xref: /linux/drivers/infiniband/hw/mlx4/qp.c (revision b9b77222d4ff6b5bb8f5d87fca20de0910618bb9)
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_array(qp->rq.wqe_cnt, sizeof(struct mlx4_ib_buf),
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 					  MLX4_IB_RX_HASH_INNER)) {
678 		pr_debug("RX Hash fields_mask has unsupported mask (0x%llx)\n",
679 			 ucmd->rx_hash_fields_mask);
680 		return (-EOPNOTSUPP);
681 	}
682 
683 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) &&
684 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
685 		rss_ctx->flags = MLX4_RSS_IPV4;
686 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) ||
687 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
688 		pr_debug("RX Hash fields_mask is not supported - both IPv4 SRC and DST must be set\n");
689 		return (-EOPNOTSUPP);
690 	}
691 
692 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) &&
693 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
694 		rss_ctx->flags |= MLX4_RSS_IPV6;
695 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) ||
696 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
697 		pr_debug("RX Hash fields_mask is not supported - both IPv6 SRC and DST must be set\n");
698 		return (-EOPNOTSUPP);
699 	}
700 
701 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) &&
702 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
703 		if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS)) {
704 			pr_debug("RX Hash fields_mask for UDP is not supported\n");
705 			return (-EOPNOTSUPP);
706 		}
707 
708 		if (rss_ctx->flags & MLX4_RSS_IPV4)
709 			rss_ctx->flags |= MLX4_RSS_UDP_IPV4;
710 		if (rss_ctx->flags & MLX4_RSS_IPV6)
711 			rss_ctx->flags |= MLX4_RSS_UDP_IPV6;
712 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
713 			pr_debug("RX Hash fields_mask is not supported - UDP must be set with IPv4 or IPv6\n");
714 			return (-EOPNOTSUPP);
715 		}
716 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) ||
717 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
718 		pr_debug("RX Hash fields_mask is not supported - both UDP SRC and DST must be set\n");
719 		return (-EOPNOTSUPP);
720 	}
721 
722 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) &&
723 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
724 		if (rss_ctx->flags & MLX4_RSS_IPV4)
725 			rss_ctx->flags |= MLX4_RSS_TCP_IPV4;
726 		if (rss_ctx->flags & MLX4_RSS_IPV6)
727 			rss_ctx->flags |= MLX4_RSS_TCP_IPV6;
728 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
729 			pr_debug("RX Hash fields_mask is not supported - TCP must be set with IPv4 or IPv6\n");
730 			return (-EOPNOTSUPP);
731 		}
732 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) ||
733 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
734 		pr_debug("RX Hash fields_mask is not supported - both TCP SRC and DST must be set\n");
735 		return (-EOPNOTSUPP);
736 	}
737 
738 	if (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_INNER) {
739 		if (dev->dev->caps.tunnel_offload_mode ==
740 		    MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
741 			/*
742 			 * Hash according to inner headers if exist, otherwise
743 			 * according to outer headers.
744 			 */
745 			rss_ctx->flags |= MLX4_RSS_BY_INNER_HEADERS_IPONLY;
746 		} else {
747 			pr_debug("RSS Hash for inner headers isn't supported\n");
748 			return (-EOPNOTSUPP);
749 		}
750 	}
751 
752 	return 0;
753 }
754 
755 static int create_qp_rss(struct mlx4_ib_dev *dev,
756 			 struct ib_qp_init_attr *init_attr,
757 			 struct mlx4_ib_create_qp_rss *ucmd,
758 			 struct mlx4_ib_qp *qp)
759 {
760 	int qpn;
761 	int err;
762 
763 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
764 
765 	err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 0, qp->mqp.usage);
766 	if (err)
767 		return err;
768 
769 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
770 	if (err)
771 		goto err_qpn;
772 
773 	mutex_init(&qp->mutex);
774 
775 	INIT_LIST_HEAD(&qp->gid_list);
776 	INIT_LIST_HEAD(&qp->steering_rules);
777 
778 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
779 	qp->state = IB_QPS_RESET;
780 
781 	/* Set dummy send resources to be compatible with HV and PRM */
782 	qp->sq_no_prefetch = 1;
783 	qp->sq.wqe_cnt = 1;
784 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
785 	qp->buf_size = qp->sq.wqe_cnt << MLX4_IB_MIN_SQ_STRIDE;
786 	qp->mtt = (to_mqp(
787 		   (struct ib_qp *)init_attr->rwq_ind_tbl->ind_tbl[0]))->mtt;
788 
789 	qp->rss_ctx = kzalloc(sizeof(*qp->rss_ctx), GFP_KERNEL);
790 	if (!qp->rss_ctx) {
791 		err = -ENOMEM;
792 		goto err_qp_alloc;
793 	}
794 
795 	err = set_qp_rss(dev, qp->rss_ctx, init_attr, ucmd);
796 	if (err)
797 		goto err;
798 
799 	return 0;
800 
801 err:
802 	kfree(qp->rss_ctx);
803 
804 err_qp_alloc:
805 	mlx4_qp_remove(dev->dev, &qp->mqp);
806 	mlx4_qp_free(dev->dev, &qp->mqp);
807 
808 err_qpn:
809 	mlx4_qp_release_range(dev->dev, qpn, 1);
810 	return err;
811 }
812 
813 static struct ib_qp *_mlx4_ib_create_qp_rss(struct ib_pd *pd,
814 					    struct ib_qp_init_attr *init_attr,
815 					    struct ib_udata *udata)
816 {
817 	struct mlx4_ib_qp *qp;
818 	struct mlx4_ib_create_qp_rss ucmd = {};
819 	size_t required_cmd_sz;
820 	int err;
821 
822 	if (!udata) {
823 		pr_debug("RSS QP with NULL udata\n");
824 		return ERR_PTR(-EINVAL);
825 	}
826 
827 	if (udata->outlen)
828 		return ERR_PTR(-EOPNOTSUPP);
829 
830 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) +
831 					sizeof(ucmd.reserved1);
832 	if (udata->inlen < required_cmd_sz) {
833 		pr_debug("invalid inlen\n");
834 		return ERR_PTR(-EINVAL);
835 	}
836 
837 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
838 		pr_debug("copy failed\n");
839 		return ERR_PTR(-EFAULT);
840 	}
841 
842 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)))
843 		return ERR_PTR(-EOPNOTSUPP);
844 
845 	if (ucmd.comp_mask || ucmd.reserved1)
846 		return ERR_PTR(-EOPNOTSUPP);
847 
848 	if (udata->inlen > sizeof(ucmd) &&
849 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
850 				 udata->inlen - sizeof(ucmd))) {
851 		pr_debug("inlen is not supported\n");
852 		return ERR_PTR(-EOPNOTSUPP);
853 	}
854 
855 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
856 		pr_debug("RSS QP with unsupported QP type %d\n",
857 			 init_attr->qp_type);
858 		return ERR_PTR(-EOPNOTSUPP);
859 	}
860 
861 	if (init_attr->create_flags) {
862 		pr_debug("RSS QP doesn't support create flags\n");
863 		return ERR_PTR(-EOPNOTSUPP);
864 	}
865 
866 	if (init_attr->send_cq || init_attr->cap.max_send_wr) {
867 		pr_debug("RSS QP with unsupported send attributes\n");
868 		return ERR_PTR(-EOPNOTSUPP);
869 	}
870 
871 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
872 	if (!qp)
873 		return ERR_PTR(-ENOMEM);
874 
875 	qp->pri.vid = 0xFFFF;
876 	qp->alt.vid = 0xFFFF;
877 
878 	err = create_qp_rss(to_mdev(pd->device), init_attr, &ucmd, qp);
879 	if (err) {
880 		kfree(qp);
881 		return ERR_PTR(err);
882 	}
883 
884 	qp->ibqp.qp_num = qp->mqp.qpn;
885 
886 	return &qp->ibqp;
887 }
888 
889 /*
890  * This function allocates a WQN from a range which is consecutive and aligned
891  * to its size. In case the range is full, then it creates a new range and
892  * allocates WQN from it. The new range will be used for following allocations.
893  */
894 static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context,
895 			     struct mlx4_ib_qp *qp, int range_size, int *wqn)
896 {
897 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
898 	struct mlx4_wqn_range *range;
899 	int err = 0;
900 
901 	mutex_lock(&context->wqn_ranges_mutex);
902 
903 	range = list_first_entry_or_null(&context->wqn_ranges_list,
904 					 struct mlx4_wqn_range, list);
905 
906 	if (!range || (range->refcount == range->size) || range->dirty) {
907 		range = kzalloc(sizeof(*range), GFP_KERNEL);
908 		if (!range) {
909 			err = -ENOMEM;
910 			goto out;
911 		}
912 
913 		err = mlx4_qp_reserve_range(dev->dev, range_size,
914 					    range_size, &range->base_wqn, 0,
915 					    qp->mqp.usage);
916 		if (err) {
917 			kfree(range);
918 			goto out;
919 		}
920 
921 		range->size = range_size;
922 		list_add(&range->list, &context->wqn_ranges_list);
923 	} else if (range_size != 1) {
924 		/*
925 		 * Requesting a new range (>1) when last range is still open, is
926 		 * not valid.
927 		 */
928 		err = -EINVAL;
929 		goto out;
930 	}
931 
932 	qp->wqn_range = range;
933 
934 	*wqn = range->base_wqn + range->refcount;
935 
936 	range->refcount++;
937 
938 out:
939 	mutex_unlock(&context->wqn_ranges_mutex);
940 
941 	return err;
942 }
943 
944 static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context,
945 				struct mlx4_ib_qp *qp, bool dirty_release)
946 {
947 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
948 	struct mlx4_wqn_range *range;
949 
950 	mutex_lock(&context->wqn_ranges_mutex);
951 
952 	range = qp->wqn_range;
953 
954 	range->refcount--;
955 	if (!range->refcount) {
956 		mlx4_qp_release_range(dev->dev, range->base_wqn,
957 				      range->size);
958 		list_del(&range->list);
959 		kfree(range);
960 	} else if (dirty_release) {
961 	/*
962 	 * A range which one of its WQNs is destroyed, won't be able to be
963 	 * reused for further WQN allocations.
964 	 * The next created WQ will allocate a new range.
965 	 */
966 		range->dirty = 1;
967 	}
968 
969 	mutex_unlock(&context->wqn_ranges_mutex);
970 }
971 
972 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
973 			    enum mlx4_ib_source_type src,
974 			    struct ib_qp_init_attr *init_attr,
975 			    struct ib_udata *udata, int sqpn,
976 			    struct mlx4_ib_qp **caller_qp)
977 {
978 	int qpn;
979 	int err;
980 	struct ib_qp_cap backup_cap;
981 	struct mlx4_ib_sqp *sqp = NULL;
982 	struct mlx4_ib_qp *qp;
983 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
984 	struct mlx4_ib_cq *mcq;
985 	unsigned long flags;
986 	int range_size = 0;
987 
988 	/* When tunneling special qps, we use a plain UD qp */
989 	if (sqpn) {
990 		if (mlx4_is_mfunc(dev->dev) &&
991 		    (!mlx4_is_master(dev->dev) ||
992 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
993 			if (init_attr->qp_type == IB_QPT_GSI)
994 				qp_type = MLX4_IB_QPT_PROXY_GSI;
995 			else {
996 				if (mlx4_is_master(dev->dev) ||
997 				    qp0_enabled_vf(dev->dev, sqpn))
998 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
999 				else
1000 					qp_type = MLX4_IB_QPT_PROXY_SMI;
1001 			}
1002 		}
1003 		qpn = sqpn;
1004 		/* add extra sg entry for tunneling */
1005 		init_attr->cap.max_recv_sge++;
1006 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
1007 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
1008 			container_of(init_attr,
1009 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
1010 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
1011 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
1012 		    !mlx4_is_master(dev->dev))
1013 			return -EINVAL;
1014 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
1015 			qp_type = MLX4_IB_QPT_TUN_GSI;
1016 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
1017 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
1018 					     tnl_init->port))
1019 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
1020 		else
1021 			qp_type = MLX4_IB_QPT_TUN_SMI;
1022 		/* we are definitely in the PPF here, since we are creating
1023 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
1024 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
1025 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
1026 		sqpn = qpn;
1027 	}
1028 
1029 	if (!*caller_qp) {
1030 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
1031 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
1032 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
1033 			sqp = kzalloc(sizeof(struct mlx4_ib_sqp), GFP_KERNEL);
1034 			if (!sqp)
1035 				return -ENOMEM;
1036 			qp = &sqp->qp;
1037 			qp->pri.vid = 0xFFFF;
1038 			qp->alt.vid = 0xFFFF;
1039 		} else {
1040 			qp = kzalloc(sizeof(struct mlx4_ib_qp), GFP_KERNEL);
1041 			if (!qp)
1042 				return -ENOMEM;
1043 			qp->pri.vid = 0xFFFF;
1044 			qp->alt.vid = 0xFFFF;
1045 		}
1046 	} else
1047 		qp = *caller_qp;
1048 
1049 	qp->mlx4_ib_qp_type = qp_type;
1050 
1051 	mutex_init(&qp->mutex);
1052 	spin_lock_init(&qp->sq.lock);
1053 	spin_lock_init(&qp->rq.lock);
1054 	INIT_LIST_HEAD(&qp->gid_list);
1055 	INIT_LIST_HEAD(&qp->steering_rules);
1056 
1057 	qp->state	 = IB_QPS_RESET;
1058 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1059 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1060 
1061 
1062 	if (pd->uobject) {
1063 		union {
1064 			struct mlx4_ib_create_qp qp;
1065 			struct mlx4_ib_create_wq wq;
1066 		} ucmd;
1067 		size_t copy_len;
1068 		int shift;
1069 		int n;
1070 
1071 		copy_len = (src == MLX4_IB_QP_SRC) ?
