xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision 1a59d1b8)
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 #include <rdma/uverbs_ioctl.h>
45 
46 #include <linux/mlx4/driver.h>
47 #include <linux/mlx4/qp.h>
48 
49 #include "mlx4_ib.h"
50 #include <rdma/mlx4-abi.h>
51 
52 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
53 			     struct mlx4_ib_cq *recv_cq);
54 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
55 			       struct mlx4_ib_cq *recv_cq);
56 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state,
57 			      struct ib_udata *udata);
58 
59 enum {
60 	MLX4_IB_ACK_REQ_FREQ	= 8,
61 };
62 
63 enum {
64 	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
65 	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
66 	MLX4_IB_LINK_TYPE_IB		= 0,
67 	MLX4_IB_LINK_TYPE_ETH		= 1
68 };
69 
70 enum {
71 	/*
72 	 * Largest possible UD header: send with GRH and immediate
73 	 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
74 	 * tag.  (LRH would only use 8 bytes, so Ethernet is the
75 	 * biggest case)
76 	 */
77 	MLX4_IB_UD_HEADER_SIZE		= 82,
78 	MLX4_IB_LSO_HEADER_SPARE	= 128,
79 };
80 
81 struct mlx4_ib_sqp {
82 	struct mlx4_ib_qp	qp;
83 	int			pkey_index;
84 	u32			qkey;
85 	u32			send_psn;
86 	struct ib_ud_header	ud_header;
87 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
88 	struct ib_qp		*roce_v2_gsi;
89 };
90 
91 enum {
92 	MLX4_IB_MIN_SQ_STRIDE	= 6,
93 	MLX4_IB_CACHE_LINE_SIZE	= 64,
94 };
95 
96 enum {
97 	MLX4_RAW_QP_MTU		= 7,
98 	MLX4_RAW_QP_MSGMAX	= 31,
99 };
100 
101 #ifndef ETH_ALEN
102 #define ETH_ALEN        6
103 #endif
104 
105 static const __be32 mlx4_ib_opcode[] = {
106 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
107 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
108 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
109 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
110 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
111 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
112 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
113 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
114 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
115 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
116 	[IB_WR_REG_MR]				= cpu_to_be32(MLX4_OPCODE_FMR),
117 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
118 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
119 };
120 
121 enum mlx4_ib_source_type {
122 	MLX4_IB_QP_SRC	= 0,
123 	MLX4_IB_RWQ_SRC	= 1,
124 };
125 
126 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
127 {
128 	return container_of(mqp, struct mlx4_ib_sqp, qp);
129 }
130 
131 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
132 {
133 	if (!mlx4_is_master(dev->dev))
134 		return 0;
135 
136 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
137 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
138 		8 * MLX4_MFUNC_MAX;
139 }
140 
141 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
142 {
143 	int proxy_sqp = 0;
144 	int real_sqp = 0;
145 	int i;
146 	/* PPF or Native -- real SQP */
147 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
148 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
149 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
150 	if (real_sqp)
151 		return 1;
152 	/* VF or PF -- proxy SQP */
153 	if (mlx4_is_mfunc(dev->dev)) {
154 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
155 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy ||
156 			    qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp1_proxy) {
157 				proxy_sqp = 1;
158 				break;
159 			}
160 		}
161 	}
162 	if (proxy_sqp)
163 		return 1;
164 
165 	return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
166 }
167 
168 /* used for INIT/CLOSE port logic */
169 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
170 {
171 	int proxy_qp0 = 0;
172 	int real_qp0 = 0;
173 	int i;
174 	/* PPF or Native -- real QP0 */
175 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
176 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
177 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
178 	if (real_qp0)
179 		return 1;
180 	/* VF or PF -- proxy QP0 */
181 	if (mlx4_is_mfunc(dev->dev)) {
182 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
183 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy) {
184 				proxy_qp0 = 1;
185 				break;
186 			}
187 		}
188 	}
189 	return proxy_qp0;
190 }
191 
192 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
193 {
194 	return mlx4_buf_offset(&qp->buf, offset);
195 }
196 
197 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
198 {
199 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
200 }
201 
202 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
203 {
204 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
205 }
206 
207 /*
208  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
209  * first four bytes of every 64 byte chunk with 0xffffffff, except for
210  * the very first chunk of the WQE.
211  */
212 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n)
213 {
214 	__be32 *wqe;
215 	int i;
216 	int s;
217 	void *buf;
218 	struct mlx4_wqe_ctrl_seg *ctrl;
219 
220 	buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
221 	ctrl = (struct mlx4_wqe_ctrl_seg *)buf;
222 	s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4;
223 	for (i = 64; i < s; i += 64) {
224 		wqe = buf + i;
225 		*wqe = cpu_to_be32(0xffffffff);
226 	}
227 }
228 
229 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
230 {
231 	struct ib_event event;
232 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
233 
234 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
235 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
236 
237 	if (ibqp->event_handler) {
238 		event.device     = ibqp->device;
239 		event.element.qp = ibqp;
240 		switch (type) {
241 		case MLX4_EVENT_TYPE_PATH_MIG:
242 			event.event = IB_EVENT_PATH_MIG;
243 			break;
244 		case MLX4_EVENT_TYPE_COMM_EST:
245 			event.event = IB_EVENT_COMM_EST;
246 			break;
247 		case MLX4_EVENT_TYPE_SQ_DRAINED:
248 			event.event = IB_EVENT_SQ_DRAINED;
249 			break;
250 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
251 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
252 			break;
253 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
254 			event.event = IB_EVENT_QP_FATAL;
255 			break;
256 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
257 			event.event = IB_EVENT_PATH_MIG_ERR;
258 			break;
259 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
260 			event.event = IB_EVENT_QP_REQ_ERR;
261 			break;
262 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
263 			event.event = IB_EVENT_QP_ACCESS_ERR;
264 			break;
265 		default:
266 			pr_warn("Unexpected event type %d "
267 			       "on QP %06x\n", type, qp->qpn);
268 			return;
269 		}
270 
271 		ibqp->event_handler(&event, ibqp->qp_context);
272 	}
273 }
274 
275 static void mlx4_ib_wq_event(struct mlx4_qp *qp, enum mlx4_event type)
276 {
277 	pr_warn_ratelimited("Unexpected event type %d on WQ 0x%06x. Events are not supported for WQs\n",
278 			    type, qp->qpn);
279 }
280 
281 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
282 {
283 	/*
284 	 * UD WQEs must have a datagram segment.
285 	 * RC and UC WQEs might have a remote address segment.
286 	 * MLX WQEs need two extra inline data segments (for the UD
287 	 * header and space for the ICRC).
288 	 */
289 	switch (type) {
290 	case MLX4_IB_QPT_UD:
291 		return sizeof (struct mlx4_wqe_ctrl_seg) +
292 			sizeof (struct mlx4_wqe_datagram_seg) +
293 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
294 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
295 	case MLX4_IB_QPT_PROXY_SMI:
296 	case MLX4_IB_QPT_PROXY_GSI:
297 		return sizeof (struct mlx4_wqe_ctrl_seg) +
298 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
299 	case MLX4_IB_QPT_TUN_SMI_OWNER:
300 	case MLX4_IB_QPT_TUN_GSI:
301 		return sizeof (struct mlx4_wqe_ctrl_seg) +
302 			sizeof (struct mlx4_wqe_datagram_seg);
303 
304 	case MLX4_IB_QPT_UC:
305 		return sizeof (struct mlx4_wqe_ctrl_seg) +
306 			sizeof (struct mlx4_wqe_raddr_seg);
307 	case MLX4_IB_QPT_RC:
308 		return sizeof (struct mlx4_wqe_ctrl_seg) +
309 			sizeof (struct mlx4_wqe_masked_atomic_seg) +
310 			sizeof (struct mlx4_wqe_raddr_seg);
311 	case MLX4_IB_QPT_SMI:
312 	case MLX4_IB_QPT_GSI:
313 		return sizeof (struct mlx4_wqe_ctrl_seg) +
314 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
315 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
316 					   MLX4_INLINE_ALIGN) *
317 			      sizeof (struct mlx4_wqe_inline_seg),
318 			      sizeof (struct mlx4_wqe_data_seg)) +
319 			ALIGN(4 +
320 			      sizeof (struct mlx4_wqe_inline_seg),
321 			      sizeof (struct mlx4_wqe_data_seg));
322 	default:
323 		return sizeof (struct mlx4_wqe_ctrl_seg);
324 	}
325 }
326 
327 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
328 		       bool is_user, int has_rq, struct mlx4_ib_qp *qp,
329 		       u32 inl_recv_sz)
330 {
331 	/* Sanity check RQ size before proceeding */
332 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
333 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
334 		return -EINVAL;
335 
336 	if (!has_rq) {
337 		if (cap->max_recv_wr || inl_recv_sz)
338 			return -EINVAL;
339 
340 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
341 	} else {
342 		u32 max_inl_recv_sz = dev->dev->caps.max_rq_sg *
343 			sizeof(struct mlx4_wqe_data_seg);
344 		u32 wqe_size;
345 
346 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
347 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge ||
348 				inl_recv_sz > max_inl_recv_sz))
349 			return -EINVAL;
350 
351 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
352 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
353 		wqe_size = qp->rq.max_gs * sizeof(struct mlx4_wqe_data_seg);
354 		qp->rq.wqe_shift = ilog2(max_t(u32, wqe_size, inl_recv_sz));
355 	}
356 
357 	/* leave userspace return values as they were, so as not to break ABI */
358 	if (is_user) {
359 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
360 		cap->max_recv_sge = qp->rq.max_gs;
361 	} else {
362 		cap->max_recv_wr  = qp->rq.max_post =
363 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
364 		cap->max_recv_sge = min(qp->rq.max_gs,
365 					min(dev->dev->caps.max_sq_sg,
366 					    dev->dev->caps.max_rq_sg));
367 	}
368 
369 	return 0;
370 }
371 
372 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
373 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
374 {
375 	int s;
376 
377 	/* Sanity check SQ size before proceeding */
378 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
379 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
380 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
381 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
382 		return -EINVAL;
383 
384 	/*
385 	 * For MLX transport we need 2 extra S/G entries:
386 	 * one for the header and one for the checksum at the end
387 	 */
388 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
389 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
390 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
391 		return -EINVAL;
392 
393 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
394 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
395 		send_wqe_overhead(type, qp->flags);
396 
397 	if (s > dev->dev->caps.max_sq_desc_sz)
398 		return -EINVAL;
399 
400 	qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
401 
402 	/*
403 	 * We need to leave 2 KB + 1 WR of headroom in the SQ to
404 	 * allow HW to prefetch.
405 	 */
406 	qp->sq_spare_wqes = MLX4_IB_SQ_HEADROOM(qp->sq.wqe_shift);
407 	qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr +
408 					    qp->sq_spare_wqes);
409 
410 	qp->sq.max_gs =
411 		(min(dev->dev->caps.max_sq_desc_sz,
412 		     (1 << qp->sq.wqe_shift)) -
413 		 send_wqe_overhead(type, qp->flags)) /
414 		sizeof (struct mlx4_wqe_data_seg);
415 
416 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
417 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
418 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
419 		qp->rq.offset = 0;
420 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
421 	} else {
422 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
423 		qp->sq.offset = 0;
424 	}
425 
426 	cap->max_send_wr  = qp->sq.max_post =
427 		qp->sq.wqe_cnt - qp->sq_spare_wqes;
428 	cap->max_send_sge = min(qp->sq.max_gs,
429 				min(dev->dev->caps.max_sq_sg,
430 				    dev->dev->caps.max_rq_sg));
431 	/* We don't support inline sends for kernel QPs (yet) */
432 	cap->max_inline_data = 0;
433 
434 	return 0;
435 }
436 
437 static int set_user_sq_size(struct mlx4_ib_dev *dev,
438 			    struct mlx4_ib_qp *qp,
439 			    struct mlx4_ib_create_qp *ucmd)
440 {
441 	/* Sanity check SQ size before proceeding */
442 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
443 	    ucmd->log_sq_stride >
444 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
445 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
446 		return -EINVAL;
447 
448 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
449 	qp->sq.wqe_shift = ucmd->log_sq_stride;
450 
451 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
452 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
453 
454 	return 0;
455 }
456 
457 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
458 {
459 	int i;
460 
461 	qp->sqp_proxy_rcv =
462 		kmalloc_array(qp->rq.wqe_cnt, sizeof(struct mlx4_ib_buf),
463 			      GFP_KERNEL);
464 	if (!qp->sqp_proxy_rcv)
465 		return -ENOMEM;
466 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
467 		qp->sqp_proxy_rcv[i].addr =
468 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
469 				GFP_KERNEL);
470 		if (!qp->sqp_proxy_rcv[i].addr)
471 			goto err;
472 		qp->sqp_proxy_rcv[i].map =
473 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
474 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
475 					  DMA_FROM_DEVICE);
476 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
477 			kfree(qp->sqp_proxy_rcv[i].addr);
478 			goto err;
479 		}
480 	}
481 	return 0;
482 
483 err:
484 	while (i > 0) {
485 		--i;
486 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
487 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
488 				    DMA_FROM_DEVICE);
489 		kfree(qp->sqp_proxy_rcv[i].addr);
490 	}
491 	kfree(qp->sqp_proxy_rcv);
492 	qp->sqp_proxy_rcv = NULL;
493 	return -ENOMEM;
494 }
495 
496 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
497 {
498 	int i;
499 
500 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
501 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
502 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
503 				    DMA_FROM_DEVICE);
504 		kfree(qp->sqp_proxy_rcv[i].addr);
505 	}
506 	kfree(qp->sqp_proxy_rcv);
507 }
508 
509 static int qp_has_rq(struct ib_qp_init_attr *attr)
510 {
511 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
512 		return 0;
513 
514 	return !attr->srq;
515 }
516 
517 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
518 {
519 	int i;
520 	for (i = 0; i < dev->caps.num_ports; i++) {
521 		if (qpn == dev->caps.spec_qps[i].qp0_proxy)
522 			return !!dev->caps.spec_qps[i].qp0_qkey;
523 	}
524 	return 0;
525 }
526 
527 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
528 				    struct mlx4_ib_qp *qp)
529 {
530 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
531 	mlx4_counter_free(dev->dev, qp->counter_index->index);
532 	list_del(&qp->counter_index->list);
533 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
534 
535 	kfree(qp->counter_index);
536 	qp->counter_index = NULL;
537 }
538 
539 static int set_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_rss *rss_ctx,
540 		      struct ib_qp_init_attr *init_attr,
541 		      struct mlx4_ib_create_qp_rss *ucmd)
542 {
543 	rss_ctx->base_qpn_tbl_sz = init_attr->rwq_ind_tbl->ind_tbl[0]->wq_num |
544 		(init_attr->rwq_ind_tbl->log_ind_tbl_size << 24);
545 
546 	if ((ucmd->rx_hash_function == MLX4_IB_RX_HASH_FUNC_TOEPLITZ) &&
547 	    (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS_TOP)) {
548 		memcpy(rss_ctx->rss_key, ucmd->rx_hash_key,
549 		       MLX4_EN_RSS_KEY_SIZE);
550 	} else {
551 		pr_debug("RX Hash function is not supported\n");
552 		return (-EOPNOTSUPP);
553 	}
554 
555 	if (ucmd->rx_hash_fields_mask & ~(MLX4_IB_RX_HASH_SRC_IPV4	|
556 					  MLX4_IB_RX_HASH_DST_IPV4	|
557 					  MLX4_IB_RX_HASH_SRC_IPV6	|
558 					  MLX4_IB_RX_HASH_DST_IPV6	|
559 					  MLX4_IB_RX_HASH_SRC_PORT_TCP	|
560 					  MLX4_IB_RX_HASH_DST_PORT_TCP	|
561 					  MLX4_IB_RX_HASH_SRC_PORT_UDP	|
562 					  MLX4_IB_RX_HASH_DST_PORT_UDP  |
563 					  MLX4_IB_RX_HASH_INNER)) {
564 		pr_debug("RX Hash fields_mask has unsupported mask (0x%llx)\n",
565 			 ucmd->rx_hash_fields_mask);
566 		return (-EOPNOTSUPP);
567 	}
568 
569 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) &&
570 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
571 		rss_ctx->flags = MLX4_RSS_IPV4;
572 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) ||
573 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
574 		pr_debug("RX Hash fields_mask is not supported - both IPv4 SRC and DST must be set\n");
575 		return (-EOPNOTSUPP);
576 	}
577 
578 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) &&
579 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
580 		rss_ctx->flags |= MLX4_RSS_IPV6;
581 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) ||
582 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
583 		pr_debug("RX Hash fields_mask is not supported - both IPv6 SRC and DST must be set\n");
584 		return (-EOPNOTSUPP);
585 	}
586 
587 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) &&
588 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
589 		if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS)) {
590 			pr_debug("RX Hash fields_mask for UDP is not supported\n");
591 			return (-EOPNOTSUPP);
592 		}
593 
594 		if (rss_ctx->flags & MLX4_RSS_IPV4)
595 			rss_ctx->flags |= MLX4_RSS_UDP_IPV4;
596 		if (rss_ctx->flags & MLX4_RSS_IPV6)
597 			rss_ctx->flags |= MLX4_RSS_UDP_IPV6;
598 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
599 			pr_debug("RX Hash fields_mask is not supported - UDP must be set with IPv4 or IPv6\n");
600 			return (-EOPNOTSUPP);
601 		}
602 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) ||
603 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
604 		pr_debug("RX Hash fields_mask is not supported - both UDP SRC and DST must be set\n");
605 		return (-EOPNOTSUPP);
606 	}
607 
608 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) &&
609 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
610 		if (rss_ctx->flags & MLX4_RSS_IPV4)
611 			rss_ctx->flags |= MLX4_RSS_TCP_IPV4;
612 		if (rss_ctx->flags & MLX4_RSS_IPV6)
613 			rss_ctx->flags |= MLX4_RSS_TCP_IPV6;
614 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
615 			pr_debug("RX Hash fields_mask is not supported - TCP must be set with IPv4 or IPv6\n");
616 			return (-EOPNOTSUPP);
617 		}
618 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) ||
619 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
620 		pr_debug("RX Hash fields_mask is not supported - both TCP SRC and DST must be set\n");
621 		return (-EOPNOTSUPP);
622 	}
623 
624 	if (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_INNER) {
625 		if (dev->dev->caps.tunnel_offload_mode ==
626 		    MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
627 			/*
628 			 * Hash according to inner headers if exist, otherwise
629 			 * according to outer headers.
