xref: /linux/drivers/infiniband/hw/mlx4/qp.c (revision 68a052239fc4b351e961f698b824f7654a346091)
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 	MLX4_IB_MIN_SQ_STRIDE	= 6,
72 	MLX4_IB_CACHE_LINE_SIZE	= 64,
73 };
74 
75 enum {
76 	MLX4_RAW_QP_MTU		= 7,
77 	MLX4_RAW_QP_MSGMAX	= 31,
78 };
79 
80 #ifndef ETH_ALEN
81 #define ETH_ALEN        6
82 #endif
83 
84 static const __be32 mlx4_ib_opcode[] = {
85 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
86 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
87 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
88 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
89 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
90 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
91 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
92 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
93 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
94 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
95 	[IB_WR_REG_MR]				= cpu_to_be32(MLX4_OPCODE_FMR),
96 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
97 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
98 };
99 
100 enum mlx4_ib_source_type {
101 	MLX4_IB_QP_SRC	= 0,
102 	MLX4_IB_RWQ_SRC	= 1,
103 };
104 
105 struct mlx4_ib_qp_event_work {
106 	struct work_struct work;
107 	struct mlx4_qp *qp;
108 	enum mlx4_event type;
109 };
110 
111 static struct workqueue_struct *mlx4_ib_qp_event_wq;
112 
113 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
114 {
115 	if (!mlx4_is_master(dev->dev))
116 		return 0;
117 
118 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
119 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
120 		8 * MLX4_MFUNC_MAX;
121 }
122 
123 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
124 {
125 	int proxy_sqp = 0;
126 	int real_sqp = 0;
127 	int i;
128 	/* PPF or Native -- real SQP */
129 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
130 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
131 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
132 	if (real_sqp)
133 		return 1;
134 	/* VF or PF -- proxy SQP */
135 	if (mlx4_is_mfunc(dev->dev)) {
136 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
137 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy ||
138 			    qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp1_proxy) {
139 				proxy_sqp = 1;
140 				break;
141 			}
142 		}
143 	}
144 	if (proxy_sqp)
145 		return 1;
146 
147 	return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
148 }
149 
150 /* used for INIT/CLOSE port logic */
151 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
152 {
153 	int proxy_qp0 = 0;
154 	int real_qp0 = 0;
155 	int i;
156 	/* PPF or Native -- real QP0 */
157 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
158 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
159 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
160 	if (real_qp0)
161 		return 1;
162 	/* VF or PF -- proxy QP0 */
163 	if (mlx4_is_mfunc(dev->dev)) {
164 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
165 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy) {
166 				proxy_qp0 = 1;
167 				break;
168 			}
169 		}
170 	}
171 	return proxy_qp0;
172 }
173 
174 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
175 {
176 	return mlx4_buf_offset(&qp->buf, offset);
177 }
178 
179 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
180 {
181 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
182 }
183 
184 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
185 {
186 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
187 }
188 
189 /*
190  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
191  * first four bytes of every 64 byte chunk with 0xffffffff, except for
192  * the very first chunk of the WQE.
193  */
194 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n)
195 {
196 	__be32 *wqe;
197 	int i;
198 	int s;
199 	void *buf;
200 	struct mlx4_wqe_ctrl_seg *ctrl;
201 
202 	buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
203 	ctrl = (struct mlx4_wqe_ctrl_seg *)buf;
204 	s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4;
205 	for (i = 64; i < s; i += 64) {
206 		wqe = buf + i;
207 		*wqe = cpu_to_be32(0xffffffff);
208 	}
209 }
210 
211 static void mlx4_ib_handle_qp_event(struct work_struct *_work)
212 {
213 	struct mlx4_ib_qp_event_work *qpe_work =
214 		container_of(_work, struct mlx4_ib_qp_event_work, work);
215 	struct ib_qp *ibqp = &to_mibqp(qpe_work->qp)->ibqp;
216 	struct ib_event event = {};
217 
218 	event.device = ibqp->device;
219 	event.element.qp = ibqp;
220 
221 	switch (qpe_work->type) {
222 	case MLX4_EVENT_TYPE_PATH_MIG:
223 		event.event = IB_EVENT_PATH_MIG;
224 		break;
225 	case MLX4_EVENT_TYPE_COMM_EST:
226 		event.event = IB_EVENT_COMM_EST;
227 		break;
228 	case MLX4_EVENT_TYPE_SQ_DRAINED:
229 		event.event = IB_EVENT_SQ_DRAINED;
230 		break;
231 	case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
232 		event.event = IB_EVENT_QP_LAST_WQE_REACHED;
233 		break;
234 	case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
235 		event.event = IB_EVENT_QP_FATAL;
236 		break;
237 	case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
238 		event.event = IB_EVENT_PATH_MIG_ERR;
239 		break;
240 	case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
241 		event.event = IB_EVENT_QP_REQ_ERR;
242 		break;
243 	case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
244 		event.event = IB_EVENT_QP_ACCESS_ERR;
245 		break;
246 	default:
247 		pr_warn("Unexpected event type %d on QP %06x\n",
248 			qpe_work->type, qpe_work->qp->qpn);
249 		goto out;
250 	}
251 
252 	ibqp->event_handler(&event, ibqp->qp_context);
253 
254 out:
255 	mlx4_put_qp(qpe_work->qp);
256 	kfree(qpe_work);
257 }
258 
259 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
260 {
261 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
262 	struct mlx4_ib_qp_event_work *qpe_work;
263 
264 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
265 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
266 
267 	if (!ibqp->event_handler)
268 		goto out_no_handler;
269 
270 	qpe_work = kzalloc(sizeof(*qpe_work), GFP_ATOMIC);
271 	if (!qpe_work)
272 		goto out_no_handler;
273 
274 	qpe_work->qp = qp;
275 	qpe_work->type = type;
276 	INIT_WORK(&qpe_work->work, mlx4_ib_handle_qp_event);
277 	queue_work(mlx4_ib_qp_event_wq, &qpe_work->work);
278 	return;
279 
280 out_no_handler:
281 	mlx4_put_qp(qp);
282 }
283 
284 static void mlx4_ib_wq_event(struct mlx4_qp *qp, enum mlx4_event type)
285 {
286 	pr_warn_ratelimited("Unexpected event type %d on WQ 0x%06x. Events are not supported for WQs\n",
287 			    type, qp->qpn);
288 }
289 
290 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
291 {
292 	/*
293 	 * UD WQEs must have a datagram segment.
294 	 * RC and UC WQEs might have a remote address segment.
295 	 * MLX WQEs need two extra inline data segments (for the UD
296 	 * header and space for the ICRC).
297 	 */
298 	switch (type) {
299 	case MLX4_IB_QPT_UD:
300 		return sizeof (struct mlx4_wqe_ctrl_seg) +
301 			sizeof (struct mlx4_wqe_datagram_seg) +
302 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
303 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
304 	case MLX4_IB_QPT_PROXY_SMI:
305 	case MLX4_IB_QPT_PROXY_GSI:
306 		return sizeof (struct mlx4_wqe_ctrl_seg) +
307 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
308 	case MLX4_IB_QPT_TUN_SMI_OWNER:
309 	case MLX4_IB_QPT_TUN_GSI:
310 		return sizeof (struct mlx4_wqe_ctrl_seg) +
311 			sizeof (struct mlx4_wqe_datagram_seg);
312 
313 	case MLX4_IB_QPT_UC:
314 		return sizeof (struct mlx4_wqe_ctrl_seg) +
315 			sizeof (struct mlx4_wqe_raddr_seg);
316 	case MLX4_IB_QPT_RC:
317 		return sizeof (struct mlx4_wqe_ctrl_seg) +
318 			sizeof (struct mlx4_wqe_masked_atomic_seg) +
319 			sizeof (struct mlx4_wqe_raddr_seg);
320 	case MLX4_IB_QPT_SMI:
321 	case MLX4_IB_QPT_GSI:
322 		return sizeof (struct mlx4_wqe_ctrl_seg) +
323 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
324 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
325 					   MLX4_INLINE_ALIGN) *
326 			      sizeof (struct mlx4_wqe_inline_seg),
327 			      sizeof (struct mlx4_wqe_data_seg)) +
328 			ALIGN(4 +
329 			      sizeof (struct mlx4_wqe_inline_seg),
330 			      sizeof (struct mlx4_wqe_data_seg));
331 	default:
332 		return sizeof (struct mlx4_wqe_ctrl_seg);
333 	}
334 }
335 
336 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
337 		       bool is_user, bool has_rq, struct mlx4_ib_qp *qp,
338 		       u32 inl_recv_sz)
339 {
340 	/* Sanity check RQ size before proceeding */
341 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
342 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
343 		return -EINVAL;
344 
345 	if (!has_rq) {
346 		if (cap->max_recv_wr || inl_recv_sz)
347 			return -EINVAL;
348 
349 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
350 	} else {
351 		u32 max_inl_recv_sz = dev->dev->caps.max_rq_sg *
352 			sizeof(struct mlx4_wqe_data_seg);
353 		u32 wqe_size;
354 
355 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
356 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge ||
357 				inl_recv_sz > max_inl_recv_sz))
358 			return -EINVAL;
359 
360 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
361 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
362 		wqe_size = qp->rq.max_gs * sizeof(struct mlx4_wqe_data_seg);
363 		qp->rq.wqe_shift = ilog2(max_t(u32, wqe_size, inl_recv_sz));
364 	}
365 
366 	/* leave userspace return values as they were, so as not to break ABI */
367 	if (is_user) {
368 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
369 		cap->max_recv_sge = qp->rq.max_gs;
370 	} else {
371 		cap->max_recv_wr  = qp->rq.max_post =
372 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
373 		cap->max_recv_sge = min(qp->rq.max_gs,
374 					min(dev->dev->caps.max_sq_sg,
375 					    dev->dev->caps.max_rq_sg));
376 	}
377 
378 	return 0;
379 }
380 
381 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
382 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
383 {
384 	int s;
385 
386 	/* Sanity check SQ size before proceeding */
387 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
388 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
389 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
390 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
391 		return -EINVAL;
392 
393 	/*
394 	 * For MLX transport we need 2 extra S/G entries:
395 	 * one for the header and one for the checksum at the end
396 	 */
397 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
398 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
399 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
400 		return -EINVAL;
401 
402 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
403 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
404 		send_wqe_overhead(type, qp->flags);
405 
406 	if (s > dev->dev->caps.max_sq_desc_sz)
407 		return -EINVAL;
408 
409 	qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
410 
411 	/*
412 	 * We need to leave 2 KB + 1 WR of headroom in the SQ to
413 	 * allow HW to prefetch.