1072 			   sizeof(struct mlx4_ib_create_qp) :
1073 			   min(sizeof(struct mlx4_ib_create_wq), udata->inlen);
1074 
1075 		if (ib_copy_from_udata(&ucmd, udata, copy_len)) {
1076 			err = -EFAULT;
1077 			goto err;
1078 		}
1079 
1080 		if (src == MLX4_IB_RWQ_SRC) {
1081 			if (ucmd.wq.comp_mask || ucmd.wq.reserved[0] ||
1082 			    ucmd.wq.reserved[1] || ucmd.wq.reserved[2]) {
1083 				pr_debug("user command isn't supported\n");
1084 				err = -EOPNOTSUPP;
1085 				goto err;
1086 			}
1087 
1088 			if (ucmd.wq.log_range_size >
1089 			    ilog2(dev->dev->caps.max_rss_tbl_sz)) {
1090 				pr_debug("WQN range size must be equal or smaller than %d\n",
1091 					 dev->dev->caps.max_rss_tbl_sz);
1092 				err = -EOPNOTSUPP;
1093 				goto err;
1094 			}
1095 			range_size = 1 << ucmd.wq.log_range_size;
1096 		} else {
1097 			qp->inl_recv_sz = ucmd.qp.inl_recv_sz;
1098 		}
1099 
1100 		if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1101 			if (!(dev->dev->caps.flags &
1102 			      MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
1103 				pr_debug("scatter FCS is unsupported\n");
1104 				err = -EOPNOTSUPP;
1105 				goto err;
1106 			}
1107 
1108 			qp->flags |= MLX4_IB_QP_SCATTER_FCS;
1109 		}
1110 
1111 		err = set_rq_size(dev, &init_attr->cap, !!pd->uobject,
1112 				  qp_has_rq(init_attr), qp, qp->inl_recv_sz);
1113 		if (err)
1114 			goto err;
1115 
1116 		if (src == MLX4_IB_QP_SRC) {
1117 			qp->sq_no_prefetch = ucmd.qp.sq_no_prefetch;
1118 
1119 			err = set_user_sq_size(dev, qp,
1120 					       (struct mlx4_ib_create_qp *)
1121 					       &ucmd);
1122 			if (err)
1123 				goto err;
1124 		} else {
1125 			qp->sq_no_prefetch = 1;
1126 			qp->sq.wqe_cnt = 1;
1127 			qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
1128 			/* Allocated buffer expects to have at least that SQ
1129 			 * size.
1130 			 */
1131 			qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
1132 				(qp->sq.wqe_cnt << qp->sq.wqe_shift);
1133 		}
1134 
1135 		qp->umem = ib_umem_get(pd->uobject->context,
1136 				(src == MLX4_IB_QP_SRC) ? ucmd.qp.buf_addr :
1137 				ucmd.wq.buf_addr, qp->buf_size, 0, 0);
1138 		if (IS_ERR(qp->umem)) {
1139 			err = PTR_ERR(qp->umem);
1140 			goto err;
1141 		}
1142 
1143 		n = ib_umem_page_count(qp->umem);
1144 		shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
1145 		err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
1146 
1147 		if (err)
1148 			goto err_buf;
1149 
1150 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
1151 		if (err)
1152 			goto err_mtt;
1153 
1154 		if (qp_has_rq(init_attr)) {
1155 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
1156 				(src == MLX4_IB_QP_SRC) ? ucmd.qp.db_addr :
1157 				ucmd.wq.db_addr, &qp->db);
1158 			if (err)
1159 				goto err_mtt;
1160 		}
1161 		qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
1162 	} else {
1163 		err = set_rq_size(dev, &init_attr->cap, !!pd->uobject,
1164 				  qp_has_rq(init_attr), qp, 0);
1165 		if (err)
1166 			goto err;
1167 
1168 		qp->sq_no_prefetch = 0;
1169 
1170 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
1171 			qp->flags |= MLX4_IB_QP_LSO;
1172 
1173 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1174 			if (dev->steering_support ==
1175 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
1176 				qp->flags |= MLX4_IB_QP_NETIF;
1177 			else
1178 				goto err;
1179 		}
1180 
1181 		memcpy(&backup_cap, &init_attr->cap, sizeof(backup_cap));
1182 		err = set_kernel_sq_size(dev, &init_attr->cap,
1183 					 qp_type, qp, true);
1184 		if (err)
1185 			goto err;
1186 
1187 		if (qp_has_rq(init_attr)) {
1188 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
1189 			if (err)
1190 				goto err;
1191 
1192 			*qp->db.db = 0;
1193 		}
1194 
1195 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, qp->buf_size,
1196 				   &qp->buf)) {
1197 			memcpy(&init_attr->cap, &backup_cap,
1198 			       sizeof(backup_cap));
1199 			err = set_kernel_sq_size(dev, &init_attr->cap, qp_type,
1200 						 qp, false);
1201 			if (err)
1202 				goto err_db;
1203 
1204 			if (mlx4_buf_alloc(dev->dev, qp->buf_size,
1205 					   PAGE_SIZE * 2, &qp->buf)) {
1206 				err = -ENOMEM;
1207 				goto err_db;
1208 			}
1209 		}
1210 
1211 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
1212 				    &qp->mtt);
1213 		if (err)
1214 			goto err_buf;
1215 
1216 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
1217 		if (err)
1218 			goto err_mtt;
1219 
1220 		qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1221 					     sizeof(u64), GFP_KERNEL);
1222 		qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1223 					     sizeof(u64), GFP_KERNEL);
1224 		if (!qp->sq.wrid || !qp->rq.wrid) {
1225 			err = -ENOMEM;
1226 			goto err_wrid;
1227 		}
1228 		qp->mqp.usage = MLX4_RES_USAGE_DRIVER;
1229 	}
1230 
1231 	if (sqpn) {
1232 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1233 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
1234 			if (alloc_proxy_bufs(pd->device, qp)) {
1235 				err = -ENOMEM;
1236 				goto err_wrid;
1237 			}
1238 		}
1239 	} else if (src == MLX4_IB_RWQ_SRC) {
1240 		err = mlx4_ib_alloc_wqn(to_mucontext(pd->uobject->context), qp,
1241 					range_size, &qpn);
1242 		if (err)
1243 			goto err_wrid;
1244 	} else {
1245 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
1246 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
1247 		 * VLAN insertion. */
1248 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
1249 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
1250 						    (init_attr->cap.max_send_wr ?
1251 						     MLX4_RESERVE_ETH_BF_QP : 0) |
1252 						    (init_attr->cap.max_recv_wr ?
1253 						     MLX4_RESERVE_A0_QP : 0),
1254 						    qp->mqp.usage);
1255 		else
1256 			if (qp->flags & MLX4_IB_QP_NETIF)
1257 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
1258 			else
1259 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
1260 							    &qpn, 0, qp->mqp.usage);
1261 		if (err)
1262 			goto err_proxy;
1263 	}
1264 
1265 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1266 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1267 
1268 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
1269 	if (err)
1270 		goto err_qpn;
1271 
1272 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1273 		qp->mqp.qpn |= (1 << 23);
1274 
1275 	/*
1276 	 * Hardware wants QPN written in big-endian order (after
1277 	 * shifting) for send doorbell.  Precompute this value to save
1278 	 * a little bit when posting sends.
1279 	 */
1280 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1281 
1282 	qp->mqp.event = (src == MLX4_IB_QP_SRC) ? mlx4_ib_qp_event :
1283 						  mlx4_ib_wq_event;
1284 
1285 	if (!*caller_qp)
1286 		*caller_qp = qp;
1287 
1288 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1289 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
1290 			 to_mcq(init_attr->recv_cq));
1291 	/* Maintain device to QPs access, needed for further handling
1292 	 * via reset flow
1293 	 */
1294 	list_add_tail(&qp->qps_list, &dev->qp_list);
1295 	/* Maintain CQ to QPs access, needed for further handling
1296 	 * via reset flow
1297 	 */
1298 	mcq = to_mcq(init_attr->send_cq);
1299 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
1300 	mcq = to_mcq(init_attr->recv_cq);
1301 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
1302 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
1303 			   to_mcq(init_attr->recv_cq));
1304 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1305 	return 0;
1306 
1307 err_qpn:
1308 	if (!sqpn) {
1309 		if (qp->flags & MLX4_IB_QP_NETIF)
1310 			mlx4_ib_steer_qp_free(dev, qpn, 1);
1311 		else if (src == MLX4_IB_RWQ_SRC)
1312 			mlx4_ib_release_wqn(to_mucontext(pd->uobject->context),
1313 					    qp, 0);
1314 		else
1315 			mlx4_qp_release_range(dev->dev, qpn, 1);
1316 	}
1317 err_proxy:
1318 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1319 		free_proxy_bufs(pd->device, qp);
1320 err_wrid:
1321 	if (pd->uobject) {
1322 		if (qp_has_rq(init_attr))
1323 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
1324 	} else {
1325 		kvfree(qp->sq.wrid);
1326 		kvfree(qp->rq.wrid);
1327 	}
1328 
1329 err_mtt:
1330 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1331 
1332 err_buf:
1333 	if (pd->uobject)
1334 		ib_umem_release(qp->umem);
1335 	else
1336 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1337 
1338 err_db:
1339 	if (!pd->uobject && qp_has_rq(init_attr))
1340 		mlx4_db_free(dev->dev, &qp->db);
1341 
1342 err:
1343 	if (sqp)
1344 		kfree(sqp);
1345 	else if (!*caller_qp)
1346 		kfree(qp);
1347 	return err;
1348 }
1349 
1350 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
1351 {
1352 	switch (state) {
1353 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
1354 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
1355 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
1356 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
1357 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
1358 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
1359 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
1360 	default:		return -1;
1361 	}
1362 }
1363 
1364 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1365 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1366 {
1367 	if (send_cq == recv_cq) {
1368 		spin_lock(&send_cq->lock);
1369 		__acquire(&recv_cq->lock);
1370 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1371 		spin_lock(&send_cq->lock);
1372 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1373 	} else {
1374 		spin_lock(&recv_cq->lock);
1375 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1376 	}
1377 }
1378 
1379 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1380 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1381 {
1382 	if (send_cq == recv_cq) {
1383 		__release(&recv_cq->lock);
1384 		spin_unlock(&send_cq->lock);
1385 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1386 		spin_unlock(&recv_cq->lock);
1387 		spin_unlock(&send_cq->lock);
1388 	} else {
1389 		spin_unlock(&send_cq->lock);
1390 		spin_unlock(&recv_cq->lock);
1391 	}
1392 }
1393 
1394 static void del_gid_entries(struct mlx4_ib_qp *qp)
1395 {
1396 	struct mlx4_ib_gid_entry *ge, *tmp;
1397 
1398 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1399 		list_del(&ge->list);
1400 		kfree(ge);
1401 	}
1402 }
1403 
1404 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
1405 {
1406 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
1407 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
1408 	else
1409 		return to_mpd(qp->ibqp.pd);
1410 }
1411 
1412 static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src,
1413 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1414 {
1415 	switch (qp->ibqp.qp_type) {
1416 	case IB_QPT_XRC_TGT:
1417 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1418 		*recv_cq = *send_cq;
1419 		break;
1420 	case IB_QPT_XRC_INI:
1421 		*send_cq = to_mcq(qp->ibqp.send_cq);
1422 		*recv_cq = *send_cq;
1423 		break;
1424 	default:
1425 		*recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) :
1426 						     to_mcq(qp->ibwq.cq);
1427 		*send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) :
1428 						     *recv_cq;
1429 		break;
1430 	}
1431 }
1432 
1433 static void destroy_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1434 {
1435 	if (qp->state != IB_QPS_RESET) {
1436 		int i;
1437 
1438 		for (i = 0; i < (1 << qp->ibqp.rwq_ind_tbl->log_ind_tbl_size);
1439 		     i++) {
1440 			struct ib_wq *ibwq = qp->ibqp.rwq_ind_tbl->ind_tbl[i];
1441 			struct mlx4_ib_qp *wq =	to_mqp((struct ib_qp *)ibwq);
1442 
1443 			mutex_lock(&wq->mutex);
1444 
1445 			wq->rss_usecnt--;
1446 
1447 			mutex_unlock(&wq->mutex);
1448 		}
1449 
1450 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1451 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1452 			pr_warn("modify QP %06x to RESET failed.\n",
1453 				qp->mqp.qpn);
1454 	}
1455 
1456 	mlx4_qp_remove(dev->dev, &qp->mqp);
1457 	mlx4_qp_free(dev->dev, &qp->mqp);
1458 	mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1459 	del_gid_entries(qp);
1460 	kfree(qp->rss_ctx);
1461 }
1462 
1463 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1464 			      enum mlx4_ib_source_type src, int is_user)
1465 {
1466 	struct mlx4_ib_cq *send_cq, *recv_cq;
1467 	unsigned long flags;
1468 
1469 	if (qp->state != IB_QPS_RESET) {
1470 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1471 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1472 			pr_warn("modify QP %06x to RESET failed.\n",
1473 			       qp->mqp.qpn);
1474 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1475 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1476 			qp->pri.smac = 0;
1477 			qp->pri.smac_port = 0;
1478 		}
1479 		if (qp->alt.smac) {
1480 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1481 			qp->alt.smac = 0;
1482 		}
1483 		if (qp->pri.vid < 0x1000) {
1484 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1485 			qp->pri.vid = 0xFFFF;
1486 			qp->pri.candidate_vid = 0xFFFF;
1487 			qp->pri.update_vid = 0;
1488 		}
1489 		if (qp->alt.vid < 0x1000) {
1490 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1491 			qp->alt.vid = 0xFFFF;
1492 			qp->alt.candidate_vid = 0xFFFF;
1493 			qp->alt.update_vid = 0;
1494 		}
1495 	}
1496 
1497 	get_cqs(qp, src, &send_cq, &recv_cq);
1498 
1499 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1500 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1501 
1502 	/* del from lists under both locks above to protect reset flow paths */
1503 	list_del(&qp->qps_list);
1504 	list_del(&qp->cq_send_list);
1505 	list_del(&qp->cq_recv_list);
1506 	if (!is_user) {
1507 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1508 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1509 		if (send_cq != recv_cq)
1510 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1511 	}
1512 
1513 	mlx4_qp_remove(dev->dev, &qp->mqp);
1514 
1515 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1516 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1517 
1518 	mlx4_qp_free(dev->dev, &qp->mqp);
1519 
1520 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1521 		if (qp->flags & MLX4_IB_QP_NETIF)
1522 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1523 		else if (src == MLX4_IB_RWQ_SRC)
1524 			mlx4_ib_release_wqn(to_mucontext(
1525 					    qp->ibwq.uobject->context), qp, 1);
1526 		else
1527 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1528 	}
1529 
1530 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1531 
1532 	if (is_user) {
1533 		if (qp->rq.wqe_cnt) {
1534 			struct mlx4_ib_ucontext *mcontext = !src ?
1535 				to_mucontext(qp->ibqp.uobject->context) :
1536 				to_mucontext(qp->ibwq.uobject->context);
1537 			mlx4_ib_db_unmap_user(mcontext, &qp->db);
1538 		}
1539 		ib_umem_release(qp->umem);
1540 	} else {
1541 		kvfree(qp->sq.wrid);
1542 		kvfree(qp->rq.wrid);
1543 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1544 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1545 			free_proxy_bufs(&dev->ib_dev, qp);
1546 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1547 		if (qp->rq.wqe_cnt)
1548 			mlx4_db_free(dev->dev, &qp->db);
1549 	}
1550 
1551 	del_gid_entries(qp);
1552 }
1553 
1554 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1555 {
1556 	/* Native or PPF */
1557 	if (!mlx4_is_mfunc(dev->dev) ||
1558 	    (mlx4_is_master(dev->dev) &&
1559 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1560 		return  dev->dev->phys_caps.base_sqpn +
1561 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1562 			attr->port_num - 1;
1563 	}
1564 	/* PF or VF -- creating proxies */
1565 	if (attr->qp_type == IB_QPT_SMI)
1566 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp0_proxy;
1567 	else
1568 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp1_proxy;
1569 }
1570 
1571 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1572 					struct ib_qp_init_attr *init_attr,
1573 					struct ib_udata *udata)
1574 {
1575 	struct mlx4_ib_qp *qp = NULL;
1576 	int err;
1577 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1578 	u16 xrcdn = 0;
1579 
1580 	if (init_attr->rwq_ind_tbl)
1581 		return _mlx4_ib_create_qp_rss(pd, init_attr, udata);
1582 
1583 	/*
1584 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1585 	 * and only for kernel UD QPs.
