock);
1387 	} else {
1388 		spin_unlock(&send_cq->lock);
1389 		spin_unlock(&recv_cq->lock);
1390 	}
1391 }
1392 
1393 static void del_gid_entries(struct mlx4_ib_qp *qp)
1394 {
1395 	struct mlx4_ib_gid_entry *ge, *tmp;
1396 
1397 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1398 		list_del(&ge->list);
1399 		kfree(ge);
1400 	}
1401 }
1402 
1403 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
1404 {
1405 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
1406 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
1407 	else
1408 		return to_mpd(qp->ibqp.pd);
1409 }
1410 
1411 static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src,
1412 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1413 {
1414 	switch (qp->ibqp.qp_type) {
1415 	case IB_QPT_XRC_TGT:
1416 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1417 		*recv_cq = *send_cq;
1418 		break;
1419 	case IB_QPT_XRC_INI:
1420 		*send_cq = to_mcq(qp->ibqp.send_cq);
1421 		*recv_cq = *send_cq;
1422 		break;
1423 	default:
1424 		*recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) :
1425 						     to_mcq(qp->ibwq.cq);
1426 		*send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) :
1427 						     *recv_cq;
1428 		break;
1429 	}
1430 }
1431 
1432 static void destroy_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1433 {
1434 	if (qp->state != IB_QPS_RESET) {
1435 		int i;
1436 
1437 		for (i = 0; i < (1 << qp->ibqp.rwq_ind_tbl->log_ind_tbl_size);
1438 		     i++) {
1439 			struct ib_wq *ibwq = qp->ibqp.rwq_ind_tbl->ind_tbl[i];
1440 			struct mlx4_ib_qp *wq =	to_mqp((struct ib_qp *)ibwq);
1441 
1442 			mutex_lock(&wq->mutex);
1443 
1444 			wq->rss_usecnt--;
1445 
1446 			mutex_unlock(&wq->mutex);
1447 		}
1448 
1449 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1450 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1451 			pr_warn("modify QP %06x to RESET failed.\n",
1452 				qp->mqp.qpn);
1453 	}
1454 
1455 	mlx4_qp_remove(dev->dev, &qp->mqp);
1456 	mlx4_qp_free(dev->dev, &qp->mqp);
1457 	mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1458 	del_gid_entries(qp);
1459 	kfree(qp->rss_ctx);
1460 }
1461 
1462 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1463 			      enum mlx4_ib_source_type src, int is_user)
1464 {
1465 	struct mlx4_ib_cq *send_cq, *recv_cq;
1466 	unsigned long flags;
1467 
1468 	if (qp->state != IB_QPS_RESET) {
1469 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1470 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1471 			pr_warn("modify QP %06x to RESET failed.\n",
1472 			       qp->mqp.qpn);
1473 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1474 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1475 			qp->pri.smac = 0;
1476 			qp->pri.smac_port = 0;
1477 		}
1478 		if (qp->alt.smac) {
1479 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1480 			qp->alt.smac = 0;
1481 		}
1482 		if (qp->pri.vid < 0x1000) {
1483 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1484 			qp->pri.vid = 0xFFFF;
1485 			qp->pri.candidate_vid = 0xFFFF;
1486 			qp->pri.update_vid = 0;
1487 		}
1488 		if (qp->alt.vid < 0x1000) {
1489 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1490 			qp->alt.vid = 0xFFFF;
1491 			qp->alt.candidate_vid = 0xFFFF;
1492 			qp->alt.update_vid = 0;
1493 		}
1494 	}
1495 
1496 	get_cqs(qp, src, &send_cq, &recv_cq);
1497 
1498 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1499 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1500 
1501 	/* del from lists under both locks above to protect reset flow paths */
1502 	list_del(&qp->qps_list);
1503 	list_del(&qp->cq_send_list);
1504 	list_del(&qp->cq_recv_list);
1505 	if (!is_user) {
1506 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1507 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1508 		if (send_cq != recv_cq)
1509 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1510 	}
1511 
1512 	mlx4_qp_remove(dev->dev, &qp->mqp);
1513 
1514 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1515 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1516 
1517 	mlx4_qp_free(dev->dev, &qp->mqp);
1518 
1519 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1520 		if (qp->flags & MLX4_IB_QP_NETIF)
1521 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1522 		else if (src == MLX4_IB_RWQ_SRC)
1523 			mlx4_ib_release_wqn(to_mucontext(
1524 					    qp->ibwq.uobject->context), qp, 1);
1525 		else
1526 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1527 	}
1528 
1529 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1530 
1531 	if (is_user) {
1532 		if (qp->rq.wqe_cnt) {
1533 			struct mlx4_ib_ucontext *mcontext = !src ?
1534 				to_mucontext(qp->ibqp.uobject->context) :
1535 				to_mucontext(qp->ibwq.uobject->context);
1536 			mlx4_ib_db_unmap_user(mcontext, &qp->db);
1537 		}
1538 		ib_umem_release(qp->umem);
1539 	} else {
1540 		kvfree(qp->sq.wrid);
1541 		kvfree(qp->rq.wrid);
1542 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1543 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1544 			free_proxy_bufs(&dev->ib_dev, qp);
1545 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1546 		if (qp->rq.wqe_cnt)
1547 			mlx4_db_free(dev->dev, &qp->db);
1548 	}
1549 
1550 	del_gid_entries(qp);
1551 }
1552 
1553 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1554 {
1555 	/* Native or PPF */
1556 	if (!mlx4_is_mfunc(dev->dev) ||
1557 	    (mlx4_is_master(dev->dev) &&
1558 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1559 		return  dev->dev->phys_caps.base_sqpn +
1560 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1561 			attr->port_num - 1;
1562 	}
1563 	/* PF or VF -- creating proxies */
1564 	if (attr->qp_type == IB_QPT_SMI)
1565 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp0_proxy;
1566 	else
1567 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp1_proxy;
1568 }
1569 
1570 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1571 					struct ib_qp_init_attr *init_attr,
1572 					struct ib_udata *udata)
1573 {
1574 	struct mlx4_ib_qp *qp = NULL;
1575 	int err;
1576 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1577 	u16 xrcdn = 0;
1578 
1579 	if (init_attr->rwq_ind_tbl)
1580 		return _mlx4_ib_create_qp_rss(pd, init_attr, udata);
1581 
1582 	/*
1583 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1584 	 * and only for kernel UD QPs.
1585 	 */
1586 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1587 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1588 					MLX4_IB_SRIOV_TUNNEL_QP |
1589 					MLX4_IB_SRIOV_SQP |
1590 					MLX4_IB_QP_NETIF |
1591 					MLX4_IB_QP_CREATE_ROCE_V2_GSI))
1592 		return ERR_PTR(-EINVAL);
1593 
1594 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1595 		if (init_attr->qp_type != IB_QPT_UD)
1596 			return ERR_PTR(-EINVAL);
1597 	}
1598 
1599 	if (init_attr->create_flags) {
1600 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1601 			return ERR_PTR(-EINVAL);
1602 
1603 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1604 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1605 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1606 		     init_attr->qp_type != IB_QPT_UD) ||
1607 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1608 		     init_attr->qp_type > IB_QPT_GSI) ||
1609 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1610 		     init_attr->qp_type != IB_QPT_GSI))
1611 			return ERR_PTR(-EINVAL);
1612 	}
1613 
1614 	switch (init_attr->qp_type) {
1615 	case IB_QPT_XRC_TGT:
1616 		pd = to_mxrcd(init_attr->xrcd)->pd;
1617 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1618 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1619 		/* fall through */
1620 	case IB_QPT_XRC_INI:
1621 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1622 			return ERR_PTR(-ENOSYS);
1623 		init_attr->recv_cq = init_attr->send_cq;
1624 		/* fall through */
1625 	case IB_QPT_RC:
1626 	case IB_QPT_UC:
1627 	case IB_QPT_RAW_PACKET:
1628 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1629 		if (!qp)
1630 			return ERR_PTR(-ENOMEM);
1631 		qp->pri.vid = 0xFFFF;
1632 		qp->alt.vid = 0xFFFF;
1633 		/* fall through */
1634 	case IB_QPT_UD:
1635 	{
1636 		err = create_qp_common(to_mdev(pd->device), pd,	MLX4_IB_QP_SRC,
1637 				       init_attr, udata, 0, &qp);
1638 		if (err) {
1639 			kfree(qp);
1640 			return ERR_PTR(err);
1641 		}
1642 
1643 		qp->ibqp.qp_num = qp->mqp.qpn;
1644 		qp->xrcdn = xrcdn;
1645 
1646 		break;
1647 	}
1648 	case IB_QPT_SMI:
1649 	case IB_QPT_GSI:
1650 	{
1651 		int sqpn;
1652 
1653 		/* Userspace is not allowed to create special QPs: */
1654 		if (udata)
1655 			return ERR_PTR(-EINVAL);
1656 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1657 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev,
1658 							1, 1, &sqpn, 0,
1659 							MLX4_RES_USAGE_DRIVER);
1660 
1661 			if (res)
1662 				return ERR_PTR(res);
1663 		} else {
1664 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1665 		}
1666 
1667 		err = create_qp_common(to_mdev(pd->device), pd, MLX4_IB_QP_SRC,
1668 				       init_attr, udata, sqpn, &qp);
1669 		if (err)
1670 			return ERR_PTR(err);
1671 
1672 		qp->port	= init_attr->port_num;
1673 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1674 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1675 		break;
1676 	}
1677 	default:
1678 		/* Don't support raw QPs */
1679 		return ERR_PTR(-EINVAL);
1680 	}
1681 
1682 	return &qp->ibqp;
1683 }
1684 
1685 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1686 				struct ib_qp_init_attr *init_attr,
1687 				struct ib_udata *udata) {
1688 	struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1689 	struct ib_qp *ibqp;
1690 	struct mlx4_ib_dev *dev = to_mdev(device);
1691 
1692 	ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1693 
1694 	if (!IS_ERR(ibqp) &&
1695 	    (init_attr->qp_type == IB_QPT_GSI) &&
1696 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1697 		struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1698 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1699 
1700 		if (is_eth &&
1701 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1702 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1703 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1704 
1705 			if (IS_ERR(sqp->roce_v2_gsi)) {
1706 				pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1707 				sqp->roce_v2_gsi = NULL;
1708 			} else {
1709 				sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1710 				sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1711 			}
1712 
1713 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1714 		}
1715 	}
1716 	return ibqp;
1717 }
1718 
1719 static int _mlx4_ib_destroy_qp(struct ib_qp *qp)
1720 {
1721 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1722 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1723 
1724 	if (is_qp0(dev, mqp))
1725 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1726 
1727 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1728 	    dev->qp1_proxy[mqp->port - 1] == mqp) {
1729 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1730 		dev->qp1_proxy[mqp->port - 1] = NULL;
1731 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1732 	}
1733 
1734 	if (mqp->counter_index)
1735 		mlx4_ib_free_qp_counter(dev, mqp);
1736 
1737 	if (qp->rwq_ind_tbl) {
1738 		destroy_qp_rss(dev, mqp);
1739 	} else {
1740 		struct mlx4_ib_pd *pd;
