xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision 68198dca)
1 /*
2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/log2.h>
35 #include <linux/etherdevice.h>
36 #include <net/ip.h>
37 #include <linux/slab.h>
38 #include <linux/netdevice.h>
39 
40 #include <rdma/ib_cache.h>
41 #include <rdma/ib_pack.h>
42 #include <rdma/ib_addr.h>
43 #include <rdma/ib_mad.h>
44 
45 #include <linux/mlx4/driver.h>
46 #include <linux/mlx4/qp.h>
47 
48 #include "mlx4_ib.h"
49 #include <rdma/mlx4-abi.h>
50 
51 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
52 			     struct mlx4_ib_cq *recv_cq);
53 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
54 			       struct mlx4_ib_cq *recv_cq);
55 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state);
56 
57 enum {
58 	MLX4_IB_ACK_REQ_FREQ	= 8,
59 };
60 
61 enum {
62 	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
63 	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
64 	MLX4_IB_LINK_TYPE_IB		= 0,
65 	MLX4_IB_LINK_TYPE_ETH		= 1
66 };
67 
68 enum {
69 	/*
70 	 * Largest possible UD header: send with GRH and immediate
71 	 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
72 	 * tag.  (LRH would only use 8 bytes, so Ethernet is the
73 	 * biggest case)
74 	 */
75 	MLX4_IB_UD_HEADER_SIZE		= 82,
76 	MLX4_IB_LSO_HEADER_SPARE	= 128,
77 };
78 
79 struct mlx4_ib_sqp {
80 	struct mlx4_ib_qp	qp;
81 	int			pkey_index;
82 	u32			qkey;
83 	u32			send_psn;
84 	struct ib_ud_header	ud_header;
85 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
86 	struct ib_qp		*roce_v2_gsi;
87 };
88 
89 enum {
90 	MLX4_IB_MIN_SQ_STRIDE	= 6,
91 	MLX4_IB_CACHE_LINE_SIZE	= 64,
92 };
93 
94 enum {
95 	MLX4_RAW_QP_MTU		= 7,
96 	MLX4_RAW_QP_MSGMAX	= 31,
97 };
98 
99 #ifndef ETH_ALEN
100 #define ETH_ALEN        6
101 #endif
102 
103 static const __be32 mlx4_ib_opcode[] = {
104 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
105 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
106 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
107 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
108 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
109 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
110 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
111 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
112 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
113 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
114 	[IB_WR_REG_MR]				= cpu_to_be32(MLX4_OPCODE_FMR),
115 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
116 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
117 };
118 
119 enum mlx4_ib_source_type {
120 	MLX4_IB_QP_SRC	= 0,
121 	MLX4_IB_RWQ_SRC	= 1,
122 };
123 
124 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
125 {
126 	return container_of(mqp, struct mlx4_ib_sqp, qp);
127 }
128 
129 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
130 {
131 	if (!mlx4_is_master(dev->dev))
132 		return 0;
133 
134 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
135 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
136 		8 * MLX4_MFUNC_MAX;
137 }
138 
139 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
140 {
141 	int proxy_sqp = 0;
142 	int real_sqp = 0;
143 	int i;
144 	/* PPF or Native -- real SQP */
145 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
146 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
147 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
148 	if (real_sqp)
149 		return 1;
150 	/* VF or PF -- proxy SQP */
151 	if (mlx4_is_mfunc(dev->dev)) {
152 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
153 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy ||
154 			    qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp1_proxy) {
155 				proxy_sqp = 1;
156 				break;
157 			}
158 		}
159 	}
160 	if (proxy_sqp)
161 		return 1;
162 
163 	return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
164 }
165 
166 /* used for INIT/CLOSE port logic */
167 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
168 {
169 	int proxy_qp0 = 0;
170 	int real_qp0 = 0;
171 	int i;
172 	/* PPF or Native -- real QP0 */
173 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
174 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
175 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
176 	if (real_qp0)
177 		return 1;
178 	/* VF or PF -- proxy QP0 */
179 	if (mlx4_is_mfunc(dev->dev)) {
180 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
181 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy) {
182 				proxy_qp0 = 1;
183 				break;
184 			}
185 		}
186 	}
187 	return proxy_qp0;
188 }
189 
190 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
191 {
192 	return mlx4_buf_offset(&qp->buf, offset);
193 }
194 
195 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
196 {
197 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
198 }
199 
200 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
201 {
202 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
203 }
204 
205 /*
206  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
207  * first four bytes of every 64 byte chunk with
208  *     0x7FFFFFF | (invalid_ownership_value << 31).
209  *
210  * When the max work request size is less than or equal to the WQE
211  * basic block size, as an optimization, we can stamp all WQEs with
212  * 0xffffffff, and skip the very first chunk of each WQE.
213  */
214 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
215 {
216 	__be32 *wqe;
217 	int i;
218 	int s;
219 	int ind;
220 	void *buf;
221 	__be32 stamp;
222 	struct mlx4_wqe_ctrl_seg *ctrl;
223 
224 	if (qp->sq_max_wqes_per_wr > 1) {
225 		s = roundup(size, 1U << qp->sq.wqe_shift);
226 		for (i = 0; i < s; i += 64) {
227 			ind = (i >> qp->sq.wqe_shift) + n;
228 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
229 						       cpu_to_be32(0xffffffff);
230 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
231 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
232 			*wqe = stamp;
233 		}
234 	} else {
235 		ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
236 		s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4;
237 		for (i = 64; i < s; i += 64) {
238 			wqe = buf + i;
239 			*wqe = cpu_to_be32(0xffffffff);
240 		}
241 	}
242 }
243 
244 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
245 {
246 	struct mlx4_wqe_ctrl_seg *ctrl;
247 	struct mlx4_wqe_inline_seg *inl;
248 	void *wqe;
249 	int s;
250 
251 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
252 	s = sizeof(struct mlx4_wqe_ctrl_seg);
253 
254 	if (qp->ibqp.qp_type == IB_QPT_UD) {
255 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
256 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
257 		memset(dgram, 0, sizeof *dgram);
258 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
259 		s += sizeof(struct mlx4_wqe_datagram_seg);
260 	}
261 
262 	/* Pad the remainder of the WQE with an inline data segment. */
263 	if (size > s) {
264 		inl = wqe + s;
265 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
266 	}
267 	ctrl->srcrb_flags = 0;
268 	ctrl->qpn_vlan.fence_size = size / 16;
269 	/*
270 	 * Make sure descriptor is fully written before setting ownership bit
271 	 * (because HW can start executing as soon as we do).
272 	 */
273 	wmb();
274 
275 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
276 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
277 
278 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
279 }
280 
281 /* Post NOP WQE to prevent wrap-around in the middle of WR */
282 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
283 {
284 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
285 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
286 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
287 		ind += s;
288 	}
289 	return ind;
290 }
291 
292 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
293 {
294 	struct ib_event event;
295 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
296 
297 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
298 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
299 
300 	if (ibqp->event_handler) {
301 		event.device     = ibqp->device;
302 		event.element.qp = ibqp;
303 		switch (type) {
304 		case MLX4_EVENT_TYPE_PATH_MIG:
305 			event.event = IB_EVENT_PATH_MIG;
306 			break;
307 		case MLX4_EVENT_TYPE_COMM_EST:
308 			event.event = IB_EVENT_COMM_EST;
309 			break;
310 		case MLX4_EVENT_TYPE_SQ_DRAINED:
311 			event.event = IB_EVENT_SQ_DRAINED;
312 			break;
313 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
314 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
315 			break;
316 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
317 			event.event = IB_EVENT_QP_FATAL;
318 			break;
319 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
320 			event.event = IB_EVENT_PATH_MIG_ERR;
321 			break;
322 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
323 			event.event = IB_EVENT_QP_REQ_ERR;
324 			break;
325 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
326 			event.event = IB_EVENT_QP_ACCESS_ERR;
327 			break;
328 		default:
329 			pr_warn("Unexpected event type %d "
330 			       "on QP %06x\n", type, qp->qpn);
331 			return;
332 		}
333 
334 		ibqp->event_handler(&event, ibqp->qp_context);
335 	}
336 }
337 
338 static void mlx4_ib_wq_event(struct mlx4_qp *qp, enum mlx4_event type)
339 {
340 	pr_warn_ratelimited("Unexpected event type %d on WQ 0x%06x. Events are not supported for WQs\n",
341 			    type, qp->qpn);
342 }
343 
344 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
345 {
346 	/*
347 	 * UD WQEs must have a datagram segment.
348 	 * RC and UC WQEs might have a remote address segment.
349 	 * MLX WQEs need two extra inline data segments (for the UD
350 	 * header and space for the ICRC).
351 	 */
352 	switch (type) {
353 	case MLX4_IB_QPT_UD:
354 		return sizeof (struct mlx4_wqe_ctrl_seg) +
355 			sizeof (struct mlx4_wqe_datagram_seg) +
356 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
357 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
358 	case MLX4_IB_QPT_PROXY_SMI:
359 	case MLX4_IB_QPT_PROXY_GSI:
360 		return sizeof (struct mlx4_wqe_ctrl_seg) +
361 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
362 	case MLX4_IB_QPT_TUN_SMI_OWNER:
363 	case MLX4_IB_QPT_TUN_GSI:
364 		return sizeof (struct mlx4_wqe_ctrl_seg) +
365 			sizeof (struct mlx4_wqe_datagram_seg);
366 
367 	case MLX4_IB_QPT_UC:
368 		return sizeof (struct mlx4_wqe_ctrl_seg) +
369 			sizeof (struct mlx4_wqe_raddr_seg);
370 	case MLX4_IB_QPT_RC:
371 		return sizeof (struct mlx4_wqe_ctrl_seg) +
372 			sizeof (struct mlx4_wqe_masked_atomic_seg) +
373 			sizeof (struct mlx4_wqe_raddr_seg);
374 	case MLX4_IB_QPT_SMI:
375 	case MLX4_IB_QPT_GSI:
376 		return sizeof (struct mlx4_wqe_ctrl_seg) +
377 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
378 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
379 					   MLX4_INLINE_ALIGN) *
380 			      sizeof (struct mlx4_wqe_inline_seg),
381 			      sizeof (struct mlx4_wqe_data_seg)) +
382 			ALIGN(4 +
383 			      sizeof (struct mlx4_wqe_inline_seg),
384 			      sizeof (struct mlx4_wqe_data_seg));
385 	default:
386 		return sizeof (struct mlx4_wqe_ctrl_seg);
387 	}
388 }
389 
390 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
391 		       int is_user, int has_rq, struct mlx4_ib_qp *qp,
392 		       u32 inl_recv_sz)
393 {
394 	/* Sanity check RQ size before proceeding */
395 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
396 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
397 		return -EINVAL;
398 
399 	if (!has_rq) {
400 		if (cap->max_recv_wr || inl_recv_sz)
401 			return -EINVAL;
402 
403 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
404 	} else {
405 		u32 max_inl_recv_sz = dev->dev->caps.max_rq_sg *
406 			sizeof(struct mlx4_wqe_data_seg);
407 		u32 wqe_size;
408 
409 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
410 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge ||
411 				inl_recv_sz > max_inl_recv_sz))
412 			return -EINVAL;
413 
414 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
415 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
416 		wqe_size = qp->rq.max_gs * sizeof(struct mlx4_wqe_data_seg);
417 		qp->rq.wqe_shift = ilog2(max_t(u32, wqe_size, inl_recv_sz));
418 	}
419 
420 	/* leave userspace return values as they were, so as not to break ABI */
421 	if (is_user) {
422 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
423 		cap->max_recv_sge = qp->rq.max_gs;
424 	} else {
425 		cap->max_recv_wr  = qp->rq.max_post =
426 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
427 		cap->max_recv_sge = min(qp->rq.max_gs,
428 					min(dev->dev->caps.max_sq_sg,
429 					    dev->dev->caps.max_rq_sg));
430 	}
431 
432 	return 0;
433 }
434 
435 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
436 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp,
437 			      bool shrink_wqe)
438 {
439 	int s;
440 
441 	/* Sanity check SQ size before proceeding */
442 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
443 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
444 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
445 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
446 		return -EINVAL;
447 
448 	/*
449 	 * For MLX transport we need 2 extra S/G entries:
450 	 * one for the header and one for the checksum at the end
451 	 */
452 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
453 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
454 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
455 		return -EINVAL;
456 
457 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
458 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
459 		send_wqe_overhead(type, qp->flags);
460 
461 	if (s > dev->dev->caps.max_sq_desc_sz)
462 		return -EINVAL;
463 
464 	/*
465 	 * Hermon supports shrinking WQEs, such that a single work
466 	 * request can include multiple units of 1 << wqe_shift.  This
467 	 * way, work requests can differ in size, and do not have to
468 	 * be a power of 2 in size, saving memory and speeding up send
469 	 * WR posting.  Unfortunately, if we do this then the
470 	 * wqe_index field in CQEs can't be used to look up the WR ID
471 	 * anymore, so we do this only if selective signaling is off.
472 	 *
473 	 * Further, on 32-bit platforms, we can't use vmap() to make
474 	 * the QP buffer virtually contiguous.  Thus we have to use
475 	 * constant-sized WRs to make sure a WR is always fully within
476 	 * a single page-sized chunk.
477 	 *
478 	 * Finally, we use NOP work requests to pad the end of the
479 	 * work queue, to avoid wrap-around in the middle of WR.  We
480 	 * set NEC bit to avoid getting completions with error for
481 	 * these NOP WRs, but since NEC is only supported starting
482 	 * with firmware 2.2.232, we use constant-sized WRs for older
483 	 * firmware.
484 	 *
485 	 * And, since MLX QPs only support SEND, we use constant-sized
486 	 * WRs in this case.
487 	 *
488 	 * We look for the smallest value of wqe_shift such that the
489 	 * resulting number of wqes does not exceed device
490 	 * capabilities.
491 	 *
492 	 * We set WQE size to at least 64 bytes, this way stamping
493 	 * invalidates each WQE.
494 	 */
495 	if (shrink_wqe && dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
496 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
497 	    type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
498 	    !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
499 		      MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
500 		qp->sq.wqe_shift = ilog2(64);
501 	else
502 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
503 
504 	for (;;) {
505 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
506 
507 		/*
508 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
509 		 * allow HW to prefetch.