414 	 */
415 	qp->sq_spare_wqes = MLX4_IB_SQ_HEADROOM(qp->sq.wqe_shift);
416 	qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr +
417 					    qp->sq_spare_wqes);
418 
419 	qp->sq.max_gs =
420 		(min(dev->dev->caps.max_sq_desc_sz,
421 		     (1 << qp->sq.wqe_shift)) -
422 		 send_wqe_overhead(type, qp->flags)) /
423 		sizeof (struct mlx4_wqe_data_seg);
424 
425 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
426 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
427 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
428 		qp->rq.offset = 0;
429 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
430 	} else {
431 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
432 		qp->sq.offset = 0;
433 	}
434 
435 	cap->max_send_wr  = qp->sq.max_post =
436 		qp->sq.wqe_cnt - qp->sq_spare_wqes;
437 	cap->max_send_sge = min(qp->sq.max_gs,
438 				min(dev->dev->caps.max_sq_sg,
439 				    dev->dev->caps.max_rq_sg));
440 	/* We don't support inline sends for kernel QPs (yet) */
441 	cap->max_inline_data = 0;
442 
443 	return 0;
444 }
445 
446 static int set_user_sq_size(struct mlx4_ib_dev *dev,
447 			    struct mlx4_ib_qp *qp,
448 			    struct mlx4_ib_create_qp *ucmd)
449 {
450 	u32 cnt;
451 
452 	/* Sanity check SQ size before proceeding */
453 	if (check_shl_overflow(1, ucmd->log_sq_bb_count, &cnt) ||
454 	    cnt > dev->dev->caps.max_wqes)
455 		return -EINVAL;
456 	if (ucmd->log_sq_stride >
457 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
458 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
459 		return -EINVAL;
460 
461 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
462 	qp->sq.wqe_shift = ucmd->log_sq_stride;
463 
464 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
465 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
466 
467 	return 0;
468 }
469 
470 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
471 {
472 	int i;
473 
474 	qp->sqp_proxy_rcv =
475 		kmalloc_array(qp->rq.wqe_cnt, sizeof(struct mlx4_ib_buf),
476 			      GFP_KERNEL);
477 	if (!qp->sqp_proxy_rcv)
478 		return -ENOMEM;
479 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
480 		qp->sqp_proxy_rcv[i].addr =
481 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
482 				GFP_KERNEL);
483 		if (!qp->sqp_proxy_rcv[i].addr)
484 			goto err;
485 		qp->sqp_proxy_rcv[i].map =
486 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
487 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
488 					  DMA_FROM_DEVICE);
489 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
490 			kfree(qp->sqp_proxy_rcv[i].addr);
491 			goto err;
492 		}
493 	}
494 	return 0;
495 
496 err:
497 	while (i > 0) {
498 		--i;
499 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
500 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
501 				    DMA_FROM_DEVICE);
502 		kfree(qp->sqp_proxy_rcv[i].addr);
503 	}
504 	kfree(qp->sqp_proxy_rcv);
505 	qp->sqp_proxy_rcv = NULL;
506 	return -ENOMEM;
507 }
508 
509 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
510 {
511 	int i;
512 
513 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
514 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
515 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
516 				    DMA_FROM_DEVICE);
517 		kfree(qp->sqp_proxy_rcv[i].addr);
518 	}
519 	kfree(qp->sqp_proxy_rcv);
520 }
521 
522 static bool qp_has_rq(struct ib_qp_init_attr *attr)
523 {
524 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
525 		return false;
526 
527 	return !attr->srq;
528 }
529 
530 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
531 {
532 	int i;
533 	for (i = 0; i < dev->caps.num_ports; i++) {
534 		if (qpn == dev->caps.spec_qps[i].qp0_proxy)
535 			return !!dev->caps.spec_qps[i].qp0_qkey;
536 	}
537 	return 0;
538 }
539 
540 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
541 				    struct mlx4_ib_qp *qp)
542 {
543 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
544 	mlx4_counter_free(dev->dev, qp->counter_index->index);
545 	list_del(&qp->counter_index->list);
546 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
547 
548 	kfree(qp->counter_index);
549 	qp->counter_index = NULL;
550 }
551 
552 static int set_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_rss *rss_ctx,
553 		      struct ib_qp_init_attr *init_attr,
554 		      struct mlx4_ib_create_qp_rss *ucmd)
555 {
556 	rss_ctx->base_qpn_tbl_sz = init_attr->rwq_ind_tbl->ind_tbl[0]->wq_num |
557 		(init_attr->rwq_ind_tbl->log_ind_tbl_size << 24);
558 
559 	if ((ucmd->rx_hash_function == MLX4_IB_RX_HASH_FUNC_TOEPLITZ) &&
560 	    (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS_TOP)) {
561 		memcpy(rss_ctx->rss_key, ucmd->rx_hash_key,
562 		       MLX4_EN_RSS_KEY_SIZE);
563 	} else {
564 		pr_debug("RX Hash function is not supported\n");
565 		return (-EOPNOTSUPP);
566 	}
567 
568 	if (ucmd->rx_hash_fields_mask & ~(u64)(MLX4_IB_RX_HASH_SRC_IPV4	|
569 					       MLX4_IB_RX_HASH_DST_IPV4	|
570 					       MLX4_IB_RX_HASH_SRC_IPV6	|
571 					       MLX4_IB_RX_HASH_DST_IPV6	|
572 					       MLX4_IB_RX_HASH_SRC_PORT_TCP |
573 					       MLX4_IB_RX_HASH_DST_PORT_TCP |
574 					       MLX4_IB_RX_HASH_SRC_PORT_UDP |
575 					       MLX4_IB_RX_HASH_DST_PORT_UDP |
576 					       MLX4_IB_RX_HASH_INNER)) {
577 		pr_debug("RX Hash fields_mask has unsupported mask (0x%llx)\n",
578 			 ucmd->rx_hash_fields_mask);
579 		return (-EOPNOTSUPP);
580 	}
581 
582 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) &&
583 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
584 		rss_ctx->flags = MLX4_RSS_IPV4;
585 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) ||
586 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
587 		pr_debug("RX Hash fields_mask is not supported - both IPv4 SRC and DST must be set\n");
588 		return (-EOPNOTSUPP);
589 	}
590 
591 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) &&
592 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
593 		rss_ctx->flags |= MLX4_RSS_IPV6;
594 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) ||
595 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
596 		pr_debug("RX Hash fields_mask is not supported - both IPv6 SRC and DST must be set\n");
597 		return (-EOPNOTSUPP);
598 	}
599 
600 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) &&
601 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
602 		if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS)) {
603 			pr_debug("RX Hash fields_mask for UDP is not supported\n");
604 			return (-EOPNOTSUPP);
605 		}
606 
607 		if (rss_ctx->flags & MLX4_RSS_IPV4)
608 			rss_ctx->flags |= MLX4_RSS_UDP_IPV4;
609 		if (rss_ctx->flags & MLX4_RSS_IPV6)
610 			rss_ctx->flags |= MLX4_RSS_UDP_IPV6;
611 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
612 			pr_debug("RX Hash fields_mask is not supported - UDP must be set with IPv4 or IPv6\n");
613 			return (-EOPNOTSUPP);
614 		}
615 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) ||
616 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
617 		pr_debug("RX Hash fields_mask is not supported - both UDP SRC and DST must be set\n");
618 		return (-EOPNOTSUPP);
619 	}
620 
621 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) &&
622 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
623 		if (rss_ctx->flags & MLX4_RSS_IPV4)
624 			rss_ctx->flags |= MLX4_RSS_TCP_IPV4;
625 		if (rss_ctx->flags & MLX4_RSS_IPV6)
626 			rss_ctx->flags |= MLX4_RSS_TCP_IPV6;
627 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
628 			pr_debug("RX Hash fields_mask is not supported - TCP must be set with IPv4 or IPv6\n");
629 			return (-EOPNOTSUPP);
630 		}
631 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) ||
632 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
633 		pr_debug("RX Hash fields_mask is not supported - both TCP SRC and DST must be set\n");
634 		return (-EOPNOTSUPP);
635 	}
636 
637 	if (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_INNER) {
638 		if (dev->dev->caps.tunnel_offload_mode ==
639 		    MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
640 			/*
641 			 * Hash according to inner headers if exist, otherwise
642 			 * according to outer headers.
643 			 */
644 			rss_ctx->flags |= MLX4_RSS_BY_INNER_HEADERS_IPONLY;
645 		} else {
646 			pr_debug("RSS Hash for inner headers isn't supported\n");
647 			return (-EOPNOTSUPP);
648 		}
649 	}
650 
651 	return 0;
652 }
653 
654 static int create_qp_rss(struct mlx4_ib_dev *dev,
655 			 struct ib_qp_init_attr *init_attr,
656 			 struct mlx4_ib_create_qp_rss *ucmd,
657 			 struct mlx4_ib_qp *qp)
658 {
659 	int qpn;
660 	int err;
661 
662 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
663 
664 	err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 0, qp->mqp.usage);
665 	if (err)
666 		return err;
667 
668 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
669 	if (err)
670 		goto err_qpn;
671 
672 	INIT_LIST_HEAD(&qp->gid_list);
673 	INIT_LIST_HEAD(&qp->steering_rules);
674 
675 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
676 	qp->state = IB_QPS_RESET;
677 
678 	/* Set dummy send resources to be compatible with HV and PRM */
679 	qp->sq_no_prefetch = 1;
680 	qp->sq.wqe_cnt = 1;
681 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
682 	qp->buf_size = qp->sq.wqe_cnt << MLX4_IB_MIN_SQ_STRIDE;
683 	qp->mtt = (to_mqp(
684 		   (struct ib_qp *)init_attr->rwq_ind_tbl->ind_tbl[0]))->mtt;
685 
686 	qp->rss_ctx = kzalloc(sizeof(*qp->rss_ctx), GFP_KERNEL);
687 	if (!qp->rss_ctx) {
688 		err = -ENOMEM;
689 		goto err_qp_alloc;
690 	}
691 
692 	err = set_qp_rss(dev, qp->rss_ctx, init_attr, ucmd);
693 	if (err)
694 		goto err;
695 
696 	return 0;
697 
698 err:
699 	kfree(qp->rss_ctx);
700 
701 err_qp_alloc:
702 	mlx4_qp_remove(dev->dev, &qp->mqp);
703 	mlx4_qp_free(dev->dev, &qp->mqp);
704 
705 err_qpn:
706 	mlx4_qp_release_range(dev->dev, qpn, 1);
707 	return err;
708 }
709 
710 static int _mlx4_ib_create_qp_rss(struct ib_pd *pd, struct mlx4_ib_qp *qp,
711 				  struct ib_qp_init_attr *init_attr,
712 				  struct ib_udata *udata)
713 {
714 	struct mlx4_ib_create_qp_rss ucmd = {};
715 	size_t required_cmd_sz;
716 	int err;
717 
718 	if (!udata) {
719 		pr_debug("RSS QP with NULL udata\n");
720 		return -EINVAL;
721 	}
722 
723 	if (udata->outlen)
724 		return -EOPNOTSUPP;
725 
726 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) +
727 					sizeof(ucmd.reserved1);
728 	if (udata->inlen < required_cmd_sz) {
729 		pr_debug("invalid inlen\n");
730 		return -EINVAL;
731 	}
732 
733 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
734 		pr_debug("copy failed\n");
735 		return -EFAULT;
736 	}
737 
738 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)))
739 		return -EOPNOTSUPP;
740 
741 	if (ucmd.comp_mask || ucmd.reserved1)
742 		return -EOPNOTSUPP;
743 
744 	if (udata->inlen > sizeof(ucmd) &&
745 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
746 				 udata->inlen - sizeof(ucmd))) {
747 		pr_debug("inlen is not supported\n");
748 		return -EOPNOTSUPP;
749 	}
750 
751 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
752 		pr_debug("RSS QP with unsupported QP type %d\n",
753 			 init_attr->qp_type);
754 		return -EOPNOTSUPP;
755 	}
756 
757 	if (init_attr->create_flags) {
758 		pr_debug("RSS QP doesn't support create flags\n");
759 		return -EOPNOTSUPP;
760 	}
761 
762 	if (init_attr->send_cq || init_attr->cap.max_send_wr) {
763 		pr_debug("RSS QP with unsupported send attributes\n");
764 		return -EOPNOTSUPP;
765 	}
766 
767 	qp->pri.vid = 0xFFFF;
768 	qp->alt.vid = 0xFFFF;
769 
770 	err = create_qp_rss(to_mdev(pd->device), init_attr, &ucmd, qp);
771 	if (err)
772 		return err;
773 
774 	qp->ibqp.qp_num = qp->mqp.qpn;
775 	return 0;
776 }
777 
778 /*
779  * This function allocates a WQN from a range which is consecutive and aligned
780  * to its size. In case the range is full, then it creates a new range and
781  * allocates WQN from it. The new range will be used for following allocations.
782  */
783 static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context,
784 			     struct mlx4_ib_qp *qp, int range_size, int *wqn)
785 {
786 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
787 	struct mlx4_wqn_range *range;
788 	int err = 0;
789 
790 	mutex_lock(&context->wqn_ranges_mutex);
791 
792 	range = list_first_entry_or_null(&context->wqn_ranges_list,
793 					 struct mlx4_wqn_range, list);
794 
795 	if (!range || (range->refcount == range->size) || range->dirty) {
796 		range = kzalloc(sizeof(*range), GFP_KERNEL);
797 		if (!range) {
798 			err = -ENOMEM;
799 			goto out;
800 		}
801 
802 		err = mlx4_qp_reserve_range(dev->dev, range_size,
803 					    range_size, &range->base_wqn, 0,
804 					    qp->mqp.usage);
805 		if (err) {
806 			kfree(range);
807 			goto out;
808 		}
809 
810 		range->size = range_size;
811 		list_add(&range->list, &context->wqn_ranges_list);
812 	} else if (range_size != 1) {
813 		/*
814 		 * Requesting a new range (>1) when last range is still open, is
815 		 * not valid.
816 		 */
817 		err = -EINVAL;
818 		goto out;
819 	}
820 
821 	qp->wqn_range = range;
822 
823 	*wqn = range->base_wqn + range->refcount;
824 
825 	range->refcount++;
826 
827 out:
828 	mutex_unlock(&context->wqn_ranges_mutex);
829 
830 	return err;
831 }
832 
833 static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context,
834 				struct mlx4_ib_qp *qp, bool dirty_release)
835 {
836 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
837 	struct mlx4_wqn_range *range;
838 
839 	mutex_lock(&context->wqn_ranges_mutex);
840 
841 	range = qp->wqn_range;
842 
843 	range->refcount--;
844 	if (!range->refcount) {
845 		mlx4_qp_release_range(dev->dev, range->base_wqn,
846 				      range->size);
847 		list_del(&range->list);
848 		kfree(range);
849 	} else if (dirty_release) {
850 	/*
851 	 * A range which one of its WQNs is destroyed, won't be able to be
852 	 * reused for further WQN allocations.
853 	 * The next created WQ will allocate a new range.