1741 
1742 		pd = get_pd(mqp);
1743 		destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, !!pd->ibpd.uobject);
1744 	}
1745 
1746 	if (is_sqp(dev, mqp))
1747 		kfree(to_msqp(mqp));
1748 	else
1749 		kfree(mqp);
1750 
1751 	return 0;
1752 }
1753 
1754 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1755 {
1756 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1757 
1758 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1759 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1760 
1761 		if (sqp->roce_v2_gsi)
1762 			ib_destroy_qp(sqp->roce_v2_gsi);
1763 	}
1764 
1765 	return _mlx4_ib_destroy_qp(qp);
1766 }
1767 
1768 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1769 {
1770 	switch (type) {
1771 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1772 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1773 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1774 	case MLX4_IB_QPT_XRC_INI:
1775 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1776 	case MLX4_IB_QPT_SMI:
1777 	case MLX4_IB_QPT_GSI:
1778 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1779 
1780 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1781 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1782 						MLX4_QP_ST_MLX : -1);
1783 	case MLX4_IB_QPT_PROXY_SMI:
1784 	case MLX4_IB_QPT_TUN_SMI:
1785 	case MLX4_IB_QPT_PROXY_GSI:
1786 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1787 						MLX4_QP_ST_UD : -1);
1788 	default:			return -1;
1789 	}
1790 }
1791 
1792 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1793 				   int attr_mask)
1794 {
1795 	u8 dest_rd_atomic;
1796 	u32 access_flags;
1797 	u32 hw_access_flags = 0;
1798 
1799 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1800 		dest_rd_atomic = attr->max_dest_rd_atomic;
1801 	else
1802 		dest_rd_atomic = qp->resp_depth;
1803 
1804 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1805 		access_flags = attr->qp_access_flags;
1806 	else
1807 		access_flags = qp->atomic_rd_en;
1808 
1809 	if (!dest_rd_atomic)
1810 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1811 
1812 	if (access_flags & IB_ACCESS_REMOTE_READ)
1813 		hw_access_flags |= MLX4_QP_BIT_RRE;
1814 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1815 		hw_access_flags |= MLX4_QP_BIT_RAE;
1816 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1817 		hw_access_flags |= MLX4_QP_BIT_RWE;
1818 
1819 	return cpu_to_be32(hw_access_flags);
1820 }
1821 
1822 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1823 			    int attr_mask)
1824 {
1825 	if (attr_mask & IB_QP_PKEY_INDEX)
1826 		sqp->pkey_index = attr->pkey_index;
1827 	if (attr_mask & IB_QP_QKEY)
1828 		sqp->qkey = attr->qkey;
1829 	if (attr_mask & IB_QP_SQ_PSN)
1830 		sqp->send_psn = attr->sq_psn;
1831 }
1832 
1833 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1834 {
1835 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1836 }
1837 
1838 static int _mlx4_set_path(struct mlx4_ib_dev *dev,
1839 			  const struct rdma_ah_attr *ah,
1840 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1841 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1842 {
1843 	int vidx;
1844 	int smac_index;
1845 	int err;
1846 
1847 	path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
1848 	path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
1849 	if (rdma_ah_get_static_rate(ah)) {
1850 		path->static_rate = rdma_ah_get_static_rate(ah) +
1851 				    MLX4_STAT_RATE_OFFSET;
1852 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1853 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1854 			--path->static_rate;
1855 	} else
1856 		path->static_rate = 0;
1857 
1858 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
1859 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
1860 		int real_sgid_index =
1861 			mlx4_ib_gid_index_to_real_index(dev, port,
1862 							grh->sgid_index);
1863 
1864 		if (real_sgid_index < 0)
1865 			return real_sgid_index;
1866 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1867 			pr_err("sgid_index (%u) too large. max is %d\n",
1868 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1869 			return -1;
1870 		}
1871 
1872 		path->grh_mylmc |= 1 << 7;
1873 		path->mgid_index = real_sgid_index;
1874 		path->hop_limit  = grh->hop_limit;
1875 		path->tclass_flowlabel =
1876 			cpu_to_be32((grh->traffic_class << 20) |
1877 				    (grh->flow_label));
1878 		memcpy(path->rgid, grh->dgid.raw, 16);
1879 	}
1880 
1881 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
1882 		if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
1883 			return -1;
1884 
1885 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1886 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
1887 
1888 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1889 		if (vlan_tag < 0x1000) {
1890 			if (smac_info->vid < 0x1000) {
1891 				/* both valid vlan ids */
1892 				if (smac_info->vid != vlan_tag) {
1893 					/* different VIDs.  unreg old and reg new */
1894 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1895 					if (err)
1896 						return err;
1897 					smac_info->candidate_vid = vlan_tag;
1898 					smac_info->candidate_vlan_index = vidx;
1899 					smac_info->candidate_vlan_port = port;
1900 					smac_info->update_vid = 1;
1901 					path->vlan_index = vidx;
1902 				} else {
1903 					path->vlan_index = smac_info->vlan_index;
1904 				}
1905 			} else {
1906 				/* no current vlan tag in qp */
1907 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1908 				if (err)
1909 					return err;
1910 				smac_info->candidate_vid = vlan_tag;
1911 				smac_info->candidate_vlan_index = vidx;
1912 				smac_info->candidate_vlan_port = port;
1913 				smac_info->update_vid = 1;
1914 				path->vlan_index = vidx;
1915 			}
1916 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1917 			path->fl = 1 << 6;
1918 		} else {
1919 			/* have current vlan tag. unregister it at modify-qp success */
1920 			if (smac_info->vid < 0x1000) {
1921 				smac_info->candidate_vid = 0xFFFF;
1922 				smac_info->update_vid = 1;
1923 			}
1924 		}
1925 
1926 		/* get smac_index for RoCE use.
1927 		 * If no smac was yet assigned, register one.
1928 		 * If one was already assigned, but the new mac differs,
1929 		 * unregister the old one and register the new one.
1930 		*/
1931 		if ((!smac_info->smac && !smac_info->smac_port) ||
1932 		    smac_info->smac != smac) {
1933 			/* register candidate now, unreg if needed, after success */
1934 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1935 			if (smac_index >= 0) {
1936 				smac_info->candidate_smac_index = smac_index;
1937 				smac_info->candidate_smac = smac;
1938 				smac_info->candidate_smac_port = port;
1939 			} else {
1940 				return -EINVAL;
1941 			}
1942 		} else {
1943 			smac_index = smac_info->smac_index;
1944 		}
1945 		memcpy(path->dmac, ah->roce.dmac, 6);
1946 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1947 		/* put MAC table smac index for IBoE */
1948 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1949 	} else {
1950 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1951 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
1952 	}
1953 
1954 	return 0;
1955 }
1956 
1957 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1958 			 enum ib_qp_attr_mask qp_attr_mask,
1959 			 struct mlx4_ib_qp *mqp,
1960 			 struct mlx4_qp_path *path, u8 port,
1961 			 u16 vlan_id, u8 *smac)
1962 {
1963 	return _mlx4_set_path(dev, &qp->ah_attr,
1964 			      mlx4_mac_to_u64(smac),
1965 			      vlan_id,
1966 			      path, &mqp->pri, port);
1967 }
1968 
1969 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1970 			     const struct ib_qp_attr *qp,
1971 			     enum ib_qp_attr_mask qp_attr_mask,
1972 			     struct mlx4_ib_qp *mqp,
1973 			     struct mlx4_qp_path *path, u8 port)
1974 {
1975 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1976 			      0,
1977 			      0xffff,
1978 			      path, &mqp->alt, port);
1979 }
1980 
1981 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1982 {
1983 	struct mlx4_ib_gid_entry *ge, *tmp;
1984 
1985 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1986 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1987 			ge->added = 1;
1988 			ge->port = qp->port;
1989 		}
1990 	}
1991 }
1992 
1993 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1994 				    struct mlx4_ib_qp *qp,
1995 				    struct mlx4_qp_context *context)
1996 {
1997 	u64 u64_mac;
1998 	int smac_index;
1999 
2000 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
2001 
2002 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
2003 	if (!qp->pri.smac && !qp->pri.smac_port) {
2004 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
2005 		if (smac_index >= 0) {
2006 			qp->pri.candidate_smac_index = smac_index;
2007 			qp->pri.candidate_smac = u64_mac;
2008 			qp->pri.candidate_smac_port = qp->port;
2009 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
2010 		} else {
2011 			return -ENOENT;
2012 		}
2013 	}
2014 	return 0;
2015 }
2016 
2017 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
2018 {
2019 	struct counter_index *new_counter_index;
2020 	int err;
2021 	u32 tmp_idx;
2022 
2023 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
2024 	    IB_LINK_LAYER_ETHERNET ||
2025 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
2026 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
2027 		return 0;
2028 
2029 	err = mlx4_counter_alloc(dev->dev, &tmp_idx, MLX4_RES_USAGE_DRIVER);
2030 	if (err)
2031 		return err;
2032 
2033 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
2034 	if (!new_counter_index) {
2035 		mlx4_counter_free(dev->dev, tmp_idx);
2036 		return -ENOMEM;
2037 	}
2038 
2039 	new_counter_index->index = tmp_idx;
2040 	new_counter_index->allocated = 1;
2041 	qp->counter_index = new_counter_index;
2042 
2043 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
2044 	list_add_tail(&new_counter_index->list,
2045 		      &dev->counters_table[qp->port - 1].counters_list);
2046 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
2047 
2048 	return 0;
2049 }
2050 
2051 enum {
2052 	MLX4_QPC_ROCE_MODE_1 = 0,
2053 	MLX4_QPC_ROCE_MODE_2 = 2,
2054 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
2055 };
2056 
2057 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
2058 {
2059 	switch (gid_type) {
2060 	case IB_GID_TYPE_ROCE:
2061 		return MLX4_QPC_ROCE_MODE_1;
2062 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
2063 		return MLX4_QPC_ROCE_MODE_2;
2064 	default:
2065 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
2066 	}
2067 }
2068 
2069 /*
2070  * Go over all RSS QP's childes (WQs) and apply their HW state according to
2071  * their logic state if the RSS QP is the first RSS QP associated for the WQ.
2072  */
2073 static int bringup_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, u8 port_num)
2074 {
2075 	int err = 0;
2076 	int i;
2077 
2078 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2079 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2080 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2081 
2082 		mutex_lock(&wq->mutex);
2083 
2084 		/* Mlx4_ib restrictions:
2085 		 * WQ's is associated to a port according to the RSS QP it is
2086 		 * associates to.
2087 		 * In case the WQ is associated to a different port by another
2088 		 * RSS QP, return a failure.