510 		 */
511 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
512 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
513 						    qp->sq_max_wqes_per_wr +
514 						    qp->sq_spare_wqes);
515 
516 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
517 			break;
518 
519 		if (qp->sq_max_wqes_per_wr <= 1)
520 			return -EINVAL;
521 
522 		++qp->sq.wqe_shift;
523 	}
524 
525 	qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
526 			     (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
527 			 send_wqe_overhead(type, qp->flags)) /
528 		sizeof (struct mlx4_wqe_data_seg);
529 
530 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
531 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
532 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
533 		qp->rq.offset = 0;
534 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
535 	} else {
536 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
537 		qp->sq.offset = 0;
538 	}
539 
540 	cap->max_send_wr  = qp->sq.max_post =
541 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
542 	cap->max_send_sge = min(qp->sq.max_gs,
543 				min(dev->dev->caps.max_sq_sg,
544 				    dev->dev->caps.max_rq_sg));
545 	/* We don't support inline sends for kernel QPs (yet) */
546 	cap->max_inline_data = 0;
547 
548 	return 0;
549 }
550 
551 static int set_user_sq_size(struct mlx4_ib_dev *dev,
552 			    struct mlx4_ib_qp *qp,
553 			    struct mlx4_ib_create_qp *ucmd)
554 {
555 	/* Sanity check SQ size before proceeding */
556 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
557 	    ucmd->log_sq_stride >
558 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
559 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
560 		return -EINVAL;
561 
562 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
563 	qp->sq.wqe_shift = ucmd->log_sq_stride;
564 
565 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
566 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
567 
568 	return 0;
569 }
570 
571 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
572 {
573 	int i;
574 
575 	qp->sqp_proxy_rcv =
576 		kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
577 			GFP_KERNEL);
578 	if (!qp->sqp_proxy_rcv)
579 		return -ENOMEM;
580 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
581 		qp->sqp_proxy_rcv[i].addr =
582 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
583 				GFP_KERNEL);
584 		if (!qp->sqp_proxy_rcv[i].addr)
585 			goto err;
586 		qp->sqp_proxy_rcv[i].map =
587 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
588 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
589 					  DMA_FROM_DEVICE);
590 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
591 			kfree(qp->sqp_proxy_rcv[i].addr);
592 			goto err;
593 		}
594 	}
595 	return 0;
596 
597 err:
598 	while (i > 0) {
599 		--i;
600 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
601 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
602 				    DMA_FROM_DEVICE);
603 		kfree(qp->sqp_proxy_rcv[i].addr);
604 	}
605 	kfree(qp->sqp_proxy_rcv);
606 	qp->sqp_proxy_rcv = NULL;
607 	return -ENOMEM;
608 }
609 
610 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
611 {
612 	int i;
613 
614 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
615 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
616 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
617 				    DMA_FROM_DEVICE);
618 		kfree(qp->sqp_proxy_rcv[i].addr);
619 	}
620 	kfree(qp->sqp_proxy_rcv);
621 }
622 
623 static int qp_has_rq(struct ib_qp_init_attr *attr)
624 {
625 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
626 		return 0;
627 
628 	return !attr->srq;
629 }
630 
631 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
632 {
633 	int i;
634 	for (i = 0; i < dev->caps.num_ports; i++) {
635 		if (qpn == dev->caps.spec_qps[i].qp0_proxy)
636 			return !!dev->caps.spec_qps[i].qp0_qkey;
637 	}
638 	return 0;
639 }
640 
641 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
642 				    struct mlx4_ib_qp *qp)
643 {
644 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
645 	mlx4_counter_free(dev->dev, qp->counter_index->index);
646 	list_del(&qp->counter_index->list);
647 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
648 
649 	kfree(qp->counter_index);
650 	qp->counter_index = NULL;
651 }
652 
653 static int set_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_rss *rss_ctx,
654 		      struct ib_qp_init_attr *init_attr,
655 		      struct mlx4_ib_create_qp_rss *ucmd)
656 {
657 	rss_ctx->base_qpn_tbl_sz = init_attr->rwq_ind_tbl->ind_tbl[0]->wq_num |
658 		(init_attr->rwq_ind_tbl->log_ind_tbl_size << 24);
659 
660 	if ((ucmd->rx_hash_function == MLX4_IB_RX_HASH_FUNC_TOEPLITZ) &&
661 	    (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS_TOP)) {
662 		memcpy(rss_ctx->rss_key, ucmd->rx_hash_key,
663 		       MLX4_EN_RSS_KEY_SIZE);
664 	} else {
665 		pr_debug("RX Hash function is not supported\n");
666 		return (-EOPNOTSUPP);
667 	}
668 
669 	if (ucmd->rx_hash_fields_mask & ~(MLX4_IB_RX_HASH_SRC_IPV4	|
670 					  MLX4_IB_RX_HASH_DST_IPV4	|
671 					  MLX4_IB_RX_HASH_SRC_IPV6	|
672 					  MLX4_IB_RX_HASH_DST_IPV6	|
673 					  MLX4_IB_RX_HASH_SRC_PORT_TCP	|
674 					  MLX4_IB_RX_HASH_DST_PORT_TCP	|
675 					  MLX4_IB_RX_HASH_SRC_PORT_UDP	|
676 					  MLX4_IB_RX_HASH_DST_PORT_UDP)) {
677 		pr_debug("RX Hash fields_mask has unsupported mask (0x%llx)\n",
678 			 ucmd->rx_hash_fields_mask);
679 		return (-EOPNOTSUPP);
680 	}
681 
682 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) &&
683 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
684 		rss_ctx->flags = MLX4_RSS_IPV4;
685 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) ||
686 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
687 		pr_debug("RX Hash fields_mask is not supported - both IPv4 SRC and DST must be set\n");
688 		return (-EOPNOTSUPP);
689 	}
690 
691 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) &&
692 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
693 		rss_ctx->flags |= MLX4_RSS_IPV6;
694 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) ||
695 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
696 		pr_debug("RX Hash fields_mask is not supported - both IPv6 SRC and DST must be set\n");
697 		return (-EOPNOTSUPP);
698 	}
699 
700 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) &&
701 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
702 		if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS)) {
703 			pr_debug("RX Hash fields_mask for UDP is not supported\n");
704 			return (-EOPNOTSUPP);
705 		}
706 
707 		if (rss_ctx->flags & MLX4_RSS_IPV4)
708 			rss_ctx->flags |= MLX4_RSS_UDP_IPV4;
709 		if (rss_ctx->flags & MLX4_RSS_IPV6)
710 			rss_ctx->flags |= MLX4_RSS_UDP_IPV6;
711 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
712 			pr_debug("RX Hash fields_mask is not supported - UDP must be set with IPv4 or IPv6\n");
713 			return (-EOPNOTSUPP);
714 		}
715 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) ||
716 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
717 		pr_debug("RX Hash fields_mask is not supported - both UDP SRC and DST must be set\n");
718 		return (-EOPNOTSUPP);
719 	}
720 
721 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) &&
722 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
723 		if (rss_ctx->flags & MLX4_RSS_IPV4)
724 			rss_ctx->flags |= MLX4_RSS_TCP_IPV4;
725 		if (rss_ctx->flags & MLX4_RSS_IPV6)
726 			rss_ctx->flags |= MLX4_RSS_TCP_IPV6;
727 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
728 			pr_debug("RX Hash fields_mask is not supported - TCP must be set with IPv4 or IPv6\n");
729 			return (-EOPNOTSUPP);
730 		}
731 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) ||
732 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
733 		pr_debug("RX Hash fields_mask is not supported - both TCP SRC and DST must be set\n");
734 		return (-EOPNOTSUPP);
735 	}
736 
737 	return 0;
738 }
739 
740 static int create_qp_rss(struct mlx4_ib_dev *dev, struct ib_pd *ibpd,
741 			 struct ib_qp_init_attr *init_attr,
742 			 struct mlx4_ib_create_qp_rss *ucmd,
743 			 struct mlx4_ib_qp *qp)
744 {
745 	int qpn;
746 	int err;
747 
748 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
749 
750 	err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 0, qp->mqp.usage);
751 	if (err)
752 		return err;
753 
754 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
755 	if (err)
756 		goto err_qpn;
757 
758 	mutex_init(&qp->mutex);
759 
760 	INIT_LIST_HEAD(&qp->gid_list);
761 	INIT_LIST_HEAD(&qp->steering_rules);
762 
763 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
764 	qp->state = IB_QPS_RESET;
765 
766 	/* Set dummy send resources to be compatible with HV and PRM */
767 	qp->sq_no_prefetch = 1;
768 	qp->sq.wqe_cnt = 1;
769 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
770 	qp->buf_size = qp->sq.wqe_cnt << MLX4_IB_MIN_SQ_STRIDE;
771 	qp->mtt = (to_mqp(
772 		   (struct ib_qp *)init_attr->rwq_ind_tbl->ind_tbl[0]))->mtt;
773 
774 	qp->rss_ctx = kzalloc(sizeof(*qp->rss_ctx), GFP_KERNEL);
775 	if (!qp->rss_ctx) {
776 		err = -ENOMEM;
777 		goto err_qp_alloc;
778 	}
779 
780 	err = set_qp_rss(dev, qp->rss_ctx, init_attr, ucmd);
781 	if (err)
782 		goto err;
783 
784 	return 0;
785 
786 err:
787 	kfree(qp->rss_ctx);
788 
789 err_qp_alloc:
790 	mlx4_qp_remove(dev->dev, &qp->mqp);
791 	mlx4_qp_free(dev->dev, &qp->mqp);
792 
793 err_qpn:
794 	mlx4_qp_release_range(dev->dev, qpn, 1);
795 	return err;
796 }
797 
798 static struct ib_qp *_mlx4_ib_create_qp_rss(struct ib_pd *pd,
799 					    struct ib_qp_init_attr *init_attr,
800 					    struct ib_udata *udata)
801 {
802 	struct mlx4_ib_qp *qp;
803 	struct mlx4_ib_create_qp_rss ucmd = {};
804 	size_t required_cmd_sz;
805 	int err;
806 
807 	if (!udata) {
808 		pr_debug("RSS QP with NULL udata\n");
809 		return ERR_PTR(-EINVAL);
810 	}
811 
812 	if (udata->outlen)
813 		return ERR_PTR(-EOPNOTSUPP);
814 
815 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) +
816 					sizeof(ucmd.reserved1);
817 	if (udata->inlen < required_cmd_sz) {
818 		pr_debug("invalid inlen\n");
819 		return ERR_PTR(-EINVAL);
820 	}
821 
822 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
823 		pr_debug("copy failed\n");
824 		return ERR_PTR(-EFAULT);
825 	}
826 
827 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)))
828 		return ERR_PTR(-EOPNOTSUPP);
829 
830 	if (ucmd.comp_mask || ucmd.reserved1)
831 		return ERR_PTR(-EOPNOTSUPP);
832 
833 	if (udata->inlen > sizeof(ucmd) &&
834 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
835 				 udata->inlen - sizeof(ucmd))) {
836 		pr_debug("inlen is not supported\n");
837 		return ERR_PTR(-EOPNOTSUPP);
838 	}
839 
840 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
841 		pr_debug("RSS QP with unsupported QP type %d\n",
842 			 init_attr->qp_type);
843 		return ERR_PTR(-EOPNOTSUPP);
844 	}
845 
846 	if (init_attr->create_flags) {
847 		pr_debug("RSS QP doesn't support create flags\n");
848 		return ERR_PTR(-EOPNOTSUPP);
849 	}
850 
851 	if (init_attr->send_cq || init_attr->cap.max_send_wr) {
852 		pr_debug("RSS QP with unsupported send attributes\n");
853 		return ERR_PTR(-EOPNOTSUPP);
854 	}
855 
856 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
857 	if (!qp)
858 		return ERR_PTR(-ENOMEM);
859 
860 	qp->pri.vid = 0xFFFF;
861 	qp->alt.vid = 0xFFFF;
862 
863 	err = create_qp_rss(to_mdev(pd->device), pd, init_attr, &ucmd, qp);
864 	if (err) {
865 		kfree(qp);
866 		return ERR_PTR(err);
867 	}
868 
869 	qp->ibqp.qp_num = qp->mqp.qpn;
870 
871 	return &qp->ibqp;
872 }
873 
874 /*
875  * This function allocates a WQN from a range which is consecutive and aligned
876  * to its size. In case the range is full, then it creates a new range and
877  * allocates WQN from it. The new range will be used for following allocations.
878  */
879 static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context,
880 			     struct mlx4_ib_qp *qp, int range_size, int *wqn)
881 {
882 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
883 	struct mlx4_wqn_range *range;
884 	int err = 0;
885 
886 	mutex_lock(&context->wqn_ranges_mutex);
887 
888 	range = list_first_entry_or_null(&context->wqn_ranges_list,
889 					 struct mlx4_wqn_range, list);
890 
891 	if (!range || (range->refcount == range->size) || range->dirty) {
892 		range = kzalloc(sizeof(*range), GFP_KERNEL);
893 		if (!range) {
894 			err = -ENOMEM;
895 			goto out;
896 		}
897 
898 		err = mlx4_qp_reserve_range(dev->dev, range_size,
899 					    range_size, &range->base_wqn, 0,
900 					    qp->mqp.usage);
901 		if (err) {
902 			kfree(range);
903 			goto out;
904 		}
905 
906 		range->size = range_size;
907 		list_add(&range->list, &context->wqn_ranges_list);
908 	} else if (range_size != 1) {
909 		/*
910 		 * Requesting a new range (>1) when last range is still open, is
911 		 * not valid.
912 		 */
913 		err = -EINVAL;
914 		goto out;
915 	}
916 
917 	qp->wqn_range = range;
918 
919 	*wqn = range->base_wqn + range->refcount;
920 
921 	range->refcount++;
922 
923 out:
924 	mutex_unlock(&context->wqn_ranges_mutex);
925 
926 	return err;
927 }
928 
929 static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context,
930 				struct mlx4_ib_qp *qp, bool dirty_release)
931 {
932 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
933 	struct mlx4_wqn_range *range;
934 
935 	mutex_lock(&context->wqn_ranges_mutex);
936 
937 	range = qp->wqn_range;
938 
939 	range->refcount--;
940 	if (!range->refcount) {
941 		mlx4_qp_release_range(dev->dev, range->base_wqn,
942 				      range->size);
943 		list_del(&range->list);
944 		kfree(range);
945 	} else if (dirty_release) {
946 	/*
947 	 * A range which one of its WQNs is destroyed, won't be able to be
948 	 * reused for further WQN allocations.
949 	 * The next created WQ will allocate a new range.
950 	 */
951 		range->dirty = 1;
952 	}
953 
954 	mutex_unlock(&context->wqn_ranges_mutex);
955 }
956 
957 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
958 			    enum mlx4_ib_source_type src,
959 			    struct ib_qp_init_attr *init_attr,
960 			    struct ib_udata *udata, int sqpn,
961 			    struct mlx4_ib_qp **caller_qp)
962 {
963 	int qpn;
964 	int err;
965 	struct ib_qp_cap backup_cap;
966 	struct mlx4_ib_sqp *sqp = NULL;
967 	struct mlx4_ib_qp *qp;
968 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
969 	struct mlx4_ib_cq *mcq;
970 	unsigned long flags;
971 	int range_size = 0;
972 
973 	/* When tunneling special qps, we use a plain UD qp */
974 	if (sqpn) {
975 		if (mlx4_is_mfunc(dev->dev) &&
976 		    (!mlx4_is_master(dev->dev) ||
977 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
978 			if (init_attr->qp_type == IB_QPT_GSI)
979 				qp_type = MLX4_IB_QPT_PROXY_GSI;
980 			else {
981 				if (mlx4_is_master(dev->dev) ||
982 				    qp0_enabled_vf(dev->dev, sqpn))
983 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
984 				else
985 					qp_type = MLX4_IB_QPT_PROXY_SMI;
986 			}
987 		}
988 		qpn = sqpn;
989 		/* add extra sg entry for tunneling */
990 		init_attr->cap.max_recv_sge++;
991 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
992 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
993 			container_of(init_attr,
994 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
995 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
996 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
997 		    !mlx4_is_master(dev->dev))
998 			return -EINVAL;
999 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
1000 			qp_type = MLX4_IB_QPT_TUN_GSI;
1001 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
1002 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
1003 					     tnl_init->port))
1004 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
1005 		else
1006 			qp_type = MLX4_IB_QPT_TUN_SMI;
1007 		/* we are definitely in the PPF here, since we are creating
1008 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
1009 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
1010 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
1011 		sqpn = qpn;
1012 	}
1013 
1014 	if (!*caller_qp) {
1015 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
1016 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
1017 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
1018 			sqp = kzalloc(sizeof(struct mlx4_ib_sqp), GFP_KERNEL);
1019 			if (!sqp)
1020 				return -ENOMEM;
1021 			qp = &sqp->qp;
1022 			qp->pri.vid = 0xFFFF;
1023 			qp->alt.vid = 0xFFFF;
1024 		} else {
1025 			qp = kzalloc(sizeof(struct mlx4_ib_qp), GFP_KERNEL);
1026 			if (!qp)
1027 				return -ENOMEM;
1028 			qp->pri.vid = 0xFFFF;
1029 			qp->alt.vid = 0xFFFF;
1030 		}
1031 	} else
1032 		qp = *caller_qp;
1033 
1034 	qp->mlx4_ib_qp_type = qp_type;
1035 
1036 	mutex_init(&qp->mutex);
1037 	spin_lock_init(&qp->sq.lock);
1038 	spin_lock_init(&qp->rq.lock);
1039 	INIT_LIST_HEAD(&qp->gid_list);
1040 	INIT_LIST_HEAD(&qp->steering_rules);
1041 
1042 	qp->state	 = IB_QPS_RESET;
1043 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1044 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1045 
1046 
1047 	if (pd->uobject) {
1048 		union {
1049 			struct mlx4_ib_create_qp qp;
1050 			struct mlx4_ib_create_wq wq;
1051 		} ucmd;
1052 		size_t copy_len;
1053 		int shift;
1054 		int n;
1055 
1056 		copy_len = (src == MLX4_IB_QP_SRC) ?
1057 			   sizeof(struct mlx4_ib_create_qp) :
1058 			   min(sizeof(struct mlx4_ib_create_wq), udata->inlen);
1059 
1060 		if (ib_copy_from_udata(&ucmd, udata, copy_len)) {
1061 			err = -EFAULT;
1062 			goto err;
1063 		}
1064 
1065 		if (src == MLX4_IB_RWQ_SRC) {
1066 			if (ucmd.wq.comp_mask || ucmd.wq.reserved[0] ||
1067 			    ucmd.wq.reserved[1] || ucmd.wq.reserved[2]) {
1068 				pr_debug("user command isn't supported\n");
1069 				err = -EOPNOTSUPP;
1070 				goto err;
1071 			}
1072 
1073 			if (ucmd.wq.log_range_size >
1074 			    ilog2(dev->dev->caps.max_rss_tbl_sz)) {
1075 				pr_debug("WQN range size must be equal or smaller than %d\n",
1076 					 dev->dev->caps.max_rss_tbl_sz);
1077 				err = -EOPNOTSUPP;
1078 				goto err;
1079 			}
1080 			range_size = 1 << ucmd.wq.log_range_size;
1081 		} else {
1082 			qp->inl_recv_sz = ucmd.qp.inl_recv_sz;
1083 		}
1084 
1085 		err = set_rq_size(dev, &init_attr->cap, !!pd->uobject,
1086 				  qp_has_rq(init_attr), qp, qp->inl_recv_sz);
1087 		if (err)
1088 			goto err;
1089 
1090 		if (src == MLX4_IB_QP_SRC) {
1091 			qp->sq_no_prefetch = ucmd.qp.sq_no_prefetch;
1092 
1093 			err = set_user_sq_size(dev, qp,
1094 					       (struct mlx4_ib_create_qp *)
1095 					       &ucmd);
1096 			if (err)
1097 				goto err;
1098 		} else {
1099 			qp->sq_no_prefetch = 1;
1100 			qp->sq.wqe_cnt = 1;
1101 			qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
1102 			/* Allocated buffer expects to have at least that SQ
1103 			 * size.