854 	 */
855 		range->dirty = true;
856 	}
857 
858 	mutex_unlock(&context->wqn_ranges_mutex);
859 }
860 
861 static int create_rq(struct ib_pd *pd, struct ib_qp_init_attr *init_attr,
862 		     struct ib_udata *udata, struct mlx4_ib_qp *qp)
863 {
864 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
865 	int qpn;
866 	int err;
867 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
868 		udata, struct mlx4_ib_ucontext, ibucontext);
869 	struct mlx4_ib_cq *mcq;
870 	unsigned long flags;
871 	int range_size;
872 	struct mlx4_ib_create_wq wq;
873 	size_t copy_len;
874 	int shift;
875 	int n;
876 
877 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
878 
879 	spin_lock_init(&qp->sq.lock);
880 	spin_lock_init(&qp->rq.lock);
881 	INIT_LIST_HEAD(&qp->gid_list);
882 	INIT_LIST_HEAD(&qp->steering_rules);
883 
884 	qp->state = IB_QPS_RESET;
885 
886 	copy_len = min(sizeof(struct mlx4_ib_create_wq), udata->inlen);
887 
888 	if (ib_copy_from_udata(&wq, udata, copy_len)) {
889 		err = -EFAULT;
890 		goto err;
891 	}
892 
893 	if (wq.comp_mask || wq.reserved[0] || wq.reserved[1] ||
894 	    wq.reserved[2]) {
895 		pr_debug("user command isn't supported\n");
896 		err = -EOPNOTSUPP;
897 		goto err;
898 	}
899 
900 	if (wq.log_range_size > ilog2(dev->dev->caps.max_rss_tbl_sz)) {
901 		pr_debug("WQN range size must be equal or smaller than %d\n",
902 			 dev->dev->caps.max_rss_tbl_sz);
903 		err = -EOPNOTSUPP;
904 		goto err;
905 	}
906 	range_size = 1 << wq.log_range_size;
907 
908 	if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS)
909 		qp->flags |= MLX4_IB_QP_SCATTER_FCS;
910 
911 	err = set_rq_size(dev, &init_attr->cap, true, true, qp, qp->inl_recv_sz);
912 	if (err)
913 		goto err;
914 
915 	qp->sq_no_prefetch = 1;
916 	qp->sq.wqe_cnt = 1;
917 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
918 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
919 		       (qp->sq.wqe_cnt << qp->sq.wqe_shift);
920 
921 	qp->umem = ib_umem_get(pd->device, wq.buf_addr, qp->buf_size, 0);
922 	if (IS_ERR(qp->umem)) {
923 		err = PTR_ERR(qp->umem);
924 		goto err;
925 	}
926 
927 	shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
928 	if (shift < 0) {
929 		err = shift;
930 		goto err_buf;
931 	}
932 
933 	err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
934 	if (err)
935 		goto err_buf;
936 
937 	err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
938 	if (err)
939 		goto err_mtt;
940 
941 	err = mlx4_ib_db_map_user(udata, wq.db_addr, &qp->db);
942 	if (err)
943 		goto err_mtt;
944 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
945 
946 	err = mlx4_ib_alloc_wqn(context, qp, range_size, &qpn);
947 	if (err)
948 		goto err_wrid;
949 
950 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
951 	if (err)
952 		goto err_qpn;
953 
954 	/*
955 	 * Hardware wants QPN written in big-endian order (after
956 	 * shifting) for send doorbell.  Precompute this value to save
957 	 * a little bit when posting sends.
958 	 */
959 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
960 
961 	qp->mqp.event = mlx4_ib_wq_event;
962 
963 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
964 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
965 			 to_mcq(init_attr->recv_cq));
966 	/* Maintain device to QPs access, needed for further handling
967 	 * via reset flow
968 	 */
969 	list_add_tail(&qp->qps_list, &dev->qp_list);
970 	/* Maintain CQ to QPs access, needed for further handling
971 	 * via reset flow
972 	 */
973 	mcq = to_mcq(init_attr->send_cq);
974 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
975 	mcq = to_mcq(init_attr->recv_cq);
976 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
977 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
978 			   to_mcq(init_attr->recv_cq));
979 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
980 	return 0;
981 
982 err_qpn:
983 	mlx4_ib_release_wqn(context, qp, 0);
984 err_wrid:
985 	mlx4_ib_db_unmap_user(context, &qp->db);
986 
987 err_mtt:
988 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
989 err_buf:
990 	ib_umem_release(qp->umem);
991 err:
992 	return err;
993 }
994 
995 static int create_qp_common(struct ib_pd *pd, struct ib_qp_init_attr *init_attr,
996 			    struct ib_udata *udata, int sqpn,
997 			    struct mlx4_ib_qp *qp)
998 {
999 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
1000 	int qpn;
1001 	int err;
1002 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
1003 		udata, struct mlx4_ib_ucontext, ibucontext);
1004 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
1005 	struct mlx4_ib_cq *mcq;
1006 	unsigned long flags;
1007 
1008 	/* When tunneling special qps, we use a plain UD qp */
1009 	if (sqpn) {
1010 		if (mlx4_is_mfunc(dev->dev) &&
1011 		    (!mlx4_is_master(dev->dev) ||
1012 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
1013 			if (init_attr->qp_type == IB_QPT_GSI)
1014 				qp_type = MLX4_IB_QPT_PROXY_GSI;
1015 			else {
1016 				if (mlx4_is_master(dev->dev) ||
1017 				    qp0_enabled_vf(dev->dev, sqpn))
1018 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
1019 				else
1020 					qp_type = MLX4_IB_QPT_PROXY_SMI;
1021 			}
1022 		}
1023 		qpn = sqpn;
1024 		/* add extra sg entry for tunneling */
1025 		init_attr->cap.max_recv_sge++;
1026 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
1027 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
1028 			container_of(init_attr,
1029 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
1030 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
1031 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
1032 		    !mlx4_is_master(dev->dev))
1033 			return -EINVAL;
1034 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
1035 			qp_type = MLX4_IB_QPT_TUN_GSI;
1036 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
1037 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
1038 					     tnl_init->port))
1039 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
1040 		else
1041 			qp_type = MLX4_IB_QPT_TUN_SMI;
1042 		/* we are definitely in the PPF here, since we are creating
1043 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
1044 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
1045 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
1046 		sqpn = qpn;
1047 	}
1048 
1049 	if (init_attr->qp_type == IB_QPT_SMI ||
1050 	    init_attr->qp_type == IB_QPT_GSI || qp_type == MLX4_IB_QPT_SMI ||
1051 	    qp_type == MLX4_IB_QPT_GSI ||
1052 	    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
1053 			MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
1054 		qp->sqp = kzalloc(sizeof(struct mlx4_ib_sqp), GFP_KERNEL);
1055 		if (!qp->sqp)
1056 			return -ENOMEM;
1057 	}
1058 
1059 	qp->mlx4_ib_qp_type = qp_type;
1060 
1061 	spin_lock_init(&qp->sq.lock);
1062 	spin_lock_init(&qp->rq.lock);
1063 	INIT_LIST_HEAD(&qp->gid_list);
1064 	INIT_LIST_HEAD(&qp->steering_rules);
1065 
1066 	qp->state = IB_QPS_RESET;
1067 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1068 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1069 
1070 	if (udata) {
1071 		struct mlx4_ib_create_qp ucmd;
1072 		size_t copy_len;
1073 		int shift;
1074 		int n;
1075 
1076 		copy_len = sizeof(struct mlx4_ib_create_qp);
1077 
1078 		if (ib_copy_from_udata(&ucmd, udata, copy_len)) {
1079 			err = -EFAULT;
1080 			goto err;
1081 		}
1082 
1083 		qp->inl_recv_sz = ucmd.inl_recv_sz;
1084 
1085 		if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1086 			if (!(dev->dev->caps.flags &
1087 			      MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
1088 				pr_debug("scatter FCS is unsupported\n");
1089 				err = -EOPNOTSUPP;
1090 				goto err;
1091 			}
1092 
1093 			qp->flags |= MLX4_IB_QP_SCATTER_FCS;
1094 		}
1095 
1096 		err = set_rq_size(dev, &init_attr->cap, udata,
1097 				  qp_has_rq(init_attr), qp, qp->inl_recv_sz);
1098 		if (err)
1099 			goto err;
1100 
1101 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
1102 
1103 		err = set_user_sq_size(dev, qp, &ucmd);
1104 		if (err)
1105 			goto err;
1106 
1107 		qp->umem =
1108 			ib_umem_get(pd->device, ucmd.buf_addr, qp->buf_size, 0);
1109 		if (IS_ERR(qp->umem)) {
1110 			err = PTR_ERR(qp->umem);
1111 			goto err;
1112 		}
1113 
1114 		shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
1115 		if (shift < 0) {
1116 			err = shift;
1117 			goto err_buf;
1118 		}
1119 
1120 		err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
1121 		if (err)
1122 			goto err_buf;
1123 
1124 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
1125 		if (err)
1126 			goto err_mtt;
1127 
1128 		if (qp_has_rq(init_attr)) {
1129 			err = mlx4_ib_db_map_user(udata, ucmd.db_addr, &qp->db);
1130 			if (err)
1131 				goto err_mtt;
1132 		}
1133 		qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
1134 	} else {
1135 		err = set_rq_size(dev, &init_attr->cap, udata,
1136 				  qp_has_rq(init_attr), qp, 0);
1137 		if (err)
1138 			goto err;
1139 
1140 		qp->sq_no_prefetch = 0;
1141 
1142 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
1143 			qp->flags |= MLX4_IB_QP_LSO;
1144 
1145 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1146 			if (dev->steering_support ==
1147 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
1148 				qp->flags |= MLX4_IB_QP_NETIF;
1149 			else {
1150 				err = -EINVAL;
1151 				goto err;
1152 			}
1153 		}
1154 
1155 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
1156 		if (err)
1157 			goto err;
1158 
1159 		if (qp_has_rq(init_attr)) {
1160 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
1161 			if (err)
1162 				goto err;
1163 
1164 			*qp->db.db = 0;
1165 		}
1166 
1167 		if (mlx4_buf_alloc(dev->dev, qp->buf_size,  PAGE_SIZE * 2,
1168 				   &qp->buf)) {
1169 			err = -ENOMEM;
1170 			goto err_db;
1171 		}
1172 
1173 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
1174 				    &qp->mtt);
1175 		if (err)
1176 			goto err_buf;
1177 
1178 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
1179 		if (err)
1180 			goto err_mtt;
1181 
1182 		qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1183 					     sizeof(u64), GFP_KERNEL);
1184 		qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1185 					     sizeof(u64), GFP_KERNEL);
1186 		if (!qp->sq.wrid || !qp->rq.wrid) {
1187 			err = -ENOMEM;
1188 			goto err_wrid;
1189 		}
1190 		qp->mqp.usage = MLX4_RES_USAGE_DRIVER;
1191 	}
1192 
1193 	if (sqpn) {
1194 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1195 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
1196 			if (alloc_proxy_bufs(pd->device, qp)) {
1197 				err = -ENOMEM;
1198 				goto err_wrid;
1199 			}
1200 		}
1201 	} else {
1202 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
1203 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
1204 		 * VLAN insertion. */
1205 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
1206 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
1207 						    (init_attr->cap.max_send_wr ?
1208 						     MLX4_RESERVE_ETH_BF_QP : 0) |
1209 						    (init_attr->cap.max_recv_wr ?
1210 						     MLX4_RESERVE_A0_QP : 0),
1211 						    qp->mqp.usage);
1212 		else
1213 			if (qp->flags & MLX4_IB_QP_NETIF)
1214 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
1215 			else
1216 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
1217 							    &qpn, 0, qp->mqp.usage);
1218 		if (err)
1219 			goto err_proxy;
1220 	}
1221 
1222 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1223 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1224 
1225 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
1226 	if (err)
1227 		goto err_qpn;
1228 
1229 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1230 		qp->mqp.qpn |= (1 << 23);
1231 
1232 	/*
1233 	 * Hardware wants QPN written in big-endian order (after
1234 	 * shifting) for send doorbell.  Precompute this value to save
1235 	 * a little bit when posting sends.