630 			 */
631 			rss_ctx->flags |= MLX4_RSS_BY_INNER_HEADERS_IPONLY;
632 		} else {
633 			pr_debug("RSS Hash for inner headers isn't supported\n");
634 			return (-EOPNOTSUPP);
635 		}
636 	}
637 
638 	return 0;
639 }
640 
641 static int create_qp_rss(struct mlx4_ib_dev *dev,
642 			 struct ib_qp_init_attr *init_attr,
643 			 struct mlx4_ib_create_qp_rss *ucmd,
644 			 struct mlx4_ib_qp *qp)
645 {
646 	int qpn;
647 	int err;
648 
649 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
650 
651 	err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 0, qp->mqp.usage);
652 	if (err)
653 		return err;
654 
655 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
656 	if (err)
657 		goto err_qpn;
658 
659 	mutex_init(&qp->mutex);
660 
661 	INIT_LIST_HEAD(&qp->gid_list);
662 	INIT_LIST_HEAD(&qp->steering_rules);
663 
664 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
665 	qp->state = IB_QPS_RESET;
666 
667 	/* Set dummy send resources to be compatible with HV and PRM */
668 	qp->sq_no_prefetch = 1;
669 	qp->sq.wqe_cnt = 1;
670 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
671 	qp->buf_size = qp->sq.wqe_cnt << MLX4_IB_MIN_SQ_STRIDE;
672 	qp->mtt = (to_mqp(
673 		   (struct ib_qp *)init_attr->rwq_ind_tbl->ind_tbl[0]))->mtt;
674 
675 	qp->rss_ctx = kzalloc(sizeof(*qp->rss_ctx), GFP_KERNEL);
676 	if (!qp->rss_ctx) {
677 		err = -ENOMEM;
678 		goto err_qp_alloc;
679 	}
680 
681 	err = set_qp_rss(dev, qp->rss_ctx, init_attr, ucmd);
682 	if (err)
683 		goto err;
684 
685 	return 0;
686 
687 err:
688 	kfree(qp->rss_ctx);
689 
690 err_qp_alloc:
691 	mlx4_qp_remove(dev->dev, &qp->mqp);
692 	mlx4_qp_free(dev->dev, &qp->mqp);
693 
694 err_qpn:
695 	mlx4_qp_release_range(dev->dev, qpn, 1);
696 	return err;
697 }
698 
699 static struct ib_qp *_mlx4_ib_create_qp_rss(struct ib_pd *pd,
700 					    struct ib_qp_init_attr *init_attr,
701 					    struct ib_udata *udata)
702 {
703 	struct mlx4_ib_qp *qp;
704 	struct mlx4_ib_create_qp_rss ucmd = {};
705 	size_t required_cmd_sz;
706 	int err;
707 
708 	if (!udata) {
709 		pr_debug("RSS QP with NULL udata\n");
710 		return ERR_PTR(-EINVAL);
711 	}
712 
713 	if (udata->outlen)
714 		return ERR_PTR(-EOPNOTSUPP);
715 
716 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) +
717 					sizeof(ucmd.reserved1);
718 	if (udata->inlen < required_cmd_sz) {
719 		pr_debug("invalid inlen\n");
720 		return ERR_PTR(-EINVAL);
721 	}
722 
723 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
724 		pr_debug("copy failed\n");
725 		return ERR_PTR(-EFAULT);
726 	}
727 
728 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)))
729 		return ERR_PTR(-EOPNOTSUPP);
730 
731 	if (ucmd.comp_mask || ucmd.reserved1)
732 		return ERR_PTR(-EOPNOTSUPP);
733 
734 	if (udata->inlen > sizeof(ucmd) &&
735 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
736 				 udata->inlen - sizeof(ucmd))) {
737 		pr_debug("inlen is not supported\n");
738 		return ERR_PTR(-EOPNOTSUPP);
739 	}
740 
741 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
742 		pr_debug("RSS QP with unsupported QP type %d\n",
743 			 init_attr->qp_type);
744 		return ERR_PTR(-EOPNOTSUPP);
745 	}
746 
747 	if (init_attr->create_flags) {
748 		pr_debug("RSS QP doesn't support create flags\n");
749 		return ERR_PTR(-EOPNOTSUPP);
750 	}
751 
752 	if (init_attr->send_cq || init_attr->cap.max_send_wr) {
753 		pr_debug("RSS QP with unsupported send attributes\n");
754 		return ERR_PTR(-EOPNOTSUPP);
755 	}
756 
757 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
758 	if (!qp)
759 		return ERR_PTR(-ENOMEM);
760 
761 	qp->pri.vid = 0xFFFF;
762 	qp->alt.vid = 0xFFFF;
763 
764 	err = create_qp_rss(to_mdev(pd->device), init_attr, &ucmd, qp);
765 	if (err) {
766 		kfree(qp);
767 		return ERR_PTR(err);
768 	}
769 
770 	qp->ibqp.qp_num = qp->mqp.qpn;
771 
772 	return &qp->ibqp;
773 }
774 
775 /*
776  * This function allocates a WQN from a range which is consecutive and aligned
777  * to its size. In case the range is full, then it creates a new range and
778  * allocates WQN from it. The new range will be used for following allocations.
779  */
780 static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context,
781 			     struct mlx4_ib_qp *qp, int range_size, int *wqn)
782 {
783 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
784 	struct mlx4_wqn_range *range;
785 	int err = 0;
786 
787 	mutex_lock(&context->wqn_ranges_mutex);
788 
789 	range = list_first_entry_or_null(&context->wqn_ranges_list,
790 					 struct mlx4_wqn_range, list);
791 
792 	if (!range || (range->refcount == range->size) || range->dirty) {
793 		range = kzalloc(sizeof(*range), GFP_KERNEL);
794 		if (!range) {
795 			err = -ENOMEM;
796 			goto out;
797 		}
798 
799 		err = mlx4_qp_reserve_range(dev->dev, range_size,
800 					    range_size, &range->base_wqn, 0,
801 					    qp->mqp.usage);
802 		if (err) {
803 			kfree(range);
804 			goto out;
805 		}
806 
807 		range->size = range_size;
808 		list_add(&range->list, &context->wqn_ranges_list);
809 	} else if (range_size != 1) {
810 		/*
811 		 * Requesting a new range (>1) when last range is still open, is
812 		 * not valid.
813 		 */
814 		err = -EINVAL;
815 		goto out;
816 	}
817 
818 	qp->wqn_range = range;
819 
820 	*wqn = range->base_wqn + range->refcount;
821 
822 	range->refcount++;
823 
824 out:
825 	mutex_unlock(&context->wqn_ranges_mutex);
826 
827 	return err;
828 }
829 
830 static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context,
831 				struct mlx4_ib_qp *qp, bool dirty_release)
832 {
833 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
834 	struct mlx4_wqn_range *range;
835 
836 	mutex_lock(&context->wqn_ranges_mutex);
837 
838 	range = qp->wqn_range;
839 
840 	range->refcount--;
841 	if (!range->refcount) {
842 		mlx4_qp_release_range(dev->dev, range->base_wqn,
843 				      range->size);
844 		list_del(&range->list);
845 		kfree(range);
846 	} else if (dirty_release) {
847 	/*
848 	 * A range which one of its WQNs is destroyed, won't be able to be
849 	 * reused for further WQN allocations.
850 	 * The next created WQ will allocate a new range.
851 	 */
852 		range->dirty = 1;
853 	}
854 
855 	mutex_unlock(&context->wqn_ranges_mutex);
856 }
857 
858 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
859 			    enum mlx4_ib_source_type src,
860 			    struct ib_qp_init_attr *init_attr,
861 			    struct ib_udata *udata, int sqpn,
862 			    struct mlx4_ib_qp **caller_qp)
863 {
864 	int qpn;
865 	int err;
866 	struct mlx4_ib_sqp *sqp = NULL;
867 	struct mlx4_ib_qp *qp;
868 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
869 		udata, struct mlx4_ib_ucontext, ibucontext);
870 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
871 	struct mlx4_ib_cq *mcq;
872 	unsigned long flags;
873 	int range_size = 0;
874 
875 	/* When tunneling special qps, we use a plain UD qp */
876 	if (sqpn) {
877 		if (mlx4_is_mfunc(dev->dev) &&
878 		    (!mlx4_is_master(dev->dev) ||
879 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
880 			if (init_attr->qp_type == IB_QPT_GSI)
881 				qp_type = MLX4_IB_QPT_PROXY_GSI;
882 			else {
883 				if (mlx4_is_master(dev->dev) ||
884 				    qp0_enabled_vf(dev->dev, sqpn))
885 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
886 				else
887 					qp_type = MLX4_IB_QPT_PROXY_SMI;
888 			}
889 		}
890 		qpn = sqpn;
891 		/* add extra sg entry for tunneling */
892 		init_attr->cap.max_recv_sge++;
893 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
894 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
895 			container_of(init_attr,
896 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
897 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
898 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
899 		    !mlx4_is_master(dev->dev))
900 			return -EINVAL;
901 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
902 			qp_type = MLX4_IB_QPT_TUN_GSI;
903 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
904 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
905 					     tnl_init->port))
906 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
907 		else
908 			qp_type = MLX4_IB_QPT_TUN_SMI;
909 		/* we are definitely in the PPF here, since we are creating
910 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
911 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
912 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
913 		sqpn = qpn;
914 	}
915 
916 	if (!*caller_qp) {
917 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
918 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
919 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
920 			sqp = kzalloc(sizeof(struct mlx4_ib_sqp), GFP_KERNEL);
921 			if (!sqp)
922 				return -ENOMEM;
923 			qp = &sqp->qp;
924 			qp->pri.vid = 0xFFFF;
925 			qp->alt.vid = 0xFFFF;
926 		} else {
927 			qp = kzalloc(sizeof(struct mlx4_ib_qp), GFP_KERNEL);
928 			if (!qp)
929 				return -ENOMEM;
930 			qp->pri.vid = 0xFFFF;
931 			qp->alt.vid = 0xFFFF;
932 		}
933 	} else
934 		qp = *caller_qp;
935 
936 	qp->mlx4_ib_qp_type = qp_type;
937 
938 	mutex_init(&qp->mutex);
939 	spin_lock_init(&qp->sq.lock);
940 	spin_lock_init(&qp->rq.lock);
941 	INIT_LIST_HEAD(&qp->gid_list);
942 	INIT_LIST_HEAD(&qp->steering_rules);
943 
944 	qp->state	 = IB_QPS_RESET;
945 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
946 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
947 
948 
949 	if (udata) {
950 		union {
951 			struct mlx4_ib_create_qp qp;
952 			struct mlx4_ib_create_wq wq;
953 		} ucmd;
954 		size_t copy_len;
955 		int shift;
956 		int n;
957 
958 		copy_len = (src == MLX4_IB_QP_SRC) ?
959 			   sizeof(struct mlx4_ib_create_qp) :
960 			   min(sizeof(struct mlx4_ib_create_wq), udata->inlen);
961 
962 		if (ib_copy_from_udata(&ucmd, udata, copy_len)) {
963 			err = -EFAULT;
964 			goto err;
965 		}
966 
967 		if (src == MLX4_IB_RWQ_SRC) {
968 			if (ucmd.wq.comp_mask || ucmd.wq.reserved[0] ||
969 			    ucmd.wq.reserved[1] || ucmd.wq.reserved[2]) {
970 				pr_debug("user command isn't supported\n");
971 				err = -EOPNOTSUPP;
972 				goto err;
973 			}
974 
975 			if (ucmd.wq.log_range_size >
976 			    ilog2(dev->dev->caps.max_rss_tbl_sz)) {
977 				pr_debug("WQN range size must be equal or smaller than %d\n",
978 					 dev->dev->caps.max_rss_tbl_sz);
979 				err = -EOPNOTSUPP;
980 				goto err;
981 			}
982 			range_size = 1 << ucmd.wq.log_range_size;
983 		} else {
984 			qp->inl_recv_sz = ucmd.qp.inl_recv_sz;
985 		}
986 
987 		if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
988 			if (!(dev->dev->caps.flags &
989 			      MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
990 				pr_debug("scatter FCS is unsupported\n");
991 				err = -EOPNOTSUPP;
992 				goto err;
993 			}
994 
995 			qp->flags |= MLX4_IB_QP_SCATTER_FCS;
996 		}
997 
998 		err = set_rq_size(dev, &init_attr->cap, udata,
999 				  qp_has_rq(init_attr), qp, qp->inl_recv_sz);
1000 		if (err)
1001 			goto err;
1002 
1003 		if (src == MLX4_IB_QP_SRC) {
1004 			qp->sq_no_prefetch = ucmd.qp.sq_no_prefetch;
1005 
1006 			err = set_user_sq_size(dev, qp,
1007 					       (struct mlx4_ib_create_qp *)
1008 					       &ucmd);
1009 			if (err)
1010 				goto err;
1011 		} else {
1012 			qp->sq_no_prefetch = 1;
1013 			qp->sq.wqe_cnt = 1;
1014 			qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
1015 			/* Allocated buffer expects to have at least that SQ
1016 			 * size.
1017 			 */
1018 			qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
1019 				(qp->sq.wqe_cnt << qp->sq.wqe_shift);
1020 		}
1021 
1022 		qp->umem =
1023 			ib_umem_get(udata,
1024 				    (src == MLX4_IB_QP_SRC) ? ucmd.qp.buf_addr :
1025 							      ucmd.wq.buf_addr,
1026 				    qp->buf_size, 0, 0);
1027 		if (IS_ERR(qp->umem)) {
1028 			err = PTR_ERR(qp->umem);
1029 			goto err;
1030 		}
1031 
1032 		n = ib_umem_page_count(qp->umem);
1033 		shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
1034 		err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
1035 
1036 		if (err)
1037 			goto err_buf;
1038 
1039 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
1040 		if (err)
1041 			goto err_mtt;
1042 
1043 		if (qp_has_rq(init_attr)) {
1044 			err = mlx4_ib_db_map_user(udata,
1045 						  (src == MLX4_IB_QP_SRC) ?
1046 							  ucmd.qp.db_addr :
1047 							  ucmd.wq.db_addr,
1048 						  &qp->db);
1049 			if (err)
1050 				goto err_mtt;
1051 		}
1052 		qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
1053 	} else {
1054 		err = set_rq_size(dev, &init_attr->cap, udata,
1055 				  qp_has_rq(init_attr), qp, 0);
1056 		if (err)
1057 			goto err;
1058 
1059 		qp->sq_no_prefetch = 0;
1060 
1061 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
1062 			qp->flags |= MLX4_IB_QP_LSO;
1063 
1064 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1065 			if (dev->steering_support ==
1066 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
1067 				qp->flags |= MLX4_IB_QP_NETIF;
1068 			else
1069 				goto err;
1070 		}
1071 
1072 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
1073 		if (err)
1074 			goto err;
1075 
1076 		if (qp_has_rq(init_attr)) {
1077 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
1078 			if (err)
1079 				goto err;
1080 
1081 			*qp->db.db = 0;
1082 		}
1083 
1084 		if (mlx4_buf_alloc(dev->dev, qp->buf_size,  PAGE_SIZE * 2,
1085 				   &qp->buf)) {
1086 			err = -ENOMEM;
1087 			goto err_db;
1088 		}
1089 
1090 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
1091 				    &qp->mtt);
1092 		if (err)
1093 			goto err_buf;
1094 
1095 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
1096 		if (err)
1097 			goto err_mtt;
1098 
1099 		qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1100 					     sizeof(u64), GFP_KERNEL);
1101 		qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1102 					     sizeof(u64), GFP_KERNEL);
1103 		if (!qp->sq.wrid || !qp->rq.wrid) {
1104 			err = -ENOMEM;
1105 			goto err_wrid;
1106 		}
1107 		qp->mqp.usage = MLX4_RES_USAGE_DRIVER;
1108 	}
1109 
1110 	if (sqpn) {
1111 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1112 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
1113 			if (alloc_proxy_bufs(pd->device, qp)) {
1114 				err = -ENOMEM;
1115 				goto err_wrid;
1116 			}
1117 		}
1118 	} else if (src == MLX4_IB_RWQ_SRC) {
1119 		err = mlx4_ib_alloc_wqn(context, qp, range_size, &qpn);
1120 		if (err)
1121 			goto err_wrid;
1122 	} else {
1123 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
1124 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
1125 		 * VLAN insertion. */
1126 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
1127 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
1128 						    (init_attr->cap.max_send_wr ?