1104 			 */
1105 			qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
1106 				(qp->sq.wqe_cnt << qp->sq.wqe_shift);
1107 		}
1108 
1109 		qp->umem = ib_umem_get(pd->uobject->context,
1110 				(src == MLX4_IB_QP_SRC) ? ucmd.qp.buf_addr :
1111 				ucmd.wq.buf_addr, qp->buf_size, 0, 0);
1112 		if (IS_ERR(qp->umem)) {
1113 			err = PTR_ERR(qp->umem);
1114 			goto err;
1115 		}
1116 
1117 		n = ib_umem_page_count(qp->umem);
1118 		shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
1119 		err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
1120 
1121 		if (err)
1122 			goto err_buf;
1123 
1124 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
1125 		if (err)
1126 			goto err_mtt;
1127 
1128 		if (qp_has_rq(init_attr)) {
1129 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
1130 				(src == MLX4_IB_QP_SRC) ? ucmd.qp.db_addr :
1131 				ucmd.wq.db_addr, &qp->db);
1132 			if (err)
1133 				goto err_mtt;
1134 		}
1135 		qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
1136 	} else {
1137 		err = set_rq_size(dev, &init_attr->cap, !!pd->uobject,
1138 				  qp_has_rq(init_attr), qp, 0);
1139 		if (err)
1140 			goto err;
1141 
1142 		qp->sq_no_prefetch = 0;
1143 
1144 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
1145 			qp->flags |= MLX4_IB_QP_LSO;
1146 
1147 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1148 			if (dev->steering_support ==
1149 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
1150 				qp->flags |= MLX4_IB_QP_NETIF;
1151 			else
1152 				goto err;
1153 		}
1154 
1155 		memcpy(&backup_cap, &init_attr->cap, sizeof(backup_cap));
1156 		err = set_kernel_sq_size(dev, &init_attr->cap,
1157 					 qp_type, qp, true);
1158 		if (err)
1159 			goto err;
1160 
1161 		if (qp_has_rq(init_attr)) {
1162 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
1163 			if (err)
1164 				goto err;
1165 
1166 			*qp->db.db = 0;
1167 		}
1168 
1169 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, qp->buf_size,
1170 				   &qp->buf)) {
1171 			memcpy(&init_attr->cap, &backup_cap,
1172 			       sizeof(backup_cap));
1173 			err = set_kernel_sq_size(dev, &init_attr->cap, qp_type,
1174 						 qp, false);
1175 			if (err)
1176 				goto err_db;
1177 
1178 			if (mlx4_buf_alloc(dev->dev, qp->buf_size,
1179 					   PAGE_SIZE * 2, &qp->buf)) {
1180 				err = -ENOMEM;
1181 				goto err_db;
1182 			}
1183 		}
1184 
1185 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
1186 				    &qp->mtt);
1187 		if (err)
1188 			goto err_buf;
1189 
1190 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
1191 		if (err)
1192 			goto err_mtt;
1193 
1194 		qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1195 					     sizeof(u64), GFP_KERNEL);
1196 		qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1197 					     sizeof(u64), GFP_KERNEL);
1198 		if (!qp->sq.wrid || !qp->rq.wrid) {
1199 			err = -ENOMEM;
1200 			goto err_wrid;
1201 		}
1202 		qp->mqp.usage = MLX4_RES_USAGE_DRIVER;
1203 	}
1204 
1205 	if (sqpn) {
1206 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1207 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
1208 			if (alloc_proxy_bufs(pd->device, qp)) {
1209 				err = -ENOMEM;
1210 				goto err_wrid;
1211 			}
1212 		}
1213 	} else if (src == MLX4_IB_RWQ_SRC) {
1214 		err = mlx4_ib_alloc_wqn(to_mucontext(pd->uobject->context), qp,
1215 					range_size, &qpn);
1216 		if (err)
1217 			goto err_wrid;
1218 	} else {
1219 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
1220 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
1221 		 * VLAN insertion. */
1222 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
1223 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
1224 						    (init_attr->cap.max_send_wr ?
1225 						     MLX4_RESERVE_ETH_BF_QP : 0) |
1226 						    (init_attr->cap.max_recv_wr ?
1227 						     MLX4_RESERVE_A0_QP : 0),
1228 						    qp->mqp.usage);
1229 		else
1230 			if (qp->flags & MLX4_IB_QP_NETIF)
1231 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
1232 			else
1233 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
1234 							    &qpn, 0, qp->mqp.usage);
1235 		if (err)
1236 			goto err_proxy;
1237 	}
1238 
1239 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1240 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1241 
1242 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
1243 	if (err)
1244 		goto err_qpn;
1245 
1246 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1247 		qp->mqp.qpn |= (1 << 23);
1248 
1249 	/*
1250 	 * Hardware wants QPN written in big-endian order (after
1251 	 * shifting) for send doorbell.  Precompute this value to save
1252 	 * a little bit when posting sends.
1253 	 */
1254 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1255 
1256 	qp->mqp.event = (src == MLX4_IB_QP_SRC) ? mlx4_ib_qp_event :
1257 						  mlx4_ib_wq_event;
1258 
1259 	if (!*caller_qp)
1260 		*caller_qp = qp;
1261 
1262 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1263 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
1264 			 to_mcq(init_attr->recv_cq));
1265 	/* Maintain device to QPs access, needed for further handling
1266 	 * via reset flow
1267 	 */
1268 	list_add_tail(&qp->qps_list, &dev->qp_list);
1269 	/* Maintain CQ to QPs access, needed for further handling
1270 	 * via reset flow
1271 	 */
1272 	mcq = to_mcq(init_attr->send_cq);
1273 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
1274 	mcq = to_mcq(init_attr->recv_cq);
1275 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
1276 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
1277 			   to_mcq(init_attr->recv_cq));
1278 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1279 	return 0;
1280 
1281 err_qpn:
1282 	if (!sqpn) {
1283 		if (qp->flags & MLX4_IB_QP_NETIF)
1284 			mlx4_ib_steer_qp_free(dev, qpn, 1);
1285 		else if (src == MLX4_IB_RWQ_SRC)
1286 			mlx4_ib_release_wqn(to_mucontext(pd->uobject->context),
1287 					    qp, 0);
1288 		else
1289 			mlx4_qp_release_range(dev->dev, qpn, 1);
1290 	}
1291 err_proxy:
1292 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1293 		free_proxy_bufs(pd->device, qp);
1294 err_wrid:
1295 	if (pd->uobject) {
1296 		if (qp_has_rq(init_attr))
1297 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
1298 	} else {
1299 		kvfree(qp->sq.wrid);
1300 		kvfree(qp->rq.wrid);
1301 	}
1302 
1303 err_mtt:
1304 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1305 
1306 err_buf:
1307 	if (pd->uobject)
1308 		ib_umem_release(qp->umem);
1309 	else
1310 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1311 
1312 err_db:
1313 	if (!pd->uobject && qp_has_rq(init_attr))
1314 		mlx4_db_free(dev->dev, &qp->db);
1315 
1316 err:
1317 	if (sqp)
1318 		kfree(sqp);
1319 	else if (!*caller_qp)
1320 		kfree(qp);
1321 	return err;
1322 }
1323 
1324 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
1325 {
1326 	switch (state) {
1327 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
1328 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
1329 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
1330 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
1331 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
1332 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
1333 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
1334 	default:		return -1;
1335 	}
1336 }
1337 
1338 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1339 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1340 {
1341 	if (send_cq == recv_cq) {
1342 		spin_lock(&send_cq->lock);
1343 		__acquire(&recv_cq->lock);
1344 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1345 		spin_lock(&send_cq->lock);
1346 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1347 	} else {
1348 		spin_lock(&recv_cq->lock);
1349 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1350 	}
1351 }
1352 
1353 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1354 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1355 {
1356 	if (send_cq == recv_cq) {
1357 		__release(&recv_cq->lock);
1358 		spin_unlock(&send_cq->lock);
1359 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1360 		spin_unlock(&recv_cq->lock);
1361 		spin_unlock(&send_cq->lock);
1362 	} else {
1363 		spin_unlock(&send_cq->lock);
1364 		spin_unlock(&recv_cq->lock);
1365 	}
1366 }
1367 
1368 static void del_gid_entries(struct mlx4_ib_qp *qp)
1369 {
1370 	struct mlx4_ib_gid_entry *ge, *tmp;
1371 
1372 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1373 		list_del(&ge->list);
1374 		kfree(ge);
1375 	}
1376 }
1377 
1378 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
1379 {
1380 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
1381 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
1382 	else
1383 		return to_mpd(qp->ibqp.pd);
1384 }
1385 
1386 static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src,
1387 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1388 {
1389 	switch (qp->ibqp.qp_type) {
1390 	case IB_QPT_XRC_TGT:
1391 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1392 		*recv_cq = *send_cq;
1393 		break;
1394 	case IB_QPT_XRC_INI:
1395 		*send_cq = to_mcq(qp->ibqp.send_cq);
1396 		*recv_cq = *send_cq;
1397 		break;
1398 	default:
1399 		*recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) :
1400 						     to_mcq(qp->ibwq.cq);
1401 		*send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) :
1402 						     *recv_cq;
1403 		break;
1404 	}
1405 }
1406 
1407 static void destroy_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1408 {
1409 	if (qp->state != IB_QPS_RESET) {
1410 		int i;
1411 
1412 		for (i = 0; i < (1 << qp->ibqp.rwq_ind_tbl->log_ind_tbl_size);
1413 		     i++) {
1414 			struct ib_wq *ibwq = qp->ibqp.rwq_ind_tbl->ind_tbl[i];
1415 			struct mlx4_ib_qp *wq =	to_mqp((struct ib_qp *)ibwq);
1416 
1417 			mutex_lock(&wq->mutex);
1418 
1419 			wq->rss_usecnt--;
1420 
1421 			mutex_unlock(&wq->mutex);
1422 		}
1423 
1424 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1425 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1426 			pr_warn("modify QP %06x to RESET failed.\n",
1427 				qp->mqp.qpn);
1428 	}
1429 
1430 	mlx4_qp_remove(dev->dev, &qp->mqp);
1431 	mlx4_qp_free(dev->dev, &qp->mqp);
1432 	mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1433 	del_gid_entries(qp);
1434 	kfree(qp->rss_ctx);
1435 }
1436 
1437 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1438 			      enum mlx4_ib_source_type src, int is_user)
1439 {
1440 	struct mlx4_ib_cq *send_cq, *recv_cq;
1441 	unsigned long flags;
1442 
1443 	if (qp->state != IB_QPS_RESET) {
1444 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1445 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1446 			pr_warn("modify QP %06x to RESET failed.\n",
1447 			       qp->mqp.qpn);
1448 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1449 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1450 			qp->pri.smac = 0;
1451 			qp->pri.smac_port = 0;
1452 		}
1453 		if (qp->alt.smac) {
1454 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1455 			qp->alt.smac = 0;
1456 		}
1457 		if (qp->pri.vid < 0x1000) {
1458 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1459 			qp->pri.vid = 0xFFFF;
1460 			qp->pri.candidate_vid = 0xFFFF;
1461 			qp->pri.update_vid = 0;
1462 		}
1463 		if (qp->alt.vid < 0x1000) {
1464 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1465 			qp->alt.vid = 0xFFFF;
1466 			qp->alt.candidate_vid = 0xFFFF;
1467 			qp->alt.update_vid = 0;
1468 		}
1469 	}
1470 
1471 	get_cqs(qp, src, &send_cq, &recv_cq);
1472 
1473 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1474 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1475 
1476 	/* del from lists under both locks above to protect reset flow paths */
1477 	list_del(&qp->qps_list);
1478 	list_del(&qp->cq_send_list);
1479 	list_del(&qp->cq_recv_list);
1480 	if (!is_user) {
1481 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1482 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1483 		if (send_cq != recv_cq)
1484 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1485 	}
1486 
1487 	mlx4_qp_remove(dev->dev, &qp->mqp);
1488 
1489 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1490 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1491 
1492 	mlx4_qp_free(dev->dev, &qp->mqp);
1493 
1494 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1495 		if (qp->flags & MLX4_IB_QP_NETIF)
1496 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1497 		else if (src == MLX4_IB_RWQ_SRC)
1498 			mlx4_ib_release_wqn(to_mucontext(
1499 					    qp->ibwq.uobject->context), qp, 1);
1500 		else
1501 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1502 	}
1503 
1504 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1505 
1506 	if (is_user) {
1507 		if (qp->rq.wqe_cnt) {
1508 			struct mlx4_ib_ucontext *mcontext = !src ?
1509 				to_mucontext(qp->ibqp.uobject->context) :
1510 				to_mucontext(qp->ibwq.uobject->context);
1511 			mlx4_ib_db_unmap_user(mcontext, &qp->db);
1512 		}
1513 		ib_umem_release(qp->umem);
1514 	} else {
1515 		kvfree(qp->sq.wrid);
1516 		kvfree(qp->rq.wrid);
1517 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1518 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1519 			free_proxy_bufs(&dev->ib_dev, qp);
1520 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1521 		if (qp->rq.wqe_cnt)
1522 			mlx4_db_free(dev->dev, &qp->db);
1523 	}
1524 
1525 	del_gid_entries(qp);
1526 }
1527 
1528 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1529 {
1530 	/* Native or PPF */
1531 	if (!mlx4_is_mfunc(dev->dev) ||
1532 	    (mlx4_is_master(dev->dev) &&
1533 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1534 		return  dev->dev->phys_caps.base_sqpn +
1535 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1536 			attr->port_num - 1;
1537 	}
1538 	/* PF or VF -- creating proxies */
1539 	if (attr->qp_type == IB_QPT_SMI)
1540 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp0_proxy;
1541 	else
1542 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp1_proxy;
1543 }
1544 
1545 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1546 					struct ib_qp_init_attr *init_attr,
1547 					struct ib_udata *udata)
1548 {
1549 	struct mlx4_ib_qp *qp = NULL;
1550 	int err;
1551 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1552 	u16 xrcdn = 0;
1553 
1554 	if (init_attr->rwq_ind_tbl)
1555 		return _mlx4_ib_create_qp_rss(pd, init_attr, udata);
1556 
1557 	/*
1558 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1559 	 * and only for kernel UD QPs.