1586 	 */
1587 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1588 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1589 					MLX4_IB_SRIOV_TUNNEL_QP |
1590 					MLX4_IB_SRIOV_SQP |
1591 					MLX4_IB_QP_NETIF |
1592 					MLX4_IB_QP_CREATE_ROCE_V2_GSI))
1593 		return ERR_PTR(-EINVAL);
1594 
1595 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1596 		if (init_attr->qp_type != IB_QPT_UD)
1597 			return ERR_PTR(-EINVAL);
1598 	}
1599 
1600 	if (init_attr->create_flags) {
1601 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1602 			return ERR_PTR(-EINVAL);
1603 
1604 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1605 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1606 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1607 		     init_attr->qp_type != IB_QPT_UD) ||
1608 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1609 		     init_attr->qp_type > IB_QPT_GSI) ||
1610 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1611 		     init_attr->qp_type != IB_QPT_GSI))
1612 			return ERR_PTR(-EINVAL);
1613 	}
1614 
1615 	switch (init_attr->qp_type) {
1616 	case IB_QPT_XRC_TGT:
1617 		pd = to_mxrcd(init_attr->xrcd)->pd;
1618 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1619 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1620 		/* fall through */
1621 	case IB_QPT_XRC_INI:
1622 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1623 			return ERR_PTR(-ENOSYS);
1624 		init_attr->recv_cq = init_attr->send_cq;
1625 		/* fall through */
1626 	case IB_QPT_RC:
1627 	case IB_QPT_UC:
1628 	case IB_QPT_RAW_PACKET:
1629 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1630 		if (!qp)
1631 			return ERR_PTR(-ENOMEM);
1632 		qp->pri.vid = 0xFFFF;
1633 		qp->alt.vid = 0xFFFF;
1634 		/* fall through */
1635 	case IB_QPT_UD:
1636 	{
1637 		err = create_qp_common(to_mdev(pd->device), pd,	MLX4_IB_QP_SRC,
1638 				       init_attr, udata, 0, &qp);
1639 		if (err) {
1640 			kfree(qp);
1641 			return ERR_PTR(err);
1642 		}
1643 
1644 		qp->ibqp.qp_num = qp->mqp.qpn;
1645 		qp->xrcdn = xrcdn;
1646 
1647 		break;
1648 	}
1649 	case IB_QPT_SMI:
1650 	case IB_QPT_GSI:
1651 	{
1652 		int sqpn;
1653 
1654 		/* Userspace is not allowed to create special QPs: */
1655 		if (udata)
1656 			return ERR_PTR(-EINVAL);
1657 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1658 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev,
1659 							1, 1, &sqpn, 0,
1660 							MLX4_RES_USAGE_DRIVER);
1661 
1662 			if (res)
1663 				return ERR_PTR(res);
1664 		} else {
1665 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1666 		}
1667 
1668 		err = create_qp_common(to_mdev(pd->device), pd, MLX4_IB_QP_SRC,
1669 				       init_attr, udata, sqpn, &qp);
1670 		if (err)
1671 			return ERR_PTR(err);
1672 
1673 		qp->port	= init_attr->port_num;
1674 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1675 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1676 		break;
1677 	}
1678 	default:
1679 		/* Don't support raw QPs */
1680 		return ERR_PTR(-EINVAL);
1681 	}
1682 
1683 	return &qp->ibqp;
1684 }
1685 
1686 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1687 				struct ib_qp_init_attr *init_attr,
1688 				struct ib_udata *udata) {
1689 	struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1690 	struct ib_qp *ibqp;
1691 	struct mlx4_ib_dev *dev = to_mdev(device);
1692 
1693 	ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1694 
1695 	if (!IS_ERR(ibqp) &&
1696 	    (init_attr->qp_type == IB_QPT_GSI) &&
1697 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1698 		struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1699 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1700 
1701 		if (is_eth &&
1702 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1703 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1704 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1705 
1706 			if (IS_ERR(sqp->roce_v2_gsi)) {
1707 				pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1708 				sqp->roce_v2_gsi = NULL;
1709 			} else {
1710 				sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1711 				sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1712 			}
1713 
1714 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1715 		}
1716 	}
1717 	return ibqp;
1718 }
1719 
1720 static int _mlx4_ib_destroy_qp(struct ib_qp *qp)
1721 {
1722 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1723 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1724 
1725 	if (is_qp0(dev, mqp))
1726 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1727 
1728 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1729 	    dev->qp1_proxy[mqp->port - 1] == mqp) {
1730 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1731 		dev->qp1_proxy[mqp->port - 1] = NULL;
1732 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1733 	}
1734 
1735 	if (mqp->counter_index)
1736 		mlx4_ib_free_qp_counter(dev, mqp);
1737 
1738 	if (qp->rwq_ind_tbl) {
1739 		destroy_qp_rss(dev, mqp);
1740 	} else {
1741 		struct mlx4_ib_pd *pd;
1742 
1743 		pd = get_pd(mqp);
1744 		destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, !!pd->ibpd.uobject);
1745 	}
1746 
1747 	if (is_sqp(dev, mqp))
1748 		kfree(to_msqp(mqp));
1749 	else
1750 		kfree(mqp);
1751 
1752 	return 0;
1753 }
1754 
1755 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1756 {
1757 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1758 
1759 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1760 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1761 
1762 		if (sqp->roce_v2_gsi)
1763 			ib_destroy_qp(sqp->roce_v2_gsi);
1764 	}
1765 
1766 	return _mlx4_ib_destroy_qp(qp);
1767 }
1768 
1769 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1770 {
1771 	switch (type) {
1772 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1773 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1774 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1775 	case MLX4_IB_QPT_XRC_INI:
1776 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1777 	case MLX4_IB_QPT_SMI:
1778 	case MLX4_IB_QPT_GSI:
1779 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1780 
1781 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1782 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1783 						MLX4_QP_ST_MLX : -1);
1784 	case MLX4_IB_QPT_PROXY_SMI:
1785 	case MLX4_IB_QPT_TUN_SMI:
1786 	case MLX4_IB_QPT_PROXY_GSI:
1787 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1788 						MLX4_QP_ST_UD : -1);
1789 	default:			return -1;
1790 	}
1791 }
1792 
1793 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1794 				   int attr_mask)
1795 {
1796 	u8 dest_rd_atomic;
1797 	u32 access_flags;
1798 	u32 hw_access_flags = 0;
1799 
1800 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1801 		dest_rd_atomic = attr->max_dest_rd_atomic;
1802 	else
1803 		dest_rd_atomic = qp->resp_depth;
1804 
1805 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1806 		access_flags = attr->qp_access_flags;
1807 	else
1808 		access_flags = qp->atomic_rd_en;
1809 
1810 	if (!dest_rd_atomic)
1811 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1812 
1813 	if (access_flags & IB_ACCESS_REMOTE_READ)
1814 		hw_access_flags |= MLX4_QP_BIT_RRE;
1815 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1816 		hw_access_flags |= MLX4_QP_BIT_RAE;
1817 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1818 		hw_access_flags |= MLX4_QP_BIT_RWE;
1819 
1820 	return cpu_to_be32(hw_access_flags);
1821 }
1822 
1823 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1824 			    int attr_mask)
1825 {
1826 	if (attr_mask & IB_QP_PKEY_INDEX)
1827 		sqp->pkey_index = attr->pkey_index;
1828 	if (attr_mask & IB_QP_QKEY)
1829 		sqp->qkey = attr->qkey;
1830 	if (attr_mask & IB_QP_SQ_PSN)
1831 		sqp->send_psn = attr->sq_psn;
1832 }
1833 
1834 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1835 {
1836 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1837 }
1838 
1839 static int _mlx4_set_path(struct mlx4_ib_dev *dev,
1840 			  const struct rdma_ah_attr *ah,
1841 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1842 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1843 {
1844 	int vidx;
1845 	int smac_index;
1846 	int err;
1847 
1848 	path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
1849 	path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
1850 	if (rdma_ah_get_static_rate(ah)) {
1851 		path->static_rate = rdma_ah_get_static_rate(ah) +
1852 				    MLX4_STAT_RATE_OFFSET;
1853 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1854 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1855 			--path->static_rate;
1856 	} else
1857 		path->static_rate = 0;
1858 
1859 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
1860 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
1861 		int real_sgid_index =
1862 			mlx4_ib_gid_index_to_real_index(dev, port,
1863 							grh->sgid_index);
1864 
1865 		if (real_sgid_index < 0)
1866 			return real_sgid_index;
1867 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1868 			pr_err("sgid_index (%u) too large. max is %d\n",
1869 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1870 			return -1;
1871 		}
1872 
1873 		path->grh_mylmc |= 1 << 7;
1874 		path->mgid_index = real_sgid_index;
1875 		path->hop_limit  = grh->hop_limit;
1876 		path->tclass_flowlabel =
1877 			cpu_to_be32((grh->traffic_class << 20) |
1878 				    (grh->flow_label));
1879 		memcpy(path->rgid, grh->dgid.raw, 16);
1880 	}
1881 
1882 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
1883 		if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
1884 			return -1;
1885 
1886 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1887 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
1888 
1889 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1890 		if (vlan_tag < 0x1000) {
1891 			if (smac_info->vid < 0x1000) {
1892 				/* both valid vlan ids */
1893 				if (smac_info->vid != vlan_tag) {
1894 					/* different VIDs.  unreg old and reg new */
1895 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1896 					if (err)
1897 						return err;
1898 					smac_info->candidate_vid = vlan_tag;
1899 					smac_info->candidate_vlan_index = vidx;
1900 					smac_info->candidate_vlan_port = port;
1901 					smac_info->update_vid = 1;
1902 					path->vlan_index = vidx;
1903 				} else {
1904 					path->vlan_index = smac_info->vlan_index;
1905 				}
1906 			} else {
1907 				/* no current vlan tag in qp */
1908 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1909 				if (err)
1910 					return err;
1911 				smac_info->candidate_vid = vlan_tag;
1912 				smac_info->candidate_vlan_index = vidx;
1913 				smac_info->candidate_vlan_port = port;
1914 				smac_info->update_vid = 1;
1915 				path->vlan_index = vidx;
1916 			}
1917 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1918 			path->fl = 1 << 6;
1919 		} else {
1920 			/* have current vlan tag. unregister it at modify-qp success */
1921 			if (smac_info->vid < 0x1000) {
1922 				smac_info->candidate_vid = 0xFFFF;
1923 				smac_info->update_vid = 1;
1924 			}
1925 		}
1926 
1927 		/* get smac_index for RoCE use.
1928 		 * If no smac was yet assigned, register one.
1929 		 * If one was already assigned, but the new mac differs,
1930 		 * unregister the old one and register the new one.
1931 		*/
1932 		if ((!smac_info->smac && !smac_info->smac_port) ||
1933 		    smac_info->smac != smac) {
1934 			/* register candidate now, unreg if needed, after success */
1935 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1936 			if (smac_index >= 0) {
1937 				smac_info->candidate_smac_index = smac_index;
1938 				smac_info->candidate_smac = smac;
1939 				smac_info->candidate_smac_port = port;
1940 			} else {
1941 				return -EINVAL;
1942 			}
1943 		} else {
1944 			smac_index = smac_info->smac_index;
1945 		}
1946 		memcpy(path->dmac, ah->roce.dmac, 6);
1947 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1948 		/* put MAC table smac index for IBoE */
1949 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1950 	} else {
1951 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1952 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
1953 	}
1954 
1955 	return 0;
1956 }
1957 
1958 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1959 			 enum ib_qp_attr_mask qp_attr_mask,
1960 			 struct mlx4_ib_qp *mqp,
1961 			 struct mlx4_qp_path *path, u8 port,
1962 			 u16 vlan_id, u8 *smac)
1963 {
1964 	return _mlx4_set_path(dev, &qp->ah_attr,
1965 			      mlx4_mac_to_u64(smac),
1966 			      vlan_id,
1967 			      path, &mqp->pri, port);
1968 }
1969 
1970 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1971 			     const struct ib_qp_attr *qp,
1972 			     enum ib_qp_attr_mask qp_attr_mask,
1973 			     struct mlx4_ib_qp *mqp,
1974 			     struct mlx4_qp_path *path, u8 port)
1975 {
1976 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1977 			      0,
1978 			      0xffff,
1979 			      path, &mqp->alt, port);
1980 }
1981 
1982 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1983 {
1984 	struct mlx4_ib_gid_entry *ge, *tmp;
1985 
1986 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1987 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1988 			ge->added = 1;
1989 			ge->port = qp->port;
1990 		}
1991 	}
1992 }
1993 
1994 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1995 				    struct mlx4_ib_qp *qp,
1996 				    struct mlx4_qp_context *context)
1997 {
1998 	u64 u64_mac;
1999 	int smac_index;
2000 
2001 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
2002 
2003 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
2004 	if (!qp->pri.smac && !qp->pri.smac_port) {
2005 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
2006 		if (smac_index >= 0) {
2007 			qp->pri.candidate_smac_index = smac_index;
2008 			qp->pri.candidate_smac = u64_mac;
2009 			qp->pri.candidate_smac_port = qp->port;
2010 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
2011 		} else {
2012 			return -ENOENT;
2013 		}
2014 	}
2015 	return 0;
2016 }
2017 
2018 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
2019 {
2020 	struct counter_index *new_counter_index;
2021 	int err;
2022 	u32 tmp_idx;
2023 
2024 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
2025 	    IB_LINK_LAYER_ETHERNET ||
2026 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
2027 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
2028 		return 0;
2029 
2030 	err = mlx4_counter_alloc(dev->dev, &tmp_idx, MLX4_RES_USAGE_DRIVER);
2031 	if (err)
2032 		return err;
2033 
2034 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
2035 	if (!new_counter_index) {
2036 		mlx4_counter_free(dev->dev, tmp_idx);
2037 		return -ENOMEM;
2038 	}
2039 
2040 	new_counter_index->index = tmp_idx;
2041 	new_counter_index->allocated = 1;
2042 	qp->counter_index = new_counter_index;
2043 
2044 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
2045 	list_add_tail(&new_counter_index->list,
2046 		      &dev->counters_table[qp->port - 1].counters_list);
2047 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
2048 
2049 	return 0;
2050 }
2051 
2052 enum {
2053 	MLX4_QPC_ROCE_MODE_1 = 0,
2054 	MLX4_QPC_ROCE_MODE_2 = 2,
2055 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
2056 };
2057 
2058 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
2059 {
2060 	switch (gid_type) {
2061 	case IB_GID_TYPE_ROCE:
2062 		return MLX4_QPC_ROCE_MODE_1;
2063 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
2064 		return MLX4_QPC_ROCE_MODE_2;
2065 	default:
2066 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
2067 	}
2068 }
2069 
2070 /*
2071  * Go over all RSS QP's childes (WQs) and apply their HW state according to
2072  * their logic state if the RSS QP is the first RSS QP associated for the WQ.