1129 						     MLX4_RESERVE_ETH_BF_QP : 0) |
1130 						    (init_attr->cap.max_recv_wr ?
1131 						     MLX4_RESERVE_A0_QP : 0),
1132 						    qp->mqp.usage);
1133 		else
1134 			if (qp->flags & MLX4_IB_QP_NETIF)
1135 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
1136 			else
1137 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
1138 							    &qpn, 0, qp->mqp.usage);
1139 		if (err)
1140 			goto err_proxy;
1141 	}
1142 
1143 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1144 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1145 
1146 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
1147 	if (err)
1148 		goto err_qpn;
1149 
1150 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1151 		qp->mqp.qpn |= (1 << 23);
1152 
1153 	/*
1154 	 * Hardware wants QPN written in big-endian order (after
1155 	 * shifting) for send doorbell.  Precompute this value to save
1156 	 * a little bit when posting sends.
1157 	 */
1158 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1159 
1160 	qp->mqp.event = (src == MLX4_IB_QP_SRC) ? mlx4_ib_qp_event :
1161 						  mlx4_ib_wq_event;
1162 
1163 	if (!*caller_qp)
1164 		*caller_qp = qp;
1165 
1166 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1167 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
1168 			 to_mcq(init_attr->recv_cq));
1169 	/* Maintain device to QPs access, needed for further handling
1170 	 * via reset flow
1171 	 */
1172 	list_add_tail(&qp->qps_list, &dev->qp_list);
1173 	/* Maintain CQ to QPs access, needed for further handling
1174 	 * via reset flow
1175 	 */
1176 	mcq = to_mcq(init_attr->send_cq);
1177 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
1178 	mcq = to_mcq(init_attr->recv_cq);
1179 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
1180 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
1181 			   to_mcq(init_attr->recv_cq));
1182 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1183 	return 0;
1184 
1185 err_qpn:
1186 	if (!sqpn) {
1187 		if (qp->flags & MLX4_IB_QP_NETIF)
1188 			mlx4_ib_steer_qp_free(dev, qpn, 1);
1189 		else if (src == MLX4_IB_RWQ_SRC)
1190 			mlx4_ib_release_wqn(context, qp, 0);
1191 		else
1192 			mlx4_qp_release_range(dev->dev, qpn, 1);
1193 	}
1194 err_proxy:
1195 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1196 		free_proxy_bufs(pd->device, qp);
1197 err_wrid:
1198 	if (udata) {
1199 		if (qp_has_rq(init_attr))
1200 			mlx4_ib_db_unmap_user(context, &qp->db);
1201 	} else {
1202 		kvfree(qp->sq.wrid);
1203 		kvfree(qp->rq.wrid);
1204 	}
1205 
1206 err_mtt:
1207 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1208 
1209 err_buf:
1210 	if (qp->umem)
1211 		ib_umem_release(qp->umem);
1212 	else
1213 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1214 
1215 err_db:
1216 	if (!udata && qp_has_rq(init_attr))
1217 		mlx4_db_free(dev->dev, &qp->db);
1218 
1219 err:
1220 	if (!sqp && !*caller_qp)
1221 		kfree(qp);
1222 	kfree(sqp);
1223 
1224 	return err;
1225 }
1226 
1227 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
1228 {
1229 	switch (state) {
1230 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
1231 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
1232 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
1233 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
1234 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
1235 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
1236 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
1237 	default:		return -1;
1238 	}
1239 }
1240 
1241 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1242 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1243 {
1244 	if (send_cq == recv_cq) {
1245 		spin_lock(&send_cq->lock);
1246 		__acquire(&recv_cq->lock);
1247 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1248 		spin_lock(&send_cq->lock);
1249 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1250 	} else {
1251 		spin_lock(&recv_cq->lock);
1252 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1253 	}
1254 }
1255 
1256 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1257 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1258 {
1259 	if (send_cq == recv_cq) {
1260 		__release(&recv_cq->lock);
1261 		spin_unlock(&send_cq->lock);
1262 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1263 		spin_unlock(&recv_cq->lock);
1264 		spin_unlock(&send_cq->lock);
1265 	} else {
1266 		spin_unlock(&send_cq->lock);
1267 		spin_unlock(&recv_cq->lock);
1268 	}
1269 }
1270 
1271 static void del_gid_entries(struct mlx4_ib_qp *qp)
1272 {
1273 	struct mlx4_ib_gid_entry *ge, *tmp;
1274 
1275 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1276 		list_del(&ge->list);
1277 		kfree(ge);
1278 	}
1279 }
1280 
1281 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
1282 {
1283 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
1284 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
1285 	else
1286 		return to_mpd(qp->ibqp.pd);
1287 }
1288 
1289 static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src,
1290 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1291 {
1292 	switch (qp->ibqp.qp_type) {
1293 	case IB_QPT_XRC_TGT:
1294 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1295 		*recv_cq = *send_cq;
1296 		break;
1297 	case IB_QPT_XRC_INI:
1298 		*send_cq = to_mcq(qp->ibqp.send_cq);
1299 		*recv_cq = *send_cq;
1300 		break;
1301 	default:
1302 		*recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) :
1303 						     to_mcq(qp->ibwq.cq);
1304 		*send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) :
1305 						     *recv_cq;
1306 		break;
1307 	}
1308 }
1309 
1310 static void destroy_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1311 {
1312 	if (qp->state != IB_QPS_RESET) {
1313 		int i;
1314 
1315 		for (i = 0; i < (1 << qp->ibqp.rwq_ind_tbl->log_ind_tbl_size);
1316 		     i++) {
1317 			struct ib_wq *ibwq = qp->ibqp.rwq_ind_tbl->ind_tbl[i];
1318 			struct mlx4_ib_qp *wq =	to_mqp((struct ib_qp *)ibwq);
1319 
1320 			mutex_lock(&wq->mutex);
1321 
1322 			wq->rss_usecnt--;
1323 
1324 			mutex_unlock(&wq->mutex);
1325 		}
1326 
1327 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1328 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1329 			pr_warn("modify QP %06x to RESET failed.\n",
1330 				qp->mqp.qpn);
1331 	}
1332 
1333 	mlx4_qp_remove(dev->dev, &qp->mqp);
1334 	mlx4_qp_free(dev->dev, &qp->mqp);
1335 	mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1336 	del_gid_entries(qp);
1337 	kfree(qp->rss_ctx);
1338 }
1339 
1340 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1341 			      enum mlx4_ib_source_type src,
1342 			      struct ib_udata *udata)
1343 {
1344 	struct mlx4_ib_cq *send_cq, *recv_cq;
1345 	unsigned long flags;
1346 
1347 	if (qp->state != IB_QPS_RESET) {
1348 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1349 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1350 			pr_warn("modify QP %06x to RESET failed.\n",
1351 			       qp->mqp.qpn);
1352 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1353 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1354 			qp->pri.smac = 0;
1355 			qp->pri.smac_port = 0;
1356 		}
1357 		if (qp->alt.smac) {
1358 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1359 			qp->alt.smac = 0;
1360 		}
1361 		if (qp->pri.vid < 0x1000) {
1362 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1363 			qp->pri.vid = 0xFFFF;
1364 			qp->pri.candidate_vid = 0xFFFF;
1365 			qp->pri.update_vid = 0;
1366 		}
1367 		if (qp->alt.vid < 0x1000) {
1368 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1369 			qp->alt.vid = 0xFFFF;
1370 			qp->alt.candidate_vid = 0xFFFF;
1371 			qp->alt.update_vid = 0;
1372 		}
1373 	}
1374 
1375 	get_cqs(qp, src, &send_cq, &recv_cq);
1376 
1377 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1378 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1379 
1380 	/* del from lists under both locks above to protect reset flow paths */
1381 	list_del(&qp->qps_list);
1382 	list_del(&qp->cq_send_list);
1383 	list_del(&qp->cq_recv_list);
1384 	if (!udata) {
1385 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1386 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1387 		if (send_cq != recv_cq)
1388 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1389 	}
1390 
1391 	mlx4_qp_remove(dev->dev, &qp->mqp);
1392 
1393 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1394 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1395 
1396 	mlx4_qp_free(dev->dev, &qp->mqp);
1397 
1398 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1399 		if (qp->flags & MLX4_IB_QP_NETIF)
1400 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1401 		else if (src == MLX4_IB_RWQ_SRC)
1402 			mlx4_ib_release_wqn(
1403 				rdma_udata_to_drv_context(
1404 					udata,
1405 					struct mlx4_ib_ucontext,
1406 					ibucontext),
1407 				qp, 1);
1408 		else
1409 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1410 	}
1411 
1412 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1413 
1414 	if (udata) {
1415 		if (qp->rq.wqe_cnt) {
1416 			struct mlx4_ib_ucontext *mcontext =
1417 				rdma_udata_to_drv_context(
1418 					udata,
1419 					struct mlx4_ib_ucontext,
1420 					ibucontext);
1421 
1422 			mlx4_ib_db_unmap_user(mcontext, &qp->db);
1423 		}
1424 		ib_umem_release(qp->umem);
1425 	} else {
1426 		kvfree(qp->sq.wrid);
1427 		kvfree(qp->rq.wrid);
1428 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1429 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1430 			free_proxy_bufs(&dev->ib_dev, qp);
1431 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1432 		if (qp->rq.wqe_cnt)
1433 			mlx4_db_free(dev->dev, &qp->db);
1434 	}
1435 
1436 	del_gid_entries(qp);
1437 }
1438 
1439 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1440 {
1441 	/* Native or PPF */
1442 	if (!mlx4_is_mfunc(dev->dev) ||
1443 	    (mlx4_is_master(dev->dev) &&
1444 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1445 		return  dev->dev->phys_caps.base_sqpn +
1446 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1447 			attr->port_num - 1;
1448 	}
1449 	/* PF or VF -- creating proxies */
1450 	if (attr->qp_type == IB_QPT_SMI)
1451 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp0_proxy;
1452 	else
1453 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp1_proxy;
1454 }
1455 
1456 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1457 					struct ib_qp_init_attr *init_attr,
1458 					struct ib_udata *udata)
1459 {
1460 	struct mlx4_ib_qp *qp = NULL;
1461 	int err;
1462 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1463 	u16 xrcdn = 0;
1464 
1465 	if (init_attr->rwq_ind_tbl)
1466 		return _mlx4_ib_create_qp_rss(pd, init_attr, udata);
1467 
1468 	/*
1469 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1470 	 * and only for kernel UD QPs.
1471 	 */
1472 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1473 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1474 					MLX4_IB_SRIOV_TUNNEL_QP |
1475 					MLX4_IB_SRIOV_SQP |
1476 					MLX4_IB_QP_NETIF |
1477 					MLX4_IB_QP_CREATE_ROCE_V2_GSI))
1478 		return ERR_PTR(-EINVAL);
1479 
1480 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1481 		if (init_attr->qp_type != IB_QPT_UD)
1482 			return ERR_PTR(-EINVAL);
1483 	}
1484 
1485 	if (init_attr->create_flags) {
1486 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1487 			return ERR_PTR(-EINVAL);
1488 
1489 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1490 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1491 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1492 		     init_attr->qp_type != IB_QPT_UD) ||
1493 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1494 		     init_attr->qp_type > IB_QPT_GSI) ||
1495 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1496 		     init_attr->qp_type != IB_QPT_GSI))
1497 			return ERR_PTR(-EINVAL);
1498 	}
1499 
1500 	switch (init_attr->qp_type) {
1501 	case IB_QPT_XRC_TGT:
1502 		pd = to_mxrcd(init_attr->xrcd)->pd;
1503 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1504 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1505 		/* fall through */
1506 	case IB_QPT_XRC_INI:
1507 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1508 			return ERR_PTR(-ENOSYS);
1509 		init_attr->recv_cq = init_attr->send_cq;
1510 		/* fall through */
1511 	case IB_QPT_RC:
1512 	case IB_QPT_UC:
1513 	case IB_QPT_RAW_PACKET:
1514 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1515 		if (!qp)
1516 			return ERR_PTR(-ENOMEM);
1517 		qp->pri.vid = 0xFFFF;
1518 		qp->alt.vid = 0xFFFF;
1519 		/* fall through */
1520 	case IB_QPT_UD:
1521 	{
1522 		err = create_qp_common(to_mdev(pd->device), pd,	MLX4_IB_QP_SRC,
1523 				       init_attr, udata, 0, &qp);
1524 		if (err) {
1525 			kfree(qp);
1526 			return ERR_PTR(err);
1527 		}
1528 
1529 		qp->ibqp.qp_num = qp->mqp.qpn;
1530 		qp->xrcdn = xrcdn;
1531 
1532 		break;
1533 	}
1534 	case IB_QPT_SMI:
1535 	case IB_QPT_GSI:
1536 	{
1537 		int sqpn;
1538 
1539 		/* Userspace is not allowed to create special QPs: */
1540 		if (udata)
1541 			return ERR_PTR(-EINVAL);
1542 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1543 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev,
1544 							1, 1, &sqpn, 0,
1545 							MLX4_RES_USAGE_DRIVER);
1546 
1547 			if (res)
1548 				return ERR_PTR(res);
1549 		} else {
1550 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1551 		}
1552 
1553 		err = create_qp_common(to_mdev(pd->device), pd, MLX4_IB_QP_SRC,
1554 				       init_attr, udata, sqpn, &qp);
1555 		if (err)
1556 			return ERR_PTR(err);
1557 
1558 		qp->port	= init_attr->port_num;
1559 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1560 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1561 		break;
1562 	}
1563 	default:
1564 		/* Don't support raw QPs */
1565 		return ERR_PTR(-EINVAL);
1566 	}
1567 
1568 	return &qp->ibqp;
1569 }
1570 
1571 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1572 				struct ib_qp_init_attr *init_attr,
1573 				struct ib_udata *udata) {
1574 	struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1575 	struct ib_qp *ibqp;
1576 	struct mlx4_ib_dev *dev = to_mdev(device);
1577 
1578 	ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1579 
1580 	if (!IS_ERR(ibqp) &&
1581 	    (init_attr->qp_type == IB_QPT_GSI) &&
1582 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1583 		struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1584 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1585 
1586 		if (is_eth &&
1587 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1588 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1589 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1590 
1591 			if (IS_ERR(sqp->roce_v2_gsi)) {
1592 				pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1593 				sqp->roce_v2_gsi = NULL;
1594 			} else {
1595 				sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1596 				sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1597 			}
1598 
1599 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1600 		}
1601 	}
1602 	return ibqp;
1603 }
1604 
1605 static int _mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
1606 {
1607 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1608 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1609 
1610 	if (is_qp0(dev, mqp))
1611 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1612 
1613 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1614 	    dev->qp1_proxy[mqp->port - 1] == mqp) {
1615 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1616 		dev->qp1_proxy[mqp->port - 1] = NULL;
1617 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1618 	}
1619 
1620 	if (mqp->counter_index)
1621 		mlx4_ib_free_qp_counter(dev, mqp);
1622 
1623 	if (qp->rwq_ind_tbl) {
1624 		destroy_qp_rss(dev, mqp);
1625 	} else {
1626 		destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, udata);
1627 	}
1628 
1629 	if (is_sqp(dev, mqp))
1630 		kfree(to_msqp(mqp));
1631 	else
1632 		kfree(mqp);
1633 
1634 	return 0;
1635 }
1636 
1637 int mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
1638 {
1639 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1640 
1641 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1642 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1643 
1644 		if (sqp->roce_v2_gsi)
1645 			ib_destroy_qp(sqp->roce_v2_gsi);
1646 	}
1647 
1648 	return _mlx4_ib_destroy_qp(qp, udata);
1649 }
1650 
1651 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1652 {
1653 	switch (type) {
1654 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1655 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1656 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1657 	case MLX4_IB_QPT_XRC_INI:
1658 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1659 	case MLX4_IB_QPT_SMI:
1660 	case MLX4_IB_QPT_GSI:
1661 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1662 
1663 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1664 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1665 						MLX4_QP_ST_MLX : -1);
1666 	case MLX4_IB_QPT_PROXY_SMI:
1667 	case MLX4_IB_QPT_TUN_SMI:
1668 	case MLX4_IB_QPT_PROXY_GSI:
1669 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1670 						MLX4_QP_ST_UD : -1);
1671 	default:			return -1;
1672 	}
1673 }
1674 
1675 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1676 				   int attr_mask)
1677 {
1678 	u8 dest_rd_atomic;
1679 	u32 access_flags;
1680 	u32 hw_access_flags = 0;
1681 
1682 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1683 		dest_rd_atomic = attr->max_dest_rd_atomic;
1684 	else
1685 		dest_rd_atomic = qp->resp_depth;
1686 
1687 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1688 		access_flags = attr->qp_access_flags;
1689 	else
1690 		access_flags = qp->atomic_rd_en;
1691 
1692 	if (!dest_rd_atomic)
1693 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1694 
1695 	if (access_flags & IB_ACCESS_REMOTE_READ)
1696 		hw_access_flags |= MLX4_QP_BIT_RRE;
1697 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1698 		hw_access_flags |= MLX4_QP_BIT_RAE;
1699 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1700 		hw_access_flags |= MLX4_QP_BIT_RWE;
1701 
1702 	return cpu_to_be32(hw_access_flags);
1703 }
1704 
1705 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1706 			    int attr_mask)
1707 {
1708 	if (attr_mask & IB_QP_PKEY_INDEX)
1709 		sqp->pkey_index = attr->pkey_index;
1710 	if (attr_mask & IB_QP_QKEY)
1711 		sqp->qkey = attr->qkey;
1712 	if (attr_mask & IB_QP_SQ_PSN)
1713 		sqp->send_psn = attr->sq_psn;
1714 }
1715 
1716 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1717 {
1718 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1719 }
1720 
1721 static int _mlx4_set_path(struct mlx4_ib_dev *dev,
1722 			  const struct rdma_ah_attr *ah,
1723 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1724 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1725 {
1726 	int vidx;
1727 	int smac_index;
1728 	int err;
1729 
1730 	path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
1731 	path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
1732 	if (rdma_ah_get_static_rate(ah)) {
1733 		path->static_rate = rdma_ah_get_static_rate(ah) +
1734 				    MLX4_STAT_RATE_OFFSET;
1735 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1736 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1737 			--path->static_rate;
1738 	} else
1739 		path->static_rate = 0;
1740 
1741 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
1742 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
1743 		int real_sgid_index =
1744 			mlx4_ib_gid_index_to_real_index(dev, grh->sgid_attr);
1745 
1746 		if (real_sgid_index < 0)
1747 			return real_sgid_index;
1748 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1749 			pr_err("sgid_index (%u) too large. max is %d\n",
1750 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1751 			return -1;
1752 		}
1753 
1754 		path->grh_mylmc |= 1 << 7;
1755 		path->mgid_index = real_sgid_index;
1756 		path->hop_limit  = grh->hop_limit;
1757 		path->tclass_flowlabel =
1758 			cpu_to_be32((grh->traffic_class << 20) |
1759 				    (grh->flow_label));
1760 		memcpy(path->rgid, grh->dgid.raw, 16);
1761 	}
1762 
1763 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
1764 		if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
1765 			return -1;
1766 
1767 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1768 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
1769 
1770 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1771 		if (vlan_tag < 0x1000) {
1772 			if (smac_info->vid < 0x1000) {
1773 				/* both valid vlan ids */
1774 				if (smac_info->vid != vlan_tag) {
1775 					/* different VIDs.  unreg old and reg new */
1776 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1777 					if (err)
1778 						return err;
1779 					smac_info->candidate_vid = vlan_tag;
1780 					smac_info->candidate_vlan_index = vidx;
1781 					smac_info->candidate_vlan_port = port;
1782 					smac_info->update_vid = 1;
1783 					path->vlan_index = vidx;
1784 				} else {
1785 					path->vlan_index = smac_info->vlan_index;
1786 				}
1787 			} else {
1788 				/* no current vlan tag in qp */
1789 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1790 				if (err)
1791 					return err;
1792 				smac_info->candidate_vid = vlan_tag;
1793 				smac_info->candidate_vlan_index = vidx;
1794 				smac_info->candidate_vlan_port = port;
1795 				smac_info->update_vid = 1;
1796 				path->vlan_index = vidx;
1797 			}
1798 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1799 			path->fl = 1 << 6;
1800 		} else {
1801 			/* have current vlan tag. unregister it at modify-qp success */
1802 			if (smac_info->vid < 0x1000) {
1803 				smac_info->candidate_vid = 0xFFFF;
1804 				smac_info->update_vid = 1;
1805 			}
1806 		}
1807 
1808 		/* get smac_index for RoCE use.