2089 		 */
2090 		if ((wq->rss_usecnt > 0) && (wq->port != port_num)) {
2091 			err = -EINVAL;
2092 			mutex_unlock(&wq->mutex);
2093 			break;
2094 		}
2095 		wq->port = port_num;
2096 		if ((wq->rss_usecnt == 0) && (ibwq->state == IB_WQS_RDY)) {
2097 			err = _mlx4_ib_modify_wq(ibwq, IB_WQS_RDY);
2098 			if (err) {
2099 				mutex_unlock(&wq->mutex);
2100 				break;
2101 			}
2102 		}
2103 		wq->rss_usecnt++;
2104 
2105 		mutex_unlock(&wq->mutex);
2106 	}
2107 
2108 	if (i && err) {
2109 		int j;
2110 
2111 		for (j = (i - 1); j >= 0; j--) {
2112 			struct ib_wq *ibwq = ind_tbl->ind_tbl[j];
2113 			struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2114 
2115 			mutex_lock(&wq->mutex);
2116 
2117 			if ((wq->rss_usecnt == 1) &&
2118 			    (ibwq->state == IB_WQS_RDY))
2119 				if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET))
2120 					pr_warn("failed to reverse WQN=0x%06x\n",
2121 						ibwq->wq_num);
2122 			wq->rss_usecnt--;
2123 
2124 			mutex_unlock(&wq->mutex);
2125 		}
2126 	}
2127 
2128 	return err;
2129 }
2130 
2131 static void bring_down_rss_rwqs(struct ib_rwq_ind_table *ind_tbl)
2132 {
2133 	int i;
2134 
2135 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2136 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2137 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2138 
2139 		mutex_lock(&wq->mutex);
2140 
2141 		if ((wq->rss_usecnt == 1) && (ibwq->state == IB_WQS_RDY))
2142 			if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET))
2143 				pr_warn("failed to reverse WQN=%x\n",
2144 					ibwq->wq_num);
2145 		wq->rss_usecnt--;
2146 
2147 		mutex_unlock(&wq->mutex);
2148 	}
2149 }
2150 
2151 static void fill_qp_rss_context(struct mlx4_qp_context *context,
2152 				struct mlx4_ib_qp *qp)
2153 {
2154 	struct mlx4_rss_context *rss_context;
2155 
2156 	rss_context = (void *)context + offsetof(struct mlx4_qp_context,
2157 			pri_path) + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
2158 
2159 	rss_context->base_qpn = cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz);
2160 	rss_context->default_qpn =
2161 		cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz & 0xffffff);
2162 	if (qp->rss_ctx->flags & (MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6))
2163 		rss_context->base_qpn_udp = rss_context->default_qpn;
2164 	rss_context->flags = qp->rss_ctx->flags;
2165 	/* Currently support just toeplitz */
2166 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
2167 
2168 	memcpy(rss_context->rss_key, qp->rss_ctx->rss_key,
2169 	       MLX4_EN_RSS_KEY_SIZE);
2170 }
2171 
2172 static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type,
2173 			       const struct ib_qp_attr *attr, int attr_mask,
2174 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
2175 {
2176 	struct ib_uobject *ibuobject;
2177 	struct ib_srq  *ibsrq;
2178 	struct ib_rwq_ind_table *rwq_ind_tbl;
2179 	enum ib_qp_type qp_type;
2180 	struct mlx4_ib_dev *dev;
2181 	struct mlx4_ib_qp *qp;
2182 	struct mlx4_ib_pd *pd;
2183 	struct mlx4_ib_cq *send_cq, *recv_cq;
2184 	struct mlx4_qp_context *context;
2185 	enum mlx4_qp_optpar optpar = 0;
2186 	int sqd_event;
2187 	int steer_qp = 0;
2188 	int err = -EINVAL;
2189 	int counter_index;
2190 
2191 	if (src_type == MLX4_IB_RWQ_SRC) {
2192 		struct ib_wq *ibwq;
2193 
2194 		ibwq	    = (struct ib_wq *)src;
2195 		ibuobject   = ibwq->uobject;
2196 		ibsrq	    = NULL;
2197 		rwq_ind_tbl = NULL;
2198 		qp_type     = IB_QPT_RAW_PACKET;
2199 		qp	    = to_mqp((struct ib_qp *)ibwq);
2200 		dev	    = to_mdev(ibwq->device);
2201 		pd	    = to_mpd(ibwq->pd);
2202 	} else {
2203 		struct ib_qp *ibqp;
2204 
2205 		ibqp	    = (struct ib_qp *)src;
2206 		ibuobject   = ibqp->uobject;
2207 		ibsrq	    = ibqp->srq;
2208 		rwq_ind_tbl = ibqp->rwq_ind_tbl;
2209 		qp_type     = ibqp->qp_type;
2210 		qp	    = to_mqp(ibqp);
2211 		dev	    = to_mdev(ibqp->device);
2212 		pd	    = get_pd(qp);
2213 	}
2214 
2215 	/* APM is not supported under RoCE */
2216 	if (attr_mask & IB_QP_ALT_PATH &&
2217 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2218 	    IB_LINK_LAYER_ETHERNET)
2219 		return -ENOTSUPP;
2220 
2221 	context = kzalloc(sizeof *context, GFP_KERNEL);
2222 	if (!context)
2223 		return -ENOMEM;
2224 
2225 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
2226 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
2227 
2228 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
2229 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2230 	else {
2231 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
2232 		switch (attr->path_mig_state) {
2233 		case IB_MIG_MIGRATED:
2234 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2235 			break;
2236 		case IB_MIG_REARM:
2237 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
2238 			break;
2239 		case IB_MIG_ARMED:
2240 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
2241 			break;
2242 		}
2243 	}
2244 
2245 	if (qp->inl_recv_sz)
2246 		context->param3 |= cpu_to_be32(1 << 25);
2247 
2248 	if (qp->flags & MLX4_IB_QP_SCATTER_FCS)
2249 		context->param3 |= cpu_to_be32(1 << 29);
2250 
2251 	if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI)
2252 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
2253 	else if (qp_type == IB_QPT_RAW_PACKET)
2254 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
2255 	else if (qp_type == IB_QPT_UD) {
2256 		if (qp->flags & MLX4_IB_QP_LSO)
2257 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
2258 					      ilog2(dev->dev->caps.max_gso_sz);
2259 		else
2260 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 13;
2261 	} else if (attr_mask & IB_QP_PATH_MTU) {
2262 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
2263 			pr_err("path MTU (%u) is invalid\n",
2264 			       attr->path_mtu);
2265 			goto out;
2266 		}
2267 		context->mtu_msgmax = (attr->path_mtu << 5) |
2268 			ilog2(dev->dev->caps.max_msg_sz);
2269 	}
2270 
2271 	if (!rwq_ind_tbl) { /* PRM RSS receive side should be left zeros */
2272 		if (qp->rq.wqe_cnt)
2273 			context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
2274 		context->rq_size_stride |= qp->rq.wqe_shift - 4;
2275 	}
2276 
2277 	if (qp->sq.wqe_cnt)
2278 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
2279 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
2280 
2281 	if (new_state == IB_QPS_RESET && qp->counter_index)
2282 		mlx4_ib_free_qp_counter(dev, qp);
2283 
2284 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2285 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
2286 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
2287 		if (qp_type == IB_QPT_RAW_PACKET)
2288 			context->param3 |= cpu_to_be32(1 << 30);
2289 	}
2290 
2291 	if (ibuobject)
2292 		context->usr_page = cpu_to_be32(
2293 			mlx4_to_hw_uar_index(dev->dev,
2294 					     to_mucontext(ibuobject->context)
2295 					     ->uar.index));
2296 	else
2297 		context->usr_page = cpu_to_be32(
2298 			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
2299 
2300 	if (attr_mask & IB_QP_DEST_QPN)
2301 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
2302 
2303 	if (attr_mask & IB_QP_PORT) {
2304 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
2305 		    !(attr_mask & IB_QP_AV)) {
2306 			mlx4_set_sched(&context->pri_path, attr->port_num);
2307 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
2308 		}
2309 	}
2310 
2311 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
2312 		err = create_qp_lb_counter(dev, qp);
2313 		if (err)
2314 			goto out;
2315 
2316 		counter_index =
2317 			dev->counters_table[qp->port - 1].default_counter;
2318 		if (qp->counter_index)
2319 			counter_index = qp->counter_index->index;
2320 
2321 		if (counter_index != -1) {
2322 			context->pri_path.counter_index = counter_index;
2323 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
2324 			if (qp->counter_index) {
2325 				context->pri_path.fl |=
2326 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
2327 				context->pri_path.vlan_control |=
2328 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
2329 			}
2330 		} else
2331 			context->pri_path.counter_index =
2332 				MLX4_SINK_COUNTER_INDEX(dev->dev);
2333 
2334 		if (qp->flags & MLX4_IB_QP_NETIF) {
2335 			mlx4_ib_steer_qp_reg(dev, qp, 1);
2336 			steer_qp = 1;
2337 		}
2338 
2339 		if (qp_type == IB_QPT_GSI) {
2340 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
2341 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
2342 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
2343 
2344 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2345 		}
2346 	}
2347 
2348 	if (attr_mask & IB_QP_PKEY_INDEX) {
2349 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2350 			context->pri_path.disable_pkey_check = 0x40;
2351 		context->pri_path.pkey_index = attr->pkey_index;
2352 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
2353 	}
2354 
2355 	if (attr_mask & IB_QP_AV) {
2356 		u8 port_num = mlx4_is_bonded(dev->dev) ? 1 :
2357 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2358 		union ib_gid gid;
2359 		struct ib_gid_attr gid_attr = {.gid_type = IB_GID_TYPE_IB};
2360 		u16 vlan = 0xffff;
2361 		u8 smac[ETH_ALEN];
2362 		int status = 0;
2363 		int is_eth =
2364 			rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
2365 			rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
2366 
2367 		if (is_eth) {
2368 			int index =
2369 				rdma_ah_read_grh(&attr->ah_attr)->sgid_index;
2370 
2371 			status = ib_get_cached_gid(&dev->ib_dev, port_num,
2372 						   index, &gid, &gid_attr);
2373 			if (!status) {
2374 				vlan = rdma_vlan_dev_vlan_id(gid_attr.ndev);
2375 				memcpy(smac, gid_attr.ndev->dev_addr, ETH_ALEN);
2376 				dev_put(gid_attr.ndev);
2377 			}
2378 		}
2379 		if (status)
2380 			goto out;
2381 
2382 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
2383 				  port_num, vlan, smac))
2384 			goto out;
2385 
2386 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
2387 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
2388 
2389 		if (is_eth &&
2390 		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
2391 			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr.gid_type);
2392 
2393 			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
2394 				err = -EINVAL;
2395 				goto out;
2396 			}
2397 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2398 		}
2399 
2400 	}
2401 
2402 	if (attr_mask & IB_QP_TIMEOUT) {
2403 		context->pri_path.ackto |= attr->timeout << 3;
2404 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
2405 	}
2406 
2407 	if (attr_mask & IB_QP_ALT_PATH) {
2408 		if (attr->alt_port_num == 0 ||
2409 		    attr->alt_port_num > dev->dev->caps.num_ports)
2410 			goto out;
2411 
2412 		if (attr->alt_pkey_index >=
2413 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
2414 			goto out;
2415 
2416 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
2417 				      &context->alt_path,
2418 				      attr->alt_port_num))
2419 			goto out;
2420 
2421 		context->alt_path.pkey_index = attr->alt_pkey_index;
2422 		context->alt_path.ackto = attr->alt_timeout << 3;
2423 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
2424 	}
2425 
2426 	context->pd = cpu_to_be32(pd->pdn);
2427 
2428 	if (!rwq_ind_tbl) {
2429 		context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
2430 		get_cqs(qp, src_type, &send_cq, &recv_cq);
2431 	} else { /* Set dummy CQs to be compatible with HV and PRM */
2432 		send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq);
2433 		recv_cq = send_cq;
2434 	}
2435 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
2436 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
2437 
2438 	/* Set "fast registration enabled" for all kernel QPs */
2439 	if (!ibuobject)
2440 		context->params1 |= cpu_to_be32(1 << 11);
2441 
2442 	if (attr_mask & IB_QP_RNR_RETRY) {
2443 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2444 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
2445 	}
2446 
2447 	if (attr_mask & IB_QP_RETRY_CNT) {
2448 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2449 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
2450 	}
2451 
2452 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2453 		if (attr->max_rd_atomic)
2454 			context->params1 |=
2455 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2456 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
2457 	}
2458 
2459 	if (attr_mask & IB_QP_SQ_PSN)
2460 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2461 
2462 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2463 		if (attr->max_dest_rd_atomic)
2464 			context->params2 |=
2465 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2466 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
2467 	}
2468 
2469 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
2470 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
2471 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
2472 	}
2473 
2474 	if (ibsrq)
2475 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
2476 
2477 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
2478 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2479 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
2480 	}
2481 	if (attr_mask & IB_QP_RQ_PSN)
2482 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2483 
2484 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
2485 	if (attr_mask & IB_QP_QKEY) {
2486 		if (qp->mlx4_ib_qp_type &
2487 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
2488 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2489 		else {
2490 			if (mlx4_is_mfunc(dev->dev) &&
2491 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
2492 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
2493 			    MLX4_RESERVED_QKEY_BASE) {
2494 				pr_err("Cannot use reserved QKEY"
2495 				       " 0x%x (range 0xffff0000..0xffffffff"
2496 				       " is reserved)\n", attr->qkey);
2497 				err = -EINVAL;
2498 				goto out;
2499 			}
2500 			context->qkey = cpu_to_be32(attr->qkey);
2501 		}
2502 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
2503 	}
2504 
2505 	if (ibsrq)
2506 		context->srqn = cpu_to_be32(1 << 24 |
2507 					    to_msrq(ibsrq)->msrq.srqn);
2508 
2509 	if (qp->rq.wqe_cnt &&
2510 	    cur_state == IB_QPS_RESET &&
2511 	    new_state == IB_QPS_INIT)
2512 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2513 
2514 	if (cur_state == IB_QPS_INIT &&
2515 	    new_state == IB_QPS_RTR  &&
2516 	    (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI ||
2517 	     qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) {
2518 		context->pri_path.sched_queue = (qp->port - 1) << 6;
2519 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2520 		    qp->mlx4_ib_qp_type &
2521 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
2522 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
2523 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
2524 				context->pri_path.fl = 0x80;
2525 		} else {
2526 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2527 				context->pri_path.fl = 0x80;
2528 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
2529 		}
2530 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2531 		    IB_LINK_LAYER_ETHERNET) {
2532 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
2533 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
2534 				context->pri_path.feup = 1 << 7; /* don't fsm */
2535 			/* handle smac_index */
2536 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
2537 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
2538 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
2539 				err = handle_eth_ud_smac_index(dev, qp, context);
2540 				if (err) {
2541 					err = -EINVAL;
2542 					goto out;
2543 				}
2544 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
2545 					dev->qp1_proxy[qp->port - 1] = qp;
2546 			}
2547 		}
2548 	}
2549 
2550 	if (qp_type == IB_QPT_RAW_PACKET) {
2551 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
2552 					MLX4_IB_LINK_TYPE_ETH;
2553 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
2554 			/* set QP to receive both tunneled & non-tunneled packets */
2555 			if (!rwq_ind_tbl)
2556 				context->srqn = cpu_to_be32(7 << 28);
2557 		}
2558 	}
2559 
2560 	if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
2561 		int is_eth = rdma_port_get_link_layer(
2562 				&dev->ib_dev, qp->port) ==
2563 				IB_LINK_LAYER_ETHERNET;
2564 		if (is_eth) {
2565 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
2566 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
2567 		}
2568 	}
2569 
2570 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2571 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2572 		sqd_event = 1;
2573 	else
2574 		sqd_event = 0;
2575 
2576 	if (!ibuobject &&
2577 	    cur_state == IB_QPS_RESET &&
2578 	    new_state == IB_QPS_INIT)
2579 		context->rlkey_roce_mode |= (1 << 4);
2580 
2581 	/*
2582 	 * Before passing a kernel QP to the HW, make sure that the
2583 	 * ownership bits of the send queue are set and the SQ
2584 	 * headroom is stamped so that the hardware doesn't start
2585 	 * processing stale work requests.