1236 	 */
1237 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1238 
1239 	qp->mqp.event = mlx4_ib_qp_event;
1240 
1241 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1242 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
1243 			 to_mcq(init_attr->recv_cq));
1244 	/* Maintain device to QPs access, needed for further handling
1245 	 * via reset flow
1246 	 */
1247 	list_add_tail(&qp->qps_list, &dev->qp_list);
1248 	/* Maintain CQ to QPs access, needed for further handling
1249 	 * via reset flow
1250 	 */
1251 	mcq = to_mcq(init_attr->send_cq);
1252 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
1253 	mcq = to_mcq(init_attr->recv_cq);
1254 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
1255 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
1256 			   to_mcq(init_attr->recv_cq));
1257 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1258 	return 0;
1259 
1260 err_qpn:
1261 	if (!sqpn) {
1262 		if (qp->flags & MLX4_IB_QP_NETIF)
1263 			mlx4_ib_steer_qp_free(dev, qpn, 1);
1264 		else
1265 			mlx4_qp_release_range(dev->dev, qpn, 1);
1266 	}
1267 err_proxy:
1268 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1269 		free_proxy_bufs(pd->device, qp);
1270 err_wrid:
1271 	if (udata) {
1272 		if (qp_has_rq(init_attr))
1273 			mlx4_ib_db_unmap_user(context, &qp->db);
1274 	} else {
1275 		kvfree(qp->sq.wrid);
1276 		kvfree(qp->rq.wrid);
1277 	}
1278 
1279 err_mtt:
1280 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1281 
1282 err_buf:
1283 	if (!qp->umem)
1284 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1285 	ib_umem_release(qp->umem);
1286 
1287 err_db:
1288 	if (!udata && qp_has_rq(init_attr))
1289 		mlx4_db_free(dev->dev, &qp->db);
1290 
1291 err:
1292 	kfree(qp->sqp);
1293 	return err;
1294 }
1295 
1296 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
1297 {
1298 	switch (state) {
1299 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
1300 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
1301 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
1302 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
1303 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
1304 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
1305 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
1306 	default:		return -1;
1307 	}
1308 }
1309 
1310 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1311 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1312 {
1313 	if (send_cq == recv_cq) {
1314 		spin_lock(&send_cq->lock);
1315 		__acquire(&recv_cq->lock);
1316 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1317 		spin_lock(&send_cq->lock);
1318 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1319 	} else {
1320 		spin_lock(&recv_cq->lock);
1321 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1322 	}
1323 }
1324 
1325 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1326 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1327 {
1328 	if (send_cq == recv_cq) {
1329 		__release(&recv_cq->lock);
1330 		spin_unlock(&send_cq->lock);
1331 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1332 		spin_unlock(&recv_cq->lock);
1333 		spin_unlock(&send_cq->lock);
1334 	} else {
1335 		spin_unlock(&send_cq->lock);
1336 		spin_unlock(&recv_cq->lock);
1337 	}
1338 }
1339 
1340 static void del_gid_entries(struct mlx4_ib_qp *qp)
1341 {
1342 	struct mlx4_ib_gid_entry *ge, *tmp;
1343 
1344 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1345 		list_del(&ge->list);
1346 		kfree(ge);
1347 	}
1348 }
1349 
1350 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
1351 {
1352 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
1353 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
1354 	else
1355 		return to_mpd(qp->ibqp.pd);
1356 }
1357 
1358 static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src,
1359 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1360 {
1361 	switch (qp->ibqp.qp_type) {
1362 	case IB_QPT_XRC_TGT:
1363 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1364 		*recv_cq = *send_cq;
1365 		break;
1366 	case IB_QPT_XRC_INI:
1367 		*send_cq = to_mcq(qp->ibqp.send_cq);
1368 		*recv_cq = *send_cq;
1369 		break;
1370 	default:
1371 		*recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) :
1372 						     to_mcq(qp->ibwq.cq);
1373 		*send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) :
1374 						     *recv_cq;
1375 		break;
1376 	}
1377 }
1378 
1379 static void destroy_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1380 {
1381 	if (qp->state != IB_QPS_RESET) {
1382 		int i;
1383 
1384 		for (i = 0; i < (1 << qp->ibqp.rwq_ind_tbl->log_ind_tbl_size);
1385 		     i++) {
1386 			struct ib_wq *ibwq = qp->ibqp.rwq_ind_tbl->ind_tbl[i];
1387 			struct mlx4_ib_qp *wq =	to_mqp((struct ib_qp *)ibwq);
1388 
1389 			mutex_lock(&wq->mutex);
1390 
1391 			wq->rss_usecnt--;
1392 
1393 			mutex_unlock(&wq->mutex);
1394 		}
1395 
1396 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1397 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1398 			pr_warn("modify QP %06x to RESET failed.\n",
1399 				qp->mqp.qpn);
1400 	}
1401 
1402 	mlx4_qp_remove(dev->dev, &qp->mqp);
1403 	mlx4_qp_free(dev->dev, &qp->mqp);
1404 	mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1405 	del_gid_entries(qp);
1406 }
1407 
1408 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1409 			      enum mlx4_ib_source_type src,
1410 			      struct ib_udata *udata)
1411 {
1412 	struct mlx4_ib_cq *send_cq, *recv_cq;
1413 	unsigned long flags;
1414 
1415 	if (qp->state != IB_QPS_RESET) {
1416 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1417 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1418 			pr_warn("modify QP %06x to RESET failed.\n",
1419 			       qp->mqp.qpn);
1420 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1421 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1422 			qp->pri.smac = 0;
1423 			qp->pri.smac_port = 0;
1424 		}
1425 		if (qp->alt.smac) {
1426 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1427 			qp->alt.smac = 0;
1428 		}
1429 		if (qp->pri.vid < 0x1000) {
1430 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1431 			qp->pri.vid = 0xFFFF;
1432 			qp->pri.candidate_vid = 0xFFFF;
1433 			qp->pri.update_vid = 0;
1434 		}
1435 		if (qp->alt.vid < 0x1000) {
1436 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1437 			qp->alt.vid = 0xFFFF;
1438 			qp->alt.candidate_vid = 0xFFFF;
1439 			qp->alt.update_vid = 0;
1440 		}
1441 	}
1442 
1443 	get_cqs(qp, src, &send_cq, &recv_cq);
1444 
1445 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1446 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1447 
1448 	/* del from lists under both locks above to protect reset flow paths */
1449 	list_del(&qp->qps_list);
1450 	list_del(&qp->cq_send_list);
1451 	list_del(&qp->cq_recv_list);
1452 	if (!udata) {
1453 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1454 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1455 		if (send_cq != recv_cq)
1456 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1457 	}
1458 
1459 	mlx4_qp_remove(dev->dev, &qp->mqp);
1460 
1461 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1462 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1463 
1464 	mlx4_qp_free(dev->dev, &qp->mqp);
1465 
1466 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1467 		if (qp->flags & MLX4_IB_QP_NETIF)
1468 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1469 		else if (src == MLX4_IB_RWQ_SRC)
1470 			mlx4_ib_release_wqn(
1471 				rdma_udata_to_drv_context(
1472 					udata,
1473 					struct mlx4_ib_ucontext,
1474 					ibucontext),
1475 				qp, 1);
1476 		else
1477 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1478 	}
1479 
1480 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1481 
1482 	if (udata) {
1483 		if (qp->rq.wqe_cnt) {
1484 			struct mlx4_ib_ucontext *mcontext =
1485 				rdma_udata_to_drv_context(
1486 					udata,
1487 					struct mlx4_ib_ucontext,
1488 					ibucontext);
1489 
1490 			mlx4_ib_db_unmap_user(mcontext, &qp->db);
1491 		}
1492 	} else {
1493 		kvfree(qp->sq.wrid);
1494 		kvfree(qp->rq.wrid);
1495 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1496 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1497 			free_proxy_bufs(&dev->ib_dev, qp);
1498 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1499 		if (qp->rq.wqe_cnt)
1500 			mlx4_db_free(dev->dev, &qp->db);
1501 	}
1502 	ib_umem_release(qp->umem);
1503 
1504 	del_gid_entries(qp);
1505 }
1506 
1507 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1508 {
1509 	/* Native or PPF */
1510 	if (!mlx4_is_mfunc(dev->dev) ||
1511 	    (mlx4_is_master(dev->dev) &&
1512 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1513 		return  dev->dev->phys_caps.base_sqpn +
1514 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1515 			attr->port_num - 1;
1516 	}
1517 	/* PF or VF -- creating proxies */
1518 	if (attr->qp_type == IB_QPT_SMI)
1519 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp0_proxy;
1520 	else
1521 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp1_proxy;
1522 }
1523 
1524 static int _mlx4_ib_create_qp(struct ib_pd *pd, struct mlx4_ib_qp *qp,
1525 			      struct ib_qp_init_attr *init_attr,
1526 			      struct ib_udata *udata)
1527 {
1528 	int err;
1529 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1530 	u16 xrcdn = 0;
1531 
1532 	if (init_attr->rwq_ind_tbl)
1533 		return _mlx4_ib_create_qp_rss(pd, qp, init_attr, udata);
1534 
1535 	/*
1536 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1537 	 * and only for kernel UD QPs.
1538 	 */
1539 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1540 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1541 					MLX4_IB_SRIOV_TUNNEL_QP |
1542 					MLX4_IB_SRIOV_SQP |
1543 					MLX4_IB_QP_NETIF |
1544 					MLX4_IB_QP_CREATE_ROCE_V2_GSI))
1545 		return -EOPNOTSUPP;
1546 
1547 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1548 		if (init_attr->qp_type != IB_QPT_UD)
1549 			return -EINVAL;
1550 	}
1551 
1552 	if (init_attr->create_flags) {
1553 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1554 			return -EINVAL;
1555 
1556 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1557 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1558 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1559 		     init_attr->qp_type != IB_QPT_UD) ||
1560 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1561 		     init_attr->qp_type > IB_QPT_GSI) ||
1562 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1563 		     init_attr->qp_type != IB_QPT_GSI))
1564 			return -EINVAL;
1565 	}
1566 
1567 	switch (init_attr->qp_type) {
1568 	case IB_QPT_XRC_TGT:
1569 		pd = to_mxrcd(init_attr->xrcd)->pd;
1570 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1571 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1572 		fallthrough;
1573 	case IB_QPT_XRC_INI:
1574 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1575 			return -ENOSYS;
1576 		init_attr->recv_cq = init_attr->send_cq;
1577 		fallthrough;
1578 	case IB_QPT_RC:
1579 	case IB_QPT_UC:
1580 	case IB_QPT_RAW_PACKET:
1581 	case IB_QPT_UD:
1582 		qp->pri.vid = 0xFFFF;
1583 		qp->alt.vid = 0xFFFF;
1584 		err = create_qp_common(pd, init_attr, udata, 0, qp);
1585 		if (err)
1586 			return err;
1587 
1588 		qp->ibqp.qp_num = qp->mqp.qpn;
1589 		qp->xrcdn = xrcdn;
1590 		break;
1591 	case IB_QPT_SMI:
1592 	case IB_QPT_GSI:
1593 	{
1594 		int sqpn;
1595 
1596 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1597 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev,
1598 							1, 1, &sqpn, 0,
1599 							MLX4_RES_USAGE_DRIVER);
1600 
1601 			if (res)
1602 				return res;
1603 		} else {
1604 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1605 		}
1606 
1607 		qp->pri.vid = 0xFFFF;
1608 		qp->alt.vid = 0xFFFF;
1609 		err = create_qp_common(pd, init_attr, udata, sqpn, qp);
1610 		if (err)
1611 			return err;
1612 
1613 		if (init_attr->create_flags &
1614 		    (MLX4_IB_SRIOV_SQP | MLX4_IB_SRIOV_TUNNEL_QP))
1615 			/* Internal QP created with ib_create_qp */
1616 			rdma_restrack_no_track(&qp->ibqp.res);
1617 
1618 		qp->port	= init_attr->port_num;
1619 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1620 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1621 		break;
1622 	}
1623 	default:
1624 		/* Don't support raw QPs */
1625 		return -EOPNOTSUPP;
1626 	}
1627 	return 0;
1628 }
1629 
1630 int mlx4_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr,
1631 		      struct ib_udata *udata)
1632 {
1633 	struct ib_device *device = ibqp->device;
1634 	struct mlx4_ib_dev *dev = to_mdev(device);
1635 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1636 	struct ib_pd *pd = ibqp->pd;
1637 	int ret;
1638 
1639 	mutex_init(&qp->mutex);
1640 	ret = _mlx4_ib_create_qp(pd, qp, init_attr, udata);