1809 		 * If no smac was yet assigned, register one.
1810 		 * If one was already assigned, but the new mac differs,
1811 		 * unregister the old one and register the new one.
1812 		*/
1813 		if ((!smac_info->smac && !smac_info->smac_port) ||
1814 		    smac_info->smac != smac) {
1815 			/* register candidate now, unreg if needed, after success */
1816 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1817 			if (smac_index >= 0) {
1818 				smac_info->candidate_smac_index = smac_index;
1819 				smac_info->candidate_smac = smac;
1820 				smac_info->candidate_smac_port = port;
1821 			} else {
1822 				return -EINVAL;
1823 			}
1824 		} else {
1825 			smac_index = smac_info->smac_index;
1826 		}
1827 		memcpy(path->dmac, ah->roce.dmac, 6);
1828 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1829 		/* put MAC table smac index for IBoE */
1830 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1831 	} else {
1832 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1833 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
1834 	}
1835 
1836 	return 0;
1837 }
1838 
1839 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1840 			 enum ib_qp_attr_mask qp_attr_mask,
1841 			 struct mlx4_ib_qp *mqp,
1842 			 struct mlx4_qp_path *path, u8 port,
1843 			 u16 vlan_id, u8 *smac)
1844 {
1845 	return _mlx4_set_path(dev, &qp->ah_attr,
1846 			      mlx4_mac_to_u64(smac),
1847 			      vlan_id,
1848 			      path, &mqp->pri, port);
1849 }
1850 
1851 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1852 			     const struct ib_qp_attr *qp,
1853 			     enum ib_qp_attr_mask qp_attr_mask,
1854 			     struct mlx4_ib_qp *mqp,
1855 			     struct mlx4_qp_path *path, u8 port)
1856 {
1857 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1858 			      0,
1859 			      0xffff,
1860 			      path, &mqp->alt, port);
1861 }
1862 
1863 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1864 {
1865 	struct mlx4_ib_gid_entry *ge, *tmp;
1866 
1867 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1868 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1869 			ge->added = 1;
1870 			ge->port = qp->port;
1871 		}
1872 	}
1873 }
1874 
1875 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1876 				    struct mlx4_ib_qp *qp,
1877 				    struct mlx4_qp_context *context)
1878 {
1879 	u64 u64_mac;
1880 	int smac_index;
1881 
1882 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1883 
1884 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1885 	if (!qp->pri.smac && !qp->pri.smac_port) {
1886 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1887 		if (smac_index >= 0) {
1888 			qp->pri.candidate_smac_index = smac_index;
1889 			qp->pri.candidate_smac = u64_mac;
1890 			qp->pri.candidate_smac_port = qp->port;
1891 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1892 		} else {
1893 			return -ENOENT;
1894 		}
1895 	}
1896 	return 0;
1897 }
1898 
1899 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1900 {
1901 	struct counter_index *new_counter_index;
1902 	int err;
1903 	u32 tmp_idx;
1904 
1905 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1906 	    IB_LINK_LAYER_ETHERNET ||
1907 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1908 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1909 		return 0;
1910 
1911 	err = mlx4_counter_alloc(dev->dev, &tmp_idx, MLX4_RES_USAGE_DRIVER);
1912 	if (err)
1913 		return err;
1914 
1915 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1916 	if (!new_counter_index) {
1917 		mlx4_counter_free(dev->dev, tmp_idx);
1918 		return -ENOMEM;
1919 	}
1920 
1921 	new_counter_index->index = tmp_idx;
1922 	new_counter_index->allocated = 1;
1923 	qp->counter_index = new_counter_index;
1924 
1925 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
1926 	list_add_tail(&new_counter_index->list,
1927 		      &dev->counters_table[qp->port - 1].counters_list);
1928 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
1929 
1930 	return 0;
1931 }
1932 
1933 enum {
1934 	MLX4_QPC_ROCE_MODE_1 = 0,
1935 	MLX4_QPC_ROCE_MODE_2 = 2,
1936 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
1937 };
1938 
1939 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
1940 {
1941 	switch (gid_type) {
1942 	case IB_GID_TYPE_ROCE:
1943 		return MLX4_QPC_ROCE_MODE_1;
1944 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
1945 		return MLX4_QPC_ROCE_MODE_2;
1946 	default:
1947 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
1948 	}
1949 }
1950 
1951 /*
1952  * Go over all RSS QP's childes (WQs) and apply their HW state according to
1953  * their logic state if the RSS QP is the first RSS QP associated for the WQ.
1954  */
1955 static int bringup_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, u8 port_num,
1956 			    struct ib_udata *udata)
1957 {
1958 	int err = 0;
1959 	int i;
1960 
1961 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
1962 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
1963 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
1964 
1965 		mutex_lock(&wq->mutex);
1966 
1967 		/* Mlx4_ib restrictions:
1968 		 * WQ's is associated to a port according to the RSS QP it is
1969 		 * associates to.
1970 		 * In case the WQ is associated to a different port by another
1971 		 * RSS QP, return a failure.
1972 		 */
1973 		if ((wq->rss_usecnt > 0) && (wq->port != port_num)) {
1974 			err = -EINVAL;
1975 			mutex_unlock(&wq->mutex);
1976 			break;
1977 		}
1978 		wq->port = port_num;
1979 		if ((wq->rss_usecnt == 0) && (ibwq->state == IB_WQS_RDY)) {
1980 			err = _mlx4_ib_modify_wq(ibwq, IB_WQS_RDY, udata);
1981 			if (err) {
1982 				mutex_unlock(&wq->mutex);
1983 				break;
1984 			}
1985 		}
1986 		wq->rss_usecnt++;
1987 
1988 		mutex_unlock(&wq->mutex);
1989 	}
1990 
1991 	if (i && err) {
1992 		int j;
1993 
1994 		for (j = (i - 1); j >= 0; j--) {
1995 			struct ib_wq *ibwq = ind_tbl->ind_tbl[j];
1996 			struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
1997 
1998 			mutex_lock(&wq->mutex);
1999 
2000 			if ((wq->rss_usecnt == 1) &&
2001 			    (ibwq->state == IB_WQS_RDY))
2002 				if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET,
2003 						       udata))
2004 					pr_warn("failed to reverse WQN=0x%06x\n",
2005 						ibwq->wq_num);
2006 			wq->rss_usecnt--;
2007 
2008 			mutex_unlock(&wq->mutex);
2009 		}
2010 	}
2011 
2012 	return err;
2013 }
2014 
2015 static void bring_down_rss_rwqs(struct ib_rwq_ind_table *ind_tbl,
2016 				struct ib_udata *udata)
2017 {
2018 	int i;
2019 
2020 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2021 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2022 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2023 
2024 		mutex_lock(&wq->mutex);
2025 
2026 		if ((wq->rss_usecnt == 1) && (ibwq->state == IB_WQS_RDY))
2027 			if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET, udata))
2028 				pr_warn("failed to reverse WQN=%x\n",
2029 					ibwq->wq_num);
2030 		wq->rss_usecnt--;
2031 
2032 		mutex_unlock(&wq->mutex);
2033 	}
2034 }
2035 
2036 static void fill_qp_rss_context(struct mlx4_qp_context *context,
2037 				struct mlx4_ib_qp *qp)
2038 {
2039 	struct mlx4_rss_context *rss_context;
2040 
2041 	rss_context = (void *)context + offsetof(struct mlx4_qp_context,
2042 			pri_path) + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
2043 
2044 	rss_context->base_qpn = cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz);
2045 	rss_context->default_qpn =
2046 		cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz & 0xffffff);
2047 	if (qp->rss_ctx->flags & (MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6))
2048 		rss_context->base_qpn_udp = rss_context->default_qpn;
2049 	rss_context->flags = qp->rss_ctx->flags;
2050 	/* Currently support just toeplitz */
2051 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
2052 
2053 	memcpy(rss_context->rss_key, qp->rss_ctx->rss_key,
2054 	       MLX4_EN_RSS_KEY_SIZE);
2055 }
2056 
2057 static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type,
2058 			       const struct ib_qp_attr *attr, int attr_mask,
2059 			       enum ib_qp_state cur_state,
2060 			       enum ib_qp_state new_state,
2061 			       struct ib_udata *udata)
2062 {
2063 	struct ib_srq  *ibsrq;
2064 	const struct ib_gid_attr *gid_attr = NULL;
2065 	struct ib_rwq_ind_table *rwq_ind_tbl;
2066 	enum ib_qp_type qp_type;
2067 	struct mlx4_ib_dev *dev;
2068 	struct mlx4_ib_qp *qp;
2069 	struct mlx4_ib_pd *pd;
2070 	struct mlx4_ib_cq *send_cq, *recv_cq;
2071 	struct mlx4_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2072 		udata, struct mlx4_ib_ucontext, ibucontext);
2073 	struct mlx4_qp_context *context;
2074 	enum mlx4_qp_optpar optpar = 0;
2075 	int sqd_event;
2076 	int steer_qp = 0;
2077 	int err = -EINVAL;
2078 	int counter_index;
2079 
2080 	if (src_type == MLX4_IB_RWQ_SRC) {
2081 		struct ib_wq *ibwq;
2082 
2083 		ibwq	    = (struct ib_wq *)src;
2084 		ibsrq	    = NULL;
2085 		rwq_ind_tbl = NULL;
2086 		qp_type     = IB_QPT_RAW_PACKET;
2087 		qp	    = to_mqp((struct ib_qp *)ibwq);
2088 		dev	    = to_mdev(ibwq->device);
2089 		pd	    = to_mpd(ibwq->pd);
2090 	} else {
2091 		struct ib_qp *ibqp;
2092 
2093 		ibqp	    = (struct ib_qp *)src;
2094 		ibsrq	    = ibqp->srq;
2095 		rwq_ind_tbl = ibqp->rwq_ind_tbl;
2096 		qp_type     = ibqp->qp_type;
2097 		qp	    = to_mqp(ibqp);
2098 		dev	    = to_mdev(ibqp->device);
2099 		pd	    = get_pd(qp);
2100 	}
2101 
2102 	/* APM is not supported under RoCE */
2103 	if (attr_mask & IB_QP_ALT_PATH &&
2104 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2105 	    IB_LINK_LAYER_ETHERNET)
2106 		return -ENOTSUPP;
2107 
2108 	context = kzalloc(sizeof *context, GFP_KERNEL);
2109 	if (!context)
2110 		return -ENOMEM;
2111 
2112 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
2113 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
2114 
2115 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
2116 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2117 	else {
2118 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
2119 		switch (attr->path_mig_state) {
2120 		case IB_MIG_MIGRATED:
2121 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2122 			break;
2123 		case IB_MIG_REARM:
2124 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
2125 			break;
2126 		case IB_MIG_ARMED:
2127 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
2128 			break;
2129 		}
2130 	}
2131 
2132 	if (qp->inl_recv_sz)
2133 		context->param3 |= cpu_to_be32(1 << 25);
2134 
2135 	if (qp->flags & MLX4_IB_QP_SCATTER_FCS)
2136 		context->param3 |= cpu_to_be32(1 << 29);
2137 
2138 	if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI)
2139 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
2140 	else if (qp_type == IB_QPT_RAW_PACKET)
2141 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
2142 	else if (qp_type == IB_QPT_UD) {
2143 		if (qp->flags & MLX4_IB_QP_LSO)
2144 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
2145 					      ilog2(dev->dev->caps.max_gso_sz);
2146 		else
2147 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 13;
2148 	} else if (attr_mask & IB_QP_PATH_MTU) {
2149 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
2150 			pr_err("path MTU (%u) is invalid\n",
2151 			       attr->path_mtu);
2152 			goto out;
2153 		}
2154 		context->mtu_msgmax = (attr->path_mtu << 5) |
2155 			ilog2(dev->dev->caps.max_msg_sz);
2156 	}
2157 
2158 	if (!rwq_ind_tbl) { /* PRM RSS receive side should be left zeros */
2159 		if (qp->rq.wqe_cnt)
2160 			context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
2161 		context->rq_size_stride |= qp->rq.wqe_shift - 4;
2162 	}
2163 
2164 	if (qp->sq.wqe_cnt)
2165 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
2166 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
2167 
2168 	if (new_state == IB_QPS_RESET && qp->counter_index)
2169 		mlx4_ib_free_qp_counter(dev, qp);
2170 
2171 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2172 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
2173 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
2174 		if (qp_type == IB_QPT_RAW_PACKET)
2175 			context->param3 |= cpu_to_be32(1 << 30);
2176 	}
2177 
2178 	if (ucontext)
2179 		context->usr_page = cpu_to_be32(
2180 			mlx4_to_hw_uar_index(dev->dev, ucontext->uar.index));
2181 	else
2182 		context->usr_page = cpu_to_be32(
2183 			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
2184 
2185 	if (attr_mask & IB_QP_DEST_QPN)
2186 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
2187 
2188 	if (attr_mask & IB_QP_PORT) {
2189 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
2190 		    !(attr_mask & IB_QP_AV)) {
2191 			mlx4_set_sched(&context->pri_path, attr->port_num);
2192 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
2193 		}
2194 	}
2195 
2196 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
2197 		err = create_qp_lb_counter(dev, qp);
2198 		if (err)
2199 			goto out;
2200 
2201 		counter_index =
2202 			dev->counters_table[qp->port - 1].default_counter;
2203 		if (qp->counter_index)
2204 			counter_index = qp->counter_index->index;
2205 
2206 		if (counter_index != -1) {
2207 			context->pri_path.counter_index = counter_index;
2208 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
2209 			if (qp->counter_index) {
2210 				context->pri_path.fl |=
2211 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
2212 				context->pri_path.vlan_control |=
2213 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
2214 			}
2215 		} else
2216 			context->pri_path.counter_index =
2217 				MLX4_SINK_COUNTER_INDEX(dev->dev);
2218 
2219 		if (qp->flags & MLX4_IB_QP_NETIF) {
2220 			mlx4_ib_steer_qp_reg(dev, qp, 1);
2221 			steer_qp = 1;
2222 		}
2223 
2224 		if (qp_type == IB_QPT_GSI) {
2225 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
2226 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
2227 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
2228 
2229 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2230 		}
2231 	}
2232 
2233 	if (attr_mask & IB_QP_PKEY_INDEX) {
2234 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2235 			context->pri_path.disable_pkey_check = 0x40;
2236 		context->pri_path.pkey_index = attr->pkey_index;
2237 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
2238 	}
2239 
2240 	if (attr_mask & IB_QP_AV) {
2241 		u8 port_num = mlx4_is_bonded(dev->dev) ? 1 :
2242 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2243 		u16 vlan = 0xffff;
2244 		u8 smac[ETH_ALEN];
2245 		int is_eth =
2246 			rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
2247 			rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
2248 
2249 		if (is_eth) {
2250 			gid_attr = attr->ah_attr.grh.sgid_attr;
2251 			err = rdma_read_gid_l2_fields(gid_attr, &vlan,
2252 						      &smac[0]);
2253 			if (err)
2254 				goto out;
2255 		}
2256 
2257 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
2258 				  port_num, vlan, smac))
2259 			goto out;
2260 
2261 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
2262 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
2263 
2264 		if (is_eth &&
2265 		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
2266 			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr->gid_type);
2267 
2268 			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
2269 				err = -EINVAL;
2270 				goto out;
2271 			}
2272 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2273 		}
2274 
2275 	}
2276 
2277 	if (attr_mask & IB_QP_TIMEOUT) {
2278 		context->pri_path.ackto |= attr->timeout << 3;
2279 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
2280 	}
2281 
2282 	if (attr_mask & IB_QP_ALT_PATH) {
2283 		if (attr->alt_port_num == 0 ||
2284 		    attr->alt_port_num > dev->dev->caps.num_ports)
2285 			goto out;
2286 
2287 		if (attr->alt_pkey_index >=
2288 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
2289 			goto out;
2290 
2291 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
2292 				      &context->alt_path,
2293 				      attr->alt_port_num))
2294 			goto out;
2295 
2296 		context->alt_path.pkey_index = attr->alt_pkey_index;
2297 		context->alt_path.ackto = attr->alt_timeout << 3;
2298 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
2299 	}
2300 
2301 	context->pd = cpu_to_be32(pd->pdn);
2302 
2303 	if (!rwq_ind_tbl) {
2304 		context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
2305 		get_cqs(qp, src_type, &send_cq, &recv_cq);
2306 	} else { /* Set dummy CQs to be compatible with HV and PRM */
2307 		send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq);
2308 		recv_cq = send_cq;
2309 	}
2310 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
2311 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
2312 
2313 	/* Set "fast registration enabled" for all kernel QPs */
2314 	if (!ucontext)
2315 		context->params1 |= cpu_to_be32(1 << 11);
2316 
2317 	if (attr_mask & IB_QP_RNR_RETRY) {
2318 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2319 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
2320 	}
2321 
2322 	if (attr_mask & IB_QP_RETRY_CNT) {
2323 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2324 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
2325 	}
2326 
2327 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2328 		if (attr->max_rd_atomic)
2329 			context->params1 |=
2330 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2331 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
2332 	}
2333 
2334 	if (attr_mask & IB_QP_SQ_PSN)
2335 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2336 
2337 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2338 		if (attr->max_dest_rd_atomic)
2339 			context->params2 |=
2340 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2341 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
2342 	}
2343 
2344 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
2345 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
2346 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
2347 	}
2348 
2349 	if (ibsrq)
2350 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
2351 
2352 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