2586 	 */
2587 	if (!ibuobject &&
2588 	    cur_state == IB_QPS_RESET &&
2589 	    new_state == IB_QPS_INIT) {
2590 		struct mlx4_wqe_ctrl_seg *ctrl;
2591 		int i;
2592 
2593 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
2594 			ctrl = get_send_wqe(qp, i);
2595 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
2596 			if (qp->sq_max_wqes_per_wr == 1)
2597 				ctrl->qpn_vlan.fence_size =
2598 						1 << (qp->sq.wqe_shift - 4);
2599 
2600 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
2601 		}
2602 	}
2603 
2604 	if (rwq_ind_tbl	&&
2605 	    cur_state == IB_QPS_RESET &&
2606 	    new_state == IB_QPS_INIT) {
2607 		fill_qp_rss_context(context, qp);
2608 		context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET);
2609 	}
2610 
2611 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2612 			     to_mlx4_state(new_state), context, optpar,
2613 			     sqd_event, &qp->mqp);
2614 	if (err)
2615 		goto out;
2616 
2617 	qp->state = new_state;
2618 
2619 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2620 		qp->atomic_rd_en = attr->qp_access_flags;
2621 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2622 		qp->resp_depth = attr->max_dest_rd_atomic;
2623 	if (attr_mask & IB_QP_PORT) {
2624 		qp->port = attr->port_num;
2625 		update_mcg_macs(dev, qp);
2626 	}
2627 	if (attr_mask & IB_QP_ALT_PATH)
2628 		qp->alt_port = attr->alt_port_num;
2629 
2630 	if (is_sqp(dev, qp))
2631 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2632 
2633 	/*
2634 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
2635 	 * QP0 to RESET or ERROR, bring the link back down.
2636 	 */
2637 	if (is_qp0(dev, qp)) {
2638 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2639 			if (mlx4_INIT_PORT(dev->dev, qp->port))
2640 				pr_warn("INIT_PORT failed for port %d\n",
2641 				       qp->port);
2642 
2643 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2644 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2645 			mlx4_CLOSE_PORT(dev->dev, qp->port);
2646 	}
2647 
2648 	/*
2649 	 * If we moved a kernel QP to RESET, clean up all old CQ
2650 	 * entries and reinitialize the QP.
2651 	 */
2652 	if (new_state == IB_QPS_RESET) {
2653 		if (!ibuobject) {
2654 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2655 					 ibsrq ? to_msrq(ibsrq) : NULL);
2656 			if (send_cq != recv_cq)
2657 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2658 
2659 			qp->rq.head = 0;
2660 			qp->rq.tail = 0;
2661 			qp->sq.head = 0;
2662 			qp->sq.tail = 0;
2663 			qp->sq_next_wqe = 0;
2664 			if (qp->rq.wqe_cnt)
2665 				*qp->db.db  = 0;
2666 
2667 			if (qp->flags & MLX4_IB_QP_NETIF)
2668 				mlx4_ib_steer_qp_reg(dev, qp, 0);
2669 		}
2670 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2671 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2672 			qp->pri.smac = 0;
2673 			qp->pri.smac_port = 0;
2674 		}
2675 		if (qp->alt.smac) {
2676 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2677 			qp->alt.smac = 0;
2678 		}
2679 		if (qp->pri.vid < 0x1000) {
2680 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2681 			qp->pri.vid = 0xFFFF;
2682 			qp->pri.candidate_vid = 0xFFFF;
2683 			qp->pri.update_vid = 0;
2684 		}
2685 
2686 		if (qp->alt.vid < 0x1000) {
2687 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2688 			qp->alt.vid = 0xFFFF;
2689 			qp->alt.candidate_vid = 0xFFFF;
2690 			qp->alt.update_vid = 0;
2691 		}
2692 	}
2693 out:
2694 	if (err && qp->counter_index)
2695 		mlx4_ib_free_qp_counter(dev, qp);
2696 	if (err && steer_qp)
2697 		mlx4_ib_steer_qp_reg(dev, qp, 0);
2698 	kfree(context);
2699 	if (qp->pri.candidate_smac ||
2700 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2701 		if (err) {
2702 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2703 		} else {
2704 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2705 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2706 			qp->pri.smac = qp->pri.candidate_smac;
2707 			qp->pri.smac_index = qp->pri.candidate_smac_index;
2708 			qp->pri.smac_port = qp->pri.candidate_smac_port;
2709 		}
2710 		qp->pri.candidate_smac = 0;
2711 		qp->pri.candidate_smac_index = 0;
2712 		qp->pri.candidate_smac_port = 0;
2713 	}
2714 	if (qp->alt.candidate_smac) {
2715 		if (err) {
2716 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2717 		} else {
2718 			if (qp->alt.smac)
2719 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2720 			qp->alt.smac = qp->alt.candidate_smac;
2721 			qp->alt.smac_index = qp->alt.candidate_smac_index;
2722 			qp->alt.smac_port = qp->alt.candidate_smac_port;
2723 		}
2724 		qp->alt.candidate_smac = 0;
2725 		qp->alt.candidate_smac_index = 0;
2726 		qp->alt.candidate_smac_port = 0;
2727 	}
2728 
2729 	if (qp->pri.update_vid) {
2730 		if (err) {
2731 			if (qp->pri.candidate_vid < 0x1000)
2732 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2733 						     qp->pri.candidate_vid);
2734 		} else {
2735 			if (qp->pri.vid < 0x1000)
2736 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2737 						     qp->pri.vid);
2738 			qp->pri.vid = qp->pri.candidate_vid;
2739 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2740 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
2741 		}
2742 		qp->pri.candidate_vid = 0xFFFF;
2743 		qp->pri.update_vid = 0;
2744 	}
2745 
2746 	if (qp->alt.update_vid) {
2747 		if (err) {
2748 			if (qp->alt.candidate_vid < 0x1000)
2749 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2750 						     qp->alt.candidate_vid);
2751 		} else {
2752 			if (qp->alt.vid < 0x1000)
2753 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2754 						     qp->alt.vid);
2755 			qp->alt.vid = qp->alt.candidate_vid;
2756 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2757 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2758 		}
2759 		qp->alt.candidate_vid = 0xFFFF;
2760 		qp->alt.update_vid = 0;
2761 	}
2762 
2763 	return err;
2764 }
2765 
2766 enum {
2767 	MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK = (IB_QP_STATE	|
2768 					      IB_QP_PORT),
2769 };
2770 
2771 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2772 			      int attr_mask, struct ib_udata *udata)
2773 {
2774 	enum rdma_link_layer ll = IB_LINK_LAYER_UNSPECIFIED;
2775 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2776 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2777 	enum ib_qp_state cur_state, new_state;
2778 	int err = -EINVAL;
2779 	mutex_lock(&qp->mutex);
2780 
2781 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2782 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2783 
2784 	if (cur_state != new_state || cur_state != IB_QPS_RESET) {
2785 		int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2786 		ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2787 	}
2788 
2789 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2790 				attr_mask, ll)) {
2791 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2792 			 "for transition %d to %d. qp_type %d,"
2793 			 " attr_mask 0x%x\n",
2794 			 ibqp->qp_num, cur_state, new_state,
2795 			 ibqp->qp_type, attr_mask);
2796 		goto out;
2797 	}
2798 
2799 	if (ibqp->rwq_ind_tbl) {
2800 		if (!(((cur_state == IB_QPS_RESET) &&
2801 		       (new_state == IB_QPS_INIT)) ||
2802 		      ((cur_state == IB_QPS_INIT)  &&
2803 		       (new_state == IB_QPS_RTR)))) {
2804 			pr_debug("qpn 0x%x: RSS QP unsupported transition %d to %d\n",
2805 				 ibqp->qp_num, cur_state, new_state);
2806 
2807 			err = -EOPNOTSUPP;
2808 			goto out;
2809 		}
2810 
2811 		if (attr_mask & ~MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK) {
2812 			pr_debug("qpn 0x%x: RSS QP unsupported attribute mask 0x%x for transition %d to %d\n",
2813 				 ibqp->qp_num, attr_mask, cur_state, new_state);
2814 
2815 			err = -EOPNOTSUPP;
2816 			goto out;
2817 		}
2818 	}
2819 
2820 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2821 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2822 			if ((ibqp->qp_type == IB_QPT_RC) ||
2823 			    (ibqp->qp_type == IB_QPT_UD) ||
2824 			    (ibqp->qp_type == IB_QPT_UC) ||
2825 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2826 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2827 				attr->port_num = mlx4_ib_bond_next_port(dev);
2828 			}
2829 		} else {
2830 			/* no sense in changing port_num
2831 			 * when ports are bonded */
2832 			attr_mask &= ~IB_QP_PORT;
2833 		}
2834 	}
2835 
2836 	if ((attr_mask & IB_QP_PORT) &&
2837 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2838 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2839 			 "for transition %d to %d. qp_type %d\n",
2840 			 ibqp->qp_num, attr->port_num, cur_state,
2841 			 new_state, ibqp->qp_type);
2842 		goto out;
2843 	}
2844 
2845 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2846 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2847 	     IB_LINK_LAYER_ETHERNET))
2848 		goto out;
2849 
2850 	if (attr_mask & IB_QP_PKEY_INDEX) {
2851 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2852 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2853 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2854 				 "for transition %d to %d. qp_type %d\n",
2855 				 ibqp->qp_num, attr->pkey_index, cur_state,
2856 				 new_state, ibqp->qp_type);
2857 			goto out;
2858 		}
2859 	}
2860 
2861 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2862 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2863 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2864 			 "Transition %d to %d. qp_type %d\n",
2865 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2866 			 new_state, ibqp->qp_type);
2867 		goto out;
2868 	}
2869 
2870 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2871 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2872 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2873 			 "Transition %d to %d. qp_type %d\n",
2874 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2875 			 new_state, ibqp->qp_type);
2876 		goto out;
2877 	}
2878 
2879 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2880 		err = 0;
2881 		goto out;
2882 	}
2883 
2884 	if (ibqp->rwq_ind_tbl && (new_state == IB_QPS_INIT)) {
2885 		err = bringup_rss_rwqs(ibqp->rwq_ind_tbl, attr->port_num);
2886 		if (err)
2887 			goto out;
2888 	}
2889 
2890 	err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask,
2891 				  cur_state, new_state);
2892 
2893 	if (ibqp->rwq_ind_tbl && err)
2894 		bring_down_rss_rwqs(ibqp->rwq_ind_tbl);
2895 
2896 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2897 		attr->port_num = 1;
2898 
2899 out:
2900 	mutex_unlock(&qp->mutex);
2901 	return err;
2902 }
2903 
2904 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2905 		      int attr_mask, struct ib_udata *udata)
2906 {
2907 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2908 	int ret;
2909 
2910 	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2911 
2912 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2913 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
2914 		int err = 0;
2915 
2916 		if (sqp->roce_v2_gsi)
2917 			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2918 		if (err)
2919 			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2920 			       err);
2921 	}
2922 	return ret;
2923 }
2924 
2925 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2926 {
2927 	int i;
2928 	for (i = 0; i < dev->caps.num_ports; i++) {
2929 		if (qpn == dev->caps.spec_qps[i].qp0_proxy ||
2930 		    qpn == dev->caps.spec_qps[i].qp0_tunnel) {
2931 			*qkey = dev->caps.spec_qps[i].qp0_qkey;
2932 			return 0;
2933 		}
2934 	}
2935 	return -EINVAL;
2936 }
2937 
2938 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2939 				  struct ib_ud_wr *wr,
2940 				  void *wqe, unsigned *mlx_seg_len)
2941 {
2942 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2943 	struct ib_device *ib_dev = &mdev->ib_dev;
2944 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2945 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2946 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2947 	u16 pkey;
2948 	u32 qkey;
2949 	int send_size;
2950 	int header_size;
2951 	int spc;
2952 	int i;
2953 
2954 	if (wr->wr.opcode != IB_WR_SEND)
2955 		return -EINVAL;
2956 
2957 	send_size = 0;
2958 
2959 	for (i = 0; i < wr->wr.num_sge; ++i)
2960 		send_size += wr->wr.sg_list[i].length;
2961 
2962 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2963 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2964 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2965 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2966 
2967 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2968 
2969 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2970 		sqp->ud_header.lrh.service_level =
2971 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2972 		sqp->ud_header.lrh.destination_lid =
2973 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2974 		sqp->ud_header.lrh.source_lid =
2975 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2976 	}
2977 
2978 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2979 
2980 	/* force loopback */
2981 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2982 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2983 
2984 	sqp->ud_header.lrh.virtual_lane    = 0;
2985 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2986 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2987 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2988 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2989 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2990 	else
2991 		sqp->ud_header.bth.destination_qpn =
2992 			cpu_to_be32(mdev->dev->caps.spec_qps[sqp->qp.port - 1].qp0_tunnel);
2993 
2994 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2995 	if (mlx4_is_master(mdev->dev)) {
2996 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2997 			return -EINVAL;
2998 	} else {
2999 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
3000 			return -EINVAL;
3001 	}
3002 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
3003 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
3004 
3005 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
3006 	sqp->ud_header.immediate_present = 0;
3007 
3008 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3009 
3010 	/*
3011 	 * Inline data segments may not cross a 64 byte boundary.  If
3012 	 * our UD header is bigger than the space available up to the
3013 	 * next 64 byte boundary in the WQE, use two inline data
3014 	 * segments to hold the UD header.