2073  */
2074 static int bringup_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, u8 port_num)
2075 {
2076 	int err = 0;
2077 	int i;
2078 
2079 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2080 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2081 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2082 
2083 		mutex_lock(&wq->mutex);
2084 
2085 		/* Mlx4_ib restrictions:
2086 		 * WQ's is associated to a port according to the RSS QP it is
2087 		 * associates to.
2088 		 * In case the WQ is associated to a different port by another
2089 		 * RSS QP, return a failure.
2090 		 */
2091 		if ((wq->rss_usecnt > 0) && (wq->port != port_num)) {
2092 			err = -EINVAL;
2093 			mutex_unlock(&wq->mutex);
2094 			break;
2095 		}
2096 		wq->port = port_num;
2097 		if ((wq->rss_usecnt == 0) && (ibwq->state == IB_WQS_RDY)) {
2098 			err = _mlx4_ib_modify_wq(ibwq, IB_WQS_RDY);
2099 			if (err) {
2100 				mutex_unlock(&wq->mutex);
2101 				break;
2102 			}
2103 		}
2104 		wq->rss_usecnt++;
2105 
2106 		mutex_unlock(&wq->mutex);
2107 	}
2108 
2109 	if (i && err) {
2110 		int j;
2111 
2112 		for (j = (i - 1); j >= 0; j--) {
2113 			struct ib_wq *ibwq = ind_tbl->ind_tbl[j];
2114 			struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2115 
2116 			mutex_lock(&wq->mutex);
2117 
2118 			if ((wq->rss_usecnt == 1) &&
2119 			    (ibwq->state == IB_WQS_RDY))
2120 				if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET))
2121 					pr_warn("failed to reverse WQN=0x%06x\n",
2122 						ibwq->wq_num);
2123 			wq->rss_usecnt--;
2124 
2125 			mutex_unlock(&wq->mutex);
2126 		}
2127 	}
2128 
2129 	return err;
2130 }
2131 
2132 static void bring_down_rss_rwqs(struct ib_rwq_ind_table *ind_tbl)
2133 {
2134 	int i;
2135 
2136 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2137 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2138 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2139 
2140 		mutex_lock(&wq->mutex);
2141 
2142 		if ((wq->rss_usecnt == 1) && (ibwq->state == IB_WQS_RDY))
2143 			if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET))
2144 				pr_warn("failed to reverse WQN=%x\n",
2145 					ibwq->wq_num);
2146 		wq->rss_usecnt--;
2147 
2148 		mutex_unlock(&wq->mutex);
2149 	}
2150 }
2151 
2152 static void fill_qp_rss_context(struct mlx4_qp_context *context,
2153 				struct mlx4_ib_qp *qp)
2154 {
2155 	struct mlx4_rss_context *rss_context;
2156 
2157 	rss_context = (void *)context + offsetof(struct mlx4_qp_context,
2158 			pri_path) + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
2159 
2160 	rss_context->base_qpn = cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz);
2161 	rss_context->default_qpn =
2162 		cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz & 0xffffff);
2163 	if (qp->rss_ctx->flags & (MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6))
2164 		rss_context->base_qpn_udp = rss_context->default_qpn;
2165 	rss_context->flags = qp->rss_ctx->flags;
2166 	/* Currently support just toeplitz */
2167 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
2168 
2169 	memcpy(rss_context->rss_key, qp->rss_ctx->rss_key,
2170 	       MLX4_EN_RSS_KEY_SIZE);
2171 }
2172 
2173 static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type,
2174 			       const struct ib_qp_attr *attr, int attr_mask,
2175 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
2176 {
2177 	struct ib_uobject *ibuobject;
2178 	struct ib_srq  *ibsrq;
2179 	struct ib_rwq_ind_table *rwq_ind_tbl;
2180 	enum ib_qp_type qp_type;
2181 	struct mlx4_ib_dev *dev;
2182 	struct mlx4_ib_qp *qp;
2183 	struct mlx4_ib_pd *pd;
2184 	struct mlx4_ib_cq *send_cq, *recv_cq;
2185 	struct mlx4_qp_context *context;
2186 	enum mlx4_qp_optpar optpar = 0;
2187 	int sqd_event;
2188 	int steer_qp = 0;
2189 	int err = -EINVAL;
2190 	int counter_index;
2191 
2192 	if (src_type == MLX4_IB_RWQ_SRC) {
2193 		struct ib_wq *ibwq;
2194 
2195 		ibwq	    = (struct ib_wq *)src;
2196 		ibuobject   = ibwq->uobject;
2197 		ibsrq	    = NULL;
2198 		rwq_ind_tbl = NULL;
2199 		qp_type     = IB_QPT_RAW_PACKET;
2200 		qp	    = to_mqp((struct ib_qp *)ibwq);
2201 		dev	    = to_mdev(ibwq->device);
2202 		pd	    = to_mpd(ibwq->pd);
2203 	} else {
2204 		struct ib_qp *ibqp;
2205 
2206 		ibqp	    = (struct ib_qp *)src;
2207 		ibuobject   = ibqp->uobject;
2208 		ibsrq	    = ibqp->srq;
2209 		rwq_ind_tbl = ibqp->rwq_ind_tbl;
2210 		qp_type     = ibqp->qp_type;
2211 		qp	    = to_mqp(ibqp);
2212 		dev	    = to_mdev(ibqp->device);
2213 		pd	    = get_pd(qp);
2214 	}
2215 
2216 	/* APM is not supported under RoCE */
2217 	if (attr_mask & IB_QP_ALT_PATH &&
2218 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2219 	    IB_LINK_LAYER_ETHERNET)
2220 		return -ENOTSUPP;
2221 
2222 	context = kzalloc(sizeof *context, GFP_KERNEL);
2223 	if (!context)
2224 		return -ENOMEM;
2225 
2226 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
2227 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
2228 
2229 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
2230 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2231 	else {
2232 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
2233 		switch (attr->path_mig_state) {
2234 		case IB_MIG_MIGRATED:
2235 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2236 			break;
2237 		case IB_MIG_REARM:
2238 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
2239 			break;
2240 		case IB_MIG_ARMED:
2241 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
2242 			break;
2243 		}
2244 	}
2245 
2246 	if (qp->inl_recv_sz)
2247 		context->param3 |= cpu_to_be32(1 << 25);
2248 
2249 	if (qp->flags & MLX4_IB_QP_SCATTER_FCS)
2250 		context->param3 |= cpu_to_be32(1 << 29);
2251 
2252 	if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI)
2253 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
2254 	else if (qp_type == IB_QPT_RAW_PACKET)
2255 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
2256 	else if (qp_type == IB_QPT_UD) {
2257 		if (qp->flags & MLX4_IB_QP_LSO)
2258 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
2259 					      ilog2(dev->dev->caps.max_gso_sz);
2260 		else
2261 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 13;
2262 	} else if (attr_mask & IB_QP_PATH_MTU) {
2263 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
2264 			pr_err("path MTU (%u) is invalid\n",
2265 			       attr->path_mtu);
2266 			goto out;
2267 		}
2268 		context->mtu_msgmax = (attr->path_mtu << 5) |
2269 			ilog2(dev->dev->caps.max_msg_sz);
2270 	}
2271 
2272 	if (!rwq_ind_tbl) { /* PRM RSS receive side should be left zeros */
2273 		if (qp->rq.wqe_cnt)
2274 			context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
2275 		context->rq_size_stride |= qp->rq.wqe_shift - 4;
2276 	}
2277 
2278 	if (qp->sq.wqe_cnt)
2279 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
2280 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
2281 
2282 	if (new_state == IB_QPS_RESET && qp->counter_index)
2283 		mlx4_ib_free_qp_counter(dev, qp);
2284 
2285 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2286 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
2287 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
2288 		if (qp_type == IB_QPT_RAW_PACKET)
2289 			context->param3 |= cpu_to_be32(1 << 30);
2290 	}
2291 
2292 	if (ibuobject)
2293 		context->usr_page = cpu_to_be32(
2294 			mlx4_to_hw_uar_index(dev->dev,
2295 					     to_mucontext(ibuobject->context)
2296 					     ->uar.index));
2297 	else
2298 		context->usr_page = cpu_to_be32(
2299 			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
2300 
2301 	if (attr_mask & IB_QP_DEST_QPN)
2302 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
2303 
2304 	if (attr_mask & IB_QP_PORT) {
2305 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
2306 		    !(attr_mask & IB_QP_AV)) {
2307 			mlx4_set_sched(&context->pri_path, attr->port_num);
2308 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
2309 		}
2310 	}
2311 
2312 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
2313 		err = create_qp_lb_counter(dev, qp);
2314 		if (err)
2315 			goto out;
2316 
2317 		counter_index =
2318 			dev->counters_table[qp->port - 1].default_counter;
2319 		if (qp->counter_index)
2320 			counter_index = qp->counter_index->index;
2321 
2322 		if (counter_index != -1) {
2323 			context->pri_path.counter_index = counter_index;
2324 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
2325 			if (qp->counter_index) {
2326 				context->pri_path.fl |=
2327 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
2328 				context->pri_path.vlan_control |=
2329 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
2330 			}
2331 		} else
2332 			context->pri_path.counter_index =
2333 				MLX4_SINK_COUNTER_INDEX(dev->dev);
2334 
2335 		if (qp->flags & MLX4_IB_QP_NETIF) {
2336 			mlx4_ib_steer_qp_reg(dev, qp, 1);
2337 			steer_qp = 1;
2338 		}
2339 
2340 		if (qp_type == IB_QPT_GSI) {
2341 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
2342 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
2343 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
2344 
2345 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2346 		}
2347 	}
2348 
2349 	if (attr_mask & IB_QP_PKEY_INDEX) {
2350 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2351 			context->pri_path.disable_pkey_check = 0x40;
2352 		context->pri_path.pkey_index = attr->pkey_index;
2353 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
2354 	}
2355 
2356 	if (attr_mask & IB_QP_AV) {
2357 		u8 port_num = mlx4_is_bonded(dev->dev) ? 1 :
2358 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2359 		union ib_gid gid;
2360 		struct ib_gid_attr gid_attr = {.gid_type = IB_GID_TYPE_IB};
2361 		u16 vlan = 0xffff;
2362 		u8 smac[ETH_ALEN];
2363 		int status = 0;
2364 		int is_eth =
2365 			rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
2366 			rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
2367 
2368 		if (is_eth) {
2369 			int index =
2370 				rdma_ah_read_grh(&attr->ah_attr)->sgid_index;
2371 
2372 			status = ib_get_cached_gid(&dev->ib_dev, port_num,
2373 						   index, &gid, &gid_attr);
2374 			if (!status) {
2375 				vlan = rdma_vlan_dev_vlan_id(gid_attr.ndev);
2376 				memcpy(smac, gid_attr.ndev->dev_addr, ETH_ALEN);
2377 				dev_put(gid_attr.ndev);
2378 			}
2379 		}
2380 		if (status)
2381 			goto out;
2382 
2383 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
2384 				  port_num, vlan, smac))
2385 			goto out;
2386 
2387 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
2388 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
2389 
2390 		if (is_eth &&
2391 		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
2392 			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr.gid_type);
2393 
2394 			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
2395 				err = -EINVAL;
2396 				goto out;
2397 			}
2398 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2399 		}
2400 
2401 	}
2402 
2403 	if (attr_mask & IB_QP_TIMEOUT) {
2404 		context->pri_path.ackto |= attr->timeout << 3;
2405 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
2406 	}
2407 
2408 	if (attr_mask & IB_QP_ALT_PATH) {
2409 		if (attr->alt_port_num == 0 ||
2410 		    attr->alt_port_num > dev->dev->caps.num_ports)
2411 			goto out;
2412 
2413 		if (attr->alt_pkey_index >=
2414 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
2415 			goto out;
2416 
2417 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
2418 				      &context->alt_path,
2419 				      attr->alt_port_num))
2420 			goto out;
2421 
2422 		context->alt_path.pkey_index = attr->alt_pkey_index;
2423 		context->alt_path.ackto = attr->alt_timeout << 3;
2424 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
2425 	}
2426 
2427 	context->pd = cpu_to_be32(pd->pdn);
2428 
2429 	if (!rwq_ind_tbl) {
2430 		context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
2431 		get_cqs(qp, src_type, &send_cq, &recv_cq);
2432 	} else { /* Set dummy CQs to be compatible with HV and PRM */
2433 		send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq);
2434 		recv_cq = send_cq;
2435 	}
2436 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
2437 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
2438 
2439 	/* Set "fast registration enabled" for all kernel QPs */
2440 	if (!ibuobject)
2441 		context->params1 |= cpu_to_be32(1 << 11);
2442 
2443 	if (attr_mask & IB_QP_RNR_RETRY) {
2444 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2445 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
2446 	}
2447 
2448 	if (attr_mask & IB_QP_RETRY_CNT) {
2449 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2450 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
2451 	}
2452 
2453 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2454 		if (attr->max_rd_atomic)
2455 			context->params1 |=
2456 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2457 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
2458 	}
2459 
2460 	if (attr_mask & IB_QP_SQ_PSN)
2461 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2462 
2463 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2464 		if (attr->max_dest_rd_atomic)
2465 			context->params2 |=
2466 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2467 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
2468 	}
2469 
2470 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
2471 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
2472 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
2473 	}
2474 
2475 	if (ibsrq)
2476 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
2477 
2478 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
2479 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2480 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
2481 	}
2482 	if (attr_mask & IB_QP_RQ_PSN)
2483 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2484 
2485 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
2486 	if (attr_mask & IB_QP_QKEY) {
2487 		if (qp->mlx4_ib_qp_type &
2488 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
2489 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2490 		else {
2491 			if (mlx4_is_mfunc(dev->dev) &&
2492 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
2493 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
2494 			    MLX4_RESERVED_QKEY_BASE) {
2495 				pr_err("Cannot use reserved QKEY"
2496 				       " 0x%x (range 0xffff0000..0xffffffff"
2497 				       " is reserved)\n", attr->qkey);
2498 				err = -EINVAL;
2499 				goto out;
2500 			}
2501 			context->qkey = cpu_to_be32(attr->qkey);
2502 		}
2503 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
2504 	}
2505 
2506 	if (ibsrq)
2507 		context->srqn = cpu_to_be32(1 << 24 |
2508 					    to_msrq(ibsrq)->msrq.srqn);
2509 
2510 	if (qp->rq.wqe_cnt &&
2511 	    cur_state == IB_QPS_RESET &&
2512 	    new_state == IB_QPS_INIT)
2513 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2514 
2515 	if (cur_state == IB_QPS_INIT &&
2516 	    new_state == IB_QPS_RTR  &&
2517 	    (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI ||
2518 	     qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) {
2519 		context->pri_path.sched_queue = (qp->port - 1) << 6;
2520 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2521 		    qp->mlx4_ib_qp_type &
2522 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
2523 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
2524 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
2525 				context->pri_path.fl = 0x80;
2526 		} else {
2527 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2528 				context->pri_path.fl = 0x80;
2529 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
2530 		}
2531 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2532 		    IB_LINK_LAYER_ETHERNET) {
2533 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
2534 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
2535 				context->pri_path.feup = 1 << 7; /* don't fsm */
2536 			/* handle smac_index */
2537 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
2538 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
2539 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
2540 				err = handle_eth_ud_smac_index(dev, qp, context);
2541 				if (err) {
2542 					err = -EINVAL;
2543 					goto out;
2544 				}
2545 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
2546 					dev->qp1_proxy[qp->port - 1] = qp;
2547 			}
2548 		}
2549 	}
2550 
2551 	if (qp_type == IB_QPT_RAW_PACKET) {
2552 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
2553 					MLX4_IB_LINK_TYPE_ETH;
2554 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
2555 			/* set QP to receive both tunneled & non-tunneled packets */
2556 			if (!rwq_ind_tbl)
2557 				context->srqn = cpu_to_be32(7 << 28);
2558 		}
2559 	}
2560 
2561 	if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
2562 		int is_eth = rdma_port_get_link_layer(
2563 				&dev->ib_dev, qp->port) ==
2564 				IB_LINK_LAYER_ETHERNET;
2565 		if (is_eth) {
2566 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
2567 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
2568 		}
2569 	}
2570 
2571 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2572 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2573 		sqd_event = 1;
2574 	else
2575 		sqd_event = 0;
2576 
2577 	if (!ibuobject &&
2578 	    cur_state == IB_QPS_RESET &&
2579 	    new_state == IB_QPS_INIT)
2580 		context->rlkey_roce_mode |= (1 << 4);
2581 
2582 	/*
2583 	 * Before passing a kernel QP to the HW, make sure that the
2584 	 * ownership bits of the send queue are set and the SQ
2585 	 * headroom is stamped so that the hardware doesn't start
2586 	 * processing stale work requests.