2353 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2354 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
2355 	}
2356 	if (attr_mask & IB_QP_RQ_PSN)
2357 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2358 
2359 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
2360 	if (attr_mask & IB_QP_QKEY) {
2361 		if (qp->mlx4_ib_qp_type &
2362 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
2363 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2364 		else {
2365 			if (mlx4_is_mfunc(dev->dev) &&
2366 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
2367 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
2368 			    MLX4_RESERVED_QKEY_BASE) {
2369 				pr_err("Cannot use reserved QKEY"
2370 				       " 0x%x (range 0xffff0000..0xffffffff"
2371 				       " is reserved)\n", attr->qkey);
2372 				err = -EINVAL;
2373 				goto out;
2374 			}
2375 			context->qkey = cpu_to_be32(attr->qkey);
2376 		}
2377 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
2378 	}
2379 
2380 	if (ibsrq)
2381 		context->srqn = cpu_to_be32(1 << 24 |
2382 					    to_msrq(ibsrq)->msrq.srqn);
2383 
2384 	if (qp->rq.wqe_cnt &&
2385 	    cur_state == IB_QPS_RESET &&
2386 	    new_state == IB_QPS_INIT)
2387 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2388 
2389 	if (cur_state == IB_QPS_INIT &&
2390 	    new_state == IB_QPS_RTR  &&
2391 	    (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI ||
2392 	     qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) {
2393 		context->pri_path.sched_queue = (qp->port - 1) << 6;
2394 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2395 		    qp->mlx4_ib_qp_type &
2396 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
2397 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
2398 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
2399 				context->pri_path.fl = 0x80;
2400 		} else {
2401 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2402 				context->pri_path.fl = 0x80;
2403 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
2404 		}
2405 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2406 		    IB_LINK_LAYER_ETHERNET) {
2407 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
2408 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
2409 				context->pri_path.feup = 1 << 7; /* don't fsm */
2410 			/* handle smac_index */
2411 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
2412 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
2413 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
2414 				err = handle_eth_ud_smac_index(dev, qp, context);
2415 				if (err) {
2416 					err = -EINVAL;
2417 					goto out;
2418 				}
2419 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
2420 					dev->qp1_proxy[qp->port - 1] = qp;
2421 			}
2422 		}
2423 	}
2424 
2425 	if (qp_type == IB_QPT_RAW_PACKET) {
2426 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
2427 					MLX4_IB_LINK_TYPE_ETH;
2428 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
2429 			/* set QP to receive both tunneled & non-tunneled packets */
2430 			if (!rwq_ind_tbl)
2431 				context->srqn = cpu_to_be32(7 << 28);
2432 		}
2433 	}
2434 
2435 	if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
2436 		int is_eth = rdma_port_get_link_layer(
2437 				&dev->ib_dev, qp->port) ==
2438 				IB_LINK_LAYER_ETHERNET;
2439 		if (is_eth) {
2440 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
2441 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
2442 		}
2443 	}
2444 
2445 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2446 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2447 		sqd_event = 1;
2448 	else
2449 		sqd_event = 0;
2450 
2451 	if (!ucontext &&
2452 	    cur_state == IB_QPS_RESET &&
2453 	    new_state == IB_QPS_INIT)
2454 		context->rlkey_roce_mode |= (1 << 4);
2455 
2456 	/*
2457 	 * Before passing a kernel QP to the HW, make sure that the
2458 	 * ownership bits of the send queue are set and the SQ
2459 	 * headroom is stamped so that the hardware doesn't start
2460 	 * processing stale work requests.
2461 	 */
2462 	if (!ucontext &&
2463 	    cur_state == IB_QPS_RESET &&
2464 	    new_state == IB_QPS_INIT) {
2465 		struct mlx4_wqe_ctrl_seg *ctrl;
2466 		int i;
2467 
2468 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
2469 			ctrl = get_send_wqe(qp, i);
2470 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
2471 			ctrl->qpn_vlan.fence_size =
2472 				1 << (qp->sq.wqe_shift - 4);
2473 			stamp_send_wqe(qp, i);
2474 		}
2475 	}
2476 
2477 	if (rwq_ind_tbl	&&
2478 	    cur_state == IB_QPS_RESET &&
2479 	    new_state == IB_QPS_INIT) {
2480 		fill_qp_rss_context(context, qp);
2481 		context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET);
2482 	}
2483 
2484 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2485 			     to_mlx4_state(new_state), context, optpar,
2486 			     sqd_event, &qp->mqp);
2487 	if (err)
2488 		goto out;
2489 
2490 	qp->state = new_state;
2491 
2492 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2493 		qp->atomic_rd_en = attr->qp_access_flags;
2494 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2495 		qp->resp_depth = attr->max_dest_rd_atomic;
2496 	if (attr_mask & IB_QP_PORT) {
2497 		qp->port = attr->port_num;
2498 		update_mcg_macs(dev, qp);
2499 	}
2500 	if (attr_mask & IB_QP_ALT_PATH)
2501 		qp->alt_port = attr->alt_port_num;
2502 
2503 	if (is_sqp(dev, qp))
2504 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2505 
2506 	/*
2507 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
2508 	 * QP0 to RESET or ERROR, bring the link back down.
2509 	 */
2510 	if (is_qp0(dev, qp)) {
2511 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2512 			if (mlx4_INIT_PORT(dev->dev, qp->port))
2513 				pr_warn("INIT_PORT failed for port %d\n",
2514 				       qp->port);
2515 
2516 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2517 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2518 			mlx4_CLOSE_PORT(dev->dev, qp->port);
2519 	}
2520 
2521 	/*
2522 	 * If we moved a kernel QP to RESET, clean up all old CQ
2523 	 * entries and reinitialize the QP.
2524 	 */
2525 	if (new_state == IB_QPS_RESET) {
2526 		if (!ucontext) {
2527 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2528 					 ibsrq ? to_msrq(ibsrq) : NULL);
2529 			if (send_cq != recv_cq)
2530 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2531 
2532 			qp->rq.head = 0;
2533 			qp->rq.tail = 0;
2534 			qp->sq.head = 0;
2535 			qp->sq.tail = 0;
2536 			qp->sq_next_wqe = 0;
2537 			if (qp->rq.wqe_cnt)
2538 				*qp->db.db  = 0;
2539 
2540 			if (qp->flags & MLX4_IB_QP_NETIF)
2541 				mlx4_ib_steer_qp_reg(dev, qp, 0);
2542 		}
2543 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2544 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2545 			qp->pri.smac = 0;
2546 			qp->pri.smac_port = 0;
2547 		}
2548 		if (qp->alt.smac) {
2549 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2550 			qp->alt.smac = 0;
2551 		}
2552 		if (qp->pri.vid < 0x1000) {
2553 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2554 			qp->pri.vid = 0xFFFF;
2555 			qp->pri.candidate_vid = 0xFFFF;
2556 			qp->pri.update_vid = 0;
2557 		}
2558 
2559 		if (qp->alt.vid < 0x1000) {
2560 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2561 			qp->alt.vid = 0xFFFF;
2562 			qp->alt.candidate_vid = 0xFFFF;
2563 			qp->alt.update_vid = 0;
2564 		}
2565 	}
2566 out:
2567 	if (err && qp->counter_index)
2568 		mlx4_ib_free_qp_counter(dev, qp);
2569 	if (err && steer_qp)
2570 		mlx4_ib_steer_qp_reg(dev, qp, 0);
2571 	kfree(context);
2572 	if (qp->pri.candidate_smac ||
2573 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2574 		if (err) {
2575 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2576 		} else {
2577 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2578 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2579 			qp->pri.smac = qp->pri.candidate_smac;
2580 			qp->pri.smac_index = qp->pri.candidate_smac_index;
2581 			qp->pri.smac_port = qp->pri.candidate_smac_port;
2582 		}
2583 		qp->pri.candidate_smac = 0;
2584 		qp->pri.candidate_smac_index = 0;
2585 		qp->pri.candidate_smac_port = 0;
2586 	}
2587 	if (qp->alt.candidate_smac) {
2588 		if (err) {
2589 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2590 		} else {
2591 			if (qp->alt.smac)
2592 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2593 			qp->alt.smac = qp->alt.candidate_smac;
2594 			qp->alt.smac_index = qp->alt.candidate_smac_index;
2595 			qp->alt.smac_port = qp->alt.candidate_smac_port;
2596 		}
2597 		qp->alt.candidate_smac = 0;
2598 		qp->alt.candidate_smac_index = 0;
2599 		qp->alt.candidate_smac_port = 0;
2600 	}
2601 
2602 	if (qp->pri.update_vid) {
2603 		if (err) {
2604 			if (qp->pri.candidate_vid < 0x1000)
2605 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2606 						     qp->pri.candidate_vid);
2607 		} else {
2608 			if (qp->pri.vid < 0x1000)
2609 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2610 						     qp->pri.vid);
2611 			qp->pri.vid = qp->pri.candidate_vid;
2612 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2613 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
2614 		}
2615 		qp->pri.candidate_vid = 0xFFFF;
2616 		qp->pri.update_vid = 0;
2617 	}
2618 
2619 	if (qp->alt.update_vid) {
2620 		if (err) {
2621 			if (qp->alt.candidate_vid < 0x1000)
2622 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2623 						     qp->alt.candidate_vid);
2624 		} else {
2625 			if (qp->alt.vid < 0x1000)
2626 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2627 						     qp->alt.vid);
2628 			qp->alt.vid = qp->alt.candidate_vid;
2629 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2630 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2631 		}
2632 		qp->alt.candidate_vid = 0xFFFF;
2633 		qp->alt.update_vid = 0;
2634 	}
2635 
2636 	return err;
2637 }
2638 
2639 enum {
2640 	MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK = (IB_QP_STATE	|
2641 					      IB_QP_PORT),
2642 };
2643 
2644 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2645 			      int attr_mask, struct ib_udata *udata)
2646 {
2647 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2648 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2649 	enum ib_qp_state cur_state, new_state;
2650 	int err = -EINVAL;
2651 	mutex_lock(&qp->mutex);
2652 
2653 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2654 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2655 
2656 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2657 				attr_mask)) {
2658 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2659 			 "for transition %d to %d. qp_type %d,"
2660 			 " attr_mask 0x%x\n",
2661 			 ibqp->qp_num, cur_state, new_state,
2662 			 ibqp->qp_type, attr_mask);
2663 		goto out;
2664 	}
2665 
2666 	if (ibqp->rwq_ind_tbl) {
2667 		if (!(((cur_state == IB_QPS_RESET) &&
2668 		       (new_state == IB_QPS_INIT)) ||
2669 		      ((cur_state == IB_QPS_INIT)  &&
2670 		       (new_state == IB_QPS_RTR)))) {
2671 			pr_debug("qpn 0x%x: RSS QP unsupported transition %d to %d\n",
2672 				 ibqp->qp_num, cur_state, new_state);
2673 
2674 			err = -EOPNOTSUPP;
2675 			goto out;
2676 		}
2677 
2678 		if (attr_mask & ~MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK) {
2679 			pr_debug("qpn 0x%x: RSS QP unsupported attribute mask 0x%x for transition %d to %d\n",
2680 				 ibqp->qp_num, attr_mask, cur_state, new_state);
2681 
2682 			err = -EOPNOTSUPP;
2683 			goto out;
2684 		}
2685 	}
2686 
2687 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2688 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2689 			if ((ibqp->qp_type == IB_QPT_RC) ||
2690 			    (ibqp->qp_type == IB_QPT_UD) ||
2691 			    (ibqp->qp_type == IB_QPT_UC) ||
2692 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2693 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2694 				attr->port_num = mlx4_ib_bond_next_port(dev);
2695 			}
2696 		} else {
2697 			/* no sense in changing port_num
2698 			 * when ports are bonded */
2699 			attr_mask &= ~IB_QP_PORT;
2700 		}
2701 	}
2702 
2703 	if ((attr_mask & IB_QP_PORT) &&
2704 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2705 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2706 			 "for transition %d to %d. qp_type %d\n",
2707 			 ibqp->qp_num, attr->port_num, cur_state,
2708 			 new_state, ibqp->qp_type);
2709 		goto out;
2710 	}
2711 
2712 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2713 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2714 	     IB_LINK_LAYER_ETHERNET))
2715 		goto out;
2716 
2717 	if (attr_mask & IB_QP_PKEY_INDEX) {
2718 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2719 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2720 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2721 				 "for transition %d to %d. qp_type %d\n",
2722 				 ibqp->qp_num, attr->pkey_index, cur_state,
2723 				 new_state, ibqp->qp_type);
2724 			goto out;
2725 		}
2726 	}
2727 
2728 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2729 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2730 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2731 			 "Transition %d to %d. qp_type %d\n",
2732 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2733 			 new_state, ibqp->qp_type);
2734 		goto out;
2735 	}
2736 
2737 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2738 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2739 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2740 			 "Transition %d to %d. qp_type %d\n",
2741 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2742 			 new_state, ibqp->qp_type);
2743 		goto out;
2744 	}
2745 
2746 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2747 		err = 0;
2748 		goto out;
2749 	}
2750 
2751 	if (ibqp->rwq_ind_tbl && (new_state == IB_QPS_INIT)) {
2752 		err = bringup_rss_rwqs(ibqp->rwq_ind_tbl, attr->port_num,
2753 				       udata);
2754 		if (err)
2755 			goto out;
2756 	}
2757 
2758 	err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask,
2759 				  cur_state, new_state, udata);
2760 
2761 	if (ibqp->rwq_ind_tbl && err)
2762 		bring_down_rss_rwqs(ibqp->rwq_ind_tbl, udata);
2763 
2764 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2765 		attr->port_num = 1;
2766 
2767 out:
2768 	mutex_unlock(&qp->mutex);
2769 	return err;
2770 }
2771 
2772 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2773 		      int attr_mask, struct ib_udata *udata)
2774 {
2775 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2776 	int ret;
2777 
2778 	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2779 
2780 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2781 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
2782 		int err = 0;
2783 
2784 		if (sqp->roce_v2_gsi)
2785 			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2786 		if (err)
2787 			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2788 			       err);
2789 	}
2790 	return ret;
2791 }
2792 
2793 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2794 {
2795 	int i;
2796 	for (i = 0; i < dev->caps.num_ports; i++) {
2797 		if (qpn == dev->caps.spec_qps[i].qp0_proxy ||
2798 		    qpn == dev->caps.spec_qps[i].qp0_tunnel) {
2799 			*qkey = dev->caps.spec_qps[i].qp0_qkey;
2800 			return 0;
2801 		}
2802 	}
2803 	return -EINVAL;
2804 }
2805 
2806 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2807 				  const struct ib_ud_wr *wr,
2808 				  void *wqe, unsigned *mlx_seg_len)
2809 {
2810 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2811 	struct ib_device *ib_dev = &mdev->ib_dev;
2812 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2813 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2814 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2815 	u16 pkey;
2816 	u32 qkey;
2817 	int send_size;
2818 	int header_size;
2819 	int spc;
2820 	int i;
2821 
2822 	if (wr->wr.opcode != IB_WR_SEND)
2823 		return -EINVAL;
2824 
2825 	send_size = 0;
2826 
2827 	for (i = 0; i < wr->wr.num_sge; ++i)
2828 		send_size += wr->wr.sg_list[i].length;
2829 
2830 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2831 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2832 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2833 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2834 
2835 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2836 
2837 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2838 		sqp->ud_header.lrh.service_level =
2839 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2840 		sqp->ud_header.lrh.destination_lid =
2841 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2842 		sqp->ud_header.lrh.source_lid =
2843 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2844 	}
2845 
2846 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2847 
2848 	/* force loopback */
2849 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2850 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2851 
2852 	sqp->ud_header.lrh.virtual_lane    = 0;
2853 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2854 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2855 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2856 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2857 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2858 	else
2859 		sqp->ud_header.bth.destination_qpn =
2860 			cpu_to_be32(mdev->dev->caps.spec_qps[sqp->qp.port - 1].qp0_tunnel);
2861 
2862 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2863 	if (mlx4_is_master(mdev->dev)) {
2864 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2865 			return -EINVAL;
2866 	} else {
2867 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2868 			return -EINVAL;
2869 	}
2870 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2871 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2872 
2873 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2874 	sqp->ud_header.immediate_present = 0;
2875 
2876 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2877 
2878 	/*
2879 	 * Inline data segments may not cross a 64 byte boundary.  If
2880 	 * our UD header is bigger than the space available up to the
2881 	 * next 64 byte boundary in the WQE, use two inline data
2882 	 * segments to hold the UD header.