3015 	 */
3016 	spc = MLX4_INLINE_ALIGN -
3017 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3018 	if (header_size <= spc) {
3019 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3020 		memcpy(inl + 1, sqp->header_buf, header_size);
3021 		i = 1;
3022 	} else {
3023 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3024 		memcpy(inl + 1, sqp->header_buf, spc);
3025 
3026 		inl = (void *) (inl + 1) + spc;
3027 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3028 		/*
3029 		 * Need a barrier here to make sure all the data is
3030 		 * visible before the byte_count field is set.
3031 		 * Otherwise the HCA prefetcher could grab the 64-byte
3032 		 * chunk with this inline segment and get a valid (!=
3033 		 * 0xffffffff) byte count but stale data, and end up
3034 		 * generating a packet with bad headers.
3035 		 *
3036 		 * The first inline segment's byte_count field doesn't
3037 		 * need a barrier, because it comes after a
3038 		 * control/MLX segment and therefore is at an offset
3039 		 * of 16 mod 64.
3040 		 */
3041 		wmb();
3042 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3043 		i = 2;
3044 	}
3045 
3046 	*mlx_seg_len =
3047 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3048 	return 0;
3049 }
3050 
3051 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
3052 {
3053 	union sl2vl_tbl_to_u64 tmp_vltab;
3054 	u8 vl;
3055 
3056 	if (sl > 15)
3057 		return 0xf;
3058 	tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
3059 	vl = tmp_vltab.sl8[sl >> 1];
3060 	if (sl & 1)
3061 		vl &= 0x0f;
3062 	else
3063 		vl >>= 4;
3064 	return vl;
3065 }
3066 
3067 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
3068 				int index, union ib_gid *gid,
3069 				enum ib_gid_type *gid_type)
3070 {
3071 	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
3072 	struct mlx4_port_gid_table *port_gid_table;
3073 	unsigned long flags;
3074 
3075 	port_gid_table = &iboe->gids[port_num - 1];
3076 	spin_lock_irqsave(&iboe->lock, flags);
3077 	memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
3078 	*gid_type = port_gid_table->gids[index].gid_type;
3079 	spin_unlock_irqrestore(&iboe->lock, flags);
3080 	if (!memcmp(gid, &zgid, sizeof(*gid)))
3081 		return -ENOENT;
3082 
3083 	return 0;
3084 }
3085 
3086 #define MLX4_ROCEV2_QP1_SPORT 0xC000
3087 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_ud_wr *wr,
3088 			    void *wqe, unsigned *mlx_seg_len)
3089 {
3090 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
3091 	struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
3092 	struct mlx4_wqe_mlx_seg *mlx = wqe;
3093 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
3094 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
3095 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3096 	union ib_gid sgid;
3097 	u16 pkey;
3098 	int send_size;
3099 	int header_size;
3100 	int spc;
3101 	int i;
3102 	int err = 0;
3103 	u16 vlan = 0xffff;
3104 	bool is_eth;
3105 	bool is_vlan = false;
3106 	bool is_grh;
3107 	bool is_udp = false;
3108 	int ip_version = 0;
3109 
3110 	send_size = 0;
3111 	for (i = 0; i < wr->wr.num_sge; ++i)
3112 		send_size += wr->wr.sg_list[i].length;
3113 
3114 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
3115 	is_grh = mlx4_ib_ah_grh_present(ah);
3116 	if (is_eth) {
3117 		enum ib_gid_type gid_type;
3118 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3119 			/* When multi-function is enabled, the ib_core gid
3120 			 * indexes don't necessarily match the hw ones, so
3121 			 * we must use our own cache */
3122 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
3123 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
3124 							   ah->av.ib.gid_index, &sgid.raw[0]);
3125 			if (err)
3126 				return err;
3127 		} else  {
3128 			err = fill_gid_by_hw_index(ibdev, sqp->qp.port,
3129 					    ah->av.ib.gid_index,
3130 					    &sgid, &gid_type);
3131 			if (!err) {
3132 				is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
3133 				if (is_udp) {
3134 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
3135 						ip_version = 4;
3136 					else
3137 						ip_version = 6;
3138 					is_grh = false;
3139 				}
3140 			} else {
3141 				return err;
3142 			}
3143 		}
3144 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
3145 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
3146 			is_vlan = 1;
3147 		}
3148 	}
3149 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
3150 			  ip_version, is_udp, 0, &sqp->ud_header);
3151 	if (err)
3152 		return err;
3153 
3154 	if (!is_eth) {
3155 		sqp->ud_header.lrh.service_level =
3156 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
3157 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
3158 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
3159 	}
3160 
3161 	if (is_grh || (ip_version == 6)) {
3162 		sqp->ud_header.grh.traffic_class =
3163 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3164 		sqp->ud_header.grh.flow_label    =
3165 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
3166 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
3167 		if (is_eth) {
3168 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
3169 		} else {
3170 			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3171 				/* When multi-function is enabled, the ib_core gid
3172 				 * indexes don't necessarily match the hw ones, so
3173 				 * we must use our own cache
3174 				 */
3175 				sqp->ud_header.grh.source_gid.global.subnet_prefix =
3176 					cpu_to_be64(atomic64_read(&(to_mdev(ib_dev)->sriov.
3177 								    demux[sqp->qp.port - 1].
3178 								    subnet_prefix)));
3179 				sqp->ud_header.grh.source_gid.global.interface_id =
3180 					to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
3181 						       guid_cache[ah->av.ib.gid_index];
3182 			} else {
3183 				ib_get_cached_gid(ib_dev,
3184 						  be32_to_cpu(ah->av.ib.port_pd) >> 24,
3185 						  ah->av.ib.gid_index,
3186 						  &sqp->ud_header.grh.source_gid, NULL);
3187 			}
3188 		}
3189 		memcpy(sqp->ud_header.grh.destination_gid.raw,
3190 		       ah->av.ib.dgid, 16);
3191 	}
3192 
3193 	if (ip_version == 4) {
3194 		sqp->ud_header.ip4.tos =
3195 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3196 		sqp->ud_header.ip4.id = 0;
3197 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
3198 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
3199 
3200 		memcpy(&sqp->ud_header.ip4.saddr,
3201 		       sgid.raw + 12, 4);
3202 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
3203 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
3204 	}
3205 
3206 	if (is_udp) {
3207 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
3208 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
3209 		sqp->ud_header.udp.csum = 0;
3210 	}
3211 
3212 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
3213 
3214 	if (!is_eth) {
3215 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
3216 					  (sqp->ud_header.lrh.destination_lid ==
3217 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
3218 					  (sqp->ud_header.lrh.service_level << 8));
3219 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
3220 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
3221 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
3222 	}
3223 
3224 	switch (wr->wr.opcode) {
3225 	case IB_WR_SEND:
3226 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
3227 		sqp->ud_header.immediate_present = 0;
3228 		break;
3229 	case IB_WR_SEND_WITH_IMM:
3230 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
3231 		sqp->ud_header.immediate_present = 1;
3232 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
3233 		break;
3234 	default:
3235 		return -EINVAL;
3236 	}
3237 
3238 	if (is_eth) {
3239 		struct in6_addr in6;
3240 		u16 ether_type;
3241 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
3242 
3243 		ether_type = (!is_udp) ? ETH_P_IBOE:
3244 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
3245 
3246 		mlx->sched_prio = cpu_to_be16(pcp);
3247 
3248 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
3249 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
3250 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
3251 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
3252 		memcpy(&in6, sgid.raw, sizeof(in6));
3253 
3254 
3255 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
3256 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
3257 		if (!is_vlan) {
3258 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
3259 		} else {
3260 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
3261 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
3262 		}
3263 	} else {
3264 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 :
3265 							sl_to_vl(to_mdev(ib_dev),
3266 								 sqp->ud_header.lrh.service_level,
3267 								 sqp->qp.port);
3268 		if (sqp->qp.ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
3269 			return -EINVAL;
3270 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
3271 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
3272 	}
3273 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
3274 	if (!sqp->qp.ibqp.qp_num)
3275 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
3276 	else
3277 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
3278 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
3279 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
3280 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
3281 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
3282 					       sqp->qkey : wr->remote_qkey);
3283 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
3284 
3285 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3286 
3287 	if (0) {
3288 		pr_err("built UD header of size %d:\n", header_size);
3289 		for (i = 0; i < header_size / 4; ++i) {
3290 			if (i % 8 == 0)
3291 				pr_err("  [%02x] ", i * 4);
3292 			pr_cont(" %08x",
3293 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
3294 			if ((i + 1) % 8 == 0)
3295 				pr_cont("\n");
3296 		}
3297 		pr_err("\n");
3298 	}
3299 
3300 	/*
3301 	 * Inline data segments may not cross a 64 byte boundary.  If
3302 	 * our UD header is bigger than the space available up to the
3303 	 * next 64 byte boundary in the WQE, use two inline data
3304 	 * segments to hold the UD header.