1641 	if (ret)
1642 		return ret;
1643 
1644 	if (init_attr->qp_type == IB_QPT_GSI &&
1645 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1646 		struct mlx4_ib_sqp *sqp = qp->sqp;
1647 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1648 
1649 		if (is_eth &&
1650 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1651 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1652 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1653 
1654 			if (IS_ERR(sqp->roce_v2_gsi)) {
1655 				pr_err("Failed to create GSI QP for RoCEv2 (%pe)\n",
1656 				       sqp->roce_v2_gsi);
1657 				sqp->roce_v2_gsi = NULL;
1658 			} else {
1659 				to_mqp(sqp->roce_v2_gsi)->flags |=
1660 					MLX4_IB_ROCE_V2_GSI_QP;
1661 			}
1662 
1663 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1664 		}
1665 	}
1666 	return 0;
1667 }
1668 
1669 static int _mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
1670 {
1671 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1672 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1673 
1674 	if (is_qp0(dev, mqp))
1675 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1676 
1677 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1678 	    dev->qp1_proxy[mqp->port - 1] == mqp) {
1679 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1680 		dev->qp1_proxy[mqp->port - 1] = NULL;
1681 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1682 	}
1683 
1684 	if (mqp->counter_index)
1685 		mlx4_ib_free_qp_counter(dev, mqp);
1686 
1687 	if (qp->rwq_ind_tbl) {
1688 		destroy_qp_rss(dev, mqp);
1689 	} else {
1690 		destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, udata);
1691 	}
1692 
1693 	kfree(mqp->sqp);
1694 	return 0;
1695 }
1696 
1697 int mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
1698 {
1699 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1700 
1701 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1702 		struct mlx4_ib_sqp *sqp = mqp->sqp;
1703 
1704 		if (sqp->roce_v2_gsi)
1705 			ib_destroy_qp(sqp->roce_v2_gsi);
1706 	}
1707 
1708 	return _mlx4_ib_destroy_qp(qp, udata);
1709 }
1710 
1711 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1712 {
1713 	switch (type) {
1714 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1715 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1716 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1717 	case MLX4_IB_QPT_XRC_INI:
1718 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1719 	case MLX4_IB_QPT_SMI:
1720 	case MLX4_IB_QPT_GSI:
1721 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1722 
1723 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1724 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1725 						MLX4_QP_ST_MLX : -1);
1726 	case MLX4_IB_QPT_PROXY_SMI:
1727 	case MLX4_IB_QPT_TUN_SMI:
1728 	case MLX4_IB_QPT_PROXY_GSI:
1729 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1730 						MLX4_QP_ST_UD : -1);
1731 	default:			return -1;
1732 	}
1733 }
1734 
1735 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1736 				   int attr_mask)
1737 {
1738 	u8 dest_rd_atomic;
1739 	u32 access_flags;
1740 	u32 hw_access_flags = 0;
1741 
1742 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1743 		dest_rd_atomic = attr->max_dest_rd_atomic;
1744 	else
1745 		dest_rd_atomic = qp->resp_depth;
1746 
1747 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1748 		access_flags = attr->qp_access_flags;
1749 	else
1750 		access_flags = qp->atomic_rd_en;
1751 
1752 	if (!dest_rd_atomic)
1753 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1754 
1755 	if (access_flags & IB_ACCESS_REMOTE_READ)
1756 		hw_access_flags |= MLX4_QP_BIT_RRE;
1757 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1758 		hw_access_flags |= MLX4_QP_BIT_RAE;
1759 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1760 		hw_access_flags |= MLX4_QP_BIT_RWE;
1761 
1762 	return cpu_to_be32(hw_access_flags);
1763 }
1764 
1765 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1766 			    int attr_mask)
1767 {
1768 	if (attr_mask & IB_QP_PKEY_INDEX)
1769 		sqp->pkey_index = attr->pkey_index;
1770 	if (attr_mask & IB_QP_QKEY)
1771 		sqp->qkey = attr->qkey;
1772 	if (attr_mask & IB_QP_SQ_PSN)
1773 		sqp->send_psn = attr->sq_psn;
1774 }
1775 
1776 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1777 {
1778 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1779 }
1780 
1781 static int _mlx4_set_path(struct mlx4_ib_dev *dev,
1782 			  const struct rdma_ah_attr *ah,
1783 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1784 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1785 {
1786 	int vidx;
1787 	int smac_index;
1788 	int err;
1789 
1790 	path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
1791 	path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
1792 	if (rdma_ah_get_static_rate(ah)) {
1793 		path->static_rate = rdma_ah_get_static_rate(ah) +
1794 				    MLX4_STAT_RATE_OFFSET;
1795 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1796 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1797 			--path->static_rate;
1798 	} else
1799 		path->static_rate = 0;
1800 
1801 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
1802 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
1803 		int real_sgid_index =
1804 			mlx4_ib_gid_index_to_real_index(dev, grh->sgid_attr);
1805 
1806 		if (real_sgid_index < 0)
1807 			return real_sgid_index;
1808 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1809 			pr_err("sgid_index (%u) too large. max is %d\n",
1810 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1811 			return -1;
1812 		}
1813 
1814 		path->grh_mylmc |= 1 << 7;
1815 		path->mgid_index = real_sgid_index;
1816 		path->hop_limit  = grh->hop_limit;
1817 		path->tclass_flowlabel =
1818 			cpu_to_be32((grh->traffic_class << 20) |
1819 				    (grh->flow_label));
1820 		memcpy(path->rgid, grh->dgid.raw, 16);
1821 	}
1822 
1823 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
1824 		if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
1825 			return -1;
1826 
1827 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1828 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
1829 
1830 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1831 		if (vlan_tag < 0x1000) {
1832 			if (smac_info->vid < 0x1000) {
1833 				/* both valid vlan ids */
1834 				if (smac_info->vid != vlan_tag) {
1835 					/* different VIDs.  unreg old and reg new */
1836 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1837 					if (err)
1838 						return err;
1839 					smac_info->candidate_vid = vlan_tag;
1840 					smac_info->candidate_vlan_index = vidx;
1841 					smac_info->candidate_vlan_port = port;
1842 					smac_info->update_vid = 1;
1843 					path->vlan_index = vidx;
1844 				} else {
1845 					path->vlan_index = smac_info->vlan_index;
1846 				}
1847 			} else {
1848 				/* no current vlan tag in qp */
1849 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1850 				if (err)
1851 					return err;
1852 				smac_info->candidate_vid = vlan_tag;
1853 				smac_info->candidate_vlan_index = vidx;
1854 				smac_info->candidate_vlan_port = port;
1855 				smac_info->update_vid = 1;
1856 				path->vlan_index = vidx;
1857 			}
1858 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1859 			path->fl = 1 << 6;
1860 		} else {
1861 			/* have current vlan tag. unregister it at modify-qp success */
1862 			if (smac_info->vid < 0x1000) {
1863 				smac_info->candidate_vid = 0xFFFF;
1864 				smac_info->update_vid = 1;
1865 			}
1866 		}
1867 
1868 		/* get smac_index for RoCE use.
1869 		 * If no smac was yet assigned, register one.
1870 		 * If one was already assigned, but the new mac differs,
1871 		 * unregister the old one and register the new one.
1872 		*/
1873 		if ((!smac_info->smac && !smac_info->smac_port) ||
1874 		    smac_info->smac != smac) {
1875 			/* register candidate now, unreg if needed, after success */
1876 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1877 			if (smac_index >= 0) {
1878 				smac_info->candidate_smac_index = smac_index;
1879 				smac_info->candidate_smac = smac;
1880 				smac_info->candidate_smac_port = port;
1881 			} else {
1882 				return -EINVAL;
1883 			}
1884 		} else {
1885 			smac_index = smac_info->smac_index;
1886 		}
1887 		memcpy(path->dmac, ah->roce.dmac, 6);
1888 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1889 		/* put MAC table smac index for IBoE */
1890 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1891 	} else {
1892 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1893 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
1894 	}
1895 
1896 	return 0;
1897 }
1898 
1899 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1900 			 enum ib_qp_attr_mask qp_attr_mask,
1901 			 struct mlx4_ib_qp *mqp,
1902 			 struct mlx4_qp_path *path, u8 port,
1903 			 u16 vlan_id, u8 *smac)
1904 {
1905 	return _mlx4_set_path(dev, &qp->ah_attr,
1906 			      ether_addr_to_u64(smac),
1907 			      vlan_id,
1908 			      path, &mqp->pri, port);
1909 }
1910 
1911 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1912 			     const struct ib_qp_attr *qp,
1913 			     enum ib_qp_attr_mask qp_attr_mask,
1914 			     struct mlx4_ib_qp *mqp,
1915 			     struct mlx4_qp_path *path, u8 port)
1916 {
1917 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1918 			      0,
1919 			      0xffff,
1920 			      path, &mqp->alt, port);
1921 }
1922 
1923 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1924 {
1925 	struct mlx4_ib_gid_entry *ge, *tmp;
1926 
1927 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1928 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1929 			ge->added = 1;
1930 			ge->port = qp->port;
1931 		}
1932 	}
1933 }
1934 
1935 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1936 				    struct mlx4_ib_qp *qp,
1937 				    struct mlx4_qp_context *context)
1938 {
1939 	u64 u64_mac;
1940 	int smac_index;
1941 
1942 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1943 
1944 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1945 	if (!qp->pri.smac && !qp->pri.smac_port) {
1946 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1947 		if (smac_index >= 0) {
1948 			qp->pri.candidate_smac_index = smac_index;
1949 			qp->pri.candidate_smac = u64_mac;
1950 			qp->pri.candidate_smac_port = qp->port;
1951 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1952 		} else {
1953 			return -ENOENT;
1954 		}
1955 	}
1956 	return 0;
1957 }
1958 
1959 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1960 {
1961 	struct counter_index *new_counter_index;
1962 	int err;
1963 	u32 tmp_idx;
1964 
1965 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1966 	    IB_LINK_LAYER_ETHERNET ||
1967 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1968 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1969 		return 0;
1970 
1971 	err = mlx4_counter_alloc(dev->dev, &tmp_idx, MLX4_RES_USAGE_DRIVER);
1972 	if (err)
1973 		return err;
1974 
1975 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1976 	if (!new_counter_index) {
1977 		mlx4_counter_free(dev->dev, tmp_idx);
1978 		return -ENOMEM;
1979 	}
1980 
1981 	new_counter_index->index = tmp_idx;
1982 	new_counter_index->allocated = 1;
1983 	qp->counter_index = new_counter_index;
1984 
1985 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
1986 	list_add_tail(&new_counter_index->list,
1987 		      &dev->counters_table[qp->port - 1].counters_list);
1988 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
1989 
1990 	return 0;
1991 }
1992 
1993 enum {
1994 	MLX4_QPC_ROCE_MODE_1 = 0,
1995 	MLX4_QPC_ROCE_MODE_2 = 2,
1996 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
1997 };
1998 
1999 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
2000 {
2001 	switch (gid_type) {
2002 	case IB_GID_TYPE_ROCE:
2003 		return MLX4_QPC_ROCE_MODE_1;
2004 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
2005 		return MLX4_QPC_ROCE_MODE_2;
2006 	default:
2007 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
2008 	}
2009 }
2010 
2011 /*
2012  * Go over all RSS QP's childes (WQs) and apply their HW state according to
2013  * their logic state if the RSS QP is the first RSS QP associated for the WQ.
2014  */
2015 static int bringup_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, u8 port_num,
2016 			    struct ib_udata *udata)
2017 {
2018 	int err = 0;
2019 	int i;
2020 
2021 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2022 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2023 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2024 
2025 		mutex_lock(&wq->mutex);
2026 
2027 		/* Mlx4_ib restrictions:
2028 		 * WQ's is associated to a port according to the RSS QP it is
2029 		 * associates to.
2030 		 * In case the WQ is associated to a different port by another
2031 		 * RSS QP, return a failure.