2587 	 */
2588 	if (!ibuobject &&
2589 	    cur_state == IB_QPS_RESET &&
2590 	    new_state == IB_QPS_INIT) {
2591 		struct mlx4_wqe_ctrl_seg *ctrl;
2592 		int i;
2593 
2594 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
2595 			ctrl = get_send_wqe(qp, i);
2596 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
2597 			if (qp->sq_max_wqes_per_wr == 1)
2598 				ctrl->qpn_vlan.fence_size =
2599 						1 << (qp->sq.wqe_shift - 4);
2600 
2601 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
2602 		}
2603 	}
2604 
2605 	if (rwq_ind_tbl	&&
2606 	    cur_state == IB_QPS_RESET &&
2607 	    new_state == IB_QPS_INIT) {
2608 		fill_qp_rss_context(context, qp);
2609 		context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET);
2610 	}
2611 
2612 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2613 			     to_mlx4_state(new_state), context, optpar,
2614 			     sqd_event, &qp->mqp);
2615 	if (err)
2616 		goto out;
2617 
2618 	qp->state = new_state;
2619 
2620 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2621 		qp->atomic_rd_en = attr->qp_access_flags;
2622 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2623 		qp->resp_depth = attr->max_dest_rd_atomic;
2624 	if (attr_mask & IB_QP_PORT) {
2625 		qp->port = attr->port_num;
2626 		update_mcg_macs(dev, qp);
2627 	}
2628 	if (attr_mask & IB_QP_ALT_PATH)
2629 		qp->alt_port = attr->alt_port_num;
2630 
2631 	if (is_sqp(dev, qp))
2632 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2633 
2634 	/*
2635 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
2636 	 * QP0 to RESET or ERROR, bring the link back down.
2637 	 */
2638 	if (is_qp0(dev, qp)) {
2639 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2640 			if (mlx4_INIT_PORT(dev->dev, qp->port))
2641 				pr_warn("INIT_PORT failed for port %d\n",
2642 				       qp->port);
2643 
2644 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2645 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2646 			mlx4_CLOSE_PORT(dev->dev, qp->port);
2647 	}
2648 
2649 	/*
2650 	 * If we moved a kernel QP to RESET, clean up all old CQ
2651 	 * entries and reinitialize the QP.
2652 	 */
2653 	if (new_state == IB_QPS_RESET) {
2654 		if (!ibuobject) {
2655 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2656 					 ibsrq ? to_msrq(ibsrq) : NULL);
2657 			if (send_cq != recv_cq)
2658 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2659 
2660 			qp->rq.head = 0;
2661 			qp->rq.tail = 0;
2662 			qp->sq.head = 0;
2663 			qp->sq.tail = 0;
2664 			qp->sq_next_wqe = 0;
2665 			if (qp->rq.wqe_cnt)
2666 				*qp->db.db  = 0;
2667 
2668 			if (qp->flags & MLX4_IB_QP_NETIF)
2669 				mlx4_ib_steer_qp_reg(dev, qp, 0);
2670 		}
2671 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2672 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2673 			qp->pri.smac = 0;
2674 			qp->pri.smac_port = 0;
2675 		}
2676 		if (qp->alt.smac) {
2677 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2678 			qp->alt.smac = 0;
2679 		}
2680 		if (qp->pri.vid < 0x1000) {
2681 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2682 			qp->pri.vid = 0xFFFF;
2683 			qp->pri.candidate_vid = 0xFFFF;
2684 			qp->pri.update_vid = 0;
2685 		}
2686 
2687 		if (qp->alt.vid < 0x1000) {
2688 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2689 			qp->alt.vid = 0xFFFF;
2690 			qp->alt.candidate_vid = 0xFFFF;
2691 			qp->alt.update_vid = 0;
2692 		}
2693 	}
2694 out:
2695 	if (err && qp->counter_index)
2696 		mlx4_ib_free_qp_counter(dev, qp);
2697 	if (err && steer_qp)
2698 		mlx4_ib_steer_qp_reg(dev, qp, 0);
2699 	kfree(context);
2700 	if (qp->pri.candidate_smac ||
2701 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2702 		if (err) {
2703 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2704 		} else {
2705 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2706 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2707 			qp->pri.smac = qp->pri.candidate_smac;
2708 			qp->pri.smac_index = qp->pri.candidate_smac_index;
2709 			qp->pri.smac_port = qp->pri.candidate_smac_port;
2710 		}
2711 		qp->pri.candidate_smac = 0;
2712 		qp->pri.candidate_smac_index = 0;
2713 		qp->pri.candidate_smac_port = 0;
2714 	}
2715 	if (qp->alt.candidate_smac) {
2716 		if (err) {
2717 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2718 		} else {
2719 			if (qp->alt.smac)
2720 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2721 			qp->alt.smac = qp->alt.candidate_smac;
2722 			qp->alt.smac_index = qp->alt.candidate_smac_index;
2723 			qp->alt.smac_port = qp->alt.candidate_smac_port;
2724 		}
2725 		qp->alt.candidate_smac = 0;
2726 		qp->alt.candidate_smac_index = 0;
2727 		qp->alt.candidate_smac_port = 0;
2728 	}
2729 
2730 	if (qp->pri.update_vid) {
2731 		if (err) {
2732 			if (qp->pri.candidate_vid < 0x1000)
2733 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2734 						     qp->pri.candidate_vid);
2735 		} else {
2736 			if (qp->pri.vid < 0x1000)
2737 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2738 						     qp->pri.vid);
2739 			qp->pri.vid = qp->pri.candidate_vid;
2740 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2741 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
2742 		}
2743 		qp->pri.candidate_vid = 0xFFFF;
2744 		qp->pri.update_vid = 0;
2745 	}
2746 
2747 	if (qp->alt.update_vid) {
2748 		if (err) {
2749 			if (qp->alt.candidate_vid < 0x1000)
2750 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2751 						     qp->alt.candidate_vid);
2752 		} else {
2753 			if (qp->alt.vid < 0x1000)
2754 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2755 						     qp->alt.vid);
2756 			qp->alt.vid = qp->alt.candidate_vid;
2757 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2758 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2759 		}
2760 		qp->alt.candidate_vid = 0xFFFF;
2761 		qp->alt.update_vid = 0;
2762 	}
2763 
2764 	return err;
2765 }
2766 
2767 enum {
2768 	MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK = (IB_QP_STATE	|
2769 					      IB_QP_PORT),
2770 };
2771 
2772 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2773 			      int attr_mask, struct ib_udata *udata)
2774 {
2775 	enum rdma_link_layer ll = IB_LINK_LAYER_UNSPECIFIED;
2776 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2777 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2778 	enum ib_qp_state cur_state, new_state;
2779 	int err = -EINVAL;
2780 	mutex_lock(&qp->mutex);
2781 
2782 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2783 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2784 
2785 	if (cur_state != new_state || cur_state != IB_QPS_RESET) {
2786 		int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2787 		ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2788 	}
2789 
2790 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2791 				attr_mask, ll)) {
2792 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2793 			 "for transition %d to %d. qp_type %d,"
2794 			 " attr_mask 0x%x\n",
2795 			 ibqp->qp_num, cur_state, new_state,
2796 			 ibqp->qp_type, attr_mask);
2797 		goto out;
2798 	}
2799 
2800 	if (ibqp->rwq_ind_tbl) {
2801 		if (!(((cur_state == IB_QPS_RESET) &&
2802 		       (new_state == IB_QPS_INIT)) ||
2803 		      ((cur_state == IB_QPS_INIT)  &&
2804 		       (new_state == IB_QPS_RTR)))) {
2805 			pr_debug("qpn 0x%x: RSS QP unsupported transition %d to %d\n",
2806 				 ibqp->qp_num, cur_state, new_state);
2807 
2808 			err = -EOPNOTSUPP;
2809 			goto out;
2810 		}
2811 
2812 		if (attr_mask & ~MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK) {
2813 			pr_debug("qpn 0x%x: RSS QP unsupported attribute mask 0x%x for transition %d to %d\n",
2814 				 ibqp->qp_num, attr_mask, cur_state, new_state);
2815 
2816 			err = -EOPNOTSUPP;
2817 			goto out;
2818 		}
2819 	}
2820 
2821 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2822 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2823 			if ((ibqp->qp_type == IB_QPT_RC) ||
2824 			    (ibqp->qp_type == IB_QPT_UD) ||
2825 			    (ibqp->qp_type == IB_QPT_UC) ||
2826 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2827 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2828 				attr->port_num = mlx4_ib_bond_next_port(dev);
2829 			}
2830 		} else {
2831 			/* no sense in changing port_num
2832 			 * when ports are bonded */
2833 			attr_mask &= ~IB_QP_PORT;
2834 		}
2835 	}
2836 
2837 	if ((attr_mask & IB_QP_PORT) &&
2838 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2839 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2840 			 "for transition %d to %d. qp_type %d\n",
2841 			 ibqp->qp_num, attr->port_num, cur_state,
2842 			 new_state, ibqp->qp_type);
2843 		goto out;
2844 	}
2845 
2846 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2847 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2848 	     IB_LINK_LAYER_ETHERNET))
2849 		goto out;
2850 
2851 	if (attr_mask & IB_QP_PKEY_INDEX) {
2852 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2853 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2854 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2855 				 "for transition %d to %d. qp_type %d\n",
2856 				 ibqp->qp_num, attr->pkey_index, cur_state,
2857 				 new_state, ibqp->qp_type);
2858 			goto out;
2859 		}
2860 	}
2861 
2862 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2863 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2864 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2865 			 "Transition %d to %d. qp_type %d\n",
2866 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2867 			 new_state, ibqp->qp_type);
2868 		goto out;
2869 	}
2870 
2871 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2872 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2873 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2874 			 "Transition %d to %d. qp_type %d\n",
2875 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2876 			 new_state, ibqp->qp_type);
2877 		goto out;
2878 	}
2879 
2880 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2881 		err = 0;
2882 		goto out;
2883 	}
2884 
2885 	if (ibqp->rwq_ind_tbl && (new_state == IB_QPS_INIT)) {
2886 		err = bringup_rss_rwqs(ibqp->rwq_ind_tbl, attr->port_num);
2887 		if (err)
2888 			goto out;
2889 	}
2890 
2891 	err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask,
2892 				  cur_state, new_state);
2893 
2894 	if (ibqp->rwq_ind_tbl && err)
2895 		bring_down_rss_rwqs(ibqp->rwq_ind_tbl);
2896 
2897 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2898 		attr->port_num = 1;
2899 
2900 out:
2901 	mutex_unlock(&qp->mutex);
2902 	return err;
2903 }
2904 
2905 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2906 		      int attr_mask, struct ib_udata *udata)
2907 {
2908 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2909 	int ret;
2910 
2911 	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2912 
2913 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2914 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
2915 		int err = 0;
2916 
2917 		if (sqp->roce_v2_gsi)
2918 			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2919 		if (err)
2920 			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2921 			       err);
2922 	}
2923 	return ret;
2924 }
2925 
2926 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2927 {
2928 	int i;
2929 	for (i = 0; i < dev->caps.num_ports; i++) {
2930 		if (qpn == dev->caps.spec_qps[i].qp0_proxy ||
2931 		    qpn == dev->caps.spec_qps[i].qp0_tunnel) {
2932 			*qkey = dev->caps.spec_qps[i].qp0_qkey;
2933 			return 0;
2934 		}
2935 	}
2936 	return -EINVAL;
2937 }
2938 
2939 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2940 				  struct ib_ud_wr *wr,
2941 				  void *wqe, unsigned *mlx_seg_len)
2942 {
2943 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2944 	struct ib_device *ib_dev = &mdev->ib_dev;
2945 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2946 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2947 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2948 	u16 pkey;
2949 	u32 qkey;
2950 	int send_size;
2951 	int header_size;
2952 	int spc;
2953 	int i;
2954 
2955 	if (wr->wr.opcode != IB_WR_SEND)
2956 		return -EINVAL;
2957 
2958 	send_size = 0;
2959 
2960 	for (i = 0; i < wr->wr.num_sge; ++i)
2961 		send_size += wr->wr.sg_list[i].length;
2962 
2963 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2964 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2965 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2966 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2967 
2968 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2969 
2970 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2971 		sqp->ud_header.lrh.service_level =
2972 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2973 		sqp->ud_header.lrh.destination_lid =
2974 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2975 		sqp->ud_header.lrh.source_lid =
2976 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2977 	}
2978 
2979 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2980 
2981 	/* force loopback */
2982 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2983 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2984 
2985 	sqp->ud_header.lrh.virtual_lane    = 0;
2986 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2987 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2988 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2989 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2990 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2991 	else
2992 		sqp->ud_header.bth.destination_qpn =
2993 			cpu_to_be32(mdev->dev->caps.spec_qps[sqp->qp.port - 1].qp0_tunnel);
2994 
2995 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2996 	if (mlx4_is_master(mdev->dev)) {
2997 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2998 			return -EINVAL;
2999 	} else {
3000 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
3001 			return -EINVAL;
3002 	}
3003 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
3004 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
3005 
3006 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
3007 	sqp->ud_header.immediate_present = 0;
3008 
3009 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3010 
3011 	/*
3012 	 * Inline data segments may not cross a 64 byte boundary.  If
3013 	 * our UD header is bigger than the space available up to the
3014 	 * next 64 byte boundary in the WQE, use two inline data
3015 	 * segments to hold the UD header.