1560 	 */
1561 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1562 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1563 					MLX4_IB_SRIOV_TUNNEL_QP |
1564 					MLX4_IB_SRIOV_SQP |
1565 					MLX4_IB_QP_NETIF |
1566 					MLX4_IB_QP_CREATE_ROCE_V2_GSI))
1567 		return ERR_PTR(-EINVAL);
1568 
1569 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1570 		if (init_attr->qp_type != IB_QPT_UD)
1571 			return ERR_PTR(-EINVAL);
1572 	}
1573 
1574 	if (init_attr->create_flags) {
1575 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1576 			return ERR_PTR(-EINVAL);
1577 
1578 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1579 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1580 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1581 		     init_attr->qp_type != IB_QPT_UD) ||
1582 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1583 		     init_attr->qp_type > IB_QPT_GSI) ||
1584 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1585 		     init_attr->qp_type != IB_QPT_GSI))
1586 			return ERR_PTR(-EINVAL);
1587 	}
1588 
1589 	switch (init_attr->qp_type) {
1590 	case IB_QPT_XRC_TGT:
1591 		pd = to_mxrcd(init_attr->xrcd)->pd;
1592 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1593 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1594 		/* fall through */
1595 	case IB_QPT_XRC_INI:
1596 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1597 			return ERR_PTR(-ENOSYS);
1598 		init_attr->recv_cq = init_attr->send_cq;
1599 		/* fall through */
1600 	case IB_QPT_RC:
1601 	case IB_QPT_UC:
1602 	case IB_QPT_RAW_PACKET:
1603 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1604 		if (!qp)
1605 			return ERR_PTR(-ENOMEM);
1606 		qp->pri.vid = 0xFFFF;
1607 		qp->alt.vid = 0xFFFF;
1608 		/* fall through */
1609 	case IB_QPT_UD:
1610 	{
1611 		err = create_qp_common(to_mdev(pd->device), pd,	MLX4_IB_QP_SRC,
1612 				       init_attr, udata, 0, &qp);
1613 		if (err) {
1614 			kfree(qp);
1615 			return ERR_PTR(err);
1616 		}
1617 
1618 		qp->ibqp.qp_num = qp->mqp.qpn;
1619 		qp->xrcdn = xrcdn;
1620 
1621 		break;
1622 	}
1623 	case IB_QPT_SMI:
1624 	case IB_QPT_GSI:
1625 	{
1626 		int sqpn;
1627 
1628 		/* Userspace is not allowed to create special QPs: */
1629 		if (udata)
1630 			return ERR_PTR(-EINVAL);
1631 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1632 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev,
1633 							1, 1, &sqpn, 0,
1634 							MLX4_RES_USAGE_DRIVER);
1635 
1636 			if (res)
1637 				return ERR_PTR(res);
1638 		} else {
1639 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1640 		}
1641 
1642 		err = create_qp_common(to_mdev(pd->device), pd, MLX4_IB_QP_SRC,
1643 				       init_attr, udata, sqpn, &qp);
1644 		if (err)
1645 			return ERR_PTR(err);
1646 
1647 		qp->port	= init_attr->port_num;
1648 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1649 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1650 		break;
1651 	}
1652 	default:
1653 		/* Don't support raw QPs */
1654 		return ERR_PTR(-EINVAL);
1655 	}
1656 
1657 	return &qp->ibqp;
1658 }
1659 
1660 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1661 				struct ib_qp_init_attr *init_attr,
1662 				struct ib_udata *udata) {
1663 	struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1664 	struct ib_qp *ibqp;
1665 	struct mlx4_ib_dev *dev = to_mdev(device);
1666 
1667 	ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1668 
1669 	if (!IS_ERR(ibqp) &&
1670 	    (init_attr->qp_type == IB_QPT_GSI) &&
1671 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1672 		struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1673 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1674 
1675 		if (is_eth &&
1676 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1677 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1678 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1679 
1680 			if (IS_ERR(sqp->roce_v2_gsi)) {
1681 				pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1682 				sqp->roce_v2_gsi = NULL;
1683 			} else {
1684 				sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1685 				sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1686 			}
1687 
1688 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1689 		}
1690 	}
1691 	return ibqp;
1692 }
1693 
1694 static int _mlx4_ib_destroy_qp(struct ib_qp *qp)
1695 {
1696 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1697 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1698 
1699 	if (is_qp0(dev, mqp))
1700 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1701 
1702 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1703 	    dev->qp1_proxy[mqp->port - 1] == mqp) {
1704 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1705 		dev->qp1_proxy[mqp->port - 1] = NULL;
1706 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1707 	}
1708 
1709 	if (mqp->counter_index)
1710 		mlx4_ib_free_qp_counter(dev, mqp);
1711 
1712 	if (qp->rwq_ind_tbl) {
1713 		destroy_qp_rss(dev, mqp);
1714 	} else {
1715 		struct mlx4_ib_pd *pd;
1716 
1717 		pd = get_pd(mqp);
1718 		destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, !!pd->ibpd.uobject);
1719 	}
1720 
1721 	if (is_sqp(dev, mqp))
1722 		kfree(to_msqp(mqp));
1723 	else
1724 		kfree(mqp);
1725 
1726 	return 0;
1727 }
1728 
1729 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1730 {
1731 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1732 
1733 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1734 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1735 
1736 		if (sqp->roce_v2_gsi)
1737 			ib_destroy_qp(sqp->roce_v2_gsi);
1738 	}
1739 
1740 	return _mlx4_ib_destroy_qp(qp);
1741 }
1742 
1743 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1744 {
1745 	switch (type) {
1746 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1747 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1748 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1749 	case MLX4_IB_QPT_XRC_INI:
1750 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1751 	case MLX4_IB_QPT_SMI:
1752 	case MLX4_IB_QPT_GSI:
1753 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1754 
1755 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1756 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1757 						MLX4_QP_ST_MLX : -1);
1758 	case MLX4_IB_QPT_PROXY_SMI:
1759 	case MLX4_IB_QPT_TUN_SMI:
1760 	case MLX4_IB_QPT_PROXY_GSI:
1761 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1762 						MLX4_QP_ST_UD : -1);
1763 	default:			return -1;
1764 	}
1765 }
1766 
1767 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1768 				   int attr_mask)
1769 {
1770 	u8 dest_rd_atomic;
1771 	u32 access_flags;
1772 	u32 hw_access_flags = 0;
1773 
1774 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1775 		dest_rd_atomic = attr->max_dest_rd_atomic;
1776 	else
1777 		dest_rd_atomic = qp->resp_depth;
1778 
1779 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1780 		access_flags = attr->qp_access_flags;
1781 	else
1782 		access_flags = qp->atomic_rd_en;
1783 
1784 	if (!dest_rd_atomic)
1785 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1786 
1787 	if (access_flags & IB_ACCESS_REMOTE_READ)
1788 		hw_access_flags |= MLX4_QP_BIT_RRE;
1789 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1790 		hw_access_flags |= MLX4_QP_BIT_RAE;
1791 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1792 		hw_access_flags |= MLX4_QP_BIT_RWE;
1793 
1794 	return cpu_to_be32(hw_access_flags);
1795 }
1796 
1797 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1798 			    int attr_mask)
1799 {
1800 	if (attr_mask & IB_QP_PKEY_INDEX)
1801 		sqp->pkey_index = attr->pkey_index;
1802 	if (attr_mask & IB_QP_QKEY)
1803 		sqp->qkey = attr->qkey;
1804 	if (attr_mask & IB_QP_SQ_PSN)
1805 		sqp->send_psn = attr->sq_psn;
1806 }
1807 
1808 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1809 {
1810 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1811 }
1812 
1813 static int _mlx4_set_path(struct mlx4_ib_dev *dev,
1814 			  const struct rdma_ah_attr *ah,
1815 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1816 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1817 {
1818 	int vidx;
1819 	int smac_index;
1820 	int err;
1821 
1822 	path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
1823 	path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
1824 	if (rdma_ah_get_static_rate(ah)) {
1825 		path->static_rate = rdma_ah_get_static_rate(ah) +
1826 				    MLX4_STAT_RATE_OFFSET;
1827 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1828 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1829 			--path->static_rate;
1830 	} else
1831 		path->static_rate = 0;
1832 
1833 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
1834 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
1835 		int real_sgid_index =
1836 			mlx4_ib_gid_index_to_real_index(dev, port,
1837 							grh->sgid_index);
1838 
1839 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1840 			pr_err("sgid_index (%u) too large. max is %d\n",
1841 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1842 			return -1;
1843 		}
1844 
1845 		path->grh_mylmc |= 1 << 7;
1846 		path->mgid_index = real_sgid_index;
1847 		path->hop_limit  = grh->hop_limit;
1848 		path->tclass_flowlabel =
1849 			cpu_to_be32((grh->traffic_class << 20) |
1850 				    (grh->flow_label));
1851 		memcpy(path->rgid, grh->dgid.raw, 16);
1852 	}
1853 
1854 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
1855 		if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
1856 			return -1;
1857 
1858 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1859 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
1860 
1861 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1862 		if (vlan_tag < 0x1000) {
1863 			if (smac_info->vid < 0x1000) {
1864 				/* both valid vlan ids */
1865 				if (smac_info->vid != vlan_tag) {
1866 					/* different VIDs.  unreg old and reg new */
1867 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1868 					if (err)
1869 						return err;
1870 					smac_info->candidate_vid = vlan_tag;
1871 					smac_info->candidate_vlan_index = vidx;
1872 					smac_info->candidate_vlan_port = port;
1873 					smac_info->update_vid = 1;
1874 					path->vlan_index = vidx;
1875 				} else {
1876 					path->vlan_index = smac_info->vlan_index;
1877 				}
1878 			} else {
1879 				/* no current vlan tag in qp */
1880 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1881 				if (err)
1882 					return err;
1883 				smac_info->candidate_vid = vlan_tag;
1884 				smac_info->candidate_vlan_index = vidx;
1885 				smac_info->candidate_vlan_port = port;
1886 				smac_info->update_vid = 1;
1887 				path->vlan_index = vidx;
1888 			}
1889 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1890 			path->fl = 1 << 6;
1891 		} else {
1892 			/* have current vlan tag. unregister it at modify-qp success */
1893 			if (smac_info->vid < 0x1000) {
1894 				smac_info->candidate_vid = 0xFFFF;
1895 				smac_info->update_vid = 1;
1896 			}
1897 		}
1898 
1899 		/* get smac_index for RoCE use.
1900 		 * If no smac was yet assigned, register one.
1901 		 * If one was already assigned, but the new mac differs,
1902 		 * unregister the old one and register the new one.
1903 		*/
1904 		if ((!smac_info->smac && !smac_info->smac_port) ||
1905 		    smac_info->smac != smac) {
1906 			/* register candidate now, unreg if needed, after success */
1907 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1908 			if (smac_index >= 0) {
1909 				smac_info->candidate_smac_index = smac_index;
1910 				smac_info->candidate_smac = smac;
1911 				smac_info->candidate_smac_port = port;
1912 			} else {
1913 				return -EINVAL;
1914 			}
1915 		} else {
1916 			smac_index = smac_info->smac_index;
1917 		}
1918 		memcpy(path->dmac, ah->roce.dmac, 6);
1919 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1920 		/* put MAC table smac index for IBoE */
1921 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1922 	} else {
1923 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1924 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
1925 	}
1926 
1927 	return 0;
1928 }
1929 
1930 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1931 			 enum ib_qp_attr_mask qp_attr_mask,
1932 			 struct mlx4_ib_qp *mqp,
1933 			 struct mlx4_qp_path *path, u8 port,
1934 			 u16 vlan_id, u8 *smac)
1935 {
1936 	return _mlx4_set_path(dev, &qp->ah_attr,
1937 			      mlx4_mac_to_u64(smac),
1938 			      vlan_id,
1939 			      path, &mqp->pri, port);
1940 }
1941 
1942 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1943 			     const struct ib_qp_attr *qp,
1944 			     enum ib_qp_attr_mask qp_attr_mask,
1945 			     struct mlx4_ib_qp *mqp,
1946 			     struct mlx4_qp_path *path, u8 port)
1947 {
1948 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1949 			      0,
1950 			      0xffff,
1951 			      path, &mqp->alt, port);
1952 }
1953 
1954 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1955 {
1956 	struct mlx4_ib_gid_entry *ge, *tmp;
1957 
1958 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1959 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1960 			ge->added = 1;
1961 			ge->port = qp->port;
1962 		}
1963 	}
1964 }
1965 
1966 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1967 				    struct mlx4_ib_qp *qp,
1968 				    struct mlx4_qp_context *context)
1969 {
1970 	u64 u64_mac;
1971 	int smac_index;
1972 
1973 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1974 
1975 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1976 	if (!qp->pri.smac && !qp->pri.smac_port) {
1977 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1978 		if (smac_index >= 0) {
1979 			qp->pri.candidate_smac_index = smac_index;
1980 			qp->pri.candidate_smac = u64_mac;
1981 			qp->pri.candidate_smac_port = qp->port;
1982 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1983 		} else {
1984 			return -ENOENT;
1985 		}
1986 	}
1987 	return 0;
1988 }
1989 
1990 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1991 {
1992 	struct counter_index *new_counter_index;
1993 	int err;
1994 	u32 tmp_idx;
1995 
1996 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1997 	    IB_LINK_LAYER_ETHERNET ||
1998 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1999 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
2000 		return 0;
2001 
2002 	err = mlx4_counter_alloc(dev->dev, &tmp_idx, MLX4_RES_USAGE_DRIVER);
2003 	if (err)
2004 		return err;
2005 
2006 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
2007 	if (!new_counter_index) {
2008 		mlx4_counter_free(dev->dev, tmp_idx);
2009 		return -ENOMEM;
2010 	}
2011 
2012 	new_counter_index->index = tmp_idx;
2013 	new_counter_index->allocated = 1;
2014 	qp->counter_index = new_counter_index;
2015 
2016 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
2017 	list_add_tail(&new_counter_index->list,
2018 		      &dev->counters_table[qp->port - 1].counters_list);
2019 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
2020 
2021 	return 0;
2022 }
2023 
2024 enum {
2025 	MLX4_QPC_ROCE_MODE_1 = 0,
2026 	MLX4_QPC_ROCE_MODE_2 = 2,
2027 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
2028 };
2029 
2030 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
2031 {
2032 	switch (gid_type) {
2033 	case IB_GID_TYPE_ROCE:
2034 		return MLX4_QPC_ROCE_MODE_1;
2035 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
2036 		return MLX4_QPC_ROCE_MODE_2;
2037 	default:
2038 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
2039 	}
2040 }
2041 
2042 /*
2043  * Go over all RSS QP's childes (WQs) and apply their HW state according to
2044  * their logic state if the RSS QP is the first RSS QP associated for the WQ.
2045  */
2046 static int bringup_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, u8 port_num)
2047 {
2048 	int err = 0;
2049 	int i;
2050 
2051 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2052 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2053 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2054 
2055 		mutex_lock(&wq->mutex);
2056 
2057 		/* Mlx4_ib restrictions:
2058 		 * WQ's is associated to a port according to the RSS QP it is
2059 		 * associates to.
2060 		 * In case the WQ is associated to a different port by another
2061 		 * RSS QP, return a failure.