2032 		 */
2033 		if ((wq->rss_usecnt > 0) && (wq->port != port_num)) {
2034 			err = -EINVAL;
2035 			mutex_unlock(&wq->mutex);
2036 			break;
2037 		}
2038 		wq->port = port_num;
2039 		if ((wq->rss_usecnt == 0) && (ibwq->state == IB_WQS_RDY)) {
2040 			err = _mlx4_ib_modify_wq(ibwq, IB_WQS_RDY, udata);
2041 			if (err) {
2042 				mutex_unlock(&wq->mutex);
2043 				break;
2044 			}
2045 		}
2046 		wq->rss_usecnt++;
2047 
2048 		mutex_unlock(&wq->mutex);
2049 	}
2050 
2051 	if (i && err) {
2052 		int j;
2053 
2054 		for (j = (i - 1); j >= 0; j--) {
2055 			struct ib_wq *ibwq = ind_tbl->ind_tbl[j];
2056 			struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2057 
2058 			mutex_lock(&wq->mutex);
2059 
2060 			if ((wq->rss_usecnt == 1) &&
2061 			    (ibwq->state == IB_WQS_RDY))
2062 				if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET,
2063 						       udata))
2064 					pr_warn("failed to reverse WQN=0x%06x\n",
2065 						ibwq->wq_num);
2066 			wq->rss_usecnt--;
2067 
2068 			mutex_unlock(&wq->mutex);
2069 		}
2070 	}
2071 
2072 	return err;
2073 }
2074 
2075 static void bring_down_rss_rwqs(struct ib_rwq_ind_table *ind_tbl,
2076 				struct ib_udata *udata)
2077 {
2078 	int i;
2079 
2080 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2081 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2082 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2083 
2084 		mutex_lock(&wq->mutex);
2085 
2086 		if ((wq->rss_usecnt == 1) && (ibwq->state == IB_WQS_RDY))
2087 			if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET, udata))
2088 				pr_warn("failed to reverse WQN=%x\n",
2089 					ibwq->wq_num);
2090 		wq->rss_usecnt--;
2091 
2092 		mutex_unlock(&wq->mutex);
2093 	}
2094 }
2095 
2096 static void fill_qp_rss_context(struct mlx4_qp_context *context,
2097 				struct mlx4_ib_qp *qp)
2098 {
2099 	struct mlx4_rss_context *rss_context;
2100 
2101 	rss_context = (void *)context + offsetof(struct mlx4_qp_context,
2102 			pri_path) + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
2103 
2104 	rss_context->base_qpn = cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz);
2105 	rss_context->default_qpn =
2106 		cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz & 0xffffff);
2107 	if (qp->rss_ctx->flags & (MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6))
2108 		rss_context->base_qpn_udp = rss_context->default_qpn;
2109 	rss_context->flags = qp->rss_ctx->flags;
2110 	/* Currently support just toeplitz */
2111 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
2112 
2113 	memcpy(rss_context->rss_key, qp->rss_ctx->rss_key,
2114 	       MLX4_EN_RSS_KEY_SIZE);
2115 }
2116 
2117 static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type,
2118 			       const struct ib_qp_attr *attr, int attr_mask,
2119 			       enum ib_qp_state cur_state,
2120 			       enum ib_qp_state new_state,
2121 			       struct ib_udata *udata)
2122 {
2123 	struct ib_srq  *ibsrq;
2124 	const struct ib_gid_attr *gid_attr = NULL;
2125 	struct ib_rwq_ind_table *rwq_ind_tbl;
2126 	enum ib_qp_type qp_type;
2127 	struct mlx4_ib_dev *dev;
2128 	struct mlx4_ib_qp *qp;
2129 	struct mlx4_ib_pd *pd;
2130 	struct mlx4_ib_cq *send_cq, *recv_cq;
2131 	struct mlx4_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2132 		udata, struct mlx4_ib_ucontext, ibucontext);
2133 	struct mlx4_qp_context *context;
2134 	enum mlx4_qp_optpar optpar = 0;
2135 	int sqd_event;
2136 	int steer_qp = 0;
2137 	int err = -EINVAL;
2138 	int counter_index;
2139 
2140 	if (src_type == MLX4_IB_RWQ_SRC) {
2141 		struct ib_wq *ibwq;
2142 
2143 		ibwq	    = (struct ib_wq *)src;
2144 		ibsrq	    = NULL;
2145 		rwq_ind_tbl = NULL;
2146 		qp_type     = IB_QPT_RAW_PACKET;
2147 		qp	    = to_mqp((struct ib_qp *)ibwq);
2148 		dev	    = to_mdev(ibwq->device);
2149 		pd	    = to_mpd(ibwq->pd);
2150 	} else {
2151 		struct ib_qp *ibqp;
2152 
2153 		ibqp	    = (struct ib_qp *)src;
2154 		ibsrq	    = ibqp->srq;
2155 		rwq_ind_tbl = ibqp->rwq_ind_tbl;
2156 		qp_type     = ibqp->qp_type;
2157 		qp	    = to_mqp(ibqp);
2158 		dev	    = to_mdev(ibqp->device);
2159 		pd	    = get_pd(qp);
2160 	}
2161 
2162 	/* APM is not supported under RoCE */
2163 	if (attr_mask & IB_QP_ALT_PATH &&
2164 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2165 	    IB_LINK_LAYER_ETHERNET)
2166 		return -ENOTSUPP;
2167 
2168 	context = kzalloc(sizeof *context, GFP_KERNEL);
2169 	if (!context)
2170 		return -ENOMEM;
2171 
2172 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
2173 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
2174 
2175 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
2176 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2177 	else {
2178 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
2179 		switch (attr->path_mig_state) {
2180 		case IB_MIG_MIGRATED:
2181 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2182 			break;
2183 		case IB_MIG_REARM:
2184 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
2185 			break;
2186 		case IB_MIG_ARMED:
2187 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
2188 			break;
2189 		}
2190 	}
2191 
2192 	if (qp->inl_recv_sz)
2193 		context->param3 |= cpu_to_be32(1 << 25);
2194 
2195 	if (qp->flags & MLX4_IB_QP_SCATTER_FCS)
2196 		context->param3 |= cpu_to_be32(1 << 29);
2197 
2198 	if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI)
2199 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
2200 	else if (qp_type == IB_QPT_RAW_PACKET)
2201 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
2202 	else if (qp_type == IB_QPT_UD) {
2203 		if (qp->flags & MLX4_IB_QP_LSO)
2204 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
2205 					      ilog2(dev->dev->caps.max_gso_sz);
2206 		else
2207 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 13;
2208 	} else if (attr_mask & IB_QP_PATH_MTU) {
2209 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
2210 			pr_err("path MTU (%u) is invalid\n",
2211 			       attr->path_mtu);
2212 			goto out;
2213 		}
2214 		context->mtu_msgmax = (attr->path_mtu << 5) |
2215 			ilog2(dev->dev->caps.max_msg_sz);
2216 	}
2217 
2218 	if (!rwq_ind_tbl) { /* PRM RSS receive side should be left zeros */
2219 		if (qp->rq.wqe_cnt)
2220 			context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
2221 		context->rq_size_stride |= qp->rq.wqe_shift - 4;
2222 	}
2223 
2224 	if (qp->sq.wqe_cnt)
2225 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
2226 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
2227 
2228 	if (new_state == IB_QPS_RESET && qp->counter_index)
2229 		mlx4_ib_free_qp_counter(dev, qp);
2230 
2231 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2232 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
2233 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
2234 		if (qp_type == IB_QPT_RAW_PACKET)
2235 			context->param3 |= cpu_to_be32(1 << 30);
2236 	}
2237 
2238 	if (ucontext)
2239 		context->usr_page = cpu_to_be32(
2240 			mlx4_to_hw_uar_index(dev->dev, ucontext->uar.index));
2241 	else
2242 		context->usr_page = cpu_to_be32(
2243 			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
2244 
2245 	if (attr_mask & IB_QP_DEST_QPN)
2246 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
2247 
2248 	if (attr_mask & IB_QP_PORT) {
2249 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
2250 		    !(attr_mask & IB_QP_AV)) {
2251 			mlx4_set_sched(&context->pri_path, attr->port_num);
2252 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
2253 		}
2254 	}
2255 
2256 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
2257 		err = create_qp_lb_counter(dev, qp);
2258 		if (err)
2259 			goto out;
2260 
2261 		counter_index =
2262 			dev->counters_table[qp->port - 1].default_counter;
2263 		if (qp->counter_index)
2264 			counter_index = qp->counter_index->index;
2265 
2266 		if (counter_index != -1) {
2267 			context->pri_path.counter_index = counter_index;
2268 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
2269 			if (qp->counter_index) {
2270 				context->pri_path.fl |=
2271 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
2272 				context->pri_path.vlan_control |=
2273 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
2274 			}
2275 		} else
2276 			context->pri_path.counter_index =
2277 				MLX4_SINK_COUNTER_INDEX(dev->dev);
2278 
2279 		if (qp->flags & MLX4_IB_QP_NETIF) {
2280 			mlx4_ib_steer_qp_reg(dev, qp, 1);
2281 			steer_qp = 1;
2282 		}
2283 
2284 		if (qp_type == IB_QPT_GSI) {
2285 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
2286 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
2287 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
2288 
2289 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2290 		}
2291 	}
2292 
2293 	if (attr_mask & IB_QP_PKEY_INDEX) {
2294 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2295 			context->pri_path.disable_pkey_check = 0x40;
2296 		context->pri_path.pkey_index = attr->pkey_index;
2297 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
2298 	}
2299 
2300 	if (attr_mask & IB_QP_AV) {
2301 		u8 port_num = mlx4_is_bonded(dev->dev) ? 1 :
2302 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2303 		u16 vlan = 0xffff;
2304 		u8 smac[ETH_ALEN];
2305 		int is_eth =
2306 			rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
2307 			rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
2308 
2309 		if (is_eth) {
2310 			gid_attr = attr->ah_attr.grh.sgid_attr;
2311 			err = rdma_read_gid_l2_fields(gid_attr, &vlan,
2312 						      &smac[0]);
2313 			if (err)
2314 				goto out;
2315 		}
2316 
2317 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
2318 				  port_num, vlan, smac))
2319 			goto out;
2320 
2321 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
2322 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
2323 
2324 		if (is_eth &&
2325 		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
2326 			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr->gid_type);
2327 
2328 			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
2329 				err = -EINVAL;
2330 				goto out;
2331 			}
2332 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2333 		}
2334 
2335 	}
2336 
2337 	if (attr_mask & IB_QP_TIMEOUT) {
2338 		context->pri_path.ackto |= attr->timeout << 3;
2339 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
2340 	}
2341 
2342 	if (attr_mask & IB_QP_ALT_PATH) {
2343 		if (attr->alt_port_num == 0 ||
2344 		    attr->alt_port_num > dev->dev->caps.num_ports)
2345 			goto out;
2346 
2347 		if (attr->alt_pkey_index >=
2348 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
2349 			goto out;
2350 
2351 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
2352 				      &context->alt_path,
2353 				      attr->alt_port_num))
2354 			goto out;
2355 
2356 		context->alt_path.pkey_index = attr->alt_pkey_index;
2357 		context->alt_path.ackto = attr->alt_timeout << 3;
2358 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
2359 	}
2360 
2361 	context->pd = cpu_to_be32(pd->pdn);
2362 
2363 	if (!rwq_ind_tbl) {
2364 		context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
2365 		get_cqs(qp, src_type, &send_cq, &recv_cq);
2366 	} else { /* Set dummy CQs to be compatible with HV and PRM */
2367 		send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq);
2368 		recv_cq = send_cq;
2369 	}
2370 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
2371 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
2372 
2373 	/* Set "fast registration enabled" for all kernel QPs */
2374 	if (!ucontext)
2375 		context->params1 |= cpu_to_be32(1 << 11);
2376 
2377 	if (attr_mask & IB_QP_RNR_RETRY) {
2378 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2379 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
2380 	}
2381 
2382 	if (attr_mask & IB_QP_RETRY_CNT) {
2383 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2384 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
2385 	}
2386 
2387 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2388 		if (attr->max_rd_atomic)
2389 			context->params1 |=
2390 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2391 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
2392 	}
2393 
2394 	if (attr_mask & IB_QP_SQ_PSN)
2395 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2396 
2397 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2398 		if (attr->max_dest_rd_atomic)
2399 			context->params2 |=
2400 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2401 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
2402 	}
2403 
2404 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
2405 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
2406 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
2407 	}
2408 
2409 	if (ibsrq)
2410 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
2411 
2412 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
2413 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2414 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
2415 	}
2416 	if (attr_mask & IB_QP_RQ_PSN)
2417 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2418 
2419 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
2420 	if (attr_mask & IB_QP_QKEY) {
2421 		if (qp->mlx4_ib_qp_type &
2422 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
2423 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2424 		else {
2425 			if (mlx4_is_mfunc(dev->dev) &&
2426 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
2427 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
2428 			    MLX4_RESERVED_QKEY_BASE) {
2429 				pr_err("Cannot use reserved QKEY"
2430 				       " 0x%x (range 0xffff0000..0xffffffff"
2431 				       " is reserved)\n", attr->qkey);
2432 				err = -EINVAL;
2433 				goto out;
2434 			}
2435 			context->qkey = cpu_to_be32(attr->qkey);
2436 		}
2437 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
2438 	}
2439 
2440 	if (ibsrq)
2441 		context->srqn = cpu_to_be32(1 << 24 |
2442 					    to_msrq(ibsrq)->msrq.srqn);
2443 
2444 	if (qp->rq.wqe_cnt &&
2445 	    cur_state == IB_QPS_RESET &&
2446 	    new_state == IB_QPS_INIT)
2447 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2448 
2449 	if (cur_state == IB_QPS_INIT &&
2450 	    new_state == IB_QPS_RTR  &&
2451 	    (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI ||
2452 	     qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) {
2453 		context->pri_path.sched_queue = (qp->port - 1) << 6;
2454 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2455 		    qp->mlx4_ib_qp_type &
2456 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
2457 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
2458 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
2459 				context->pri_path.fl = 0x80;
2460 		} else {
2461 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2462 				context->pri_path.fl = 0x80;
2463 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
2464 		}
2465 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2466 		    IB_LINK_LAYER_ETHERNET) {
2467 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
2468 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
2469 				context->pri_path.feup = 1 << 7; /* don't fsm */
2470 			/* handle smac_index */
2471 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
2472 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
2473 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
2474 				err = handle_eth_ud_smac_index(dev, qp, context);
2475 				if (err) {
2476 					err = -EINVAL;
2477 					goto out;
2478 				}
2479 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
2480 					dev->qp1_proxy[qp->port - 1] = qp;
2481 			}
2482 		}
2483 	}
2484 
2485 	if (qp_type == IB_QPT_RAW_PACKET) {
2486 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
2487 					MLX4_IB_LINK_TYPE_ETH;
2488 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
2489 			/* set QP to receive both tunneled & non-tunneled packets */
2490 			if (!rwq_ind_tbl)
2491 				context->srqn = cpu_to_be32(7 << 28);
2492 		}
2493 	}
2494 
2495 	if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
2496 		int is_eth = rdma_port_get_link_layer(
2497 				&dev->ib_dev, qp->port) ==
2498 				IB_LINK_LAYER_ETHERNET;
2499 		if (is_eth) {
2500 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
2501 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
2502 		}
2503 	}
2504 
2505 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2506 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2507 		sqd_event = 1;
2508 	else
2509 		sqd_event = 0;
2510 
2511 	if (!ucontext &&
2512 	    cur_state == IB_QPS_RESET &&
2513 	    new_state == IB_QPS_INIT)
2514 		context->rlkey_roce_mode |= (1 << 4);
2515 
2516 	/*
2517 	 * Before passing a kernel QP to the HW, make sure that the
2518 	 * ownership bits of the send queue are set and the SQ
2519 	 * headroom is stamped so that the hardware doesn't start
2520 	 * processing stale work requests.