3016 	 */
3017 	spc = MLX4_INLINE_ALIGN -
3018 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3019 	if (header_size <= spc) {
3020 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3021 		memcpy(inl + 1, sqp->header_buf, header_size);
3022 		i = 1;
3023 	} else {
3024 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3025 		memcpy(inl + 1, sqp->header_buf, spc);
3026 
3027 		inl = (void *) (inl + 1) + spc;
3028 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3029 		/*
3030 		 * Need a barrier here to make sure all the data is
3031 		 * visible before the byte_count field is set.
3032 		 * Otherwise the HCA prefetcher could grab the 64-byte
3033 		 * chunk with this inline segment and get a valid (!=
3034 		 * 0xffffffff) byte count but stale data, and end up
3035 		 * generating a packet with bad headers.
3036 		 *
3037 		 * The first inline segment's byte_count field doesn't
3038 		 * need a barrier, because it comes after a
3039 		 * control/MLX segment and therefore is at an offset
3040 		 * of 16 mod 64.
3041 		 */
3042 		wmb();
3043 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3044 		i = 2;
3045 	}
3046 
3047 	*mlx_seg_len =
3048 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3049 	return 0;
3050 }
3051 
3052 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
3053 {
3054 	union sl2vl_tbl_to_u64 tmp_vltab;
3055 	u8 vl;
3056 
3057 	if (sl > 15)
3058 		return 0xf;
3059 	tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
3060 	vl = tmp_vltab.sl8[sl >> 1];
3061 	if (sl & 1)
3062 		vl &= 0x0f;
3063 	else
3064 		vl >>= 4;
3065 	return vl;
3066 }
3067 
3068 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
3069 				int index, union ib_gid *gid,
3070 				enum ib_gid_type *gid_type)
3071 {
3072 	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
3073 	struct mlx4_port_gid_table *port_gid_table;
3074 	unsigned long flags;
3075 
3076 	port_gid_table = &iboe->gids[port_num - 1];
3077 	spin_lock_irqsave(&iboe->lock, flags);
3078 	memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
3079 	*gid_type = port_gid_table->gids[index].gid_type;
3080 	spin_unlock_irqrestore(&iboe->lock, flags);
3081 	if (rdma_is_zero_gid(gid))
3082 		return -ENOENT;
3083 
3084 	return 0;
3085 }
3086 
3087 #define MLX4_ROCEV2_QP1_SPORT 0xC000
3088 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_ud_wr *wr,
3089 			    void *wqe, unsigned *mlx_seg_len)
3090 {
3091 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
3092 	struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
3093 	struct mlx4_wqe_mlx_seg *mlx = wqe;
3094 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
3095 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
3096 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3097 	union ib_gid sgid;
3098 	u16 pkey;
3099 	int send_size;
3100 	int header_size;
3101 	int spc;
3102 	int i;
3103 	int err = 0;
3104 	u16 vlan = 0xffff;
3105 	bool is_eth;
3106 	bool is_vlan = false;
3107 	bool is_grh;
3108 	bool is_udp = false;
3109 	int ip_version = 0;
3110 
3111 	send_size = 0;
3112 	for (i = 0; i < wr->wr.num_sge; ++i)
3113 		send_size += wr->wr.sg_list[i].length;
3114 
3115 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
3116 	is_grh = mlx4_ib_ah_grh_present(ah);
3117 	if (is_eth) {
3118 		enum ib_gid_type gid_type;
3119 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3120 			/* When multi-function is enabled, the ib_core gid
3121 			 * indexes don't necessarily match the hw ones, so
3122 			 * we must use our own cache */
3123 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
3124 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
3125 							   ah->av.ib.gid_index, &sgid.raw[0]);
3126 			if (err)
3127 				return err;
3128 		} else  {
3129 			err = fill_gid_by_hw_index(ibdev, sqp->qp.port,
3130 					    ah->av.ib.gid_index,
3131 					    &sgid, &gid_type);
3132 			if (!err) {
3133 				is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
3134 				if (is_udp) {
3135 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
3136 						ip_version = 4;
3137 					else
3138 						ip_version = 6;
3139 					is_grh = false;
3140 				}
3141 			} else {
3142 				return err;
3143 			}
3144 		}
3145 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
3146 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
3147 			is_vlan = 1;
3148 		}
3149 	}
3150 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
3151 			  ip_version, is_udp, 0, &sqp->ud_header);
3152 	if (err)
3153 		return err;
3154 
3155 	if (!is_eth) {
3156 		sqp->ud_header.lrh.service_level =
3157 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
3158 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
3159 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
3160 	}
3161 
3162 	if (is_grh || (ip_version == 6)) {
3163 		sqp->ud_header.grh.traffic_class =
3164 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3165 		sqp->ud_header.grh.flow_label    =
3166 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
3167 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
3168 		if (is_eth) {
3169 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
3170 		} else {
3171 			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3172 				/* When multi-function is enabled, the ib_core gid
3173 				 * indexes don't necessarily match the hw ones, so
3174 				 * we must use our own cache
3175 				 */
3176 				sqp->ud_header.grh.source_gid.global.subnet_prefix =
3177 					cpu_to_be64(atomic64_read(&(to_mdev(ib_dev)->sriov.
3178 								    demux[sqp->qp.port - 1].
3179 								    subnet_prefix)));
3180 				sqp->ud_header.grh.source_gid.global.interface_id =
3181 					to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
3182 						       guid_cache[ah->av.ib.gid_index];
3183 			} else {
3184 				ib_get_cached_gid(ib_dev,
3185 						  be32_to_cpu(ah->av.ib.port_pd) >> 24,
3186 						  ah->av.ib.gid_index,
3187 						  &sqp->ud_header.grh.source_gid, NULL);
3188 			}
3189 		}
3190 		memcpy(sqp->ud_header.grh.destination_gid.raw,
3191 		       ah->av.ib.dgid, 16);
3192 	}
3193 
3194 	if (ip_version == 4) {
3195 		sqp->ud_header.ip4.tos =
3196 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3197 		sqp->ud_header.ip4.id = 0;
3198 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
3199 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
3200 
3201 		memcpy(&sqp->ud_header.ip4.saddr,
3202 		       sgid.raw + 12, 4);
3203 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
3204 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
3205 	}
3206 
3207 	if (is_udp) {
3208 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
3209 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
3210 		sqp->ud_header.udp.csum = 0;
3211 	}
3212 
3213 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
3214 
3215 	if (!is_eth) {
3216 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
3217 					  (sqp->ud_header.lrh.destination_lid ==
3218 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
3219 					  (sqp->ud_header.lrh.service_level << 8));
3220 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
3221 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
3222 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
3223 	}
3224 
3225 	switch (wr->wr.opcode) {
3226 	case IB_WR_SEND:
3227 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
3228 		sqp->ud_header.immediate_present = 0;
3229 		break;
3230 	case IB_WR_SEND_WITH_IMM:
3231 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
3232 		sqp->ud_header.immediate_present = 1;
3233 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
3234 		break;
3235 	default:
3236 		return -EINVAL;
3237 	}
3238 
3239 	if (is_eth) {
3240 		struct in6_addr in6;
3241 		u16 ether_type;
3242 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
3243 
3244 		ether_type = (!is_udp) ? ETH_P_IBOE:
3245 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
3246 
3247 		mlx->sched_prio = cpu_to_be16(pcp);
3248 
3249 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
3250 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
3251 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
3252 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
3253 		memcpy(&in6, sgid.raw, sizeof(in6));
3254 
3255 
3256 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
3257 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
3258 		if (!is_vlan) {
3259 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
3260 		} else {
3261 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
3262 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
3263 		}
3264 	} else {
3265 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 :
3266 							sl_to_vl(to_mdev(ib_dev),
3267 								 sqp->ud_header.lrh.service_level,
3268 								 sqp->qp.port);
3269 		if (sqp->qp.ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
3270 			return -EINVAL;
3271 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
3272 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
3273 	}
3274 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
3275 	if (!sqp->qp.ibqp.qp_num)
3276 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
3277 	else
3278 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
3279 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
3280 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
3281 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
3282 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
3283 					       sqp->qkey : wr->remote_qkey);
3284 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
3285 
3286 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3287 
3288 	if (0) {
3289 		pr_err("built UD header of size %d:\n", header_size);
3290 		for (i = 0; i < header_size / 4; ++i) {
3291 			if (i % 8 == 0)
3292 				pr_err("  [%02x] ", i * 4);
3293 			pr_cont(" %08x",
3294 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
3295 			if ((i + 1) % 8 == 0)
3296 				pr_cont("\n");
3297 		}
3298 		pr_err("\n");
3299 	}
3300 
3301 	/*
3302 	 * Inline data segments may not cross a 64 byte boundary.  If
3303 	 * our UD header is bigger than the space available up to the
3304 	 * next 64 byte boundary in the WQE, use two inline data
3305 	 * segments to hold the UD header.
3306 	 */
3307 	spc = MLX4_INLINE_ALIGN -
3308 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3309 	if (header_size <= spc) {
3310 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3311 		memcpy(inl + 1, sqp->header_buf, header_size);
3312 		i = 1;
3313 	} else {
3314 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3315 		memcpy(inl + 1, sqp->header_buf, spc);
3316 
3317 		inl = (void *) (inl + 1) + spc;
3318 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3319 		/*
3320 		 * Need a barrier here to make sure all the data is
3321 		 * visible before the byte_count field is set.
3322 		 * Otherwise the HCA prefetcher could grab the 64-byte
3323 		 * chunk with this inline segment and get a valid (!=
3324 		 * 0xffffffff) byte count but stale data, and end up
3325 		 * generating a packet with bad headers.
3326 		 *
3327 		 * The first inline segment's byte_count field doesn't
3328 		 * need a barrier, because it comes after a
3329 		 * control/MLX segment and therefore is at an offset
3330 		 * of 16 mod 64.
3331 		 */
3332 		wmb();
3333 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3334 		i = 2;
3335 	}
3336 
3337 	*mlx_seg_len =
3338 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3339 	return 0;
3340 }
3341 
3342 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3343 {
3344 	unsigned cur;
3345 	struct mlx4_ib_cq *cq;
3346 
3347 	cur = wq->head - wq->tail;
3348 	if (likely(cur + nreq < wq->max_post))
3349 		return 0;
3350 
3351 	cq = to_mcq(ib_cq);
3352 	spin_lock(&cq->lock);
3353 	cur = wq->head - wq->tail;
3354 	spin_unlock(&cq->lock);
3355 
3356 	return cur + nreq >= wq->max_post;
3357 }
3358 
3359 static __be32 convert_access(int acc)
3360 {
3361 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
3362 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
3363 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
3364 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
3365 	       (acc & IB_ACCESS_REMOTE_READ   ?
3366 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
3367 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
3368 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
3369 }
3370 
3371 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
3372 			struct ib_reg_wr *wr)
3373 {
3374 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
3375 
3376 	fseg->flags		= convert_access(wr->access);
3377 	fseg->mem_key		= cpu_to_be32(wr->key);
3378 	fseg->buf_list		= cpu_to_be64(mr->page_map);
3379 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
3380 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
3381 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
3382 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
3383 	fseg->reserved[0]	= 0;
3384 	fseg->reserved[1]	= 0;
3385 }
3386 
3387 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
3388 {
3389 	memset(iseg, 0, sizeof(*iseg));
3390 	iseg->mem_key = cpu_to_be32(rkey);
3391 }
3392 
3393 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
3394 					  u64 remote_addr, u32 rkey)
3395 {
3396 	rseg->raddr    = cpu_to_be64(remote_addr);
3397 	rseg->rkey     = cpu_to_be32(rkey);
3398 	rseg->reserved = 0;
3399 }
3400 
3401 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
3402 		struct ib_atomic_wr *wr)
3403 {
3404 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
3405 		aseg->swap_add = cpu_to_be64(wr->swap);
3406 		aseg->compare  = cpu_to_be64(wr->compare_add);
3407 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
3408 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3409 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
3410 	} else {
3411 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3412 		aseg->compare  = 0;
3413 	}
3414 
3415 }
3416 
3417 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
3418 				  struct ib_atomic_wr *wr)
3419 {
3420 	aseg->swap_add		= cpu_to_be64(wr->swap);
3421 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
3422 	aseg->compare		= cpu_to_be64(wr->compare_add);
3423 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
3424 }
3425 
3426 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
3427 			     struct ib_ud_wr *wr)
3428 {
3429 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
3430 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
3431 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
3432 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
3433 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
3434 }
3435 
3436 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
3437 				    struct mlx4_wqe_datagram_seg *dseg,
3438 				    struct ib_ud_wr *wr,
3439 				    enum mlx4_ib_qp_type qpt)
3440 {
3441 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
3442 	struct mlx4_av sqp_av = {0};
3443 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
3444 
3445 	/* force loopback */
3446 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
3447 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
3448 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
3449 			cpu_to_be32(0xf0000000);
3450 
3451 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
3452 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
3453 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp1_tunnel);
3454 	else
3455 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp0_tunnel);
3456 	/* Use QKEY from the QP context, which is set by master */
3457 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
3458 }
3459 
3460 static void build_tunnel_header(struct ib_ud_wr *wr, void *wqe, unsigned *mlx_seg_len)
3461 {
3462 	struct mlx4_wqe_inline_seg *inl = wqe;
3463 	struct mlx4_ib_tunnel_header hdr;
3464 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3465 	int spc;
3466 	int i;
3467 
3468 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
3469 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
3470 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
3471 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
3472 	memcpy(hdr.mac, ah->av.eth.mac, 6);
3473 	hdr.vlan = ah->av.eth.vlan;
3474 
3475 	spc = MLX4_INLINE_ALIGN -
3476 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3477 	if (sizeof (hdr) <= spc) {
3478 		memcpy(inl + 1, &hdr, sizeof (hdr));
3479 		wmb();
3480 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
3481 		i = 1;
3482 	} else {
3483 		memcpy(inl + 1, &hdr, spc);
3484 		wmb();
3485 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3486 
3487 		inl = (void *) (inl + 1) + spc;
3488 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
3489 		wmb();
3490 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
3491 		i = 2;
3492 	}
3493 
3494 	*mlx_seg_len =
3495 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
3496 }
3497 
3498 static void set_mlx_icrc_seg(void *dseg)
3499 {
3500 	u32 *t = dseg;
3501 	struct mlx4_wqe_inline_seg *iseg = dseg;
3502 
3503 	t[1] = 0;
3504 
3505 	/*
3506 	 * Need a barrier here before writing the byte_count field to
3507 	 * make sure that all the data is visible before the
3508 	 * byte_count field is set.  Otherwise, if the segment begins
3509 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3510 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3511 	 * stale data, and end up sending the wrong data.