3305 	 */
3306 	spc = MLX4_INLINE_ALIGN -
3307 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3308 	if (header_size <= spc) {
3309 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3310 		memcpy(inl + 1, sqp->header_buf, header_size);
3311 		i = 1;
3312 	} else {
3313 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3314 		memcpy(inl + 1, sqp->header_buf, spc);
3315 
3316 		inl = (void *) (inl + 1) + spc;
3317 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3318 		/*
3319 		 * Need a barrier here to make sure all the data is
3320 		 * visible before the byte_count field is set.
3321 		 * Otherwise the HCA prefetcher could grab the 64-byte
3322 		 * chunk with this inline segment and get a valid (!=
3323 		 * 0xffffffff) byte count but stale data, and end up
3324 		 * generating a packet with bad headers.
3325 		 *
3326 		 * The first inline segment's byte_count field doesn't
3327 		 * need a barrier, because it comes after a
3328 		 * control/MLX segment and therefore is at an offset
3329 		 * of 16 mod 64.
3330 		 */
3331 		wmb();
3332 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3333 		i = 2;
3334 	}
3335 
3336 	*mlx_seg_len =
3337 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3338 	return 0;
3339 }
3340 
3341 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3342 {
3343 	unsigned cur;
3344 	struct mlx4_ib_cq *cq;
3345 
3346 	cur = wq->head - wq->tail;
3347 	if (likely(cur + nreq < wq->max_post))
3348 		return 0;
3349 
3350 	cq = to_mcq(ib_cq);
3351 	spin_lock(&cq->lock);
3352 	cur = wq->head - wq->tail;
3353 	spin_unlock(&cq->lock);
3354 
3355 	return cur + nreq >= wq->max_post;
3356 }
3357 
3358 static __be32 convert_access(int acc)
3359 {
3360 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
3361 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
3362 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
3363 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
3364 	       (acc & IB_ACCESS_REMOTE_READ   ?
3365 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
3366 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
3367 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
3368 }
3369 
3370 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
3371 			struct ib_reg_wr *wr)
3372 {
3373 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
3374 
3375 	fseg->flags		= convert_access(wr->access);
3376 	fseg->mem_key		= cpu_to_be32(wr->key);
3377 	fseg->buf_list		= cpu_to_be64(mr->page_map);
3378 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
3379 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
3380 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
3381 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
3382 	fseg->reserved[0]	= 0;
3383 	fseg->reserved[1]	= 0;
3384 }
3385 
3386 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
3387 {
3388 	memset(iseg, 0, sizeof(*iseg));
3389 	iseg->mem_key = cpu_to_be32(rkey);
3390 }
3391 
3392 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
3393 					  u64 remote_addr, u32 rkey)
3394 {
3395 	rseg->raddr    = cpu_to_be64(remote_addr);
3396 	rseg->rkey     = cpu_to_be32(rkey);
3397 	rseg->reserved = 0;
3398 }
3399 
3400 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
3401 		struct ib_atomic_wr *wr)
3402 {
3403 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
3404 		aseg->swap_add = cpu_to_be64(wr->swap);
3405 		aseg->compare  = cpu_to_be64(wr->compare_add);
3406 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
3407 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3408 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
3409 	} else {
3410 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3411 		aseg->compare  = 0;
3412 	}
3413 
3414 }
3415 
3416 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
3417 				  struct ib_atomic_wr *wr)
3418 {
3419 	aseg->swap_add		= cpu_to_be64(wr->swap);
3420 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
3421 	aseg->compare		= cpu_to_be64(wr->compare_add);
3422 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
3423 }
3424 
3425 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
3426 			     struct ib_ud_wr *wr)
3427 {
3428 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
3429 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
3430 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
3431 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
3432 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
3433 }
3434 
3435 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
3436 				    struct mlx4_wqe_datagram_seg *dseg,
3437 				    struct ib_ud_wr *wr,
3438 				    enum mlx4_ib_qp_type qpt)
3439 {
3440 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
3441 	struct mlx4_av sqp_av = {0};
3442 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
3443 
3444 	/* force loopback */
3445 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
3446 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
3447 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
3448 			cpu_to_be32(0xf0000000);
3449 
3450 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
3451 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
3452 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp1_tunnel);
3453 	else
3454 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp0_tunnel);
3455 	/* Use QKEY from the QP context, which is set by master */
3456 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
3457 }
3458 
3459 static void build_tunnel_header(struct ib_ud_wr *wr, void *wqe, unsigned *mlx_seg_len)
3460 {
3461 	struct mlx4_wqe_inline_seg *inl = wqe;
3462 	struct mlx4_ib_tunnel_header hdr;
3463 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3464 	int spc;
3465 	int i;
3466 
3467 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
3468 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
3469 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
3470 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
3471 	memcpy(hdr.mac, ah->av.eth.mac, 6);
3472 	hdr.vlan = ah->av.eth.vlan;
3473 
3474 	spc = MLX4_INLINE_ALIGN -
3475 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3476 	if (sizeof (hdr) <= spc) {
3477 		memcpy(inl + 1, &hdr, sizeof (hdr));
3478 		wmb();
3479 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
3480 		i = 1;
3481 	} else {
3482 		memcpy(inl + 1, &hdr, spc);
3483 		wmb();
3484 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3485 
3486 		inl = (void *) (inl + 1) + spc;
3487 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
3488 		wmb();
3489 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
3490 		i = 2;
3491 	}
3492 
3493 	*mlx_seg_len =
3494 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
3495 }
3496 
3497 static void set_mlx_icrc_seg(void *dseg)
3498 {
3499 	u32 *t = dseg;
3500 	struct mlx4_wqe_inline_seg *iseg = dseg;
3501 
3502 	t[1] = 0;
3503 
3504 	/*
3505 	 * Need a barrier here before writing the byte_count field to
3506 	 * make sure that all the data is visible before the
3507 	 * byte_count field is set.  Otherwise, if the segment begins
3508 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3509 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3510 	 * stale data, and end up sending the wrong data.
3511 	 */
3512 	wmb();
3513 
3514 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
3515 }
3516 
3517 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3518 {
3519 	dseg->lkey       = cpu_to_be32(sg->lkey);
3520 	dseg->addr       = cpu_to_be64(sg->addr);
3521 
3522 	/*
3523 	 * Need a barrier here before writing the byte_count field to
3524 	 * make sure that all the data is visible before the
3525 	 * byte_count field is set.  Otherwise, if the segment begins
3526 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3527 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3528 	 * stale data, and end up sending the wrong data.
3529 	 */
3530 	wmb();
3531 
3532 	dseg->byte_count = cpu_to_be32(sg->length);
3533 }
3534 
3535 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3536 {
3537 	dseg->byte_count = cpu_to_be32(sg->length);
3538 	dseg->lkey       = cpu_to_be32(sg->lkey);
3539 	dseg->addr       = cpu_to_be64(sg->addr);
3540 }
3541 
3542 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_ud_wr *wr,
3543 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
3544 			 __be32 *lso_hdr_sz, __be32 *blh)
3545 {
3546 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
3547 
3548 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
3549 		*blh = cpu_to_be32(1 << 6);
3550 
3551 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
3552 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
3553 		return -EINVAL;
3554 
3555 	memcpy(wqe->header, wr->header, wr->hlen);
3556 
3557 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
3558 	*lso_seg_len = halign;
3559 	return 0;
3560 }
3561 
3562 static __be32 send_ieth(struct ib_send_wr *wr)
3563 {
3564 	switch (wr->opcode) {
3565 	case IB_WR_SEND_WITH_IMM:
3566 	case IB_WR_RDMA_WRITE_WITH_IMM:
3567 		return wr->ex.imm_data;
3568 
3569 	case IB_WR_SEND_WITH_INV:
3570 		return cpu_to_be32(wr->ex.invalidate_rkey);
3571 
3572 	default:
3573 		return 0;
3574 	}
3575 }
3576 
3577 static void add_zero_len_inline(void *wqe)
3578 {
3579 	struct mlx4_wqe_inline_seg *inl = wqe;
3580 	memset(wqe, 0, 16);
3581 	inl->byte_count = cpu_to_be32(1 << 31);
3582 }
3583 
3584 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
3585 		      struct ib_send_wr **bad_wr)
3586 {
3587 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3588 	void *wqe;
3589 	struct mlx4_wqe_ctrl_seg *ctrl;
3590 	struct mlx4_wqe_data_seg *dseg;
3591 	unsigned long flags;
3592 	int nreq;
3593 	int err = 0;
3594 	unsigned ind;
3595 	int uninitialized_var(stamp);
3596 	int uninitialized_var(size);
3597 	unsigned uninitialized_var(seglen);
3598 	__be32 dummy;
3599 	__be32 *lso_wqe;
3600 	__be32 uninitialized_var(lso_hdr_sz);
3601 	__be32 blh;
3602 	int i;
3603 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3604 
3605 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
3606 		struct mlx4_ib_sqp *sqp = to_msqp(qp);
3607 
3608 		if (sqp->roce_v2_gsi) {
3609 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
3610 			enum ib_gid_type gid_type;
3611 			union ib_gid gid;
3612 
3613 			if (!fill_gid_by_hw_index(mdev, sqp->qp.port,
3614 					   ah->av.ib.gid_index,
3615 					   &gid, &gid_type))
3616 				qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
3617 						to_mqp(sqp->roce_v2_gsi) : qp;
3618 			else
3619 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
3620 				       ah->av.ib.gid_index);
3621 		}
3622 	}
3623 
3624 	spin_lock_irqsave(&qp->sq.lock, flags);
3625 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3626 		err = -EIO;
3627 		*bad_wr = wr;
3628 		nreq = 0;
3629 		goto out;
3630 	}
3631 
3632 	ind = qp->sq_next_wqe;
3633 
3634 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3635 		lso_wqe = &dummy;
3636 		blh = 0;
3637 
3638 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
3639 			err = -ENOMEM;
3640 			*bad_wr = wr;
3641 			goto out;
3642 		}
3643 
3644 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3645 			err = -EINVAL;
3646 			*bad_wr = wr;
3647 			goto out;
3648 		}
3649 
3650 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3651 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3652 
3653 		ctrl->srcrb_flags =
3654 			(wr->send_flags & IB_SEND_SIGNALED ?
3655 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3656 			(wr->send_flags & IB_SEND_SOLICITED ?
3657 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3658 			((wr->send_flags & IB_SEND_IP_CSUM) ?