2521 	 */
2522 	if (!ucontext &&
2523 	    cur_state == IB_QPS_RESET &&
2524 	    new_state == IB_QPS_INIT) {
2525 		struct mlx4_wqe_ctrl_seg *ctrl;
2526 		int i;
2527 
2528 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
2529 			ctrl = get_send_wqe(qp, i);
2530 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
2531 			ctrl->qpn_vlan.fence_size =
2532 				1 << (qp->sq.wqe_shift - 4);
2533 			stamp_send_wqe(qp, i);
2534 		}
2535 	}
2536 
2537 	if (rwq_ind_tbl	&&
2538 	    cur_state == IB_QPS_RESET &&
2539 	    new_state == IB_QPS_INIT) {
2540 		fill_qp_rss_context(context, qp);
2541 		context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET);
2542 	}
2543 
2544 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2545 			     to_mlx4_state(new_state), context, optpar,
2546 			     sqd_event, &qp->mqp);
2547 	if (err)
2548 		goto out;
2549 
2550 	qp->state = new_state;
2551 
2552 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2553 		qp->atomic_rd_en = attr->qp_access_flags;
2554 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2555 		qp->resp_depth = attr->max_dest_rd_atomic;
2556 	if (attr_mask & IB_QP_PORT) {
2557 		qp->port = attr->port_num;
2558 		update_mcg_macs(dev, qp);
2559 	}
2560 	if (attr_mask & IB_QP_ALT_PATH)
2561 		qp->alt_port = attr->alt_port_num;
2562 
2563 	if (is_sqp(dev, qp))
2564 		store_sqp_attrs(qp->sqp, attr, attr_mask);
2565 
2566 	/*
2567 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
2568 	 * QP0 to RESET or ERROR, bring the link back down.
2569 	 */
2570 	if (is_qp0(dev, qp)) {
2571 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2572 			if (mlx4_INIT_PORT(dev->dev, qp->port))
2573 				pr_warn("INIT_PORT failed for port %d\n",
2574 				       qp->port);
2575 
2576 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2577 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2578 			mlx4_CLOSE_PORT(dev->dev, qp->port);
2579 	}
2580 
2581 	/*
2582 	 * If we moved a kernel QP to RESET, clean up all old CQ
2583 	 * entries and reinitialize the QP.
2584 	 */
2585 	if (new_state == IB_QPS_RESET) {
2586 		if (!ucontext) {
2587 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2588 					 ibsrq ? to_msrq(ibsrq) : NULL);
2589 			if (send_cq != recv_cq)
2590 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2591 
2592 			qp->rq.head = 0;
2593 			qp->rq.tail = 0;
2594 			qp->sq.head = 0;
2595 			qp->sq.tail = 0;
2596 			qp->sq_next_wqe = 0;
2597 			if (qp->rq.wqe_cnt)
2598 				*qp->db.db  = 0;
2599 
2600 			if (qp->flags & MLX4_IB_QP_NETIF)
2601 				mlx4_ib_steer_qp_reg(dev, qp, 0);
2602 		}
2603 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2604 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2605 			qp->pri.smac = 0;
2606 			qp->pri.smac_port = 0;
2607 		}
2608 		if (qp->alt.smac) {
2609 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2610 			qp->alt.smac = 0;
2611 		}
2612 		if (qp->pri.vid < 0x1000) {
2613 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2614 			qp->pri.vid = 0xFFFF;
2615 			qp->pri.candidate_vid = 0xFFFF;
2616 			qp->pri.update_vid = 0;
2617 		}
2618 
2619 		if (qp->alt.vid < 0x1000) {
2620 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2621 			qp->alt.vid = 0xFFFF;
2622 			qp->alt.candidate_vid = 0xFFFF;
2623 			qp->alt.update_vid = 0;
2624 		}
2625 	}
2626 out:
2627 	if (err && qp->counter_index)
2628 		mlx4_ib_free_qp_counter(dev, qp);
2629 	if (err && steer_qp)
2630 		mlx4_ib_steer_qp_reg(dev, qp, 0);
2631 	kfree(context);
2632 	if (qp->pri.candidate_smac ||
2633 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2634 		if (err) {
2635 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2636 		} else {
2637 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2638 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2639 			qp->pri.smac = qp->pri.candidate_smac;
2640 			qp->pri.smac_index = qp->pri.candidate_smac_index;
2641 			qp->pri.smac_port = qp->pri.candidate_smac_port;
2642 		}
2643 		qp->pri.candidate_smac = 0;
2644 		qp->pri.candidate_smac_index = 0;
2645 		qp->pri.candidate_smac_port = 0;
2646 	}
2647 	if (qp->alt.candidate_smac) {
2648 		if (err) {
2649 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2650 		} else {
2651 			if (qp->alt.smac)
2652 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2653 			qp->alt.smac = qp->alt.candidate_smac;
2654 			qp->alt.smac_index = qp->alt.candidate_smac_index;
2655 			qp->alt.smac_port = qp->alt.candidate_smac_port;
2656 		}
2657 		qp->alt.candidate_smac = 0;
2658 		qp->alt.candidate_smac_index = 0;
2659 		qp->alt.candidate_smac_port = 0;
2660 	}
2661 
2662 	if (qp->pri.update_vid) {
2663 		if (err) {
2664 			if (qp->pri.candidate_vid < 0x1000)
2665 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2666 						     qp->pri.candidate_vid);
2667 		} else {
2668 			if (qp->pri.vid < 0x1000)
2669 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2670 						     qp->pri.vid);
2671 			qp->pri.vid = qp->pri.candidate_vid;
2672 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2673 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
2674 		}
2675 		qp->pri.candidate_vid = 0xFFFF;
2676 		qp->pri.update_vid = 0;
2677 	}
2678 
2679 	if (qp->alt.update_vid) {
2680 		if (err) {
2681 			if (qp->alt.candidate_vid < 0x1000)
2682 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2683 						     qp->alt.candidate_vid);
2684 		} else {
2685 			if (qp->alt.vid < 0x1000)
2686 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2687 						     qp->alt.vid);
2688 			qp->alt.vid = qp->alt.candidate_vid;
2689 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2690 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2691 		}
2692 		qp->alt.candidate_vid = 0xFFFF;
2693 		qp->alt.update_vid = 0;
2694 	}
2695 
2696 	return err;
2697 }
2698 
2699 enum {
2700 	MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK = (IB_QP_STATE	|
2701 					      IB_QP_PORT),
2702 };
2703 
2704 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2705 			      int attr_mask, struct ib_udata *udata)
2706 {
2707 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2708 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2709 	enum ib_qp_state cur_state, new_state;
2710 	int err = -EINVAL;
2711 	mutex_lock(&qp->mutex);
2712 
2713 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2714 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2715 
2716 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2717 				attr_mask)) {
2718 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2719 			 "for transition %d to %d. qp_type %d,"
2720 			 " attr_mask 0x%x\n",
2721 			 ibqp->qp_num, cur_state, new_state,
2722 			 ibqp->qp_type, attr_mask);
2723 		goto out;
2724 	}
2725 
2726 	if (ibqp->rwq_ind_tbl) {
2727 		if (!(((cur_state == IB_QPS_RESET) &&
2728 		       (new_state == IB_QPS_INIT)) ||
2729 		      ((cur_state == IB_QPS_INIT)  &&
2730 		       (new_state == IB_QPS_RTR)))) {
2731 			pr_debug("qpn 0x%x: RSS QP unsupported transition %d to %d\n",
2732 				 ibqp->qp_num, cur_state, new_state);
2733 
2734 			err = -EOPNOTSUPP;
2735 			goto out;
2736 		}
2737 
2738 		if (attr_mask & ~MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK) {
2739 			pr_debug("qpn 0x%x: RSS QP unsupported attribute mask 0x%x for transition %d to %d\n",
2740 				 ibqp->qp_num, attr_mask, cur_state, new_state);
2741 
2742 			err = -EOPNOTSUPP;
2743 			goto out;
2744 		}
2745 	}
2746 
2747 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2748 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2749 			if ((ibqp->qp_type == IB_QPT_RC) ||
2750 			    (ibqp->qp_type == IB_QPT_UD) ||
2751 			    (ibqp->qp_type == IB_QPT_UC) ||
2752 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2753 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2754 				attr->port_num = mlx4_ib_bond_next_port(dev);
2755 			}
2756 		} else {
2757 			/* no sense in changing port_num
2758 			 * when ports are bonded */
2759 			attr_mask &= ~IB_QP_PORT;
2760 		}
2761 	}
2762 
2763 	if ((attr_mask & IB_QP_PORT) &&
2764 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2765 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2766 			 "for transition %d to %d. qp_type %d\n",
2767 			 ibqp->qp_num, attr->port_num, cur_state,
2768 			 new_state, ibqp->qp_type);
2769 		goto out;
2770 	}
2771 
2772 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2773 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2774 	     IB_LINK_LAYER_ETHERNET))
2775 		goto out;
2776 
2777 	if (attr_mask & IB_QP_PKEY_INDEX) {
2778 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2779 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2780 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2781 				 "for transition %d to %d. qp_type %d\n",
2782 				 ibqp->qp_num, attr->pkey_index, cur_state,
2783 				 new_state, ibqp->qp_type);
2784 			goto out;
2785 		}
2786 	}
2787 
2788 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2789 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2790 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2791 			 "Transition %d to %d. qp_type %d\n",
2792 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2793 			 new_state, ibqp->qp_type);
2794 		goto out;
2795 	}
2796 
2797 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2798 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2799 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2800 			 "Transition %d to %d. qp_type %d\n",
2801 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2802 			 new_state, ibqp->qp_type);
2803 		goto out;
2804 	}
2805 
2806 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2807 		err = 0;
2808 		goto out;
2809 	}
2810 
2811 	if (ibqp->rwq_ind_tbl && (new_state == IB_QPS_INIT)) {
2812 		err = bringup_rss_rwqs(ibqp->rwq_ind_tbl, attr->port_num,
2813 				       udata);
2814 		if (err)
2815 			goto out;
2816 	}
2817 
2818 	err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask,
2819 				  cur_state, new_state, udata);
2820 
2821 	if (ibqp->rwq_ind_tbl && err)
2822 		bring_down_rss_rwqs(ibqp->rwq_ind_tbl, udata);
2823 
2824 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2825 		attr->port_num = 1;
2826 
2827 out:
2828 	mutex_unlock(&qp->mutex);
2829 	return err;
2830 }
2831 
2832 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2833 		      int attr_mask, struct ib_udata *udata)
2834 {
2835 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2836 	int ret;
2837 
2838 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
2839 		return -EOPNOTSUPP;
2840 
2841 	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2842 
2843 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2844 		struct mlx4_ib_sqp *sqp = mqp->sqp;
2845 		int err = 0;
2846 
2847 		if (sqp->roce_v2_gsi)
2848 			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2849 		if (err)
2850 			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2851 			       err);
2852 	}
2853 	return ret;
2854 }
2855 
2856 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2857 {
2858 	int i;
2859 	for (i = 0; i < dev->caps.num_ports; i++) {
2860 		if (qpn == dev->caps.spec_qps[i].qp0_proxy ||
2861 		    qpn == dev->caps.spec_qps[i].qp0_tunnel) {
2862 			*qkey = dev->caps.spec_qps[i].qp0_qkey;
2863 			return 0;
2864 		}
2865 	}
2866 	return -EINVAL;
2867 }
2868 
2869 static int build_sriov_qp0_header(struct mlx4_ib_qp *qp,
2870 				  const struct ib_ud_wr *wr,
2871 				  void *wqe, unsigned *mlx_seg_len)
2872 {
2873 	struct mlx4_ib_dev *mdev = to_mdev(qp->ibqp.device);
2874 	struct mlx4_ib_sqp *sqp = qp->sqp;
2875 	struct ib_device *ib_dev = qp->ibqp.device;
2876 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2877 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2878 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2879 	u16 pkey;
2880 	u32 qkey;
2881 	int send_size;
2882 	int header_size;
2883 	int spc;
2884 	int err;
2885 	int i;
2886 
2887 	if (wr->wr.opcode != IB_WR_SEND)
2888 		return -EINVAL;
2889 
2890 	send_size = 0;
2891 
2892 	for (i = 0; i < wr->wr.num_sge; ++i)
2893 		send_size += wr->wr.sg_list[i].length;
2894 
2895 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2896 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2897 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2898 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2899 
2900 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2901 
2902 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2903 		sqp->ud_header.lrh.service_level =
2904 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2905 		sqp->ud_header.lrh.destination_lid =
2906 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2907 		sqp->ud_header.lrh.source_lid =
2908 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2909 	}
2910 
2911 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2912 
2913 	/* force loopback */
2914 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2915 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2916 
2917 	sqp->ud_header.lrh.virtual_lane    = 0;
2918 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2919 	err = ib_get_cached_pkey(ib_dev, qp->port, 0, &pkey);
2920 	if (err)
2921 		return err;
2922 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2923 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2924 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2925 	else
2926 		sqp->ud_header.bth.destination_qpn =
2927 			cpu_to_be32(mdev->dev->caps.spec_qps[qp->port - 1].qp0_tunnel);
2928 
2929 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2930 	if (mlx4_is_master(mdev->dev)) {
2931 		if (mlx4_get_parav_qkey(mdev->dev, qp->mqp.qpn, &qkey))
2932 			return -EINVAL;
2933 	} else {
2934 		if (vf_get_qp0_qkey(mdev->dev, qp->mqp.qpn, &qkey))
2935 			return -EINVAL;
2936 	}
2937 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2938 	sqp->ud_header.deth.source_qpn = cpu_to_be32(qp->mqp.qpn);
2939 
2940 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2941 	sqp->ud_header.immediate_present = 0;
2942 
2943 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2944 
2945 	/*
2946 	 * Inline data segments may not cross a 64 byte boundary.  If
2947 	 * our UD header is bigger than the space available up to the
2948 	 * next 64 byte boundary in the WQE, use two inline data
2949 	 * segments to hold the UD header.