2883 	 */
2884 	spc = MLX4_INLINE_ALIGN -
2885 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2886 	if (header_size <= spc) {
2887 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2888 		memcpy(inl + 1, sqp->header_buf, header_size);
2889 		i = 1;
2890 	} else {
2891 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2892 		memcpy(inl + 1, sqp->header_buf, spc);
2893 
2894 		inl = (void *) (inl + 1) + spc;
2895 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2896 		/*
2897 		 * Need a barrier here to make sure all the data is
2898 		 * visible before the byte_count field is set.
2899 		 * Otherwise the HCA prefetcher could grab the 64-byte
2900 		 * chunk with this inline segment and get a valid (!=
2901 		 * 0xffffffff) byte count but stale data, and end up
2902 		 * generating a packet with bad headers.
2903 		 *
2904 		 * The first inline segment's byte_count field doesn't
2905 		 * need a barrier, because it comes after a
2906 		 * control/MLX segment and therefore is at an offset
2907 		 * of 16 mod 64.
2908 		 */
2909 		wmb();
2910 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2911 		i = 2;
2912 	}
2913 
2914 	*mlx_seg_len =
2915 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2916 	return 0;
2917 }
2918 
2919 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
2920 {
2921 	union sl2vl_tbl_to_u64 tmp_vltab;
2922 	u8 vl;
2923 
2924 	if (sl > 15)
2925 		return 0xf;
2926 	tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
2927 	vl = tmp_vltab.sl8[sl >> 1];
2928 	if (sl & 1)
2929 		vl &= 0x0f;
2930 	else
2931 		vl >>= 4;
2932 	return vl;
2933 }
2934 
2935 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
2936 				int index, union ib_gid *gid,
2937 				enum ib_gid_type *gid_type)
2938 {
2939 	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
2940 	struct mlx4_port_gid_table *port_gid_table;
2941 	unsigned long flags;
2942 
2943 	port_gid_table = &iboe->gids[port_num - 1];
2944 	spin_lock_irqsave(&iboe->lock, flags);
2945 	memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
2946 	*gid_type = port_gid_table->gids[index].gid_type;
2947 	spin_unlock_irqrestore(&iboe->lock, flags);
2948 	if (rdma_is_zero_gid(gid))
2949 		return -ENOENT;
2950 
2951 	return 0;
2952 }
2953 
2954 #define MLX4_ROCEV2_QP1_SPORT 0xC000
2955 static int build_mlx_header(struct mlx4_ib_sqp *sqp, const struct ib_ud_wr *wr,
2956 			    void *wqe, unsigned *mlx_seg_len)
2957 {
2958 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
2959 	struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
2960 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2961 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2962 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2963 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2964 	union ib_gid sgid;
2965 	u16 pkey;
2966 	int send_size;
2967 	int header_size;
2968 	int spc;
2969 	int i;
2970 	int err = 0;
2971 	u16 vlan = 0xffff;
2972 	bool is_eth;
2973 	bool is_vlan = false;
2974 	bool is_grh;
2975 	bool is_udp = false;
2976 	int ip_version = 0;
2977 
2978 	send_size = 0;
2979 	for (i = 0; i < wr->wr.num_sge; ++i)
2980 		send_size += wr->wr.sg_list[i].length;
2981 
2982 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2983 	is_grh = mlx4_ib_ah_grh_present(ah);
2984 	if (is_eth) {
2985 		enum ib_gid_type gid_type;
2986 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2987 			/* When multi-function is enabled, the ib_core gid
2988 			 * indexes don't necessarily match the hw ones, so
2989 			 * we must use our own cache */
2990 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2991 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
2992 							   ah->av.ib.gid_index, &sgid.raw[0]);
2993 			if (err)
2994 				return err;
2995 		} else  {
2996 			err = fill_gid_by_hw_index(ibdev, sqp->qp.port,
2997 					    ah->av.ib.gid_index,
2998 					    &sgid, &gid_type);
2999 			if (!err) {
3000 				is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
3001 				if (is_udp) {
3002 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
3003 						ip_version = 4;
3004 					else
3005 						ip_version = 6;
3006 					is_grh = false;
3007 				}
3008 			} else {
3009 				return err;
3010 			}
3011 		}
3012 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
3013 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
3014 			is_vlan = 1;
3015 		}
3016 	}
3017 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
3018 			  ip_version, is_udp, 0, &sqp->ud_header);
3019 	if (err)
3020 		return err;
3021 
3022 	if (!is_eth) {
3023 		sqp->ud_header.lrh.service_level =
3024 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
3025 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
3026 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
3027 	}
3028 
3029 	if (is_grh || (ip_version == 6)) {
3030 		sqp->ud_header.grh.traffic_class =
3031 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3032 		sqp->ud_header.grh.flow_label    =
3033 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
3034 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
3035 		if (is_eth) {
3036 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
3037 		} else {
3038 			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3039 				/* When multi-function is enabled, the ib_core gid
3040 				 * indexes don't necessarily match the hw ones, so
3041 				 * we must use our own cache
3042 				 */
3043 				sqp->ud_header.grh.source_gid.global.subnet_prefix =
3044 					cpu_to_be64(atomic64_read(&(to_mdev(ib_dev)->sriov.
3045 								    demux[sqp->qp.port - 1].
3046 								    subnet_prefix)));
3047 				sqp->ud_header.grh.source_gid.global.interface_id =
3048 					to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
3049 						       guid_cache[ah->av.ib.gid_index];
3050 			} else {
3051 				sqp->ud_header.grh.source_gid =
3052 					ah->ibah.sgid_attr->gid;
3053 			}
3054 		}
3055 		memcpy(sqp->ud_header.grh.destination_gid.raw,
3056 		       ah->av.ib.dgid, 16);
3057 	}
3058 
3059 	if (ip_version == 4) {
3060 		sqp->ud_header.ip4.tos =
3061 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3062 		sqp->ud_header.ip4.id = 0;
3063 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
3064 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
3065 
3066 		memcpy(&sqp->ud_header.ip4.saddr,
3067 		       sgid.raw + 12, 4);
3068 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
3069 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
3070 	}
3071 
3072 	if (is_udp) {
3073 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
3074 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
3075 		sqp->ud_header.udp.csum = 0;
3076 	}
3077 
3078 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
3079 
3080 	if (!is_eth) {
3081 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
3082 					  (sqp->ud_header.lrh.destination_lid ==
3083 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
3084 					  (sqp->ud_header.lrh.service_level << 8));
3085 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
3086 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
3087 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
3088 	}
3089 
3090 	switch (wr->wr.opcode) {
3091 	case IB_WR_SEND:
3092 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
3093 		sqp->ud_header.immediate_present = 0;
3094 		break;
3095 	case IB_WR_SEND_WITH_IMM:
3096 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
3097 		sqp->ud_header.immediate_present = 1;
3098 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
3099 		break;
3100 	default:
3101 		return -EINVAL;
3102 	}
3103 
3104 	if (is_eth) {
3105 		struct in6_addr in6;
3106 		u16 ether_type;
3107 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
3108 
3109 		ether_type = (!is_udp) ? ETH_P_IBOE:
3110 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
3111 
3112 		mlx->sched_prio = cpu_to_be16(pcp);
3113 
3114 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
3115 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
3116 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
3117 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
3118 		memcpy(&in6, sgid.raw, sizeof(in6));
3119 
3120 
3121 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
3122 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
3123 		if (!is_vlan) {
3124 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
3125 		} else {
3126 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
3127 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
3128 		}
3129 	} else {
3130 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 :
3131 							sl_to_vl(to_mdev(ib_dev),
3132 								 sqp->ud_header.lrh.service_level,
3133 								 sqp->qp.port);
3134 		if (sqp->qp.ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
3135 			return -EINVAL;
3136 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
3137 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
3138 	}
3139 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
3140 	if (!sqp->qp.ibqp.qp_num)
3141 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
3142 	else
3143 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
3144 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
3145 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
3146 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
3147 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
3148 					       sqp->qkey : wr->remote_qkey);
3149 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
3150 
3151 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3152 
3153 	if (0) {
3154 		pr_err("built UD header of size %d:\n", header_size);
3155 		for (i = 0; i < header_size / 4; ++i) {
3156 			if (i % 8 == 0)
3157 				pr_err("  [%02x] ", i * 4);
3158 			pr_cont(" %08x",
3159 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
3160 			if ((i + 1) % 8 == 0)
3161 				pr_cont("\n");
3162 		}
3163 		pr_err("\n");
3164 	}
3165 
3166 	/*
3167 	 * Inline data segments may not cross a 64 byte boundary.  If
3168 	 * our UD header is bigger than the space available up to the
3169 	 * next 64 byte boundary in the WQE, use two inline data
3170 	 * segments to hold the UD header.
3171 	 */
3172 	spc = MLX4_INLINE_ALIGN -
3173 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3174 	if (header_size <= spc) {
3175 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3176 		memcpy(inl + 1, sqp->header_buf, header_size);
3177 		i = 1;
3178 	} else {
3179 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3180 		memcpy(inl + 1, sqp->header_buf, spc);
3181 
3182 		inl = (void *) (inl + 1) + spc;
3183 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3184 		/*
3185 		 * Need a barrier here to make sure all the data is
3186 		 * visible before the byte_count field is set.
3187 		 * Otherwise the HCA prefetcher could grab the 64-byte
3188 		 * chunk with this inline segment and get a valid (!=
3189 		 * 0xffffffff) byte count but stale data, and end up
3190 		 * generating a packet with bad headers.
3191 		 *
3192 		 * The first inline segment's byte_count field doesn't
3193 		 * need a barrier, because it comes after a
3194 		 * control/MLX segment and therefore is at an offset
3195 		 * of 16 mod 64.
3196 		 */
3197 		wmb();
3198 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3199 		i = 2;
3200 	}
3201 
3202 	*mlx_seg_len =
3203 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3204 	return 0;
3205 }
3206 
3207 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3208 {
3209 	unsigned cur;
3210 	struct mlx4_ib_cq *cq;
3211 
3212 	cur = wq->head - wq->tail;
3213 	if (likely(cur + nreq < wq->max_post))
3214 		return 0;
3215 
3216 	cq = to_mcq(ib_cq);
3217 	spin_lock(&cq->lock);
3218 	cur = wq->head - wq->tail;
3219 	spin_unlock(&cq->lock);
3220 
3221 	return cur + nreq >= wq->max_post;
3222 }
3223 
3224 static __be32 convert_access(int acc)
3225 {
3226 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
3227 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
3228 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
3229 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
3230 	       (acc & IB_ACCESS_REMOTE_READ   ?
3231 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
3232 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
3233 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
3234 }
3235 
3236 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
3237 			const struct ib_reg_wr *wr)
3238 {
3239 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
3240 
3241 	fseg->flags		= convert_access(wr->access);
3242 	fseg->mem_key		= cpu_to_be32(wr->key);
3243 	fseg->buf_list		= cpu_to_be64(mr->page_map);
3244 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
3245 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
3246 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
3247 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
3248 	fseg->reserved[0]	= 0;
3249 	fseg->reserved[1]	= 0;
3250 }
3251 
3252 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
3253 {
3254 	memset(iseg, 0, sizeof(*iseg));
3255 	iseg->mem_key = cpu_to_be32(rkey);
3256 }
3257 
3258 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
3259 					  u64 remote_addr, u32 rkey)
3260 {
3261 	rseg->raddr    = cpu_to_be64(remote_addr);
3262 	rseg->rkey     = cpu_to_be32(rkey);
3263 	rseg->reserved = 0;
3264 }
3265 
3266 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
3267 			   const struct ib_atomic_wr *wr)
3268 {
3269 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
3270 		aseg->swap_add = cpu_to_be64(wr->swap);
3271 		aseg->compare  = cpu_to_be64(wr->compare_add);
3272 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
3273 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3274 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
3275 	} else {
3276 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3277 		aseg->compare  = 0;
3278 	}
3279 
3280 }
3281 
3282 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
3283 				  const struct ib_atomic_wr *wr)
3284 {
3285 	aseg->swap_add		= cpu_to_be64(wr->swap);
3286 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
3287 	aseg->compare		= cpu_to_be64(wr->compare_add);
3288 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
3289 }
3290 
3291 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
3292 			     const struct ib_ud_wr *wr)
3293 {
3294 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
3295 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
3296 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
3297 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
3298 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
3299 }
3300 
3301 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
3302 				    struct mlx4_wqe_datagram_seg *dseg,
3303 				    const struct ib_ud_wr *wr,
3304 				    enum mlx4_ib_qp_type qpt)
3305 {
3306 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
3307 	struct mlx4_av sqp_av = {0};
3308 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
3309 
3310 	/* force loopback */
3311 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
3312 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
3313 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
3314 			cpu_to_be32(0xf0000000);
3315 
3316 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
3317 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
3318 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp1_tunnel);
3319 	else
3320 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp0_tunnel);
3321 	/* Use QKEY from the QP context, which is set by master */
3322 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
3323 }
3324 
3325 static void build_tunnel_header(const struct ib_ud_wr *wr, void *wqe,
3326 				unsigned *mlx_seg_len)
3327 {
3328 	struct mlx4_wqe_inline_seg *inl = wqe;
3329 	struct mlx4_ib_tunnel_header hdr;
3330 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3331 	int spc;
3332 	int i;
3333 
3334 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
3335 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
3336 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
3337 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
3338 	memcpy(hdr.mac, ah->av.eth.mac, 6);
3339 	hdr.vlan = ah->av.eth.vlan;
3340 
3341 	spc = MLX4_INLINE_ALIGN -
3342 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3343 	if (sizeof (hdr) <= spc) {
3344 		memcpy(inl + 1, &hdr, sizeof (hdr));
3345 		wmb();
3346 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
3347 		i = 1;
3348 	} else {
3349 		memcpy(inl + 1, &hdr, spc);
3350 		wmb();
3351 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3352 
3353 		inl = (void *) (inl + 1) + spc;
3354 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
3355 		wmb();
3356 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
3357 		i = 2;
3358 	}
3359 
3360 	*mlx_seg_len =
3361 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
3362 }
3363 
3364 static void set_mlx_icrc_seg(void *dseg)
3365 {
3366 	u32 *t = dseg;
3367 	struct mlx4_wqe_inline_seg *iseg = dseg;
3368 
3369 	t[1] = 0;
3370 
3371 	/*
3372 	 * Need a barrier here before writing the byte_count field to
3373 	 * make sure that all the data is visible before the
3374 	 * byte_count field is set.  Otherwise, if the segment begins
3375 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3376 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3377 	 * stale data, and end up sending the wrong data.