2062 		 */
2063 		if ((wq->rss_usecnt > 0) && (wq->port != port_num)) {
2064 			err = -EINVAL;
2065 			mutex_unlock(&wq->mutex);
2066 			break;
2067 		}
2068 		wq->port = port_num;
2069 		if ((wq->rss_usecnt == 0) && (ibwq->state == IB_WQS_RDY)) {
2070 			err = _mlx4_ib_modify_wq(ibwq, IB_WQS_RDY);
2071 			if (err) {
2072 				mutex_unlock(&wq->mutex);
2073 				break;
2074 			}
2075 		}
2076 		wq->rss_usecnt++;
2077 
2078 		mutex_unlock(&wq->mutex);
2079 	}
2080 
2081 	if (i && err) {
2082 		int j;
2083 
2084 		for (j = (i - 1); j >= 0; j--) {
2085 			struct ib_wq *ibwq = ind_tbl->ind_tbl[j];
2086 			struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2087 
2088 			mutex_lock(&wq->mutex);
2089 
2090 			if ((wq->rss_usecnt == 1) &&
2091 			    (ibwq->state == IB_WQS_RDY))
2092 				if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET))
2093 					pr_warn("failed to reverse WQN=0x%06x\n",
2094 						ibwq->wq_num);
2095 			wq->rss_usecnt--;
2096 
2097 			mutex_unlock(&wq->mutex);
2098 		}
2099 	}
2100 
2101 	return err;
2102 }
2103 
2104 static void bring_down_rss_rwqs(struct ib_rwq_ind_table *ind_tbl)
2105 {
2106 	int i;
2107 
2108 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2109 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2110 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2111 
2112 		mutex_lock(&wq->mutex);
2113 
2114 		if ((wq->rss_usecnt == 1) && (ibwq->state == IB_WQS_RDY))
2115 			if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET))
2116 				pr_warn("failed to reverse WQN=%x\n",
2117 					ibwq->wq_num);
2118 		wq->rss_usecnt--;
2119 
2120 		mutex_unlock(&wq->mutex);
2121 	}
2122 }
2123 
2124 static void fill_qp_rss_context(struct mlx4_qp_context *context,
2125 				struct mlx4_ib_qp *qp)
2126 {
2127 	struct mlx4_rss_context *rss_context;
2128 
2129 	rss_context = (void *)context + offsetof(struct mlx4_qp_context,
2130 			pri_path) + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
2131 
2132 	rss_context->base_qpn = cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz);
2133 	rss_context->default_qpn =
2134 		cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz & 0xffffff);
2135 	if (qp->rss_ctx->flags & (MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6))
2136 		rss_context->base_qpn_udp = rss_context->default_qpn;
2137 	rss_context->flags = qp->rss_ctx->flags;
2138 	/* Currently support just toeplitz */
2139 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
2140 
2141 	memcpy(rss_context->rss_key, qp->rss_ctx->rss_key,
2142 	       MLX4_EN_RSS_KEY_SIZE);
2143 }
2144 
2145 static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type,
2146 			       const struct ib_qp_attr *attr, int attr_mask,
2147 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
2148 {
2149 	struct ib_uobject *ibuobject;
2150 	struct ib_srq  *ibsrq;
2151 	struct ib_rwq_ind_table *rwq_ind_tbl;
2152 	enum ib_qp_type qp_type;
2153 	struct mlx4_ib_dev *dev;
2154 	struct mlx4_ib_qp *qp;
2155 	struct mlx4_ib_pd *pd;
2156 	struct mlx4_ib_cq *send_cq, *recv_cq;
2157 	struct mlx4_qp_context *context;
2158 	enum mlx4_qp_optpar optpar = 0;
2159 	int sqd_event;
2160 	int steer_qp = 0;
2161 	int err = -EINVAL;
2162 	int counter_index;
2163 
2164 	if (src_type == MLX4_IB_RWQ_SRC) {
2165 		struct ib_wq *ibwq;
2166 
2167 		ibwq	    = (struct ib_wq *)src;
2168 		ibuobject   = ibwq->uobject;
2169 		ibsrq	    = NULL;
2170 		rwq_ind_tbl = NULL;
2171 		qp_type     = IB_QPT_RAW_PACKET;
2172 		qp	    = to_mqp((struct ib_qp *)ibwq);
2173 		dev	    = to_mdev(ibwq->device);
2174 		pd	    = to_mpd(ibwq->pd);
2175 	} else {
2176 		struct ib_qp *ibqp;
2177 
2178 		ibqp	    = (struct ib_qp *)src;
2179 		ibuobject   = ibqp->uobject;
2180 		ibsrq	    = ibqp->srq;
2181 		rwq_ind_tbl = ibqp->rwq_ind_tbl;
2182 		qp_type     = ibqp->qp_type;
2183 		qp	    = to_mqp(ibqp);
2184 		dev	    = to_mdev(ibqp->device);
2185 		pd	    = get_pd(qp);
2186 	}
2187 
2188 	/* APM is not supported under RoCE */
2189 	if (attr_mask & IB_QP_ALT_PATH &&
2190 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2191 	    IB_LINK_LAYER_ETHERNET)
2192 		return -ENOTSUPP;
2193 
2194 	context = kzalloc(sizeof *context, GFP_KERNEL);
2195 	if (!context)
2196 		return -ENOMEM;
2197 
2198 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
2199 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
2200 
2201 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
2202 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2203 	else {
2204 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
2205 		switch (attr->path_mig_state) {
2206 		case IB_MIG_MIGRATED:
2207 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2208 			break;
2209 		case IB_MIG_REARM:
2210 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
2211 			break;
2212 		case IB_MIG_ARMED:
2213 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
2214 			break;
2215 		}
2216 	}
2217 
2218 	if (qp->inl_recv_sz)
2219 		context->param3 |= cpu_to_be32(1 << 25);
2220 
2221 	if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI)
2222 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
2223 	else if (qp_type == IB_QPT_RAW_PACKET)
2224 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
2225 	else if (qp_type == IB_QPT_UD) {
2226 		if (qp->flags & MLX4_IB_QP_LSO)
2227 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
2228 					      ilog2(dev->dev->caps.max_gso_sz);
2229 		else
2230 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 13;
2231 	} else if (attr_mask & IB_QP_PATH_MTU) {
2232 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
2233 			pr_err("path MTU (%u) is invalid\n",
2234 			       attr->path_mtu);
2235 			goto out;
2236 		}
2237 		context->mtu_msgmax = (attr->path_mtu << 5) |
2238 			ilog2(dev->dev->caps.max_msg_sz);
2239 	}
2240 
2241 	if (!rwq_ind_tbl) { /* PRM RSS receive side should be left zeros */
2242 		if (qp->rq.wqe_cnt)
2243 			context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
2244 		context->rq_size_stride |= qp->rq.wqe_shift - 4;
2245 	}
2246 
2247 	if (qp->sq.wqe_cnt)
2248 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
2249 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
2250 
2251 	if (new_state == IB_QPS_RESET && qp->counter_index)
2252 		mlx4_ib_free_qp_counter(dev, qp);
2253 
2254 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2255 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
2256 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
2257 		if (qp_type == IB_QPT_RAW_PACKET)
2258 			context->param3 |= cpu_to_be32(1 << 30);
2259 	}
2260 
2261 	if (ibuobject)
2262 		context->usr_page = cpu_to_be32(
2263 			mlx4_to_hw_uar_index(dev->dev,
2264 					     to_mucontext(ibuobject->context)
2265 					     ->uar.index));
2266 	else
2267 		context->usr_page = cpu_to_be32(
2268 			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
2269 
2270 	if (attr_mask & IB_QP_DEST_QPN)
2271 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
2272 
2273 	if (attr_mask & IB_QP_PORT) {
2274 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
2275 		    !(attr_mask & IB_QP_AV)) {
2276 			mlx4_set_sched(&context->pri_path, attr->port_num);
2277 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
2278 		}
2279 	}
2280 
2281 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
2282 		err = create_qp_lb_counter(dev, qp);
2283 		if (err)
2284 			goto out;
2285 
2286 		counter_index =
2287 			dev->counters_table[qp->port - 1].default_counter;
2288 		if (qp->counter_index)
2289 			counter_index = qp->counter_index->index;
2290 
2291 		if (counter_index != -1) {
2292 			context->pri_path.counter_index = counter_index;
2293 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
2294 			if (qp->counter_index) {
2295 				context->pri_path.fl |=
2296 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
2297 				context->pri_path.vlan_control |=
2298 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
2299 			}
2300 		} else
2301 			context->pri_path.counter_index =
2302 				MLX4_SINK_COUNTER_INDEX(dev->dev);
2303 
2304 		if (qp->flags & MLX4_IB_QP_NETIF) {
2305 			mlx4_ib_steer_qp_reg(dev, qp, 1);
2306 			steer_qp = 1;
2307 		}
2308 
2309 		if (qp_type == IB_QPT_GSI) {
2310 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
2311 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
2312 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
2313 
2314 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2315 		}
2316 	}
2317 
2318 	if (attr_mask & IB_QP_PKEY_INDEX) {
2319 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2320 			context->pri_path.disable_pkey_check = 0x40;
2321 		context->pri_path.pkey_index = attr->pkey_index;
2322 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
2323 	}
2324 
2325 	if (attr_mask & IB_QP_AV) {
2326 		u8 port_num = mlx4_is_bonded(dev->dev) ? 1 :
2327 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2328 		union ib_gid gid;
2329 		struct ib_gid_attr gid_attr = {.gid_type = IB_GID_TYPE_IB};
2330 		u16 vlan = 0xffff;
2331 		u8 smac[ETH_ALEN];
2332 		int status = 0;
2333 		int is_eth =
2334 			rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
2335 			rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
2336 
2337 		if (is_eth) {
2338 			int index =
2339 				rdma_ah_read_grh(&attr->ah_attr)->sgid_index;
2340 
2341 			status = ib_get_cached_gid(&dev->ib_dev, port_num,
2342 						   index, &gid, &gid_attr);
2343 			if (!status && !memcmp(&gid, &zgid, sizeof(gid)))
2344 				status = -ENOENT;
2345 			if (!status && gid_attr.ndev) {
2346 				vlan = rdma_vlan_dev_vlan_id(gid_attr.ndev);
2347 				memcpy(smac, gid_attr.ndev->dev_addr, ETH_ALEN);
2348 				dev_put(gid_attr.ndev);
2349 			}
2350 		}
2351 		if (status)
2352 			goto out;
2353 
2354 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
2355 				  port_num, vlan, smac))
2356 			goto out;
2357 
2358 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
2359 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
2360 
2361 		if (is_eth &&
2362 		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
2363 			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr.gid_type);
2364 
2365 			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
2366 				err = -EINVAL;
2367 				goto out;
2368 			}
2369 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2370 		}
2371 
2372 	}
2373 
2374 	if (attr_mask & IB_QP_TIMEOUT) {
2375 		context->pri_path.ackto |= attr->timeout << 3;
2376 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
2377 	}
2378 
2379 	if (attr_mask & IB_QP_ALT_PATH) {
2380 		if (attr->alt_port_num == 0 ||
2381 		    attr->alt_port_num > dev->dev->caps.num_ports)
2382 			goto out;
2383 
2384 		if (attr->alt_pkey_index >=
2385 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
2386 			goto out;
2387 
2388 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
2389 				      &context->alt_path,
2390 				      attr->alt_port_num))
2391 			goto out;
2392 
2393 		context->alt_path.pkey_index = attr->alt_pkey_index;
2394 		context->alt_path.ackto = attr->alt_timeout << 3;
2395 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
2396 	}
2397 
2398 	context->pd = cpu_to_be32(pd->pdn);
2399 
2400 	if (!rwq_ind_tbl) {
2401 		context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
2402 		get_cqs(qp, src_type, &send_cq, &recv_cq);
2403 	} else { /* Set dummy CQs to be compatible with HV and PRM */
2404 		send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq);
2405 		recv_cq = send_cq;
2406 	}
2407 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
2408 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
2409 
2410 	/* Set "fast registration enabled" for all kernel QPs */
2411 	if (!ibuobject)
2412 		context->params1 |= cpu_to_be32(1 << 11);
2413 
2414 	if (attr_mask & IB_QP_RNR_RETRY) {
2415 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2416 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
2417 	}
2418 
2419 	if (attr_mask & IB_QP_RETRY_CNT) {
2420 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2421 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
2422 	}
2423 
2424 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2425 		if (attr->max_rd_atomic)
2426 			context->params1 |=
2427 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2428 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
2429 	}
2430 
2431 	if (attr_mask & IB_QP_SQ_PSN)
2432 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2433 
2434 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2435 		if (attr->max_dest_rd_atomic)
2436 			context->params2 |=
2437 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2438 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
2439 	}
2440 
2441 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
2442 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
2443 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
2444 	}
2445 
2446 	if (ibsrq)
2447 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
2448 
2449 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
2450 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2451 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
2452 	}
2453 	if (attr_mask & IB_QP_RQ_PSN)
2454 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2455 
2456 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
2457 	if (attr_mask & IB_QP_QKEY) {
2458 		if (qp->mlx4_ib_qp_type &
2459 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
2460 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2461 		else {
2462 			if (mlx4_is_mfunc(dev->dev) &&
2463 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
2464 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
2465 			    MLX4_RESERVED_QKEY_BASE) {
2466 				pr_err("Cannot use reserved QKEY"
2467 				       " 0x%x (range 0xffff0000..0xffffffff"
2468 				       " is reserved)\n", attr->qkey);
2469 				err = -EINVAL;
2470 				goto out;
2471 			}
2472 			context->qkey = cpu_to_be32(attr->qkey);
2473 		}
2474 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
2475 	}
2476 
2477 	if (ibsrq)
2478 		context->srqn = cpu_to_be32(1 << 24 |
2479 					    to_msrq(ibsrq)->msrq.srqn);
2480 
2481 	if (qp->rq.wqe_cnt &&
2482 	    cur_state == IB_QPS_RESET &&
2483 	    new_state == IB_QPS_INIT)
2484 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2485 
2486 	if (cur_state == IB_QPS_INIT &&
2487 	    new_state == IB_QPS_RTR  &&
2488 	    (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI ||
2489 	     qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) {
2490 		context->pri_path.sched_queue = (qp->port - 1) << 6;
2491 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2492 		    qp->mlx4_ib_qp_type &
2493 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
2494 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
2495 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
2496 				context->pri_path.fl = 0x80;
2497 		} else {
2498 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2499 				context->pri_path.fl = 0x80;
2500 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
2501 		}
2502 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2503 		    IB_LINK_LAYER_ETHERNET) {
2504 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
2505 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
2506 				context->pri_path.feup = 1 << 7; /* don't fsm */
2507 			/* handle smac_index */
2508 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
2509 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
2510 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
2511 				err = handle_eth_ud_smac_index(dev, qp, context);
2512 				if (err) {
2513 					err = -EINVAL;
2514 					goto out;
2515 				}
2516 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
2517 					dev->qp1_proxy[qp->port - 1] = qp;
2518 			}
2519 		}
2520 	}
2521 
2522 	if (qp_type == IB_QPT_RAW_PACKET) {
2523 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
2524 					MLX4_IB_LINK_TYPE_ETH;
2525 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
2526 			/* set QP to receive both tunneled & non-tunneled packets */
2527 			if (!rwq_ind_tbl)
2528 				context->srqn = cpu_to_be32(7 << 28);
2529 		}
2530 	}
2531 
2532 	if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
2533 		int is_eth = rdma_port_get_link_layer(
2534 				&dev->ib_dev, qp->port) ==
2535 				IB_LINK_LAYER_ETHERNET;
2536 		if (is_eth) {
2537 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
2538 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
2539 		}
2540 	}
2541 
2542 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2543 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2544 		sqd_event = 1;
2545 	else
2546 		sqd_event = 0;
2547 
2548 	if (!ibuobject &&
2549 	    cur_state == IB_QPS_RESET &&
2550 	    new_state == IB_QPS_INIT)
2551 		context->rlkey_roce_mode |= (1 << 4);
2552 
2553 	/*
2554 	 * Before passing a kernel QP to the HW, make sure that the
2555 	 * ownership bits of the send queue are set and the SQ
2556 	 * headroom is stamped so that the hardware doesn't start
2557 	 * processing stale work requests.
2558 	 */
2559 	if (!ibuobject &&
2560 	    cur_state == IB_QPS_RESET &&
2561 	    new_state == IB_QPS_INIT) {
2562 		struct mlx4_wqe_ctrl_seg *ctrl;
2563 		int i;
2564 
2565 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
2566 			ctrl = get_send_wqe(qp, i);
2567 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
2568 			if (qp->sq_max_wqes_per_wr == 1)
2569 				ctrl->qpn_vlan.fence_size =
2570 						1 << (qp->sq.wqe_shift - 4);
2571 
2572 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
2573 		}
2574 	}
2575 
2576 	if (rwq_ind_tbl	&&
2577 	    cur_state == IB_QPS_RESET &&
2578 	    new_state == IB_QPS_INIT) {
2579 		fill_qp_rss_context(context, qp);
2580 		context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET);
2581 	}
2582 
2583 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2584 			     to_mlx4_state(new_state), context, optpar,
2585 			     sqd_event, &qp->mqp);
2586 	if (err)
2587 		goto out;
2588 
2589 	qp->state = new_state;
2590 
2591 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2592 		qp->atomic_rd_en = attr->qp_access_flags;
2593 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2594 		qp->resp_depth = attr->max_dest_rd_atomic;
2595 	if (attr_mask & IB_QP_PORT) {
2596 		qp->port = attr->port_num;
2597 		update_mcg_macs(dev, qp);
2598 	}
2599 	if (attr_mask & IB_QP_ALT_PATH)
2600 		qp->alt_port = attr->alt_port_num;
2601 
2602 	if (is_sqp(dev, qp))
2603 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2604 
2605 	/*
2606 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
2607 	 * QP0 to RESET or ERROR, bring the link back down.
2608 	 */
2609 	if (is_qp0(dev, qp)) {
2610 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2611 			if (mlx4_INIT_PORT(dev->dev, qp->port))
2612 				pr_warn("INIT_PORT failed for port %d\n",
2613 				       qp->port);
2614 
2615 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2616 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2617 			mlx4_CLOSE_PORT(dev->dev, qp->port);
2618 	}
2619 
2620 	/*
2621 	 * If we moved a kernel QP to RESET, clean up all old CQ
2622 	 * entries and reinitialize the QP.