3659 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3660 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3661 			qp->sq_signal_bits;
3662 
3663 		ctrl->imm = send_ieth(wr);
3664 
3665 		wqe += sizeof *ctrl;
3666 		size = sizeof *ctrl / 16;
3667 
3668 		switch (qp->mlx4_ib_qp_type) {
3669 		case MLX4_IB_QPT_RC:
3670 		case MLX4_IB_QPT_UC:
3671 			switch (wr->opcode) {
3672 			case IB_WR_ATOMIC_CMP_AND_SWP:
3673 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3674 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3675 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3676 					      atomic_wr(wr)->rkey);
3677 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3678 
3679 				set_atomic_seg(wqe, atomic_wr(wr));
3680 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
3681 
3682 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3683 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3684 
3685 				break;
3686 
3687 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3688 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3689 					      atomic_wr(wr)->rkey);
3690 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3691 
3692 				set_masked_atomic_seg(wqe, atomic_wr(wr));
3693 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
3694 
3695 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3696 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3697 
3698 				break;
3699 
3700 			case IB_WR_RDMA_READ:
3701 			case IB_WR_RDMA_WRITE:
3702 			case IB_WR_RDMA_WRITE_WITH_IMM:
3703 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3704 					      rdma_wr(wr)->rkey);
3705 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3706 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3707 				break;
3708 
3709 			case IB_WR_LOCAL_INV:
3710 				ctrl->srcrb_flags |=
3711 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3712 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3713 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3714 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3715 				break;
3716 
3717 			case IB_WR_REG_MR:
3718 				ctrl->srcrb_flags |=
3719 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3720 				set_reg_seg(wqe, reg_wr(wr));
3721 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3722 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3723 				break;
3724 
3725 			default:
3726 				/* No extra segments required for sends */
3727 				break;
3728 			}
3729 			break;
3730 
3731 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3732 			err =  build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3733 					ctrl, &seglen);
3734 			if (unlikely(err)) {
3735 				*bad_wr = wr;
3736 				goto out;
3737 			}
3738 			wqe  += seglen;
3739 			size += seglen / 16;
3740 			break;
3741 		case MLX4_IB_QPT_TUN_SMI:
3742 		case MLX4_IB_QPT_TUN_GSI:
3743 			/* this is a UD qp used in MAD responses to slaves. */
3744 			set_datagram_seg(wqe, ud_wr(wr));
3745 			/* set the forced-loopback bit in the data seg av */
3746 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3747 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3748 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3749 			break;
3750 		case MLX4_IB_QPT_UD:
3751 			set_datagram_seg(wqe, ud_wr(wr));
3752 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3753 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3754 
3755 			if (wr->opcode == IB_WR_LSO) {
3756 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3757 						&lso_hdr_sz, &blh);
3758 				if (unlikely(err)) {
3759 					*bad_wr = wr;
3760 					goto out;
3761 				}
3762 				lso_wqe = (__be32 *) wqe;
3763 				wqe  += seglen;
3764 				size += seglen / 16;
3765 			}
3766 			break;
3767 
3768 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3769 			err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3770 					ctrl, &seglen);
3771 			if (unlikely(err)) {
3772 				*bad_wr = wr;
3773 				goto out;
3774 			}
3775 			wqe  += seglen;
3776 			size += seglen / 16;
3777 			/* to start tunnel header on a cache-line boundary */
3778 			add_zero_len_inline(wqe);
3779 			wqe += 16;
3780 			size++;
3781 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3782 			wqe  += seglen;
3783 			size += seglen / 16;
3784 			break;
3785 		case MLX4_IB_QPT_PROXY_SMI:
3786 		case MLX4_IB_QPT_PROXY_GSI:
3787 			/* If we are tunneling special qps, this is a UD qp.
3788 			 * In this case we first add a UD segment targeting
3789 			 * the tunnel qp, and then add a header with address
3790 			 * information */
3791 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3792 						ud_wr(wr),
3793 						qp->mlx4_ib_qp_type);
3794 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3795 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3796 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3797 			wqe  += seglen;
3798 			size += seglen / 16;
3799 			break;
3800 
3801 		case MLX4_IB_QPT_SMI:
3802 		case MLX4_IB_QPT_GSI:
3803 			err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
3804 					&seglen);
3805 			if (unlikely(err)) {
3806 				*bad_wr = wr;
3807 				goto out;
3808 			}
3809 			wqe  += seglen;
3810 			size += seglen / 16;
3811 			break;
3812 
3813 		default:
3814 			break;
3815 		}
3816 
3817 		/*
3818 		 * Write data segments in reverse order, so as to
3819 		 * overwrite cacheline stamp last within each
3820 		 * cacheline.  This avoids issues with WQE
3821 		 * prefetching.
3822 		 */
3823 
3824 		dseg = wqe;
3825 		dseg += wr->num_sge - 1;
3826 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3827 
3828 		/* Add one more inline data segment for ICRC for MLX sends */
3829 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3830 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3831 			     qp->mlx4_ib_qp_type &
3832 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3833 			set_mlx_icrc_seg(dseg + 1);
3834 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3835 		}
3836 
3837 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3838 			set_data_seg(dseg, wr->sg_list + i);
3839 
3840 		/*
3841 		 * Possibly overwrite stamping in cacheline with LSO
3842 		 * segment only after making sure all data segments
3843 		 * are written.
3844 		 */
3845 		wmb();
3846 		*lso_wqe = lso_hdr_sz;
3847 
3848 		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3849 					     MLX4_WQE_CTRL_FENCE : 0) | size;
3850 
3851 		/*
3852 		 * Make sure descriptor is fully written before
3853 		 * setting ownership bit (because HW can start
3854 		 * executing as soon as we do).
3855 		 */
3856 		wmb();
3857 
3858 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3859 			*bad_wr = wr;
3860 			err = -EINVAL;
3861 			goto out;
3862 		}
3863 
3864 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3865 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3866 
3867 		stamp = ind + qp->sq_spare_wqes;
3868 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
3869 
3870 		/*
3871 		 * We can improve latency by not stamping the last
3872 		 * send queue WQE until after ringing the doorbell, so
3873 		 * only stamp here if there are still more WQEs to post.
3874 		 *
3875 		 * Same optimization applies to padding with NOP wqe
3876 		 * in case of WQE shrinking (used to prevent wrap-around
3877 		 * in the middle of WR).
3878 		 */
3879 		if (wr->next) {
3880 			stamp_send_wqe(qp, stamp, size * 16);
3881 			ind = pad_wraparound(qp, ind);
3882 		}
3883 	}
3884 
3885 out:
3886 	if (likely(nreq)) {
3887 		qp->sq.head += nreq;
3888 
3889 		/*
3890 		 * Make sure that descriptors are written before
3891 		 * doorbell record.
3892 		 */
3893 		wmb();
3894 
3895 		writel_relaxed(qp->doorbell_qpn,
3896 			to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3897 
3898 		/*
3899 		 * Make sure doorbells don't leak out of SQ spinlock
3900 		 * and reach the HCA out of order.
3901 		 */
3902 		mmiowb();
3903 
3904 		stamp_send_wqe(qp, stamp, size * 16);
3905 
3906 		ind = pad_wraparound(qp, ind);
3907 		qp->sq_next_wqe = ind;
3908 	}
3909 
3910 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3911 
3912 	return err;
3913 }
3914 
3915 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3916 		      struct ib_recv_wr **bad_wr)
3917 {
3918 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3919 	struct mlx4_wqe_data_seg *scat;
3920 	unsigned long flags;
3921 	int err = 0;
3922 	int nreq;
3923 	int ind;
3924 	int max_gs;
3925 	int i;
3926 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3927 
3928 	max_gs = qp->rq.max_gs;
3929 	spin_lock_irqsave(&qp->rq.lock, flags);
3930 
3931 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3932 		err = -EIO;
3933 		*bad_wr = wr;
3934 		nreq = 0;
3935 		goto out;
3936 	}
3937 
3938 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3939 
3940 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3941 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3942 			err = -ENOMEM;
3943 			*bad_wr = wr;
3944 			goto out;
3945 		}
3946 
3947 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3948 			err = -EINVAL;
3949 			*bad_wr = wr;
3950 			goto out;
3951 		}
3952 
3953 		scat = get_recv_wqe(qp, ind);
3954 
3955 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3956 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3957 			ib_dma_sync_single_for_device(ibqp->device,
3958 						      qp->sqp_proxy_rcv[ind].map,
3959 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3960 						      DMA_FROM_DEVICE);
3961 			scat->byte_count =
3962 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3963 			/* use dma lkey from upper layer entry */
3964 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3965 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3966 			scat++;
3967 			max_gs--;
3968 		}
3969 
3970 		for (i = 0; i < wr->num_sge; ++i)
3971 			__set_data_seg(scat + i, wr->sg_list + i);
3972 
3973 		if (i < max_gs) {
3974 			scat[i].byte_count = 0;
3975 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3976 			scat[i].addr       = 0;
3977 		}
3978 
3979 		qp->rq.wrid[ind] = wr->wr_id;
3980 
3981 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3982 	}
3983 
3984 out:
3985 	if (likely(nreq)) {
3986 		qp->rq.head += nreq;
3987 
3988 		/*
3989 		 * Make sure that descriptors are written before
3990 		 * doorbell record.
3991 		 */
3992 		wmb();
3993 
3994 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3995 	}
3996 
3997 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3998 
3999 	return err;
4000 }
4001 
4002 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
4003 {
4004 	switch (mlx4_state) {
4005 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
4006 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
4007 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
4008 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
4009 	case MLX4_QP_STATE_SQ_DRAINING:
4010 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
4011 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
4012 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
4013 	default:		     return -1;
4014 	}
4015 }
4016 
4017 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
4018 {
4019 	switch (mlx4_mig_state) {
4020 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
4021 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
4022 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
4023 	default: return -1;
4024 	}
4025 }
4026 
4027 static int to_ib_qp_access_flags(int mlx4_flags)
4028 {
4029 	int ib_flags = 0;
4030 
4031 	if (mlx4_flags & MLX4_QP_BIT_RRE)
4032 		ib_flags |= IB_ACCESS_REMOTE_READ;
4033 	if (mlx4_flags & MLX4_QP_BIT_RWE)
4034 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
4035 	if (mlx4_flags & MLX4_QP_BIT_RAE)
4036 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
4037 
4038 	return ib_flags;
4039 }
4040 
4041 static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
4042 			    struct rdma_ah_attr *ah_attr,
4043 			    struct mlx4_qp_path *path)
4044 {
4045 	struct mlx4_dev *dev = ibdev->dev;
4046 	u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
4047 
4048 	memset(ah_attr, 0, sizeof(*ah_attr));
4049 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
4050 	if (port_num == 0 || port_num > dev->caps.num_ports)
4051 		return;
4052 
4053 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
4054 		rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
4055 			       ((path->sched_queue & 4) << 1));
4056 	else
4057 		rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
4058 	rdma_ah_set_port_num(ah_attr, port_num);
4059 
4060 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
4061 	rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
4062 	rdma_ah_set_static_rate(ah_attr,
4063 				path->static_rate ? path->static_rate - 5 : 0);
4064 	if (path->grh_mylmc & (1 << 7)) {
4065 		rdma_ah_set_grh(ah_attr, NULL,
4066 				be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
4067 				path->mgid_index,
4068 				path->hop_limit,
4069 				(be32_to_cpu(path->tclass_flowlabel)
4070 				 >> 20) & 0xff);
4071 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
4072 	}
4073 }
4074 
4075 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
4076 		     struct ib_qp_init_attr *qp_init_attr)
4077 {
4078 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
4079 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
4080 	struct mlx4_qp_context context;
4081 	int mlx4_state;
4082 	int err = 0;
4083 
4084 	if (ibqp->rwq_ind_tbl)
4085 		return -EOPNOTSUPP;
4086 
4087 	mutex_lock(&qp->mutex);
4088 
4089 	if (qp->state == IB_QPS_RESET) {
4090 		qp_attr->qp_state = IB_QPS_RESET;
4091 		goto done;
4092 	}
4093 
4094 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
4095 	if (err) {
4096 		err = -EINVAL;
4097 		goto out;
4098 	}
4099 
4100 	mlx4_state = be32_to_cpu(context.flags) >> 28;
4101 
4102 	qp->state		     = to_ib_qp_state(mlx4_state);
4103 	qp_attr->qp_state	     = qp->state;
4104 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
4105 	qp_attr->path_mig_state	     =
4106 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
4107 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
4108 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
4109 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
4110 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
4111 	qp_attr->qp_access_flags     =
4112 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
4113 
4114 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4115 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
4116 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
4117 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
4118 		qp_attr->alt_port_num	=
4119 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
4120 	}
4121 
4122 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
4123 	if (qp_attr->qp_state == IB_QPS_INIT)
4124 		qp_attr->port_num = qp->port;
4125 	else
4126 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
4127 
4128 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4129 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
4130 
4131 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
4132 
4133 	qp_attr->max_dest_rd_atomic =
4134 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
4135 	qp_attr->min_rnr_timer	    =
4136 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
4137 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
4138 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
4139 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
4140 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
4141 
4142 done:
4143 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4144 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4145 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4146 
4147 	if (!ibqp->uobject) {
4148 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
4149 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4150 	} else {
4151 		qp_attr->cap.max_send_wr  = 0;
4152 		qp_attr->cap.max_send_sge = 0;
4153 	}
4154 
4155 	/*
4156 	 * We don't support inline sends for kernel QPs (yet), and we
4157 	 * don't know what userspace's value should be.