3512 	 */
3513 	wmb();
3514 
3515 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
3516 }
3517 
3518 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3519 {
3520 	dseg->lkey       = cpu_to_be32(sg->lkey);
3521 	dseg->addr       = cpu_to_be64(sg->addr);
3522 
3523 	/*
3524 	 * Need a barrier here before writing the byte_count field to
3525 	 * make sure that all the data is visible before the
3526 	 * byte_count field is set.  Otherwise, if the segment begins
3527 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3528 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3529 	 * stale data, and end up sending the wrong data.
3530 	 */
3531 	wmb();
3532 
3533 	dseg->byte_count = cpu_to_be32(sg->length);
3534 }
3535 
3536 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3537 {
3538 	dseg->byte_count = cpu_to_be32(sg->length);
3539 	dseg->lkey       = cpu_to_be32(sg->lkey);
3540 	dseg->addr       = cpu_to_be64(sg->addr);
3541 }
3542 
3543 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_ud_wr *wr,
3544 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
3545 			 __be32 *lso_hdr_sz, __be32 *blh)
3546 {
3547 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
3548 
3549 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
3550 		*blh = cpu_to_be32(1 << 6);
3551 
3552 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
3553 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
3554 		return -EINVAL;
3555 
3556 	memcpy(wqe->header, wr->header, wr->hlen);
3557 
3558 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
3559 	*lso_seg_len = halign;
3560 	return 0;
3561 }
3562 
3563 static __be32 send_ieth(struct ib_send_wr *wr)
3564 {
3565 	switch (wr->opcode) {
3566 	case IB_WR_SEND_WITH_IMM:
3567 	case IB_WR_RDMA_WRITE_WITH_IMM:
3568 		return wr->ex.imm_data;
3569 
3570 	case IB_WR_SEND_WITH_INV:
3571 		return cpu_to_be32(wr->ex.invalidate_rkey);
3572 
3573 	default:
3574 		return 0;
3575 	}
3576 }
3577 
3578 static void add_zero_len_inline(void *wqe)
3579 {
3580 	struct mlx4_wqe_inline_seg *inl = wqe;
3581 	memset(wqe, 0, 16);
3582 	inl->byte_count = cpu_to_be32(1 << 31);
3583 }
3584 
3585 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
3586 		      struct ib_send_wr **bad_wr)
3587 {
3588 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3589 	void *wqe;
3590 	struct mlx4_wqe_ctrl_seg *ctrl;
3591 	struct mlx4_wqe_data_seg *dseg;
3592 	unsigned long flags;
3593 	int nreq;
3594 	int err = 0;
3595 	unsigned ind;
3596 	int uninitialized_var(stamp);
3597 	int uninitialized_var(size);
3598 	unsigned uninitialized_var(seglen);
3599 	__be32 dummy;
3600 	__be32 *lso_wqe;
3601 	__be32 uninitialized_var(lso_hdr_sz);
3602 	__be32 blh;
3603 	int i;
3604 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3605 
3606 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
3607 		struct mlx4_ib_sqp *sqp = to_msqp(qp);
3608 
3609 		if (sqp->roce_v2_gsi) {
3610 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
3611 			enum ib_gid_type gid_type;
3612 			union ib_gid gid;
3613 
3614 			if (!fill_gid_by_hw_index(mdev, sqp->qp.port,
3615 					   ah->av.ib.gid_index,
3616 					   &gid, &gid_type))
3617 				qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
3618 						to_mqp(sqp->roce_v2_gsi) : qp;
3619 			else
3620 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
3621 				       ah->av.ib.gid_index);
3622 		}
3623 	}
3624 
3625 	spin_lock_irqsave(&qp->sq.lock, flags);
3626 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3627 		err = -EIO;
3628 		*bad_wr = wr;
3629 		nreq = 0;
3630 		goto out;
3631 	}
3632 
3633 	ind = qp->sq_next_wqe;
3634 
3635 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3636 		lso_wqe = &dummy;
3637 		blh = 0;
3638 
3639 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
3640 			err = -ENOMEM;
3641 			*bad_wr = wr;
3642 			goto out;
3643 		}
3644 
3645 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3646 			err = -EINVAL;
3647 			*bad_wr = wr;
3648 			goto out;
3649 		}
3650 
3651 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3652 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3653 
3654 		ctrl->srcrb_flags =
3655 			(wr->send_flags & IB_SEND_SIGNALED ?
3656 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3657 			(wr->send_flags & IB_SEND_SOLICITED ?
3658 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3659 			((wr->send_flags & IB_SEND_IP_CSUM) ?
3660 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3661 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3662 			qp->sq_signal_bits;
3663 
3664 		ctrl->imm = send_ieth(wr);
3665 
3666 		wqe += sizeof *ctrl;
3667 		size = sizeof *ctrl / 16;
3668 
3669 		switch (qp->mlx4_ib_qp_type) {
3670 		case MLX4_IB_QPT_RC:
3671 		case MLX4_IB_QPT_UC:
3672 			switch (wr->opcode) {
3673 			case IB_WR_ATOMIC_CMP_AND_SWP:
3674 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3675 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3676 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3677 					      atomic_wr(wr)->rkey);
3678 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3679 
3680 				set_atomic_seg(wqe, atomic_wr(wr));
3681 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
3682 
3683 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3684 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3685 
3686 				break;
3687 
3688 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3689 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3690 					      atomic_wr(wr)->rkey);
3691 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3692 
3693 				set_masked_atomic_seg(wqe, atomic_wr(wr));
3694 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
3695 
3696 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3697 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3698 
3699 				break;
3700 
3701 			case IB_WR_RDMA_READ:
3702 			case IB_WR_RDMA_WRITE:
3703 			case IB_WR_RDMA_WRITE_WITH_IMM:
3704 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3705 					      rdma_wr(wr)->rkey);
3706 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3707 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3708 				break;
3709 
3710 			case IB_WR_LOCAL_INV:
3711 				ctrl->srcrb_flags |=
3712 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3713 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3714 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3715 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3716 				break;
3717 
3718 			case IB_WR_REG_MR:
3719 				ctrl->srcrb_flags |=
3720 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3721 				set_reg_seg(wqe, reg_wr(wr));
3722 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3723 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3724 				break;
3725 
3726 			default:
3727 				/* No extra segments required for sends */
3728 				break;
3729 			}
3730 			break;
3731 
3732 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3733 			err =  build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3734 					ctrl, &seglen);
3735 			if (unlikely(err)) {
3736 				*bad_wr = wr;
3737 				goto out;
3738 			}
3739 			wqe  += seglen;
3740 			size += seglen / 16;
3741 			break;
3742 		case MLX4_IB_QPT_TUN_SMI:
3743 		case MLX4_IB_QPT_TUN_GSI:
3744 			/* this is a UD qp used in MAD responses to slaves. */
3745 			set_datagram_seg(wqe, ud_wr(wr));
3746 			/* set the forced-loopback bit in the data seg av */
3747 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3748 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3749 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3750 			break;
3751 		case MLX4_IB_QPT_UD:
3752 			set_datagram_seg(wqe, ud_wr(wr));
3753 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3754 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3755 
3756 			if (wr->opcode == IB_WR_LSO) {
3757 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3758 						&lso_hdr_sz, &blh);
3759 				if (unlikely(err)) {
3760 					*bad_wr = wr;
3761 					goto out;
3762 				}
3763 				lso_wqe = (__be32 *) wqe;
3764 				wqe  += seglen;
3765 				size += seglen / 16;
3766 			}
3767 			break;
3768 
3769 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3770 			err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3771 					ctrl, &seglen);
3772 			if (unlikely(err)) {
3773 				*bad_wr = wr;
3774 				goto out;
3775 			}
3776 			wqe  += seglen;
3777 			size += seglen / 16;
3778 			/* to start tunnel header on a cache-line boundary */
3779 			add_zero_len_inline(wqe);
3780 			wqe += 16;
3781 			size++;
3782 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3783 			wqe  += seglen;
3784 			size += seglen / 16;
3785 			break;
3786 		case MLX4_IB_QPT_PROXY_SMI:
3787 		case MLX4_IB_QPT_PROXY_GSI:
3788 			/* If we are tunneling special qps, this is a UD qp.
3789 			 * In this case we first add a UD segment targeting
3790 			 * the tunnel qp, and then add a header with address
3791 			 * information */
3792 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3793 						ud_wr(wr),
3794 						qp->mlx4_ib_qp_type);
3795 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3796 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3797 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3798 			wqe  += seglen;
3799 			size += seglen / 16;
3800 			break;
3801 
3802 		case MLX4_IB_QPT_SMI:
3803 		case MLX4_IB_QPT_GSI:
3804 			err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
3805 					&seglen);
3806 			if (unlikely(err)) {
3807 				*bad_wr = wr;
3808 				goto out;
3809 			}
3810 			wqe  += seglen;
3811 			size += seglen / 16;
3812 			break;
3813 
3814 		default:
3815 			break;
3816 		}
3817 
3818 		/*
3819 		 * Write data segments in reverse order, so as to
3820 		 * overwrite cacheline stamp last within each
3821 		 * cacheline.  This avoids issues with WQE
3822 		 * prefetching.
3823 		 */
3824 
3825 		dseg = wqe;
3826 		dseg += wr->num_sge - 1;
3827 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3828 
3829 		/* Add one more inline data segment for ICRC for MLX sends */
3830 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3831 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3832 			     qp->mlx4_ib_qp_type &
3833 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3834 			set_mlx_icrc_seg(dseg + 1);
3835 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3836 		}
3837 
3838 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3839 			set_data_seg(dseg, wr->sg_list + i);
3840 
3841 		/*
3842 		 * Possibly overwrite stamping in cacheline with LSO
3843 		 * segment only after making sure all data segments
3844 		 * are written.
3845 		 */
3846 		wmb();
3847 		*lso_wqe = lso_hdr_sz;
3848 
3849 		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3850 					     MLX4_WQE_CTRL_FENCE : 0) | size;
3851 
3852 		/*
3853 		 * Make sure descriptor is fully written before
3854 		 * setting ownership bit (because HW can start
3855 		 * executing as soon as we do).
3856 		 */
3857 		wmb();
3858 
3859 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3860 			*bad_wr = wr;
3861 			err = -EINVAL;
3862 			goto out;
3863 		}
3864 
3865 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3866 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3867 
3868 		stamp = ind + qp->sq_spare_wqes;
3869 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
3870 
3871 		/*
3872 		 * We can improve latency by not stamping the last
3873 		 * send queue WQE until after ringing the doorbell, so
3874 		 * only stamp here if there are still more WQEs to post.
3875 		 *
3876 		 * Same optimization applies to padding with NOP wqe
3877 		 * in case of WQE shrinking (used to prevent wrap-around
3878 		 * in the middle of WR).
3879 		 */
3880 		if (wr->next) {
3881 			stamp_send_wqe(qp, stamp, size * 16);
3882 			ind = pad_wraparound(qp, ind);
3883 		}
3884 	}
3885 
3886 out:
3887 	if (likely(nreq)) {
3888 		qp->sq.head += nreq;
3889 
3890 		/*
3891 		 * Make sure that descriptors are written before
3892 		 * doorbell record.
3893 		 */
3894 		wmb();
3895 
3896 		writel_relaxed(qp->doorbell_qpn,
3897 			to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3898 
3899 		/*
3900 		 * Make sure doorbells don't leak out of SQ spinlock
3901 		 * and reach the HCA out of order.
3902 		 */
3903 		mmiowb();
3904 
3905 		stamp_send_wqe(qp, stamp, size * 16);
3906 
3907 		ind = pad_wraparound(qp, ind);
3908 		qp->sq_next_wqe = ind;
3909 	}
3910 
3911 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3912 
3913 	return err;
3914 }
3915 
3916 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3917 		      struct ib_recv_wr **bad_wr)
3918 {
3919 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3920 	struct mlx4_wqe_data_seg *scat;
3921 	unsigned long flags;
3922 	int err = 0;
3923 	int nreq;
3924 	int ind;
3925 	int max_gs;
3926 	int i;
3927 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3928 
3929 	max_gs = qp->rq.max_gs;
3930 	spin_lock_irqsave(&qp->rq.lock, flags);
3931 
3932 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3933 		err = -EIO;
3934 		*bad_wr = wr;
3935 		nreq = 0;
3936 		goto out;
3937 	}
3938 
3939 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3940 
3941 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3942 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3943 			err = -ENOMEM;
3944 			*bad_wr = wr;
3945 			goto out;
3946 		}
3947 
3948 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3949 			err = -EINVAL;
3950 			*bad_wr = wr;
3951 			goto out;
3952 		}
3953 
3954 		scat = get_recv_wqe(qp, ind);
3955 
3956 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3957 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3958 			ib_dma_sync_single_for_device(ibqp->device,
3959 						      qp->sqp_proxy_rcv[ind].map,
3960 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3961 						      DMA_FROM_DEVICE);
3962 			scat->byte_count =
3963 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3964 			/* use dma lkey from upper layer entry */
3965 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3966 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3967 			scat++;
3968 			max_gs--;
3969 		}
3970 
3971 		for (i = 0; i < wr->num_sge; ++i)
3972 			__set_data_seg(scat + i, wr->sg_list + i);
3973 
3974 		if (i < max_gs) {
3975 			scat[i].byte_count = 0;
3976 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3977 			scat[i].addr       = 0;
3978 		}
3979 
3980 		qp->rq.wrid[ind] = wr->wr_id;
3981 
3982 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3983 	}
3984 
3985 out:
3986 	if (likely(nreq)) {
3987 		qp->rq.head += nreq;
3988 
3989 		/*
3990 		 * Make sure that descriptors are written before
3991 		 * doorbell record.