3378 	 */
3379 	wmb();
3380 
3381 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
3382 }
3383 
3384 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3385 {
3386 	dseg->lkey       = cpu_to_be32(sg->lkey);
3387 	dseg->addr       = cpu_to_be64(sg->addr);
3388 
3389 	/*
3390 	 * Need a barrier here before writing the byte_count field to
3391 	 * make sure that all the data is visible before the
3392 	 * byte_count field is set.  Otherwise, if the segment begins
3393 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3394 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3395 	 * stale data, and end up sending the wrong data.
3396 	 */
3397 	wmb();
3398 
3399 	dseg->byte_count = cpu_to_be32(sg->length);
3400 }
3401 
3402 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3403 {
3404 	dseg->byte_count = cpu_to_be32(sg->length);
3405 	dseg->lkey       = cpu_to_be32(sg->lkey);
3406 	dseg->addr       = cpu_to_be64(sg->addr);
3407 }
3408 
3409 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe,
3410 			 const struct ib_ud_wr *wr, struct mlx4_ib_qp *qp,
3411 			 unsigned *lso_seg_len, __be32 *lso_hdr_sz, __be32 *blh)
3412 {
3413 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
3414 
3415 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
3416 		*blh = cpu_to_be32(1 << 6);
3417 
3418 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
3419 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
3420 		return -EINVAL;
3421 
3422 	memcpy(wqe->header, wr->header, wr->hlen);
3423 
3424 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
3425 	*lso_seg_len = halign;
3426 	return 0;
3427 }
3428 
3429 static __be32 send_ieth(const struct ib_send_wr *wr)
3430 {
3431 	switch (wr->opcode) {
3432 	case IB_WR_SEND_WITH_IMM:
3433 	case IB_WR_RDMA_WRITE_WITH_IMM:
3434 		return wr->ex.imm_data;
3435 
3436 	case IB_WR_SEND_WITH_INV:
3437 		return cpu_to_be32(wr->ex.invalidate_rkey);
3438 
3439 	default:
3440 		return 0;
3441 	}
3442 }
3443 
3444 static void add_zero_len_inline(void *wqe)
3445 {
3446 	struct mlx4_wqe_inline_seg *inl = wqe;
3447 	memset(wqe, 0, 16);
3448 	inl->byte_count = cpu_to_be32(1 << 31);
3449 }
3450 
3451 static int _mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3452 			      const struct ib_send_wr **bad_wr, bool drain)
3453 {
3454 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3455 	void *wqe;
3456 	struct mlx4_wqe_ctrl_seg *ctrl;
3457 	struct mlx4_wqe_data_seg *dseg;
3458 	unsigned long flags;
3459 	int nreq;
3460 	int err = 0;
3461 	unsigned ind;
3462 	int uninitialized_var(size);
3463 	unsigned uninitialized_var(seglen);
3464 	__be32 dummy;
3465 	__be32 *lso_wqe;
3466 	__be32 uninitialized_var(lso_hdr_sz);
3467 	__be32 blh;
3468 	int i;
3469 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3470 
3471 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
3472 		struct mlx4_ib_sqp *sqp = to_msqp(qp);
3473 
3474 		if (sqp->roce_v2_gsi) {
3475 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
3476 			enum ib_gid_type gid_type;
3477 			union ib_gid gid;
3478 
3479 			if (!fill_gid_by_hw_index(mdev, sqp->qp.port,
3480 					   ah->av.ib.gid_index,
3481 					   &gid, &gid_type))
3482 				qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
3483 						to_mqp(sqp->roce_v2_gsi) : qp;
3484 			else
3485 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
3486 				       ah->av.ib.gid_index);
3487 		}
3488 	}
3489 
3490 	spin_lock_irqsave(&qp->sq.lock, flags);
3491 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3492 	    !drain) {
3493 		err = -EIO;
3494 		*bad_wr = wr;
3495 		nreq = 0;
3496 		goto out;
3497 	}
3498 
3499 	ind = qp->sq_next_wqe;
3500 
3501 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3502 		lso_wqe = &dummy;
3503 		blh = 0;
3504 
3505 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
3506 			err = -ENOMEM;
3507 			*bad_wr = wr;
3508 			goto out;
3509 		}
3510 
3511 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3512 			err = -EINVAL;
3513 			*bad_wr = wr;
3514 			goto out;
3515 		}
3516 
3517 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3518 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3519 
3520 		ctrl->srcrb_flags =
3521 			(wr->send_flags & IB_SEND_SIGNALED ?
3522 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3523 			(wr->send_flags & IB_SEND_SOLICITED ?
3524 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3525 			((wr->send_flags & IB_SEND_IP_CSUM) ?
3526 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3527 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3528 			qp->sq_signal_bits;
3529 
3530 		ctrl->imm = send_ieth(wr);
3531 
3532 		wqe += sizeof *ctrl;
3533 		size = sizeof *ctrl / 16;
3534 
3535 		switch (qp->mlx4_ib_qp_type) {
3536 		case MLX4_IB_QPT_RC:
3537 		case MLX4_IB_QPT_UC:
3538 			switch (wr->opcode) {
3539 			case IB_WR_ATOMIC_CMP_AND_SWP:
3540 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3541 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3542 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3543 					      atomic_wr(wr)->rkey);
3544 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3545 
3546 				set_atomic_seg(wqe, atomic_wr(wr));
3547 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
3548 
3549 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3550 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3551 
3552 				break;
3553 
3554 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3555 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3556 					      atomic_wr(wr)->rkey);
3557 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3558 
3559 				set_masked_atomic_seg(wqe, atomic_wr(wr));
3560 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
3561 
3562 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3563 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3564 
3565 				break;
3566 
3567 			case IB_WR_RDMA_READ:
3568 			case IB_WR_RDMA_WRITE:
3569 			case IB_WR_RDMA_WRITE_WITH_IMM:
3570 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3571 					      rdma_wr(wr)->rkey);
3572 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3573 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3574 				break;
3575 
3576 			case IB_WR_LOCAL_INV:
3577 				ctrl->srcrb_flags |=
3578 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3579 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3580 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3581 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3582 				break;
3583 
3584 			case IB_WR_REG_MR:
3585 				ctrl->srcrb_flags |=
3586 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3587 				set_reg_seg(wqe, reg_wr(wr));
3588 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3589 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3590 				break;
3591 
3592 			default:
3593 				/* No extra segments required for sends */
3594 				break;
3595 			}
3596 			break;
3597 
3598 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3599 			err =  build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3600 					ctrl, &seglen);
3601 			if (unlikely(err)) {
3602 				*bad_wr = wr;
3603 				goto out;
3604 			}
3605 			wqe  += seglen;
3606 			size += seglen / 16;
3607 			break;
3608 		case MLX4_IB_QPT_TUN_SMI:
3609 		case MLX4_IB_QPT_TUN_GSI:
3610 			/* this is a UD qp used in MAD responses to slaves. */
3611 			set_datagram_seg(wqe, ud_wr(wr));
3612 			/* set the forced-loopback bit in the data seg av */
3613 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3614 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3615 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3616 			break;
3617 		case MLX4_IB_QPT_UD:
3618 			set_datagram_seg(wqe, ud_wr(wr));
3619 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3620 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3621 
3622 			if (wr->opcode == IB_WR_LSO) {
3623 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3624 						&lso_hdr_sz, &blh);
3625 				if (unlikely(err)) {
3626 					*bad_wr = wr;
3627 					goto out;
3628 				}
3629 				lso_wqe = (__be32 *) wqe;
3630 				wqe  += seglen;
3631 				size += seglen / 16;
3632 			}
3633 			break;
3634 
3635 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3636 			err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3637 					ctrl, &seglen);
3638 			if (unlikely(err)) {
3639 				*bad_wr = wr;
3640 				goto out;
3641 			}
3642 			wqe  += seglen;
3643 			size += seglen / 16;
3644 			/* to start tunnel header on a cache-line boundary */
3645 			add_zero_len_inline(wqe);
3646 			wqe += 16;
3647 			size++;
3648 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3649 			wqe  += seglen;
3650 			size += seglen / 16;
3651 			break;
3652 		case MLX4_IB_QPT_PROXY_SMI:
3653 		case MLX4_IB_QPT_PROXY_GSI:
3654 			/* If we are tunneling special qps, this is a UD qp.
3655 			 * In this case we first add a UD segment targeting
3656 			 * the tunnel qp, and then add a header with address
3657 			 * information */
3658 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3659 						ud_wr(wr),
3660 						qp->mlx4_ib_qp_type);
3661 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3662 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3663 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3664 			wqe  += seglen;
3665 			size += seglen / 16;
3666 			break;
3667 
3668 		case MLX4_IB_QPT_SMI:
3669 		case MLX4_IB_QPT_GSI:
3670 			err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
3671 					&seglen);
3672 			if (unlikely(err)) {
3673 				*bad_wr = wr;
3674 				goto out;
3675 			}
3676 			wqe  += seglen;
3677 			size += seglen / 16;
3678 			break;
3679 
3680 		default:
3681 			break;
3682 		}
3683 
3684 		/*
3685 		 * Write data segments in reverse order, so as to
3686 		 * overwrite cacheline stamp last within each
3687 		 * cacheline.  This avoids issues with WQE
3688 		 * prefetching.
3689 		 */
3690 
3691 		dseg = wqe;
3692 		dseg += wr->num_sge - 1;
3693 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3694 
3695 		/* Add one more inline data segment for ICRC for MLX sends */
3696 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3697 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3698 			     qp->mlx4_ib_qp_type &
3699 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3700 			set_mlx_icrc_seg(dseg + 1);
3701 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3702 		}
3703 
3704 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3705 			set_data_seg(dseg, wr->sg_list + i);
3706 
3707 		/*
3708 		 * Possibly overwrite stamping in cacheline with LSO
3709 		 * segment only after making sure all data segments
3710 		 * are written.
3711 		 */
3712 		wmb();
3713 		*lso_wqe = lso_hdr_sz;
3714 
3715 		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3716 					     MLX4_WQE_CTRL_FENCE : 0) | size;
3717 
3718 		/*
3719 		 * Make sure descriptor is fully written before
3720 		 * setting ownership bit (because HW can start
3721 		 * executing as soon as we do).
3722 		 */
3723 		wmb();
3724 
3725 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3726 			*bad_wr = wr;
3727 			err = -EINVAL;
3728 			goto out;
3729 		}
3730 
3731 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3732 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3733 
3734 		/*
3735 		 * We can improve latency by not stamping the last
3736 		 * send queue WQE until after ringing the doorbell, so
3737 		 * only stamp here if there are still more WQEs to post.
3738 		 */
3739 		if (wr->next)
3740 			stamp_send_wqe(qp, ind + qp->sq_spare_wqes);
3741 		ind++;
3742 	}
3743 
3744 out:
3745 	if (likely(nreq)) {
3746 		qp->sq.head += nreq;
3747 
3748 		/*
3749 		 * Make sure that descriptors are written before
3750 		 * doorbell record.
3751 		 */
3752 		wmb();
3753 
3754 		writel_relaxed(qp->doorbell_qpn,
3755 			to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3756 
3757 		stamp_send_wqe(qp, ind + qp->sq_spare_wqes - 1);
3758 
3759 		qp->sq_next_wqe = ind;
3760 	}
3761 
3762 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3763 
3764 	return err;
3765 }
3766 
3767 int mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3768 		      const struct ib_send_wr **bad_wr)
3769 {
3770 	return _mlx4_ib_post_send(ibqp, wr, bad_wr, false);
3771 }
3772 
3773 static int _mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3774 			      const struct ib_recv_wr **bad_wr, bool drain)
3775 {
3776 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3777 	struct mlx4_wqe_data_seg *scat;
3778 	unsigned long flags;
3779 	int err = 0;
3780 	int nreq;
3781 	int ind;
3782 	int max_gs;
3783 	int i;
3784 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3785 
3786 	max_gs = qp->rq.max_gs;
3787 	spin_lock_irqsave(&qp->rq.lock, flags);
3788 
3789 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3790 	    !drain) {
3791 		err = -EIO;
3792 		*bad_wr = wr;
3793 		nreq = 0;
3794 		goto out;
3795 	}
3796 
3797 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3798 
3799 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3800 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3801 			err = -ENOMEM;
3802 			*bad_wr = wr;
3803 			goto out;
3804 		}
3805 
3806 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3807 			err = -EINVAL;
3808 			*bad_wr = wr;
3809 			goto out;
3810 		}
3811 
3812 		scat = get_recv_wqe(qp, ind);
3813 
3814 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3815 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3816 			ib_dma_sync_single_for_device(ibqp->device,
3817 						      qp->sqp_proxy_rcv[ind].map,
3818 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3819 						      DMA_FROM_DEVICE);
3820 			scat->byte_count =
3821 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3822 			/* use dma lkey from upper layer entry */
3823 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3824 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3825 			scat++;
3826 			max_gs--;
3827 		}
3828 
3829 		for (i = 0; i < wr->num_sge; ++i)
3830 			__set_data_seg(scat + i, wr->sg_list + i);
3831 
3832 		if (i < max_gs) {
3833 			scat[i].byte_count = 0;
3834 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3835 			scat[i].addr       = 0;
3836 		}
3837 
3838 		qp->rq.wrid[ind] = wr->wr_id;
3839 
3840 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3841 	}
3842 
3843 out:
3844 	if (likely(nreq)) {
3845 		qp->rq.head += nreq;
3846 
3847 		/*
3848 		 * Make sure that descriptors are written before
3849 		 * doorbell record.
3850 		 */
3851 		wmb();
3852 
3853 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3854 	}
3855 
3856 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3857 
3858 	return err;
3859 }
3860 
3861 int mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3862 		      const struct ib_recv_wr **bad_wr)
3863 {
3864 	return _mlx4_ib_post_recv(ibqp, wr, bad_wr, false);
3865 }
3866 
3867 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3868 {
3869 	switch (mlx4_state) {
3870 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3871 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3872 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3873 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3874 	case MLX4_QP_STATE_SQ_DRAINING:
3875 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3876 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3877 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3878 	default:		     return -1;
3879 	}
3880 }
3881 
3882 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3883 {
3884 	switch (mlx4_mig_state) {
3885 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3886 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3887 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3888 	default: return -1;
3889 	}
3890 }
3891 
3892 static int to_ib_qp_access_flags(int mlx4_flags)
3893 {
3894 	int ib_flags = 0;
3895 
3896 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3897 		ib_flags |= IB_ACCESS_REMOTE_READ;
3898 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3899 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3900 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3901 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3902 
3903 	return ib_flags;
3904 }
3905 
3906 static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
3907 			    struct rdma_ah_attr *ah_attr,
3908 			    struct mlx4_qp_path *path)
3909 {
3910 	struct mlx4_dev *dev = ibdev->dev;
3911 	u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
3912 
3913 	memset(ah_attr, 0, sizeof(*ah_attr));
3914 	if (port_num == 0 || port_num > dev->caps.num_ports)
3915 		return;
3916 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
3917 
3918 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
3919 		rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
3920 			       ((path->sched_queue & 4) << 1));
3921 	else
3922 		rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
3923 	rdma_ah_set_port_num(ah_attr, port_num);
3924 
3925 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
3926 	rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
3927 	rdma_ah_set_static_rate(ah_attr,
3928 				path->static_rate ? path->static_rate - 5 : 0);
3929 	if (path->grh_mylmc & (1 << 7)) {
3930 		rdma_ah_set_grh(ah_attr, NULL,
3931 				be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
3932 				path->mgid_index,
3933 				path->hop_limit,
3934 				(be32_to_cpu(path->tclass_flowlabel)
3935 				 >> 20) & 0xff);
3936 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
3937 	}
3938 }
3939 
3940 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3941 		     struct ib_qp_init_attr *qp_init_attr)
3942 {
3943 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3944 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3945 	struct mlx4_qp_context context;
3946 	int mlx4_state;
3947 	int err = 0;
3948 
3949 	if (ibqp->rwq_ind_tbl)
3950 		return -EOPNOTSUPP;
3951 
3952 	mutex_lock(&qp->mutex);
3953 
3954 	if (qp->state == IB_QPS_RESET) {
3955 		qp_attr->qp_state = IB_QPS_RESET;
3956 		goto done;
3957 	}
3958 
3959 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3960 	if (err) {
3961 		err = -EINVAL;
3962 		goto out;
3963 	}
3964 
3965 	mlx4_state = be32_to_cpu(context.flags) >> 28;
3966 
3967 	qp->state		     = to_ib_qp_state(mlx4_state);
3968 	qp_attr->qp_state	     = qp->state;
3969 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
3970 	qp_attr->path_mig_state	     =
3971 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3972 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
3973 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3974 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
3975 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
3976 	qp_attr->qp_access_flags     =
3977 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
3978 
3979 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3980 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3981 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3982 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3983 		qp_attr->alt_port_num	=
3984 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
3985 	}
3986 
3987 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3988 	if (qp_attr->qp_state == IB_QPS_INIT)
3989 		qp_attr->port_num = qp->port;
3990 	else
3991 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3992 
3993 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3994 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3995 
3996 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3997 
3998 	qp_attr->max_dest_rd_atomic =
3999 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
4000 	qp_attr->min_rnr_timer	    =
4001 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
4002 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
4003 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
4004 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
4005 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
4006 
4007 done:
4008 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4009 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4010 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4011 
4012 	if (!ibqp->uobject) {
4013 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
4014 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4015 	} else {
4016 		qp_attr->cap.max_send_wr  = 0;
4017 		qp_attr->cap.max_send_sge = 0;
4018 	}
4019 
4020 	/*
4021 	 * We don't support inline sends for kernel QPs (yet), and we
4022 	 * don't know what userspace's value should be.