2623 	 */
2624 	if (new_state == IB_QPS_RESET) {
2625 		if (!ibuobject) {
2626 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2627 					 ibsrq ? to_msrq(ibsrq) : NULL);
2628 			if (send_cq != recv_cq)
2629 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2630 
2631 			qp->rq.head = 0;
2632 			qp->rq.tail = 0;
2633 			qp->sq.head = 0;
2634 			qp->sq.tail = 0;
2635 			qp->sq_next_wqe = 0;
2636 			if (qp->rq.wqe_cnt)
2637 				*qp->db.db  = 0;
2638 
2639 			if (qp->flags & MLX4_IB_QP_NETIF)
2640 				mlx4_ib_steer_qp_reg(dev, qp, 0);
2641 		}
2642 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2643 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2644 			qp->pri.smac = 0;
2645 			qp->pri.smac_port = 0;
2646 		}
2647 		if (qp->alt.smac) {
2648 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2649 			qp->alt.smac = 0;
2650 		}
2651 		if (qp->pri.vid < 0x1000) {
2652 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2653 			qp->pri.vid = 0xFFFF;
2654 			qp->pri.candidate_vid = 0xFFFF;
2655 			qp->pri.update_vid = 0;
2656 		}
2657 
2658 		if (qp->alt.vid < 0x1000) {
2659 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2660 			qp->alt.vid = 0xFFFF;
2661 			qp->alt.candidate_vid = 0xFFFF;
2662 			qp->alt.update_vid = 0;
2663 		}
2664 	}
2665 out:
2666 	if (err && qp->counter_index)
2667 		mlx4_ib_free_qp_counter(dev, qp);
2668 	if (err && steer_qp)
2669 		mlx4_ib_steer_qp_reg(dev, qp, 0);
2670 	kfree(context);
2671 	if (qp->pri.candidate_smac ||
2672 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2673 		if (err) {
2674 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2675 		} else {
2676 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2677 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2678 			qp->pri.smac = qp->pri.candidate_smac;
2679 			qp->pri.smac_index = qp->pri.candidate_smac_index;
2680 			qp->pri.smac_port = qp->pri.candidate_smac_port;
2681 		}
2682 		qp->pri.candidate_smac = 0;
2683 		qp->pri.candidate_smac_index = 0;
2684 		qp->pri.candidate_smac_port = 0;
2685 	}
2686 	if (qp->alt.candidate_smac) {
2687 		if (err) {
2688 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2689 		} else {
2690 			if (qp->alt.smac)
2691 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2692 			qp->alt.smac = qp->alt.candidate_smac;
2693 			qp->alt.smac_index = qp->alt.candidate_smac_index;
2694 			qp->alt.smac_port = qp->alt.candidate_smac_port;
2695 		}
2696 		qp->alt.candidate_smac = 0;
2697 		qp->alt.candidate_smac_index = 0;
2698 		qp->alt.candidate_smac_port = 0;
2699 	}
2700 
2701 	if (qp->pri.update_vid) {
2702 		if (err) {
2703 			if (qp->pri.candidate_vid < 0x1000)
2704 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2705 						     qp->pri.candidate_vid);
2706 		} else {
2707 			if (qp->pri.vid < 0x1000)
2708 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2709 						     qp->pri.vid);
2710 			qp->pri.vid = qp->pri.candidate_vid;
2711 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2712 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
2713 		}
2714 		qp->pri.candidate_vid = 0xFFFF;
2715 		qp->pri.update_vid = 0;
2716 	}
2717 
2718 	if (qp->alt.update_vid) {
2719 		if (err) {
2720 			if (qp->alt.candidate_vid < 0x1000)
2721 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2722 						     qp->alt.candidate_vid);
2723 		} else {
2724 			if (qp->alt.vid < 0x1000)
2725 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2726 						     qp->alt.vid);
2727 			qp->alt.vid = qp->alt.candidate_vid;
2728 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2729 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2730 		}
2731 		qp->alt.candidate_vid = 0xFFFF;
2732 		qp->alt.update_vid = 0;
2733 	}
2734 
2735 	return err;
2736 }
2737 
2738 enum {
2739 	MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK = (IB_QP_STATE	|
2740 					      IB_QP_PORT),
2741 };
2742 
2743 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2744 			      int attr_mask, struct ib_udata *udata)
2745 {
2746 	enum rdma_link_layer ll = IB_LINK_LAYER_UNSPECIFIED;
2747 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2748 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2749 	enum ib_qp_state cur_state, new_state;
2750 	int err = -EINVAL;
2751 	mutex_lock(&qp->mutex);
2752 
2753 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2754 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2755 
2756 	if (cur_state != new_state || cur_state != IB_QPS_RESET) {
2757 		int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2758 		ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2759 	}
2760 
2761 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2762 				attr_mask, ll)) {
2763 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2764 			 "for transition %d to %d. qp_type %d,"
2765 			 " attr_mask 0x%x\n",
2766 			 ibqp->qp_num, cur_state, new_state,
2767 			 ibqp->qp_type, attr_mask);
2768 		goto out;
2769 	}
2770 
2771 	if (ibqp->rwq_ind_tbl) {
2772 		if (!(((cur_state == IB_QPS_RESET) &&
2773 		       (new_state == IB_QPS_INIT)) ||
2774 		      ((cur_state == IB_QPS_INIT)  &&
2775 		       (new_state == IB_QPS_RTR)))) {
2776 			pr_debug("qpn 0x%x: RSS QP unsupported transition %d to %d\n",
2777 				 ibqp->qp_num, cur_state, new_state);
2778 
2779 			err = -EOPNOTSUPP;
2780 			goto out;
2781 		}
2782 
2783 		if (attr_mask & ~MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK) {
2784 			pr_debug("qpn 0x%x: RSS QP unsupported attribute mask 0x%x for transition %d to %d\n",
2785 				 ibqp->qp_num, attr_mask, cur_state, new_state);
2786 
2787 			err = -EOPNOTSUPP;
2788 			goto out;
2789 		}
2790 	}
2791 
2792 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2793 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2794 			if ((ibqp->qp_type == IB_QPT_RC) ||
2795 			    (ibqp->qp_type == IB_QPT_UD) ||
2796 			    (ibqp->qp_type == IB_QPT_UC) ||
2797 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2798 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2799 				attr->port_num = mlx4_ib_bond_next_port(dev);
2800 			}
2801 		} else {
2802 			/* no sense in changing port_num
2803 			 * when ports are bonded */
2804 			attr_mask &= ~IB_QP_PORT;
2805 		}
2806 	}
2807 
2808 	if ((attr_mask & IB_QP_PORT) &&
2809 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2810 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2811 			 "for transition %d to %d. qp_type %d\n",
2812 			 ibqp->qp_num, attr->port_num, cur_state,
2813 			 new_state, ibqp->qp_type);
2814 		goto out;
2815 	}
2816 
2817 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2818 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2819 	     IB_LINK_LAYER_ETHERNET))
2820 		goto out;
2821 
2822 	if (attr_mask & IB_QP_PKEY_INDEX) {
2823 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2824 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2825 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2826 				 "for transition %d to %d. qp_type %d\n",
2827 				 ibqp->qp_num, attr->pkey_index, cur_state,
2828 				 new_state, ibqp->qp_type);
2829 			goto out;
2830 		}
2831 	}
2832 
2833 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2834 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2835 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2836 			 "Transition %d to %d. qp_type %d\n",
2837 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2838 			 new_state, ibqp->qp_type);
2839 		goto out;
2840 	}
2841 
2842 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2843 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2844 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2845 			 "Transition %d to %d. qp_type %d\n",
2846 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2847 			 new_state, ibqp->qp_type);
2848 		goto out;
2849 	}
2850 
2851 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2852 		err = 0;
2853 		goto out;
2854 	}
2855 
2856 	if (ibqp->rwq_ind_tbl && (new_state == IB_QPS_INIT)) {
2857 		err = bringup_rss_rwqs(ibqp->rwq_ind_tbl, attr->port_num);
2858 		if (err)
2859 			goto out;
2860 	}
2861 
2862 	err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask,
2863 				  cur_state, new_state);
2864 
2865 	if (ibqp->rwq_ind_tbl && err)
2866 		bring_down_rss_rwqs(ibqp->rwq_ind_tbl);
2867 
2868 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2869 		attr->port_num = 1;
2870 
2871 out:
2872 	mutex_unlock(&qp->mutex);
2873 	return err;
2874 }
2875 
2876 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2877 		      int attr_mask, struct ib_udata *udata)
2878 {
2879 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2880 	int ret;
2881 
2882 	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2883 
2884 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2885 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
2886 		int err = 0;
2887 
2888 		if (sqp->roce_v2_gsi)
2889 			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2890 		if (err)
2891 			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2892 			       err);
2893 	}
2894 	return ret;
2895 }
2896 
2897 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2898 {
2899 	int i;
2900 	for (i = 0; i < dev->caps.num_ports; i++) {
2901 		if (qpn == dev->caps.spec_qps[i].qp0_proxy ||
2902 		    qpn == dev->caps.spec_qps[i].qp0_tunnel) {
2903 			*qkey = dev->caps.spec_qps[i].qp0_qkey;
2904 			return 0;
2905 		}
2906 	}
2907 	return -EINVAL;
2908 }
2909 
2910 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2911 				  struct ib_ud_wr *wr,
2912 				  void *wqe, unsigned *mlx_seg_len)
2913 {
2914 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2915 	struct ib_device *ib_dev = &mdev->ib_dev;
2916 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2917 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2918 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2919 	u16 pkey;
2920 	u32 qkey;
2921 	int send_size;
2922 	int header_size;
2923 	int spc;
2924 	int i;
2925 
2926 	if (wr->wr.opcode != IB_WR_SEND)
2927 		return -EINVAL;
2928 
2929 	send_size = 0;
2930 
2931 	for (i = 0; i < wr->wr.num_sge; ++i)
2932 		send_size += wr->wr.sg_list[i].length;
2933 
2934 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2935 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2936 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2937 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2938 
2939 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2940 
2941 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2942 		sqp->ud_header.lrh.service_level =
2943 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2944 		sqp->ud_header.lrh.destination_lid =
2945 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2946 		sqp->ud_header.lrh.source_lid =
2947 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2948 	}
2949 
2950 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2951 
2952 	/* force loopback */
2953 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2954 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2955 
2956 	sqp->ud_header.lrh.virtual_lane    = 0;
2957 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2958 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2959 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2960 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2961 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2962 	else
2963 		sqp->ud_header.bth.destination_qpn =
2964 			cpu_to_be32(mdev->dev->caps.spec_qps[sqp->qp.port - 1].qp0_tunnel);
2965 
2966 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2967 	if (mlx4_is_master(mdev->dev)) {
2968 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2969 			return -EINVAL;
2970 	} else {
2971 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2972 			return -EINVAL;
2973 	}
2974 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2975 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2976 
2977 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2978 	sqp->ud_header.immediate_present = 0;
2979 
2980 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2981 
2982 	/*
2983 	 * Inline data segments may not cross a 64 byte boundary.  If
2984 	 * our UD header is bigger than the space available up to the
2985 	 * next 64 byte boundary in the WQE, use two inline data
2986 	 * segments to hold the UD header.
2987 	 */
2988 	spc = MLX4_INLINE_ALIGN -
2989 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2990 	if (header_size <= spc) {
2991 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2992 		memcpy(inl + 1, sqp->header_buf, header_size);
2993 		i = 1;
2994 	} else {
2995 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2996 		memcpy(inl + 1, sqp->header_buf, spc);
2997 
2998 		inl = (void *) (inl + 1) + spc;
2999 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3000 		/*
3001 		 * Need a barrier here to make sure all the data is
3002 		 * visible before the byte_count field is set.
3003 		 * Otherwise the HCA prefetcher could grab the 64-byte
3004 		 * chunk with this inline segment and get a valid (!=
3005 		 * 0xffffffff) byte count but stale data, and end up
3006 		 * generating a packet with bad headers.
3007 		 *
3008 		 * The first inline segment's byte_count field doesn't
3009 		 * need a barrier, because it comes after a
3010 		 * control/MLX segment and therefore is at an offset
3011 		 * of 16 mod 64.
3012 		 */
3013 		wmb();
3014 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3015 		i = 2;
3016 	}
3017 
3018 	*mlx_seg_len =
3019 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3020 	return 0;
3021 }
3022 
3023 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
3024 {
3025 	union sl2vl_tbl_to_u64 tmp_vltab;
3026 	u8 vl;
3027 
3028 	if (sl > 15)
3029 		return 0xf;
3030 	tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
3031 	vl = tmp_vltab.sl8[sl >> 1];
3032 	if (sl & 1)
3033 		vl &= 0x0f;
3034 	else
3035 		vl >>= 4;
3036 	return vl;
3037 }
3038 
3039 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
3040 				int index, union ib_gid *gid,
3041 				enum ib_gid_type *gid_type)
3042 {
3043 	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
3044 	struct mlx4_port_gid_table *port_gid_table;
3045 	unsigned long flags;
3046 
3047 	port_gid_table = &iboe->gids[port_num - 1];
3048 	spin_lock_irqsave(&iboe->lock, flags);
3049 	memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
3050 	*gid_type = port_gid_table->gids[index].gid_type;
3051 	spin_unlock_irqrestore(&iboe->lock, flags);
3052 	if (!memcmp(gid, &zgid, sizeof(*gid)))
3053 		return -ENOENT;
3054 
3055 	return 0;
3056 }
3057 
3058 #define MLX4_ROCEV2_QP1_SPORT 0xC000
3059 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_ud_wr *wr,
3060 			    void *wqe, unsigned *mlx_seg_len)
3061 {
3062 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
3063 	struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
3064 	struct mlx4_wqe_mlx_seg *mlx = wqe;
3065 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
3066 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
3067 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3068 	union ib_gid sgid;
3069 	u16 pkey;
3070 	int send_size;
3071 	int header_size;
3072 	int spc;
3073 	int i;
3074 	int err = 0;
3075 	u16 vlan = 0xffff;
3076 	bool is_eth;
3077 	bool is_vlan = false;
3078 	bool is_grh;
3079 	bool is_udp = false;
3080 	int ip_version = 0;
3081 
3082 	send_size = 0;
3083 	for (i = 0; i < wr->wr.num_sge; ++i)
3084 		send_size += wr->wr.sg_list[i].length;
3085 
3086 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
3087 	is_grh = mlx4_ib_ah_grh_present(ah);
3088 	if (is_eth) {
3089 		enum ib_gid_type gid_type;
3090 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3091 			/* When multi-function is enabled, the ib_core gid
3092 			 * indexes don't necessarily match the hw ones, so
3093 			 * we must use our own cache */
3094 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
3095 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
3096 							   ah->av.ib.gid_index, &sgid.raw[0]);
3097 			if (err)
3098 				return err;
3099 		} else  {
3100 			err = fill_gid_by_hw_index(ibdev, sqp->qp.port,
3101 					    ah->av.ib.gid_index,
3102 					    &sgid, &gid_type);
3103 			if (!err) {
3104 				is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
3105 				if (is_udp) {
3106 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
3107 						ip_version = 4;
3108 					else
3109 						ip_version = 6;
3110 					is_grh = false;
3111 				}
3112 			} else {
3113 				return err;
3114 			}
3115 		}
3116 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
3117 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
3118 			is_vlan = 1;
3119 		}
3120 	}
3121 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
3122 			  ip_version, is_udp, 0, &sqp->ud_header);
3123 	if (err)
3124 		return err;
3125 
3126 	if (!is_eth) {
3127 		sqp->ud_header.lrh.service_level =
3128 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
3129 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
3130 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
3131 	}
3132 
3133 	if (is_grh || (ip_version == 6)) {
3134 		sqp->ud_header.grh.traffic_class =
3135 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3136 		sqp->ud_header.grh.flow_label    =
3137 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
3138 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
3139 		if (is_eth) {
3140 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
3141 		} else {
3142 			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3143 				/* When multi-function is enabled, the ib_core gid
3144 				 * indexes don't necessarily match the hw ones, so
3145 				 * we must use our own cache
3146 				 */
3147 				sqp->ud_header.grh.source_gid.global.subnet_prefix =
3148 					cpu_to_be64(atomic64_read(&(to_mdev(ib_dev)->sriov.
3149 								    demux[sqp->qp.port - 1].
3150 								    subnet_prefix)));
3151 				sqp->ud_header.grh.source_gid.global.interface_id =
3152 					to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
3153 						       guid_cache[ah->av.ib.gid_index];
3154 			} else {
3155 				ib_get_cached_gid(ib_dev,
3156 						  be32_to_cpu(ah->av.ib.port_pd) >> 24,
3157 						  ah->av.ib.gid_index,
3158 						  &sqp->ud_header.grh.source_gid, NULL);
3159 			}
3160 		}
3161 		memcpy(sqp->ud_header.grh.destination_gid.raw,
3162 		       ah->av.ib.dgid, 16);
3163 	}
3164 
3165 	if (ip_version == 4) {
3166 		sqp->ud_header.ip4.tos =
3167 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3168 		sqp->ud_header.ip4.id = 0;
3169 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
3170 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
3171 
3172 		memcpy(&sqp->ud_header.ip4.saddr,
3173 		       sgid.raw + 12, 4);
3174 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
3175 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
3176 	}
3177 
3178 	if (is_udp) {
3179 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
3180 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
3181 		sqp->ud_header.udp.csum = 0;
3182 	}
3183 
3184 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
3185 
3186 	if (!is_eth) {
3187 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
3188 					  (sqp->ud_header.lrh.destination_lid ==
3189 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
3190 					  (sqp->ud_header.lrh.service_level << 8));
3191 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
3192 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
3193 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
3194 	}
3195 
3196 	switch (wr->wr.opcode) {
3197 	case IB_WR_SEND:
3198 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
3199 		sqp->ud_header.immediate_present = 0;
3200 		break;
3201 	case IB_WR_SEND_WITH_IMM:
3202 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
3203 		sqp->ud_header.immediate_present = 1;
3204 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
3205 		break;
3206 	default:
3207 		return -EINVAL;
3208 	}
3209 
3210 	if (is_eth) {
3211 		struct in6_addr in6;
3212 		u16 ether_type;
3213 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
3214 
3215 		ether_type = (!is_udp) ? ETH_P_IBOE:
3216 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
3217 
3218 		mlx->sched_prio = cpu_to_be16(pcp);
3219 
3220 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
3221 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
3222 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
3223 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
3224 		memcpy(&in6, sgid.raw, sizeof(in6));
3225 
3226 
3227 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
3228 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
3229 		if (!is_vlan) {
3230 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
3231 		} else {
3232 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
3233 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
3234 		}
3235 	} else {
3236 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 :
3237 							sl_to_vl(to_mdev(ib_dev),
3238 								 sqp->ud_header.lrh.service_level,
3239 								 sqp->qp.port);
3240 		if (sqp->qp.ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
3241 			return -EINVAL;
3242 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
3243 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
3244 	}
3245 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
3246 	if (!sqp->qp.ibqp.qp_num)
3247 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
3248 	else
3249 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
3250 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
3251 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
3252 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
3253 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
3254 					       sqp->qkey : wr->remote_qkey);
3255 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
3256 
3257 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3258 
3259 	if (0) {
3260 		pr_err("built UD header of size %d:\n", header_size);
3261 		for (i = 0; i < header_size / 4; ++i) {
3262 			if (i % 8 == 0)
3263 				pr_err("  [%02x] ", i * 4);
3264 			pr_cont(" %08x",
3265 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
3266 			if ((i + 1) % 8 == 0)
3267 				pr_cont("\n");
3268 		}
3269 		pr_err("\n");
3270 	}
3271 
3272 	/*
3273 	 * Inline data segments may not cross a 64 byte boundary.  If
3274 	 * our UD header is bigger than the space available up to the
3275 	 * next 64 byte boundary in the WQE, use two inline data
3276 	 * segments to hold the UD header.
3277 	 */
3278 	spc = MLX4_INLINE_ALIGN -
3279 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3280 	if (header_size <= spc) {
3281 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3282 		memcpy(inl + 1, sqp->header_buf, header_size);
3283 		i = 1;
3284 	} else {
3285 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3286 		memcpy(inl + 1, sqp->header_buf, spc);
3287 
3288 		inl = (void *) (inl + 1) + spc;
3289 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3290 		/*
3291 		 * Need a barrier here to make sure all the data is
3292 		 * visible before the byte_count field is set.
3293 		 * Otherwise the HCA prefetcher could grab the 64-byte
3294 		 * chunk with this inline segment and get a valid (!=
3295 		 * 0xffffffff) byte count but stale data, and end up
3296 		 * generating a packet with bad headers.
3297 		 *
3298 		 * The first inline segment's byte_count field doesn't
3299 		 * need a barrier, because it comes after a
3300 		 * control/MLX segment and therefore is at an offset
3301 		 * of 16 mod 64.