3992 		 */
3993 		wmb();
3994 
3995 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3996 	}
3997 
3998 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3999 
4000 	return err;
4001 }
4002 
4003 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
4004 {
4005 	switch (mlx4_state) {
4006 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
4007 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
4008 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
4009 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
4010 	case MLX4_QP_STATE_SQ_DRAINING:
4011 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
4012 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
4013 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
4014 	default:		     return -1;
4015 	}
4016 }
4017 
4018 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
4019 {
4020 	switch (mlx4_mig_state) {
4021 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
4022 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
4023 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
4024 	default: return -1;
4025 	}
4026 }
4027 
4028 static int to_ib_qp_access_flags(int mlx4_flags)
4029 {
4030 	int ib_flags = 0;
4031 
4032 	if (mlx4_flags & MLX4_QP_BIT_RRE)
4033 		ib_flags |= IB_ACCESS_REMOTE_READ;
4034 	if (mlx4_flags & MLX4_QP_BIT_RWE)
4035 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
4036 	if (mlx4_flags & MLX4_QP_BIT_RAE)
4037 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
4038 
4039 	return ib_flags;
4040 }
4041 
4042 static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
4043 			    struct rdma_ah_attr *ah_attr,
4044 			    struct mlx4_qp_path *path)
4045 {
4046 	struct mlx4_dev *dev = ibdev->dev;
4047 	u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
4048 
4049 	memset(ah_attr, 0, sizeof(*ah_attr));
4050 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
4051 	if (port_num == 0 || port_num > dev->caps.num_ports)
4052 		return;
4053 
4054 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
4055 		rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
4056 			       ((path->sched_queue & 4) << 1));
4057 	else
4058 		rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
4059 	rdma_ah_set_port_num(ah_attr, port_num);
4060 
4061 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
4062 	rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
4063 	rdma_ah_set_static_rate(ah_attr,
4064 				path->static_rate ? path->static_rate - 5 : 0);
4065 	if (path->grh_mylmc & (1 << 7)) {
4066 		rdma_ah_set_grh(ah_attr, NULL,
4067 				be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
4068 				path->mgid_index,
4069 				path->hop_limit,
4070 				(be32_to_cpu(path->tclass_flowlabel)
4071 				 >> 20) & 0xff);
4072 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
4073 	}
4074 }
4075 
4076 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
4077 		     struct ib_qp_init_attr *qp_init_attr)
4078 {
4079 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
4080 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
4081 	struct mlx4_qp_context context;
4082 	int mlx4_state;
4083 	int err = 0;
4084 
4085 	if (ibqp->rwq_ind_tbl)
4086 		return -EOPNOTSUPP;
4087 
4088 	mutex_lock(&qp->mutex);
4089 
4090 	if (qp->state == IB_QPS_RESET) {
4091 		qp_attr->qp_state = IB_QPS_RESET;
4092 		goto done;
4093 	}
4094 
4095 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
4096 	if (err) {
4097 		err = -EINVAL;
4098 		goto out;
4099 	}
4100 
4101 	mlx4_state = be32_to_cpu(context.flags) >> 28;
4102 
4103 	qp->state		     = to_ib_qp_state(mlx4_state);
4104 	qp_attr->qp_state	     = qp->state;
4105 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
4106 	qp_attr->path_mig_state	     =
4107 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
4108 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
4109 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
4110 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
4111 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
4112 	qp_attr->qp_access_flags     =
4113 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
4114 
4115 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4116 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
4117 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
4118 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
4119 		qp_attr->alt_port_num	=
4120 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
4121 	}
4122 
4123 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
4124 	if (qp_attr->qp_state == IB_QPS_INIT)
4125 		qp_attr->port_num = qp->port;
4126 	else
4127 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
4128 
4129 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4130 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
4131 
4132 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
4133 
4134 	qp_attr->max_dest_rd_atomic =
4135 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
4136 	qp_attr->min_rnr_timer	    =
4137 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
4138 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
4139 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
4140 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
4141 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
4142 
4143 done:
4144 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4145 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4146 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4147 
4148 	if (!ibqp->uobject) {
4149 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
4150 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4151 	} else {
4152 		qp_attr->cap.max_send_wr  = 0;
4153 		qp_attr->cap.max_send_sge = 0;
4154 	}
4155 
4156 	/*
4157 	 * We don't support inline sends for kernel QPs (yet), and we
4158 	 * don't know what userspace's value should be.
4159 	 */
4160 	qp_attr->cap.max_inline_data = 0;
4161 
4162 	qp_init_attr->cap	     = qp_attr->cap;
4163 
4164 	qp_init_attr->create_flags = 0;
4165 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4166 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4167 
4168 	if (qp->flags & MLX4_IB_QP_LSO)
4169 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
4170 
4171 	if (qp->flags & MLX4_IB_QP_NETIF)
4172 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
4173 
4174 	qp_init_attr->sq_sig_type =
4175 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
4176 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4177 
4178 out:
4179 	mutex_unlock(&qp->mutex);
4180 	return err;
4181 }
4182 
4183 struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
4184 				struct ib_wq_init_attr *init_attr,
4185 				struct ib_udata *udata)
4186 {
4187 	struct mlx4_ib_dev *dev;
4188 	struct ib_qp_init_attr ib_qp_init_attr;
4189 	struct mlx4_ib_qp *qp;
4190 	struct mlx4_ib_create_wq ucmd;
4191 	int err, required_cmd_sz;
4192 
4193 	if (!(udata && pd->uobject))
4194 		return ERR_PTR(-EINVAL);
4195 
4196 	required_cmd_sz = offsetof(typeof(ucmd), comp_mask) +
4197 			  sizeof(ucmd.comp_mask);
4198 	if (udata->inlen < required_cmd_sz) {
4199 		pr_debug("invalid inlen\n");
4200 		return ERR_PTR(-EINVAL);
4201 	}
4202 
4203 	if (udata->inlen > sizeof(ucmd) &&
4204 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4205 				 udata->inlen - sizeof(ucmd))) {
4206 		pr_debug("inlen is not supported\n");
4207 		return ERR_PTR(-EOPNOTSUPP);
4208 	}
4209 
4210 	if (udata->outlen)
4211 		return ERR_PTR(-EOPNOTSUPP);
4212 
4213 	dev = to_mdev(pd->device);
4214 
4215 	if (init_attr->wq_type != IB_WQT_RQ) {
4216 		pr_debug("unsupported wq type %d\n", init_attr->wq_type);
4217 		return ERR_PTR(-EOPNOTSUPP);
4218 	}
4219 
4220 	if (init_attr->create_flags & ~IB_WQ_FLAGS_SCATTER_FCS) {
4221 		pr_debug("unsupported create_flags %u\n",
4222 			 init_attr->create_flags);
4223 		return ERR_PTR(-EOPNOTSUPP);
4224 	}
4225 
4226 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
4227 	if (!qp)
4228 		return ERR_PTR(-ENOMEM);
4229 
4230 	qp->pri.vid = 0xFFFF;
4231 	qp->alt.vid = 0xFFFF;
4232 
4233 	memset(&ib_qp_init_attr, 0, sizeof(ib_qp_init_attr));
4234 	ib_qp_init_attr.qp_context = init_attr->wq_context;
4235 	ib_qp_init_attr.qp_type = IB_QPT_RAW_PACKET;
4236 	ib_qp_init_attr.cap.max_recv_wr = init_attr->max_wr;
4237 	ib_qp_init_attr.cap.max_recv_sge = init_attr->max_sge;
4238 	ib_qp_init_attr.recv_cq = init_attr->cq;
4239 	ib_qp_init_attr.send_cq = ib_qp_init_attr.recv_cq; /* Dummy CQ */
4240 
4241 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS)
4242 		ib_qp_init_attr.create_flags |= IB_QP_CREATE_SCATTER_FCS;
4243 
4244 	err = create_qp_common(dev, pd, MLX4_IB_RWQ_SRC, &ib_qp_init_attr,
4245 			       udata, 0, &qp);
4246 	if (err) {
4247 		kfree(qp);
4248 		return ERR_PTR(err);
4249 	}
4250 
4251 	qp->ibwq.event_handler = init_attr->event_handler;
4252 	qp->ibwq.wq_num = qp->mqp.qpn;
4253 	qp->ibwq.state = IB_WQS_RESET;
4254 
4255 	return &qp->ibwq;
4256 }
4257 
4258 static int ib_wq2qp_state(enum ib_wq_state state)
4259 {
4260 	switch (state) {
4261 	case IB_WQS_RESET:
4262 		return IB_QPS_RESET;
4263 	case IB_WQS_RDY:
4264 		return IB_QPS_RTR;
4265 	default:
4266 		return IB_QPS_ERR;
4267 	}
4268 }
4269 
4270 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state)
4271 {
4272 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4273 	enum ib_qp_state qp_cur_state;
4274 	enum ib_qp_state qp_new_state;
4275 	int attr_mask;
4276 	int err;
4277 
4278 	/* ib_qp.state represents the WQ HW state while ib_wq.state represents
4279 	 * the WQ logic state.
4280 	 */
4281 	qp_cur_state = qp->state;
4282 	qp_new_state = ib_wq2qp_state(new_state);
4283 
4284 	if (ib_wq2qp_state(new_state) == qp_cur_state)
4285 		return 0;
4286 
4287 	if (new_state == IB_WQS_RDY) {
4288 		struct ib_qp_attr attr = {};
4289 
4290 		attr.port_num = qp->port;
4291 		attr_mask = IB_QP_PORT;
4292 
4293 		err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, &attr,
4294 					  attr_mask, IB_QPS_RESET, IB_QPS_INIT);
4295 		if (err) {
4296 			pr_debug("WQN=0x%06x failed to apply RST->INIT on the HW QP\n",
4297 				 ibwq->wq_num);
4298 			return err;
4299 		}
4300 
4301 		qp_cur_state = IB_QPS_INIT;
4302 	}
4303 
4304 	attr_mask = 0;
4305 	err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, attr_mask,
4306 				  qp_cur_state,  qp_new_state);
4307 
4308 	if (err && (qp_cur_state == IB_QPS_INIT)) {
4309 		qp_new_state = IB_QPS_RESET;
4310 		if (__mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL,
4311 					attr_mask, IB_QPS_INIT, IB_QPS_RESET)) {
4312 			pr_warn("WQN=0x%06x failed with reverting HW's resources failure\n",
4313 				ibwq->wq_num);
4314 			qp_new_state = IB_QPS_INIT;
4315 		}
4316 	}
4317 
4318 	qp->state = qp_new_state;
4319 
4320 	return err;
4321 }
4322 
4323 int mlx4_ib_modify_wq(struct ib_wq *ibwq, struct ib_wq_attr *wq_attr,
4324 		      u32 wq_attr_mask, struct ib_udata *udata)
4325 {
4326 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4327 	struct mlx4_ib_modify_wq ucmd = {};
4328 	size_t required_cmd_sz;
4329 	enum ib_wq_state cur_state, new_state;
4330 	int err = 0;
4331 
4332 	required_cmd_sz = offsetof(typeof(ucmd), reserved) +
4333 				   sizeof(ucmd.reserved);
4334 	if (udata->inlen < required_cmd_sz)
4335 		return -EINVAL;
4336 
4337 	if (udata->inlen > sizeof(ucmd) &&
4338 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4339 				 udata->inlen - sizeof(ucmd)))
4340 		return -EOPNOTSUPP;
4341 
4342 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4343 		return -EFAULT;
4344 
4345 	if (ucmd.comp_mask || ucmd.reserved)
4346 		return -EOPNOTSUPP;
4347 
4348 	if (wq_attr_mask & IB_WQ_FLAGS)
4349 		return -EOPNOTSUPP;
4350 
4351 	cur_state = wq_attr_mask & IB_WQ_CUR_STATE ? wq_attr->curr_wq_state :
4352 						     ibwq->state;
4353 	new_state = wq_attr_mask & IB_WQ_STATE ? wq_attr->wq_state : cur_state;
4354 
4355 	if (cur_state  < IB_WQS_RESET || cur_state  > IB_WQS_ERR ||
4356 	    new_state < IB_WQS_RESET || new_state > IB_WQS_ERR)
4357 		return -EINVAL;
4358 
4359 	if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR))
4360 		return -EINVAL;
4361 
4362 	if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET))
4363 		return -EINVAL;
4364 
4365 	/* Need to protect against the parent RSS which also may modify WQ
4366 	 * state.
4367 	 */
4368 	mutex_lock(&qp->mutex);
4369 
4370 	/* Can update HW state only if a RSS QP has already associated to this
4371 	 * WQ, so we can apply its port on the WQ.
4372 	 */
4373 	if (qp->rss_usecnt)
4374 		err = _mlx4_ib_modify_wq(ibwq, new_state);
4375 
4376 	if (!err)
4377 		ibwq->state = new_state;
4378 
4379 	mutex_unlock(&qp->mutex);
4380 
4381 	return err;
4382 }
4383 
4384 int mlx4_ib_destroy_wq(struct ib_wq *ibwq)
4385 {
4386 	struct mlx4_ib_dev *dev = to_mdev(ibwq->device);
4387 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4388 
4389 	if (qp->counter_index)
4390 		mlx4_ib_free_qp_counter(dev, qp);
4391 
4392 	destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, 1);
4393 
4394 	kfree(qp);
4395 
4396 	return 0;
4397 }
4398 
4399 struct ib_rwq_ind_table
4400 *mlx4_ib_create_rwq_ind_table(struct ib_device *device,
4401 			      struct ib_rwq_ind_table_init_attr *init_attr,
4402 			      struct ib_udata *udata)
4403 {
4404 	struct ib_rwq_ind_table *rwq_ind_table;
4405 	struct mlx4_ib_create_rwq_ind_tbl_resp resp = {};
4406 	unsigned int ind_tbl_size = 1 << init_attr->log_ind_tbl_size;
4407 	unsigned int base_wqn;
4408 	size_t min_resp_len;
4409 	int i;
4410 	int err;
4411 
4412 	if (udata->inlen > 0 &&
4413 	    !ib_is_udata_cleared(udata, 0,
4414 				 udata->inlen))
4415 		return ERR_PTR(-EOPNOTSUPP);
4416 
4417 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4418 	if (udata->outlen && udata->outlen < min_resp_len)
4419 		return ERR_PTR(-EINVAL);
4420 
4421 	if (ind_tbl_size >
4422 	    device->attrs.rss_caps.max_rwq_indirection_table_size) {
4423 		pr_debug("log_ind_tbl_size = %d is bigger than supported = %d\n",
4424 			 ind_tbl_size,
4425 			 device->attrs.rss_caps.max_rwq_indirection_table_size);
4426 		return ERR_PTR(-EINVAL);
4427 	}
4428 
4429 	base_wqn = init_attr->ind_tbl[0]->wq_num;
4430 
4431 	if (base_wqn % ind_tbl_size) {
4432 		pr_debug("WQN=0x%x isn't aligned with indirection table size\n",
4433 			 base_wqn);
4434 		return ERR_PTR(-EINVAL);
4435 	}
4436 
4437 	for (i = 1; i < ind_tbl_size; i++) {
4438 		if (++base_wqn != init_attr->ind_tbl[i]->wq_num) {
4439 			pr_debug("indirection table's WQNs aren't consecutive\n");
4440 			return ERR_PTR(-EINVAL);
4441 		}
4442 	}
4443 
4444 	rwq_ind_table = kzalloc(sizeof(*rwq_ind_table), GFP_KERNEL);
4445 	if (!rwq_ind_table)
4446 		return ERR_PTR(-ENOMEM);
4447 
4448 	if (udata->outlen) {
4449 		resp.response_length = offsetof(typeof(resp), response_length) +
4450 					sizeof(resp.response_length);
4451 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4452 		if (err)
4453 			goto err;
4454 	}
4455 
4456 	return rwq_ind_table;
4457 
4458 err:
4459 	kfree(rwq_ind_table);
4460 	return ERR_PTR(err);
4461 }
4462 
4463 int mlx4_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4464 {
4465 	kfree(ib_rwq_ind_tbl);
4466 	return 0;
4467 }
4468