4023 	 */
4024 	qp_attr->cap.max_inline_data = 0;
4025 
4026 	qp_init_attr->cap	     = qp_attr->cap;
4027 
4028 	qp_init_attr->create_flags = 0;
4029 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4030 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4031 
4032 	if (qp->flags & MLX4_IB_QP_LSO)
4033 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
4034 
4035 	if (qp->flags & MLX4_IB_QP_NETIF)
4036 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
4037 
4038 	qp_init_attr->sq_sig_type =
4039 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
4040 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4041 
4042 out:
4043 	mutex_unlock(&qp->mutex);
4044 	return err;
4045 }
4046 
4047 struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
4048 				struct ib_wq_init_attr *init_attr,
4049 				struct ib_udata *udata)
4050 {
4051 	struct mlx4_ib_dev *dev;
4052 	struct ib_qp_init_attr ib_qp_init_attr;
4053 	struct mlx4_ib_qp *qp;
4054 	struct mlx4_ib_create_wq ucmd;
4055 	int err, required_cmd_sz;
4056 
4057 	if (!udata)
4058 		return ERR_PTR(-EINVAL);
4059 
4060 	required_cmd_sz = offsetof(typeof(ucmd), comp_mask) +
4061 			  sizeof(ucmd.comp_mask);
4062 	if (udata->inlen < required_cmd_sz) {
4063 		pr_debug("invalid inlen\n");
4064 		return ERR_PTR(-EINVAL);
4065 	}
4066 
4067 	if (udata->inlen > sizeof(ucmd) &&
4068 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4069 				 udata->inlen - sizeof(ucmd))) {
4070 		pr_debug("inlen is not supported\n");
4071 		return ERR_PTR(-EOPNOTSUPP);
4072 	}
4073 
4074 	if (udata->outlen)
4075 		return ERR_PTR(-EOPNOTSUPP);
4076 
4077 	dev = to_mdev(pd->device);
4078 
4079 	if (init_attr->wq_type != IB_WQT_RQ) {
4080 		pr_debug("unsupported wq type %d\n", init_attr->wq_type);
4081 		return ERR_PTR(-EOPNOTSUPP);
4082 	}
4083 
4084 	if (init_attr->create_flags & ~IB_WQ_FLAGS_SCATTER_FCS) {
4085 		pr_debug("unsupported create_flags %u\n",
4086 			 init_attr->create_flags);
4087 		return ERR_PTR(-EOPNOTSUPP);
4088 	}
4089 
4090 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
4091 	if (!qp)
4092 		return ERR_PTR(-ENOMEM);
4093 
4094 	qp->pri.vid = 0xFFFF;
4095 	qp->alt.vid = 0xFFFF;
4096 
4097 	memset(&ib_qp_init_attr, 0, sizeof(ib_qp_init_attr));
4098 	ib_qp_init_attr.qp_context = init_attr->wq_context;
4099 	ib_qp_init_attr.qp_type = IB_QPT_RAW_PACKET;
4100 	ib_qp_init_attr.cap.max_recv_wr = init_attr->max_wr;
4101 	ib_qp_init_attr.cap.max_recv_sge = init_attr->max_sge;
4102 	ib_qp_init_attr.recv_cq = init_attr->cq;
4103 	ib_qp_init_attr.send_cq = ib_qp_init_attr.recv_cq; /* Dummy CQ */
4104 
4105 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS)
4106 		ib_qp_init_attr.create_flags |= IB_QP_CREATE_SCATTER_FCS;
4107 
4108 	err = create_qp_common(dev, pd, MLX4_IB_RWQ_SRC, &ib_qp_init_attr,
4109 			       udata, 0, &qp);
4110 	if (err) {
4111 		kfree(qp);
4112 		return ERR_PTR(err);
4113 	}
4114 
4115 	qp->ibwq.event_handler = init_attr->event_handler;
4116 	qp->ibwq.wq_num = qp->mqp.qpn;
4117 	qp->ibwq.state = IB_WQS_RESET;
4118 
4119 	return &qp->ibwq;
4120 }
4121 
4122 static int ib_wq2qp_state(enum ib_wq_state state)
4123 {
4124 	switch (state) {
4125 	case IB_WQS_RESET:
4126 		return IB_QPS_RESET;
4127 	case IB_WQS_RDY:
4128 		return IB_QPS_RTR;
4129 	default:
4130 		return IB_QPS_ERR;
4131 	}
4132 }
4133 
4134 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state,
4135 			      struct ib_udata *udata)
4136 {
4137 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4138 	enum ib_qp_state qp_cur_state;
4139 	enum ib_qp_state qp_new_state;
4140 	int attr_mask;
4141 	int err;
4142 
4143 	/* ib_qp.state represents the WQ HW state while ib_wq.state represents
4144 	 * the WQ logic state.
4145 	 */
4146 	qp_cur_state = qp->state;
4147 	qp_new_state = ib_wq2qp_state(new_state);
4148 
4149 	if (ib_wq2qp_state(new_state) == qp_cur_state)
4150 		return 0;
4151 
4152 	if (new_state == IB_WQS_RDY) {
4153 		struct ib_qp_attr attr = {};
4154 
4155 		attr.port_num = qp->port;
4156 		attr_mask = IB_QP_PORT;
4157 
4158 		err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, &attr,
4159 					  attr_mask, IB_QPS_RESET, IB_QPS_INIT,
4160 					  udata);
4161 		if (err) {
4162 			pr_debug("WQN=0x%06x failed to apply RST->INIT on the HW QP\n",
4163 				 ibwq->wq_num);
4164 			return err;
4165 		}
4166 
4167 		qp_cur_state = IB_QPS_INIT;
4168 	}
4169 
4170 	attr_mask = 0;
4171 	err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, attr_mask,
4172 				  qp_cur_state,  qp_new_state, udata);
4173 
4174 	if (err && (qp_cur_state == IB_QPS_INIT)) {
4175 		qp_new_state = IB_QPS_RESET;
4176 		if (__mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL,
4177 					attr_mask, IB_QPS_INIT, IB_QPS_RESET,
4178 					udata)) {
4179 			pr_warn("WQN=0x%06x failed with reverting HW's resources failure\n",
4180 				ibwq->wq_num);
4181 			qp_new_state = IB_QPS_INIT;
4182 		}
4183 	}
4184 
4185 	qp->state = qp_new_state;
4186 
4187 	return err;
4188 }
4189 
4190 int mlx4_ib_modify_wq(struct ib_wq *ibwq, struct ib_wq_attr *wq_attr,
4191 		      u32 wq_attr_mask, struct ib_udata *udata)
4192 {
4193 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4194 	struct mlx4_ib_modify_wq ucmd = {};
4195 	size_t required_cmd_sz;
4196 	enum ib_wq_state cur_state, new_state;
4197 	int err = 0;
4198 
4199 	required_cmd_sz = offsetof(typeof(ucmd), reserved) +
4200 				   sizeof(ucmd.reserved);
4201 	if (udata->inlen < required_cmd_sz)
4202 		return -EINVAL;
4203 
4204 	if (udata->inlen > sizeof(ucmd) &&
4205 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4206 				 udata->inlen - sizeof(ucmd)))
4207 		return -EOPNOTSUPP;
4208 
4209 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4210 		return -EFAULT;
4211 
4212 	if (ucmd.comp_mask || ucmd.reserved)
4213 		return -EOPNOTSUPP;
4214 
4215 	if (wq_attr_mask & IB_WQ_FLAGS)
4216 		return -EOPNOTSUPP;
4217 
4218 	cur_state = wq_attr_mask & IB_WQ_CUR_STATE ? wq_attr->curr_wq_state :
4219 						     ibwq->state;
4220 	new_state = wq_attr_mask & IB_WQ_STATE ? wq_attr->wq_state : cur_state;
4221 
4222 	if (cur_state  < IB_WQS_RESET || cur_state  > IB_WQS_ERR ||
4223 	    new_state < IB_WQS_RESET || new_state > IB_WQS_ERR)
4224 		return -EINVAL;
4225 
4226 	if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR))
4227 		return -EINVAL;
4228 
4229 	if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET))
4230 		return -EINVAL;
4231 
4232 	/* Need to protect against the parent RSS which also may modify WQ
4233 	 * state.
4234 	 */
4235 	mutex_lock(&qp->mutex);
4236 
4237 	/* Can update HW state only if a RSS QP has already associated to this
4238 	 * WQ, so we can apply its port on the WQ.
4239 	 */
4240 	if (qp->rss_usecnt)
4241 		err = _mlx4_ib_modify_wq(ibwq, new_state, udata);
4242 
4243 	if (!err)
4244 		ibwq->state = new_state;
4245 
4246 	mutex_unlock(&qp->mutex);
4247 
4248 	return err;
4249 }
4250 
4251 int mlx4_ib_destroy_wq(struct ib_wq *ibwq, struct ib_udata *udata)
4252 {
4253 	struct mlx4_ib_dev *dev = to_mdev(ibwq->device);
4254 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4255 
4256 	if (qp->counter_index)
4257 		mlx4_ib_free_qp_counter(dev, qp);
4258 
4259 	destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, udata);
4260 
4261 	kfree(qp);
4262 
4263 	return 0;
4264 }
4265 
4266 struct ib_rwq_ind_table
4267 *mlx4_ib_create_rwq_ind_table(struct ib_device *device,
4268 			      struct ib_rwq_ind_table_init_attr *init_attr,
4269 			      struct ib_udata *udata)
4270 {
4271 	struct ib_rwq_ind_table *rwq_ind_table;
4272 	struct mlx4_ib_create_rwq_ind_tbl_resp resp = {};
4273 	unsigned int ind_tbl_size = 1 << init_attr->log_ind_tbl_size;
4274 	unsigned int base_wqn;
4275 	size_t min_resp_len;
4276 	int i;
4277 	int err;
4278 
4279 	if (udata->inlen > 0 &&
4280 	    !ib_is_udata_cleared(udata, 0,
4281 				 udata->inlen))
4282 		return ERR_PTR(-EOPNOTSUPP);
4283 
4284 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4285 	if (udata->outlen && udata->outlen < min_resp_len)
4286 		return ERR_PTR(-EINVAL);
4287 
4288 	if (ind_tbl_size >
4289 	    device->attrs.rss_caps.max_rwq_indirection_table_size) {
4290 		pr_debug("log_ind_tbl_size = %d is bigger than supported = %d\n",
4291 			 ind_tbl_size,
4292 			 device->attrs.rss_caps.max_rwq_indirection_table_size);
4293 		return ERR_PTR(-EINVAL);
4294 	}
4295 
4296 	base_wqn = init_attr->ind_tbl[0]->wq_num;
4297 
4298 	if (base_wqn % ind_tbl_size) {
4299 		pr_debug("WQN=0x%x isn't aligned with indirection table size\n",
4300 			 base_wqn);
4301 		return ERR_PTR(-EINVAL);
4302 	}
4303 
4304 	for (i = 1; i < ind_tbl_size; i++) {
4305 		if (++base_wqn != init_attr->ind_tbl[i]->wq_num) {
4306 			pr_debug("indirection table's WQNs aren't consecutive\n");
4307 			return ERR_PTR(-EINVAL);
4308 		}
4309 	}
4310 
4311 	rwq_ind_table = kzalloc(sizeof(*rwq_ind_table), GFP_KERNEL);
4312 	if (!rwq_ind_table)
4313 		return ERR_PTR(-ENOMEM);
4314 
4315 	if (udata->outlen) {
4316 		resp.response_length = offsetof(typeof(resp), response_length) +
4317 					sizeof(resp.response_length);
4318 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4319 		if (err)
4320 			goto err;
4321 	}
4322 
4323 	return rwq_ind_table;
4324 
4325 err:
4326 	kfree(rwq_ind_table);
4327 	return ERR_PTR(err);
4328 }
4329 
4330 int mlx4_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4331 {
4332 	kfree(ib_rwq_ind_tbl);
4333 	return 0;
4334 }
4335 
4336 struct mlx4_ib_drain_cqe {
4337 	struct ib_cqe cqe;
4338 	struct completion done;
4339 };
4340 
4341 static void mlx4_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
4342 {
4343 	struct mlx4_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
4344 						     struct mlx4_ib_drain_cqe,
4345 						     cqe);
4346 
4347 	complete(&cqe->done);
4348 }
4349 
4350 /* This function returns only once the drained WR was completed */
4351 static void handle_drain_completion(struct ib_cq *cq,
4352 				    struct mlx4_ib_drain_cqe *sdrain,
4353 				    struct mlx4_ib_dev *dev)
4354 {
4355 	struct mlx4_dev *mdev = dev->dev;
4356 
4357 	if (cq->poll_ctx == IB_POLL_DIRECT) {
4358 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
4359 			ib_process_cq_direct(cq, -1);
4360 		return;
4361 	}
4362 
4363 	if (mdev->persist->state == MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4364 		struct mlx4_ib_cq *mcq = to_mcq(cq);
4365 		bool triggered = false;
4366 		unsigned long flags;
4367 
4368 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
4369 		/* Make sure that the CQ handler won't run if wasn't run yet */
4370 		if (!mcq->mcq.reset_notify_added)
4371 			mcq->mcq.reset_notify_added = 1;
4372 		else
4373 			triggered = true;
4374 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
4375 
4376 		if (triggered) {
4377 			/* Wait for any scheduled/running task to be ended */
4378 			switch (cq->poll_ctx) {
4379 			case IB_POLL_SOFTIRQ:
4380 				irq_poll_disable(&cq->iop);
4381 				irq_poll_enable(&cq->iop);
4382 				break;
4383 			case IB_POLL_WORKQUEUE:
4384 				cancel_work_sync(&cq->work);
4385 				break;
4386 			default:
4387 				WARN_ON_ONCE(1);
4388 			}
4389 		}
4390 
4391 		/* Run the CQ handler - this makes sure that the drain WR will
4392 		 * be processed if wasn't processed yet.
4393 		 */
4394 		mcq->mcq.comp(&mcq->mcq);
4395 	}
4396 
4397 	wait_for_completion(&sdrain->done);
4398 }
4399 
4400 void mlx4_ib_drain_sq(struct ib_qp *qp)
4401 {
4402 	struct ib_cq *cq = qp->send_cq;
4403 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4404 	struct mlx4_ib_drain_cqe sdrain;
4405 	const struct ib_send_wr *bad_swr;
4406 	struct ib_rdma_wr swr = {
4407 		.wr = {
4408 			.next = NULL,
4409 			{ .wr_cqe	= &sdrain.cqe, },
4410 			.opcode	= IB_WR_RDMA_WRITE,
4411 		},
4412 	};
4413 	int ret;
4414 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
4415 	struct mlx4_dev *mdev = dev->dev;
4416 
4417 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4418 	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4419 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4420 		return;
4421 	}
4422 
4423 	sdrain.cqe.done = mlx4_ib_drain_qp_done;
4424 	init_completion(&sdrain.done);
4425 
4426 	ret = _mlx4_ib_post_send(qp, &swr.wr, &bad_swr, true);
4427 	if (ret) {
4428 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4429 		return;
4430 	}
4431 
4432 	handle_drain_completion(cq, &sdrain, dev);
4433 }
4434 
4435 void mlx4_ib_drain_rq(struct ib_qp *qp)
4436 {
4437 	struct ib_cq *cq = qp->recv_cq;
4438 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4439 	struct mlx4_ib_drain_cqe rdrain;
4440 	struct ib_recv_wr rwr = {};
4441 	const struct ib_recv_wr *bad_rwr;
4442 	int ret;
4443 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
4444 	struct mlx4_dev *mdev = dev->dev;
4445 
4446 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4447 	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4448 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4449 		return;
4450 	}
4451 
4452 	rwr.wr_cqe = &rdrain.cqe;
4453 	rdrain.cqe.done = mlx4_ib_drain_qp_done;
4454 	init_completion(&rdrain.done);
4455 
4456 	ret = _mlx4_ib_post_recv(qp, &rwr, &bad_rwr, true);
4457 	if (ret) {
4458 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4459 		return;
4460 	}
4461 
4462 	handle_drain_completion(cq, &rdrain, dev);
4463 }
4464