4158 	 */
4159 	qp_attr->cap.max_inline_data = 0;
4160 
4161 	qp_init_attr->cap	     = qp_attr->cap;
4162 
4163 	qp_init_attr->create_flags = 0;
4164 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4165 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4166 
4167 	if (qp->flags & MLX4_IB_QP_LSO)
4168 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
4169 
4170 	if (qp->flags & MLX4_IB_QP_NETIF)
4171 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
4172 
4173 	qp_init_attr->sq_sig_type =
4174 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
4175 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4176 
4177 out:
4178 	mutex_unlock(&qp->mutex);
4179 	return err;
4180 }
4181 
4182 struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
4183 				struct ib_wq_init_attr *init_attr,
4184 				struct ib_udata *udata)
4185 {
4186 	struct mlx4_ib_dev *dev;
4187 	struct ib_qp_init_attr ib_qp_init_attr;
4188 	struct mlx4_ib_qp *qp;
4189 	struct mlx4_ib_create_wq ucmd;
4190 	int err, required_cmd_sz;
4191 
4192 	if (!(udata && pd->uobject))
4193 		return ERR_PTR(-EINVAL);
4194 
4195 	required_cmd_sz = offsetof(typeof(ucmd), comp_mask) +
4196 			  sizeof(ucmd.comp_mask);
4197 	if (udata->inlen < required_cmd_sz) {
4198 		pr_debug("invalid inlen\n");
4199 		return ERR_PTR(-EINVAL);
4200 	}
4201 
4202 	if (udata->inlen > sizeof(ucmd) &&
4203 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4204 				 udata->inlen - sizeof(ucmd))) {
4205 		pr_debug("inlen is not supported\n");
4206 		return ERR_PTR(-EOPNOTSUPP);
4207 	}
4208 
4209 	if (udata->outlen)
4210 		return ERR_PTR(-EOPNOTSUPP);
4211 
4212 	dev = to_mdev(pd->device);
4213 
4214 	if (init_attr->wq_type != IB_WQT_RQ) {
4215 		pr_debug("unsupported wq type %d\n", init_attr->wq_type);
4216 		return ERR_PTR(-EOPNOTSUPP);
4217 	}
4218 
4219 	if (init_attr->create_flags & ~IB_WQ_FLAGS_SCATTER_FCS) {
4220 		pr_debug("unsupported create_flags %u\n",
4221 			 init_attr->create_flags);
4222 		return ERR_PTR(-EOPNOTSUPP);
4223 	}
4224 
4225 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
4226 	if (!qp)
4227 		return ERR_PTR(-ENOMEM);
4228 
4229 	qp->pri.vid = 0xFFFF;
4230 	qp->alt.vid = 0xFFFF;
4231 
4232 	memset(&ib_qp_init_attr, 0, sizeof(ib_qp_init_attr));
4233 	ib_qp_init_attr.qp_context = init_attr->wq_context;
4234 	ib_qp_init_attr.qp_type = IB_QPT_RAW_PACKET;
4235 	ib_qp_init_attr.cap.max_recv_wr = init_attr->max_wr;
4236 	ib_qp_init_attr.cap.max_recv_sge = init_attr->max_sge;
4237 	ib_qp_init_attr.recv_cq = init_attr->cq;
4238 	ib_qp_init_attr.send_cq = ib_qp_init_attr.recv_cq; /* Dummy CQ */
4239 
4240 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS)
4241 		ib_qp_init_attr.create_flags |= IB_QP_CREATE_SCATTER_FCS;
4242 
4243 	err = create_qp_common(dev, pd, MLX4_IB_RWQ_SRC, &ib_qp_init_attr,
4244 			       udata, 0, &qp);
4245 	if (err) {
4246 		kfree(qp);
4247 		return ERR_PTR(err);
4248 	}
4249 
4250 	qp->ibwq.event_handler = init_attr->event_handler;
4251 	qp->ibwq.wq_num = qp->mqp.qpn;
4252 	qp->ibwq.state = IB_WQS_RESET;
4253 
4254 	return &qp->ibwq;
4255 }
4256 
4257 static int ib_wq2qp_state(enum ib_wq_state state)
4258 {
4259 	switch (state) {
4260 	case IB_WQS_RESET:
4261 		return IB_QPS_RESET;
4262 	case IB_WQS_RDY:
4263 		return IB_QPS_RTR;
4264 	default:
4265 		return IB_QPS_ERR;
4266 	}
4267 }
4268 
4269 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state)
4270 {
4271 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4272 	enum ib_qp_state qp_cur_state;
4273 	enum ib_qp_state qp_new_state;
4274 	int attr_mask;
4275 	int err;
4276 
4277 	/* ib_qp.state represents the WQ HW state while ib_wq.state represents
4278 	 * the WQ logic state.
4279 	 */
4280 	qp_cur_state = qp->state;
4281 	qp_new_state = ib_wq2qp_state(new_state);
4282 
4283 	if (ib_wq2qp_state(new_state) == qp_cur_state)
4284 		return 0;
4285 
4286 	if (new_state == IB_WQS_RDY) {
4287 		struct ib_qp_attr attr = {};
4288 
4289 		attr.port_num = qp->port;
4290 		attr_mask = IB_QP_PORT;
4291 
4292 		err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, &attr,
4293 					  attr_mask, IB_QPS_RESET, IB_QPS_INIT);
4294 		if (err) {
4295 			pr_debug("WQN=0x%06x failed to apply RST->INIT on the HW QP\n",
4296 				 ibwq->wq_num);
4297 			return err;
4298 		}
4299 
4300 		qp_cur_state = IB_QPS_INIT;
4301 	}
4302 
4303 	attr_mask = 0;
4304 	err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, attr_mask,
4305 				  qp_cur_state,  qp_new_state);
4306 
4307 	if (err && (qp_cur_state == IB_QPS_INIT)) {
4308 		qp_new_state = IB_QPS_RESET;
4309 		if (__mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL,
4310 					attr_mask, IB_QPS_INIT, IB_QPS_RESET)) {
4311 			pr_warn("WQN=0x%06x failed with reverting HW's resources failure\n",
4312 				ibwq->wq_num);
4313 			qp_new_state = IB_QPS_INIT;
4314 		}
4315 	}
4316 
4317 	qp->state = qp_new_state;
4318 
4319 	return err;
4320 }
4321 
4322 int mlx4_ib_modify_wq(struct ib_wq *ibwq, struct ib_wq_attr *wq_attr,
4323 		      u32 wq_attr_mask, struct ib_udata *udata)
4324 {
4325 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4326 	struct mlx4_ib_modify_wq ucmd = {};
4327 	size_t required_cmd_sz;
4328 	enum ib_wq_state cur_state, new_state;
4329 	int err = 0;
4330 
4331 	required_cmd_sz = offsetof(typeof(ucmd), reserved) +
4332 				   sizeof(ucmd.reserved);
4333 	if (udata->inlen < required_cmd_sz)
4334 		return -EINVAL;
4335 
4336 	if (udata->inlen > sizeof(ucmd) &&
4337 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4338 				 udata->inlen - sizeof(ucmd)))
4339 		return -EOPNOTSUPP;
4340 
4341 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4342 		return -EFAULT;
4343 
4344 	if (ucmd.comp_mask || ucmd.reserved)
4345 		return -EOPNOTSUPP;
4346 
4347 	if (wq_attr_mask & IB_WQ_FLAGS)
4348 		return -EOPNOTSUPP;
4349 
4350 	cur_state = wq_attr_mask & IB_WQ_CUR_STATE ? wq_attr->curr_wq_state :
4351 						     ibwq->state;
4352 	new_state = wq_attr_mask & IB_WQ_STATE ? wq_attr->wq_state : cur_state;
4353 
4354 	if (cur_state  < IB_WQS_RESET || cur_state  > IB_WQS_ERR ||
4355 	    new_state < IB_WQS_RESET || new_state > IB_WQS_ERR)
4356 		return -EINVAL;
4357 
4358 	if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR))
4359 		return -EINVAL;
4360 
4361 	if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET))
4362 		return -EINVAL;
4363 
4364 	/* Need to protect against the parent RSS which also may modify WQ
4365 	 * state.
4366 	 */
4367 	mutex_lock(&qp->mutex);
4368 
4369 	/* Can update HW state only if a RSS QP has already associated to this
4370 	 * WQ, so we can apply its port on the WQ.
4371 	 */
4372 	if (qp->rss_usecnt)
4373 		err = _mlx4_ib_modify_wq(ibwq, new_state);
4374 
4375 	if (!err)
4376 		ibwq->state = new_state;
4377 
4378 	mutex_unlock(&qp->mutex);
4379 
4380 	return err;
4381 }
4382 
4383 int mlx4_ib_destroy_wq(struct ib_wq *ibwq)
4384 {
4385 	struct mlx4_ib_dev *dev = to_mdev(ibwq->device);
4386 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4387 
4388 	if (qp->counter_index)
4389 		mlx4_ib_free_qp_counter(dev, qp);
4390 
4391 	destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, 1);
4392 
4393 	kfree(qp);
4394 
4395 	return 0;
4396 }
4397 
4398 struct ib_rwq_ind_table
4399 *mlx4_ib_create_rwq_ind_table(struct ib_device *device,
4400 			      struct ib_rwq_ind_table_init_attr *init_attr,
4401 			      struct ib_udata *udata)
4402 {
4403 	struct ib_rwq_ind_table *rwq_ind_table;
4404 	struct mlx4_ib_create_rwq_ind_tbl_resp resp = {};
4405 	unsigned int ind_tbl_size = 1 << init_attr->log_ind_tbl_size;
4406 	unsigned int base_wqn;
4407 	size_t min_resp_len;
4408 	int i;
4409 	int err;
4410 
4411 	if (udata->inlen > 0 &&
4412 	    !ib_is_udata_cleared(udata, 0,
4413 				 udata->inlen))
4414 		return ERR_PTR(-EOPNOTSUPP);
4415 
4416 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4417 	if (udata->outlen && udata->outlen < min_resp_len)
4418 		return ERR_PTR(-EINVAL);
4419 
4420 	if (ind_tbl_size >
4421 	    device->attrs.rss_caps.max_rwq_indirection_table_size) {
4422 		pr_debug("log_ind_tbl_size = %d is bigger than supported = %d\n",
4423 			 ind_tbl_size,
4424 			 device->attrs.rss_caps.max_rwq_indirection_table_size);
4425 		return ERR_PTR(-EINVAL);
4426 	}
4427 
4428 	base_wqn = init_attr->ind_tbl[0]->wq_num;
4429 
4430 	if (base_wqn % ind_tbl_size) {
4431 		pr_debug("WQN=0x%x isn't aligned with indirection table size\n",
4432 			 base_wqn);
4433 		return ERR_PTR(-EINVAL);
4434 	}
4435 
4436 	for (i = 1; i < ind_tbl_size; i++) {
4437 		if (++base_wqn != init_attr->ind_tbl[i]->wq_num) {
4438 			pr_debug("indirection table's WQNs aren't consecutive\n");
4439 			return ERR_PTR(-EINVAL);
4440 		}
4441 	}
4442 
4443 	rwq_ind_table = kzalloc(sizeof(*rwq_ind_table), GFP_KERNEL);
4444 	if (!rwq_ind_table)
4445 		return ERR_PTR(-ENOMEM);
4446 
4447 	if (udata->outlen) {
4448 		resp.response_length = offsetof(typeof(resp), response_length) +
4449 					sizeof(resp.response_length);
4450 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4451 		if (err)
4452 			goto err;
4453 	}
4454 
4455 	return rwq_ind_table;
4456 
4457 err:
4458 	kfree(rwq_ind_table);
4459 	return ERR_PTR(err);
4460 }
4461 
4462 int mlx4_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4463 {
4464 	kfree(ib_rwq_ind_tbl);
4465 	return 0;
4466 }
4467