3302 		 */
3303 		wmb();
3304 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3305 		i = 2;
3306 	}
3307 
3308 	*mlx_seg_len =
3309 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3310 	return 0;
3311 }
3312 
3313 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3314 {
3315 	unsigned cur;
3316 	struct mlx4_ib_cq *cq;
3317 
3318 	cur = wq->head - wq->tail;
3319 	if (likely(cur + nreq < wq->max_post))
3320 		return 0;
3321 
3322 	cq = to_mcq(ib_cq);
3323 	spin_lock(&cq->lock);
3324 	cur = wq->head - wq->tail;
3325 	spin_unlock(&cq->lock);
3326 
3327 	return cur + nreq >= wq->max_post;
3328 }
3329 
3330 static __be32 convert_access(int acc)
3331 {
3332 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
3333 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
3334 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
3335 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
3336 	       (acc & IB_ACCESS_REMOTE_READ   ?
3337 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
3338 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
3339 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
3340 }
3341 
3342 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
3343 			struct ib_reg_wr *wr)
3344 {
3345 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
3346 
3347 	fseg->flags		= convert_access(wr->access);
3348 	fseg->mem_key		= cpu_to_be32(wr->key);
3349 	fseg->buf_list		= cpu_to_be64(mr->page_map);
3350 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
3351 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
3352 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
3353 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
3354 	fseg->reserved[0]	= 0;
3355 	fseg->reserved[1]	= 0;
3356 }
3357 
3358 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
3359 {
3360 	memset(iseg, 0, sizeof(*iseg));
3361 	iseg->mem_key = cpu_to_be32(rkey);
3362 }
3363 
3364 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
3365 					  u64 remote_addr, u32 rkey)
3366 {
3367 	rseg->raddr    = cpu_to_be64(remote_addr);
3368 	rseg->rkey     = cpu_to_be32(rkey);
3369 	rseg->reserved = 0;
3370 }
3371 
3372 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
3373 		struct ib_atomic_wr *wr)
3374 {
3375 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
3376 		aseg->swap_add = cpu_to_be64(wr->swap);
3377 		aseg->compare  = cpu_to_be64(wr->compare_add);
3378 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
3379 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3380 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
3381 	} else {
3382 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3383 		aseg->compare  = 0;
3384 	}
3385 
3386 }
3387 
3388 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
3389 				  struct ib_atomic_wr *wr)
3390 {
3391 	aseg->swap_add		= cpu_to_be64(wr->swap);
3392 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
3393 	aseg->compare		= cpu_to_be64(wr->compare_add);
3394 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
3395 }
3396 
3397 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
3398 			     struct ib_ud_wr *wr)
3399 {
3400 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
3401 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
3402 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
3403 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
3404 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
3405 }
3406 
3407 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
3408 				    struct mlx4_wqe_datagram_seg *dseg,
3409 				    struct ib_ud_wr *wr,
3410 				    enum mlx4_ib_qp_type qpt)
3411 {
3412 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
3413 	struct mlx4_av sqp_av = {0};
3414 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
3415 
3416 	/* force loopback */
3417 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
3418 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
3419 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
3420 			cpu_to_be32(0xf0000000);
3421 
3422 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
3423 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
3424 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp1_tunnel);
3425 	else
3426 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp0_tunnel);
3427 	/* Use QKEY from the QP context, which is set by master */
3428 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
3429 }
3430 
3431 static void build_tunnel_header(struct ib_ud_wr *wr, void *wqe, unsigned *mlx_seg_len)
3432 {
3433 	struct mlx4_wqe_inline_seg *inl = wqe;
3434 	struct mlx4_ib_tunnel_header hdr;
3435 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3436 	int spc;
3437 	int i;
3438 
3439 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
3440 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
3441 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
3442 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
3443 	memcpy(hdr.mac, ah->av.eth.mac, 6);
3444 	hdr.vlan = ah->av.eth.vlan;
3445 
3446 	spc = MLX4_INLINE_ALIGN -
3447 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3448 	if (sizeof (hdr) <= spc) {
3449 		memcpy(inl + 1, &hdr, sizeof (hdr));
3450 		wmb();
3451 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
3452 		i = 1;
3453 	} else {
3454 		memcpy(inl + 1, &hdr, spc);
3455 		wmb();
3456 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3457 
3458 		inl = (void *) (inl + 1) + spc;
3459 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
3460 		wmb();
3461 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
3462 		i = 2;
3463 	}
3464 
3465 	*mlx_seg_len =
3466 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
3467 }
3468 
3469 static void set_mlx_icrc_seg(void *dseg)
3470 {
3471 	u32 *t = dseg;
3472 	struct mlx4_wqe_inline_seg *iseg = dseg;
3473 
3474 	t[1] = 0;
3475 
3476 	/*
3477 	 * Need a barrier here before writing the byte_count field to
3478 	 * make sure that all the data is visible before the
3479 	 * byte_count field is set.  Otherwise, if the segment begins
3480 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3481 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3482 	 * stale data, and end up sending the wrong data.
3483 	 */
3484 	wmb();
3485 
3486 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
3487 }
3488 
3489 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3490 {
3491 	dseg->lkey       = cpu_to_be32(sg->lkey);
3492 	dseg->addr       = cpu_to_be64(sg->addr);
3493 
3494 	/*
3495 	 * Need a barrier here before writing the byte_count field to
3496 	 * make sure that all the data is visible before the
3497 	 * byte_count field is set.  Otherwise, if the segment begins
3498 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3499 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3500 	 * stale data, and end up sending the wrong data.
3501 	 */
3502 	wmb();
3503 
3504 	dseg->byte_count = cpu_to_be32(sg->length);
3505 }
3506 
3507 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3508 {
3509 	dseg->byte_count = cpu_to_be32(sg->length);
3510 	dseg->lkey       = cpu_to_be32(sg->lkey);
3511 	dseg->addr       = cpu_to_be64(sg->addr);
3512 }
3513 
3514 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_ud_wr *wr,
3515 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
3516 			 __be32 *lso_hdr_sz, __be32 *blh)
3517 {
3518 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
3519 
3520 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
3521 		*blh = cpu_to_be32(1 << 6);
3522 
3523 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
3524 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
3525 		return -EINVAL;
3526 
3527 	memcpy(wqe->header, wr->header, wr->hlen);
3528 
3529 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
3530 	*lso_seg_len = halign;
3531 	return 0;
3532 }
3533 
3534 static __be32 send_ieth(struct ib_send_wr *wr)
3535 {
3536 	switch (wr->opcode) {
3537 	case IB_WR_SEND_WITH_IMM:
3538 	case IB_WR_RDMA_WRITE_WITH_IMM:
3539 		return wr->ex.imm_data;
3540 
3541 	case IB_WR_SEND_WITH_INV:
3542 		return cpu_to_be32(wr->ex.invalidate_rkey);
3543 
3544 	default:
3545 		return 0;
3546 	}
3547 }
3548 
3549 static void add_zero_len_inline(void *wqe)
3550 {
3551 	struct mlx4_wqe_inline_seg *inl = wqe;
3552 	memset(wqe, 0, 16);
3553 	inl->byte_count = cpu_to_be32(1 << 31);
3554 }
3555 
3556 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
3557 		      struct ib_send_wr **bad_wr)
3558 {
3559 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3560 	void *wqe;
3561 	struct mlx4_wqe_ctrl_seg *ctrl;
3562 	struct mlx4_wqe_data_seg *dseg;
3563 	unsigned long flags;
3564 	int nreq;
3565 	int err = 0;
3566 	unsigned ind;
3567 	int uninitialized_var(stamp);
3568 	int uninitialized_var(size);
3569 	unsigned uninitialized_var(seglen);
3570 	__be32 dummy;
3571 	__be32 *lso_wqe;
3572 	__be32 uninitialized_var(lso_hdr_sz);
3573 	__be32 blh;
3574 	int i;
3575 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3576 
3577 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
3578 		struct mlx4_ib_sqp *sqp = to_msqp(qp);
3579 
3580 		if (sqp->roce_v2_gsi) {
3581 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
3582 			enum ib_gid_type gid_type;
3583 			union ib_gid gid;
3584 
3585 			if (!fill_gid_by_hw_index(mdev, sqp->qp.port,
3586 					   ah->av.ib.gid_index,
3587 					   &gid, &gid_type))
3588 				qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
3589 						to_mqp(sqp->roce_v2_gsi) : qp;
3590 			else
3591 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
3592 				       ah->av.ib.gid_index);
3593 		}
3594 	}
3595 
3596 	spin_lock_irqsave(&qp->sq.lock, flags);
3597 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3598 		err = -EIO;
3599 		*bad_wr = wr;
3600 		nreq = 0;
3601 		goto out;
3602 	}
3603 
3604 	ind = qp->sq_next_wqe;
3605 
3606 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3607 		lso_wqe = &dummy;
3608 		blh = 0;
3609 
3610 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
3611 			err = -ENOMEM;
3612 			*bad_wr = wr;
3613 			goto out;
3614 		}
3615 
3616 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3617 			err = -EINVAL;
3618 			*bad_wr = wr;
3619 			goto out;
3620 		}
3621 
3622 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3623 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3624 
3625 		ctrl->srcrb_flags =
3626 			(wr->send_flags & IB_SEND_SIGNALED ?
3627 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3628 			(wr->send_flags & IB_SEND_SOLICITED ?
3629 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3630 			((wr->send_flags & IB_SEND_IP_CSUM) ?
3631 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3632 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3633 			qp->sq_signal_bits;
3634 
3635 		ctrl->imm = send_ieth(wr);
3636 
3637 		wqe += sizeof *ctrl;
3638 		size = sizeof *ctrl / 16;
3639 
3640 		switch (qp->mlx4_ib_qp_type) {
3641 		case MLX4_IB_QPT_RC:
3642 		case MLX4_IB_QPT_UC:
3643 			switch (wr->opcode) {
3644 			case IB_WR_ATOMIC_CMP_AND_SWP:
3645 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3646 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3647 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3648 					      atomic_wr(wr)->rkey);
3649 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3650 
3651 				set_atomic_seg(wqe, atomic_wr(wr));
3652 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
3653 
3654 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3655 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3656 
3657 				break;
3658 
3659 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3660 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3661 					      atomic_wr(wr)->rkey);
3662 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3663 
3664 				set_masked_atomic_seg(wqe, atomic_wr(wr));
3665 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
3666 
3667 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3668 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3669 
3670 				break;
3671 
3672 			case IB_WR_RDMA_READ:
3673 			case IB_WR_RDMA_WRITE:
3674 			case IB_WR_RDMA_WRITE_WITH_IMM:
3675 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3676 					      rdma_wr(wr)->rkey);
3677 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3678 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3679 				break;
3680 
3681 			case IB_WR_LOCAL_INV:
3682 				ctrl->srcrb_flags |=
3683 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3684 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3685 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3686 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3687 				break;
3688 
3689 			case IB_WR_REG_MR:
3690 				ctrl->srcrb_flags |=
3691 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3692 				set_reg_seg(wqe, reg_wr(wr));
3693 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3694 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3695 				break;
3696 
3697 			default:
3698 				/* No extra segments required for sends */
3699 				break;
3700 			}
3701 			break;
3702 
3703 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3704 			err =  build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3705 					ctrl, &seglen);
3706 			if (unlikely(err)) {
3707 				*bad_wr = wr;
3708 				goto out;
3709 			}
3710 			wqe  += seglen;
3711 			size += seglen / 16;
3712 			break;
3713 		case MLX4_IB_QPT_TUN_SMI:
3714 		case MLX4_IB_QPT_TUN_GSI:
3715 			/* this is a UD qp used in MAD responses to slaves. */
3716 			set_datagram_seg(wqe, ud_wr(wr));
3717 			/* set the forced-loopback bit in the data seg av */
3718 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3719 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3720 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3721 			break;
3722 		case MLX4_IB_QPT_UD:
3723 			set_datagram_seg(wqe, ud_wr(wr));
3724 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3725 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3726 
3727 			if (wr->opcode == IB_WR_LSO) {
3728 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3729 						&lso_hdr_sz, &blh);
3730 				if (unlikely(err)) {
3731 					*bad_wr = wr;
3732 					goto out;
3733 				}
3734 				lso_wqe = (__be32 *) wqe;
3735 				wqe  += seglen;
3736 				size += seglen / 16;
3737 			}
3738 			break;
3739 
3740 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3741 			err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3742 					ctrl, &seglen);
3743 			if (unlikely(err)) {
3744 				*bad_wr = wr;
3745 				goto out;
3746 			}
3747 			wqe  += seglen;
3748 			size += seglen / 16;
3749 			/* to start tunnel header on a cache-line boundary */
3750 			add_zero_len_inline(wqe);
3751 			wqe += 16;
3752 			size++;
3753 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3754 			wqe  += seglen;
3755 			size += seglen / 16;
3756 			break;
3757 		case MLX4_IB_QPT_PROXY_SMI:
3758 		case MLX4_IB_QPT_PROXY_GSI:
3759 			/* If we are tunneling special qps, this is a UD qp.
3760 			 * In this case we first add a UD segment targeting
3761 			 * the tunnel qp, and then add a header with address
3762 			 * information */
3763 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3764 						ud_wr(wr),
3765 						qp->mlx4_ib_qp_type);
3766 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3767 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3768 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3769 			wqe  += seglen;
3770 			size += seglen / 16;
3771 			break;
3772 
3773 		case MLX4_IB_QPT_SMI:
3774 		case MLX4_IB_QPT_GSI:
3775 			err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
3776 					&seglen);
3777 			if (unlikely(err)) {
3778 				*bad_wr = wr;
3779 				goto out;
3780 			}
3781 			wqe  += seglen;
3782 			size += seglen / 16;
3783 			break;
3784 
3785 		default:
3786 			break;
3787 		}
3788 
3789 		/*
3790 		 * Write data segments in reverse order, so as to
3791 		 * overwrite cacheline stamp last within each
3792 		 * cacheline.  This avoids issues with WQE
3793 		 * prefetching.
3794 		 */
3795 
3796 		dseg = wqe;
3797 		dseg += wr->num_sge - 1;
3798 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3799 
3800 		/* Add one more inline data segment for ICRC for MLX sends */
3801 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3802 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3803 			     qp->mlx4_ib_qp_type &
3804 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3805 			set_mlx_icrc_seg(dseg + 1);
3806 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3807 		}
3808 
3809 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3810 			set_data_seg(dseg, wr->sg_list + i);
3811 
3812 		/*
3813 		 * Possibly overwrite stamping in cacheline with LSO
3814 		 * segment only after making sure all data segments
3815 		 * are written.
3816 		 */
3817 		wmb();
3818 		*lso_wqe = lso_hdr_sz;
3819 
3820 		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3821 					     MLX4_WQE_CTRL_FENCE : 0) | size;
3822 
3823 		/*
3824 		 * Make sure descriptor is fully written before
3825 		 * setting ownership bit (because HW can start
3826 		 * executing as soon as we do).
3827 		 */
3828 		wmb();
3829 
3830 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3831 			*bad_wr = wr;
3832 			err = -EINVAL;
3833 			goto out;
3834 		}
3835 
3836 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3837 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3838 
3839 		stamp = ind + qp->sq_spare_wqes;
3840 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
3841 
3842 		/*
3843 		 * We can improve latency by not stamping the last
3844 		 * send queue WQE until after ringing the doorbell, so
3845 		 * only stamp here if there are still more WQEs to post.
3846 		 *
3847 		 * Same optimization applies to padding with NOP wqe
3848 		 * in case of WQE shrinking (used to prevent wrap-around
3849 		 * in the middle of WR).
3850 		 */
3851 		if (wr->next) {
3852 			stamp_send_wqe(qp, stamp, size * 16);
3853 			ind = pad_wraparound(qp, ind);
3854 		}
3855 	}
3856 
3857 out:
3858 	if (likely(nreq)) {
3859 		qp->sq.head += nreq;
3860 
3861 		/*
3862 		 * Make sure that descriptors are written before
3863 		 * doorbell record.
3864 		 */
3865 		wmb();
3866 
3867 		writel(qp->doorbell_qpn,
3868 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3869 
3870 		/*
3871 		 * Make sure doorbells don't leak out of SQ spinlock
3872 		 * and reach the HCA out of order.
3873 		 */
3874 		mmiowb();
3875 
3876 		stamp_send_wqe(qp, stamp, size * 16);
3877 
3878 		ind = pad_wraparound(qp, ind);
3879 		qp->sq_next_wqe = ind;
3880 	}
3881 
3882 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3883 
3884 	return err;
3885 }
3886 
3887 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3888 		      struct ib_recv_wr **bad_wr)
3889 {
3890 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3891 	struct mlx4_wqe_data_seg *scat;
3892 	unsigned long flags;
3893 	int err = 0;
3894 	int nreq;
3895 	int ind;
3896 	int max_gs;
3897 	int i;
3898 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3899 
3900 	max_gs = qp->rq.max_gs;
3901 	spin_lock_irqsave(&qp->rq.lock, flags);
3902 
3903 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3904 		err = -EIO;
3905 		*bad_wr = wr;
3906 		nreq = 0;
3907 		goto out;
3908 	}
3909 
3910 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3911 
3912 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3913 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3914 			err = -ENOMEM;
3915 			*bad_wr = wr;
3916 			goto out;
3917 		}
3918 
3919 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3920 			err = -EINVAL;
3921 			*bad_wr = wr;
3922 			goto out;
3923 		}
3924 
3925 		scat = get_recv_wqe(qp, ind);
3926 
3927 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3928 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3929 			ib_dma_sync_single_for_device(ibqp->device,
3930 						      qp->sqp_proxy_rcv[ind].map,
3931 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3932 						      DMA_FROM_DEVICE);
3933 			scat->byte_count =
3934 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3935 			/* use dma lkey from upper layer entry */
3936 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3937 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3938 			scat++;
3939 			max_gs--;
3940 		}
3941 
3942 		for (i = 0; i < wr->num_sge; ++i)
3943 			__set_data_seg(scat + i, wr->sg_list + i);
3944 
3945 		if (i < max_gs) {
3946 			scat[i].byte_count = 0;
3947 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3948 			scat[i].addr       = 0;
3949 		}
3950 
3951 		qp->rq.wrid[ind] = wr->wr_id;
3952 
3953 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3954 	}
3955 
3956 out:
3957 	if (likely(nreq)) {
3958 		qp->rq.head += nreq;
3959 
3960 		/*
3961 		 * Make sure that descriptors are written before
3962 		 * doorbell record.
3963 		 */
3964 		wmb();
3965 
3966 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3967 	}
3968 
3969 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3970 
3971 	return err;
3972 }
3973 
3974 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3975 {
3976 	switch (mlx4_state) {
3977 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3978 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3979 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3980 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3981 	case MLX4_QP_STATE_SQ_DRAINING:
3982 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3983 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3984 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3985 	default:		     return -1;
3986 	}
3987 }
3988 
3989 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3990 {
3991 	switch (mlx4_mig_state) {
3992 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3993 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3994 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3995 	default: return -1;
3996 	}
3997 }
3998 
3999 static int to_ib_qp_access_flags(int mlx4_flags)
4000 {
4001 	int ib_flags = 0;
4002 
4003 	if (mlx4_flags & MLX4_QP_BIT_RRE)
4004 		ib_flags |= IB_ACCESS_REMOTE_READ;
4005 	if (mlx4_flags & MLX4_QP_BIT_RWE)
4006 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
4007 	if (mlx4_flags & MLX4_QP_BIT_RAE)
4008 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
4009 
4010 	return ib_flags;
4011 }
4012 
4013 static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
4014 			    struct rdma_ah_attr *ah_attr,
4015 			    struct mlx4_qp_path *path)
4016 {
4017 	struct mlx4_dev *dev = ibdev->dev;
4018 	u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
4019 
4020 	memset(ah_attr, 0, sizeof(*ah_attr));
4021 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
4022 	if (port_num == 0 || port_num > dev->caps.num_ports)
4023 		return;
4024 
4025 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
4026 		rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
4027 			       ((path->sched_queue & 4) << 1));
4028 	else
4029 		rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
4030 	rdma_ah_set_port_num(ah_attr, port_num);
4031 
4032 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
4033 	rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
4034 	rdma_ah_set_static_rate(ah_attr,
4035 				path->static_rate ? path->static_rate - 5 : 0);
4036 	if (path->grh_mylmc & (1 << 7)) {
4037 		rdma_ah_set_grh(ah_attr, NULL,
4038 				be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
4039 				path->mgid_index,
4040 				path->hop_limit,
4041 				(be32_to_cpu(path->tclass_flowlabel)
4042 				 >> 20) & 0xff);
4043 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
4044 	}
4045 }
4046 
4047 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
4048 		     struct ib_qp_init_attr *qp_init_attr)
4049 {
4050 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
4051 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
4052 	struct mlx4_qp_context context;
4053 	int mlx4_state;
4054 	int err = 0;
4055 
4056 	if (ibqp->rwq_ind_tbl)
4057 		return -EOPNOTSUPP;
4058 
4059 	mutex_lock(&qp->mutex);
4060 
4061 	if (qp->state == IB_QPS_RESET) {
4062 		qp_attr->qp_state = IB_QPS_RESET;
4063 		goto done;
4064 	}
4065 
4066 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
4067 	if (err) {
4068 		err = -EINVAL;
4069 		goto out;
4070 	}
4071 
4072 	mlx4_state = be32_to_cpu(context.flags) >> 28;
4073 
4074 	qp->state		     = to_ib_qp_state(mlx4_state);
4075 	qp_attr->qp_state	     = qp->state;
4076 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
4077 	qp_attr->path_mig_state	     =
4078 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
4079 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
4080 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
4081 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
4082 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
4083 	qp_attr->qp_access_flags     =
4084 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
4085 
4086 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4087 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
4088 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
4089 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
4090 		qp_attr->alt_port_num	=
4091 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
4092 	}
4093 
4094 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
4095 	if (qp_attr->qp_state == IB_QPS_INIT)
4096 		qp_attr->port_num = qp->port;
4097 	else
4098 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
4099 
4100 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4101 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
4102 
4103 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
4104 
4105 	qp_attr->max_dest_rd_atomic =
4106 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
4107 	qp_attr->min_rnr_timer	    =
4108 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
4109 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
4110 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
4111 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
4112 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
4113 
4114 done:
4115 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4116 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4117 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4118 
4119 	if (!ibqp->uobject) {
4120 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
4121 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4122 	} else {
4123 		qp_attr->cap.max_send_wr  = 0;
4124 		qp_attr->cap.max_send_sge = 0;
4125 	}
4126 
4127 	/*
4128 	 * We don't support inline sends for kernel QPs (yet), and we
4129 	 * don't know what userspace's value should be.
4130 	 */
4131 	qp_attr->cap.max_inline_data = 0;
4132 
4133 	qp_init_attr->cap	     = qp_attr->cap;
4134 
4135 	qp_init_attr->create_flags = 0;
4136 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4137 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4138 
4139 	if (qp->flags & MLX4_IB_QP_LSO)
4140 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
4141 
4142 	if (qp->flags & MLX4_IB_QP_NETIF)
4143 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
4144 
4145 	qp_init_attr->sq_sig_type =
4146 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
4147 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4148 
4149 out:
4150 	mutex_unlock(&qp->mutex);
4151 	return err;
4152 }
4153 
4154 struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
4155 				struct ib_wq_init_attr *init_attr,
4156 				struct ib_udata *udata)
4157 {
4158 	struct mlx4_ib_dev *dev;
4159 	struct ib_qp_init_attr ib_qp_init_attr;
4160 	struct mlx4_ib_qp *qp;
4161 	struct mlx4_ib_create_wq ucmd;
4162 	int err, required_cmd_sz;
4163 
4164 	if (!(udata && pd->uobject))
4165 		return ERR_PTR(-EINVAL);
4166 
4167 	required_cmd_sz = offsetof(typeof(ucmd), comp_mask) +
4168 			  sizeof(ucmd.comp_mask);
4169 	if (udata->inlen < required_cmd_sz) {
4170 		pr_debug("invalid inlen\n");
4171 		return ERR_PTR(-EINVAL);
4172 	}
4173 
4174 	if (udata->inlen > sizeof(ucmd) &&
4175 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4176 				 udata->inlen - sizeof(ucmd))) {
4177 		pr_debug("inlen is not supported\n");
4178 		return ERR_PTR(-EOPNOTSUPP);
4179 	}
4180 
4181 	if (udata->outlen)
4182 		return ERR_PTR(-EOPNOTSUPP);
4183 
4184 	dev = to_mdev(pd->device);
4185 
4186 	if (init_attr->wq_type != IB_WQT_RQ) {
4187 		pr_debug("unsupported wq type %d\n", init_attr->wq_type);
4188 		return ERR_PTR(-EOPNOTSUPP);
4189 	}
4190 
4191 	if (init_attr->create_flags) {
4192 		pr_debug("unsupported create_flags %u\n",
4193 			 init_attr->create_flags);
4194 		return ERR_PTR(-EOPNOTSUPP);
4195 	}
4196 
4197 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
4198 	if (!qp)
4199 		return ERR_PTR(-ENOMEM);
4200 
4201 	qp->pri.vid = 0xFFFF;
4202 	qp->alt.vid = 0xFFFF;
4203 
4204 	memset(&ib_qp_init_attr, 0, sizeof(ib_qp_init_attr));
4205 	ib_qp_init_attr.qp_context = init_attr->wq_context;
4206 	ib_qp_init_attr.qp_type = IB_QPT_RAW_PACKET;
4207 	ib_qp_init_attr.cap.max_recv_wr = init_attr->max_wr;
4208 	ib_qp_init_attr.cap.max_recv_sge = init_attr->max_sge;
4209 	ib_qp_init_attr.recv_cq = init_attr->cq;
4210 	ib_qp_init_attr.send_cq = ib_qp_init_attr.recv_cq; /* Dummy CQ */
4211 
4212 	err = create_qp_common(dev, pd, MLX4_IB_RWQ_SRC, &ib_qp_init_attr,
4213 			       udata, 0, &qp);
4214 	if (err) {
4215 		kfree(qp);
4216 		return ERR_PTR(err);
4217 	}
4218 
4219 	qp->ibwq.event_handler = init_attr->event_handler;
4220 	qp->ibwq.wq_num = qp->mqp.qpn;
4221 	qp->ibwq.state = IB_WQS_RESET;
4222 
4223 	return &qp->ibwq;
4224 }
4225 
4226 static int ib_wq2qp_state(enum ib_wq_state state)
4227 {
4228 	switch (state) {
4229 	case IB_WQS_RESET:
4230 		return IB_QPS_RESET;
4231 	case IB_WQS_RDY:
4232 		return IB_QPS_RTR;
4233 	default:
4234 		return IB_QPS_ERR;
4235 	}
4236 }
4237 
4238 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state)
4239 {
4240 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4241 	enum ib_qp_state qp_cur_state;
4242 	enum ib_qp_state qp_new_state;
4243 	int attr_mask;
4244 	int err;
4245 
4246 	/* ib_qp.state represents the WQ HW state while ib_wq.state represents
4247 	 * the WQ logic state.
4248 	 */
4249 	qp_cur_state = qp->state;
4250 	qp_new_state = ib_wq2qp_state(new_state);
4251 
4252 	if (ib_wq2qp_state(new_state) == qp_cur_state)
4253 		return 0;
4254 
4255 	if (new_state == IB_WQS_RDY) {
4256 		struct ib_qp_attr attr = {};
4257 
4258 		attr.port_num = qp->port;
4259 		attr_mask = IB_QP_PORT;
4260 
4261 		err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, &attr,
4262 					  attr_mask, IB_QPS_RESET, IB_QPS_INIT);
4263 		if (err) {
4264 			pr_debug("WQN=0x%06x failed to apply RST->INIT on the HW QP\n",
4265 				 ibwq->wq_num);
4266 			return err;
4267 		}
4268 
4269 		qp_cur_state = IB_QPS_INIT;
4270 	}
4271 
4272 	attr_mask = 0;
4273 	err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, attr_mask,
4274 				  qp_cur_state,  qp_new_state);
4275 
4276 	if (err && (qp_cur_state == IB_QPS_INIT)) {
4277 		qp_new_state = IB_QPS_RESET;
4278 		if (__mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL,
4279 					attr_mask, IB_QPS_INIT, IB_QPS_RESET)) {
4280 			pr_warn("WQN=0x%06x failed with reverting HW's resources failure\n",
4281 				ibwq->wq_num);
4282 			qp_new_state = IB_QPS_INIT;
4283 		}
4284 	}
4285 
4286 	qp->state = qp_new_state;
4287 
4288 	return err;
4289 }
4290 
4291 int mlx4_ib_modify_wq(struct ib_wq *ibwq, struct ib_wq_attr *wq_attr,
4292 		      u32 wq_attr_mask, struct ib_udata *udata)
4293 {
4294 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4295 	struct mlx4_ib_modify_wq ucmd = {};
4296 	size_t required_cmd_sz;
4297 	enum ib_wq_state cur_state, new_state;
4298 	int err = 0;
4299 
4300 	required_cmd_sz = offsetof(typeof(ucmd), reserved) +
4301 				   sizeof(ucmd.reserved);
4302 	if (udata->inlen < required_cmd_sz)
4303 		return -EINVAL;
4304 
4305 	if (udata->inlen > sizeof(ucmd) &&
4306 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4307 				 udata->inlen - sizeof(ucmd)))
4308 		return -EOPNOTSUPP;
4309 
4310 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4311 		return -EFAULT;
4312 
4313 	if (ucmd.comp_mask || ucmd.reserved)
4314 		return -EOPNOTSUPP;
4315 
4316 	if (wq_attr_mask & IB_WQ_FLAGS)
4317 		return -EOPNOTSUPP;
4318 
4319 	cur_state = wq_attr_mask & IB_WQ_CUR_STATE ? wq_attr->curr_wq_state :
4320 						     ibwq->state;
4321 	new_state = wq_attr_mask & IB_WQ_STATE ? wq_attr->wq_state : cur_state;
4322 
4323 	if (cur_state  < IB_WQS_RESET || cur_state  > IB_WQS_ERR ||
4324 	    new_state < IB_WQS_RESET || new_state > IB_WQS_ERR)
4325 		return -EINVAL;
4326 
4327 	if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR))
4328 		return -EINVAL;
4329 
4330 	if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET))
4331 		return -EINVAL;
4332 
4333 	/* Need to protect against the parent RSS which also may modify WQ
4334 	 * state.
4335 	 */
4336 	mutex_lock(&qp->mutex);
4337 
4338 	/* Can update HW state only if a RSS QP has already associated to this
4339 	 * WQ, so we can apply its port on the WQ.
4340 	 */
4341 	if (qp->rss_usecnt)
4342 		err = _mlx4_ib_modify_wq(ibwq, new_state);
4343 
4344 	if (!err)
4345 		ibwq->state = new_state;
4346 
4347 	mutex_unlock(&qp->mutex);
4348 
4349 	return err;
4350 }
4351 
4352 int mlx4_ib_destroy_wq(struct ib_wq *ibwq)
4353 {
4354 	struct mlx4_ib_dev *dev = to_mdev(ibwq->device);
4355 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4356 
4357 	if (qp->counter_index)
4358 		mlx4_ib_free_qp_counter(dev, qp);
4359 
4360 	destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, 1);
4361 
4362 	kfree(qp);
4363 
4364 	return 0;
4365 }
4366 
4367 struct ib_rwq_ind_table
4368 *mlx4_ib_create_rwq_ind_table(struct ib_device *device,
4369 			      struct ib_rwq_ind_table_init_attr *init_attr,
4370 			      struct ib_udata *udata)
4371 {
4372 	struct ib_rwq_ind_table *rwq_ind_table;
4373 	struct mlx4_ib_create_rwq_ind_tbl_resp resp = {};
4374 	unsigned int ind_tbl_size = 1 << init_attr->log_ind_tbl_size;
4375 	unsigned int base_wqn;
4376 	size_t min_resp_len;
4377 	int i;
4378 	int err;
4379 
4380 	if (udata->inlen > 0 &&
4381 	    !ib_is_udata_cleared(udata, 0,
4382 				 udata->inlen))
4383 		return ERR_PTR(-EOPNOTSUPP);
4384 
4385 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4386 	if (udata->outlen && udata->outlen < min_resp_len)
4387 		return ERR_PTR(-EINVAL);
4388 
4389 	if (ind_tbl_size >
4390 	    device->attrs.rss_caps.max_rwq_indirection_table_size) {
4391 		pr_debug("log_ind_tbl_size = %d is bigger than supported = %d\n",
4392 			 ind_tbl_size,
4393 			 device->attrs.rss_caps.max_rwq_indirection_table_size);
4394 		return ERR_PTR(-EINVAL);
4395 	}
4396 
4397 	base_wqn = init_attr->ind_tbl[0]->wq_num;
4398 
4399 	if (base_wqn % ind_tbl_size) {
4400 		pr_debug("WQN=0x%x isn't aligned with indirection table size\n",
4401 			 base_wqn);
4402 		return ERR_PTR(-EINVAL);
4403 	}
4404 
4405 	for (i = 1; i < ind_tbl_size; i++) {
4406 		if (++base_wqn != init_attr->ind_tbl[i]->wq_num) {
4407 			pr_debug("indirection table's WQNs aren't consecutive\n");
4408 			return ERR_PTR(-EINVAL);
4409 		}
4410 	}
4411 
4412 	rwq_ind_table = kzalloc(sizeof(*rwq_ind_table), GFP_KERNEL);
4413 	if (!rwq_ind_table)
4414 		return ERR_PTR(-ENOMEM);
4415 
4416 	if (udata->outlen) {
4417 		resp.response_length = offsetof(typeof(resp), response_length) +
4418 					sizeof(resp.response_length);
4419 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4420 		if (err)
4421 			goto err;
4422 	}
4423 
4424 	return rwq_ind_table;
4425 
4426 err:
4427 	kfree(rwq_ind_table);
4428 	return ERR_PTR(err);
4429 }
4430 
4431 int mlx4_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4432 {
4433 	kfree(ib_rwq_ind_tbl);
4434 	return 0;
4435 }
4436