2950 	 */
2951 	spc = MLX4_INLINE_ALIGN -
2952 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2953 	if (header_size <= spc) {
2954 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2955 		memcpy(inl + 1, sqp->header_buf, header_size);
2956 		i = 1;
2957 	} else {
2958 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2959 		memcpy(inl + 1, sqp->header_buf, spc);
2960 
2961 		inl = (void *) (inl + 1) + spc;
2962 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2963 		/*
2964 		 * Need a barrier here to make sure all the data is
2965 		 * visible before the byte_count field is set.
2966 		 * Otherwise the HCA prefetcher could grab the 64-byte
2967 		 * chunk with this inline segment and get a valid (!=
2968 		 * 0xffffffff) byte count but stale data, and end up
2969 		 * generating a packet with bad headers.
2970 		 *
2971 		 * The first inline segment's byte_count field doesn't
2972 		 * need a barrier, because it comes after a
2973 		 * control/MLX segment and therefore is at an offset
2974 		 * of 16 mod 64.
2975 		 */
2976 		wmb();
2977 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2978 		i = 2;
2979 	}
2980 
2981 	*mlx_seg_len =
2982 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2983 	return 0;
2984 }
2985 
2986 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
2987 {
2988 	union sl2vl_tbl_to_u64 tmp_vltab;
2989 	u8 vl;
2990 
2991 	if (sl > 15)
2992 		return 0xf;
2993 	tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
2994 	vl = tmp_vltab.sl8[sl >> 1];
2995 	if (sl & 1)
2996 		vl &= 0x0f;
2997 	else
2998 		vl >>= 4;
2999 	return vl;
3000 }
3001 
3002 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
3003 				int index, union ib_gid *gid,
3004 				enum ib_gid_type *gid_type)
3005 {
3006 	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
3007 	struct mlx4_port_gid_table *port_gid_table;
3008 	unsigned long flags;
3009 
3010 	port_gid_table = &iboe->gids[port_num - 1];
3011 	spin_lock_irqsave(&iboe->lock, flags);
3012 	memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
3013 	*gid_type = port_gid_table->gids[index].gid_type;
3014 	spin_unlock_irqrestore(&iboe->lock, flags);
3015 	if (rdma_is_zero_gid(gid))
3016 		return -ENOENT;
3017 
3018 	return 0;
3019 }
3020 
3021 #define MLX4_ROCEV2_QP1_SPORT 0xC000
3022 static int build_mlx_header(struct mlx4_ib_qp *qp, const struct ib_ud_wr *wr,
3023 			    void *wqe, unsigned *mlx_seg_len)
3024 {
3025 	struct mlx4_ib_sqp *sqp = qp->sqp;
3026 	struct ib_device *ib_dev = qp->ibqp.device;
3027 	struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
3028 	struct mlx4_wqe_mlx_seg *mlx = wqe;
3029 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
3030 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
3031 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3032 	union ib_gid sgid;
3033 	u16 pkey;
3034 	int send_size;
3035 	int header_size;
3036 	int spc;
3037 	int i;
3038 	int err = 0;
3039 	u16 vlan = 0xffff;
3040 	bool is_eth;
3041 	bool is_vlan = false;
3042 	bool is_grh;
3043 	bool is_udp = false;
3044 	int ip_version = 0;
3045 
3046 	send_size = 0;
3047 	for (i = 0; i < wr->wr.num_sge; ++i)
3048 		send_size += wr->wr.sg_list[i].length;
3049 
3050 	is_eth = rdma_port_get_link_layer(qp->ibqp.device, qp->port) == IB_LINK_LAYER_ETHERNET;
3051 	is_grh = mlx4_ib_ah_grh_present(ah);
3052 	if (is_eth) {
3053 		enum ib_gid_type gid_type;
3054 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3055 			/* When multi-function is enabled, the ib_core gid
3056 			 * indexes don't necessarily match the hw ones, so
3057 			 * we must use our own cache */
3058 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
3059 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
3060 							   ah->av.ib.gid_index, &sgid.raw[0]);
3061 			if (err)
3062 				return err;
3063 		} else  {
3064 			err = fill_gid_by_hw_index(ibdev, qp->port,
3065 						   ah->av.ib.gid_index, &sgid,
3066 						   &gid_type);
3067 			if (!err) {
3068 				is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
3069 				if (is_udp) {
3070 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
3071 						ip_version = 4;
3072 					else
3073 						ip_version = 6;
3074 					is_grh = false;
3075 				}
3076 			} else {
3077 				return err;
3078 			}
3079 		}
3080 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
3081 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
3082 			is_vlan = true;
3083 		}
3084 	}
3085 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
3086 			  ip_version, is_udp, 0, &sqp->ud_header);
3087 	if (err)
3088 		return err;
3089 
3090 	if (!is_eth) {
3091 		sqp->ud_header.lrh.service_level =
3092 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
3093 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
3094 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
3095 	}
3096 
3097 	if (is_grh || (ip_version == 6)) {
3098 		sqp->ud_header.grh.traffic_class =
3099 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3100 		sqp->ud_header.grh.flow_label    =
3101 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
3102 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
3103 		if (is_eth) {
3104 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
3105 		} else {
3106 			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3107 				/* When multi-function is enabled, the ib_core gid
3108 				 * indexes don't necessarily match the hw ones, so
3109 				 * we must use our own cache
3110 				 */
3111 				sqp->ud_header.grh.source_gid.global
3112 					.subnet_prefix =
3113 					cpu_to_be64(atomic64_read(
3114 						&(to_mdev(ib_dev)
3115 							  ->sriov
3116 							  .demux[qp->port - 1]
3117 							  .subnet_prefix)));
3118 				sqp->ud_header.grh.source_gid.global
3119 					.interface_id =
3120 					to_mdev(ib_dev)
3121 						->sriov.demux[qp->port - 1]
3122 						.guid_cache[ah->av.ib.gid_index];
3123 			} else {
3124 				sqp->ud_header.grh.source_gid =
3125 					ah->ibah.sgid_attr->gid;
3126 			}
3127 		}
3128 		memcpy(sqp->ud_header.grh.destination_gid.raw,
3129 		       ah->av.ib.dgid, 16);
3130 	}
3131 
3132 	if (ip_version == 4) {
3133 		sqp->ud_header.ip4.tos =
3134 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3135 		sqp->ud_header.ip4.id = 0;
3136 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
3137 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
3138 
3139 		memcpy(&sqp->ud_header.ip4.saddr,
3140 		       sgid.raw + 12, 4);
3141 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
3142 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
3143 	}
3144 
3145 	if (is_udp) {
3146 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
3147 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
3148 		sqp->ud_header.udp.csum = 0;
3149 	}
3150 
3151 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
3152 
3153 	if (!is_eth) {
3154 		mlx->flags |=
3155 			cpu_to_be32((!qp->ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
3156 				    (sqp->ud_header.lrh.destination_lid ==
3157 						     IB_LID_PERMISSIVE ?
3158 					     MLX4_WQE_MLX_SLR :
3159 					     0) |
3160 				    (sqp->ud_header.lrh.service_level << 8));
3161 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
3162 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
3163 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
3164 	}
3165 
3166 	switch (wr->wr.opcode) {
3167 	case IB_WR_SEND:
3168 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
3169 		sqp->ud_header.immediate_present = 0;
3170 		break;
3171 	case IB_WR_SEND_WITH_IMM:
3172 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
3173 		sqp->ud_header.immediate_present = 1;
3174 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
3175 		break;
3176 	default:
3177 		return -EINVAL;
3178 	}
3179 
3180 	if (is_eth) {
3181 		u16 ether_type;
3182 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
3183 
3184 		ether_type = (!is_udp) ? ETH_P_IBOE:
3185 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
3186 
3187 		mlx->sched_prio = cpu_to_be16(pcp);
3188 
3189 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
3190 		ether_addr_copy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac);
3191 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
3192 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
3193 
3194 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
3195 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
3196 		if (!is_vlan) {
3197 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
3198 		} else {
3199 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
3200 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
3201 		}
3202 	} else {
3203 		sqp->ud_header.lrh.virtual_lane =
3204 			!qp->ibqp.qp_num ?
3205 				15 :
3206 				sl_to_vl(to_mdev(ib_dev),
3207 					 sqp->ud_header.lrh.service_level,
3208 					 qp->port);
3209 		if (qp->ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
3210 			return -EINVAL;
3211 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
3212 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
3213 	}
3214 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
3215 	if (!qp->ibqp.qp_num)
3216 		err = ib_get_cached_pkey(ib_dev, qp->port, sqp->pkey_index,
3217 					 &pkey);
3218 	else
3219 		err = ib_get_cached_pkey(ib_dev, qp->port, wr->pkey_index,
3220 					 &pkey);
3221 	if (err)
3222 		return err;
3223 
3224 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
3225 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
3226 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
3227 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
3228 					       sqp->qkey : wr->remote_qkey);
3229 	sqp->ud_header.deth.source_qpn = cpu_to_be32(qp->ibqp.qp_num);
3230 
3231 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3232 
3233 	if (0) {
3234 		pr_err("built UD header of size %d:\n", header_size);
3235 		for (i = 0; i < header_size / 4; ++i) {
3236 			if (i % 8 == 0)
3237 				pr_err("  [%02x] ", i * 4);
3238 			pr_cont(" %08x",
3239 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
3240 			if ((i + 1) % 8 == 0)
3241 				pr_cont("\n");
3242 		}
3243 		pr_err("\n");
3244 	}
3245 
3246 	/*
3247 	 * Inline data segments may not cross a 64 byte boundary.  If
3248 	 * our UD header is bigger than the space available up to the
3249 	 * next 64 byte boundary in the WQE, use two inline data
3250 	 * segments to hold the UD header.
3251 	 */
3252 	spc = MLX4_INLINE_ALIGN -
3253 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3254 	if (header_size <= spc) {
3255 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3256 		memcpy(inl + 1, sqp->header_buf, header_size);
3257 		i = 1;
3258 	} else {
3259 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3260 		memcpy(inl + 1, sqp->header_buf, spc);
3261 
3262 		inl = (void *) (inl + 1) + spc;
3263 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3264 		/*
3265 		 * Need a barrier here to make sure all the data is
3266 		 * visible before the byte_count field is set.
3267 		 * Otherwise the HCA prefetcher could grab the 64-byte
3268 		 * chunk with this inline segment and get a valid (!=
3269 		 * 0xffffffff) byte count but stale data, and end up
3270 		 * generating a packet with bad headers.
3271 		 *
3272 		 * The first inline segment's byte_count field doesn't
3273 		 * need a barrier, because it comes after a
3274 		 * control/MLX segment and therefore is at an offset
3275 		 * of 16 mod 64.
3276 		 */
3277 		wmb();
3278 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3279 		i = 2;
3280 	}
3281 
3282 	*mlx_seg_len =
3283 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3284 	return 0;
3285 }
3286 
3287 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3288 {
3289 	unsigned cur;
3290 	struct mlx4_ib_cq *cq;
3291 
3292 	cur = wq->head - wq->tail;
3293 	if (likely(cur + nreq < wq->max_post))
3294 		return 0;
3295 
3296 	cq = to_mcq(ib_cq);
3297 	spin_lock(&cq->lock);
3298 	cur = wq->head - wq->tail;
3299 	spin_unlock(&cq->lock);
3300 
3301 	return cur + nreq >= wq->max_post;
3302 }
3303 
3304 static __be32 convert_access(int acc)
3305 {
3306 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
3307 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
3308 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
3309 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
3310 	       (acc & IB_ACCESS_REMOTE_READ   ?
3311 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
3312 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
3313 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
3314 }
3315 
3316 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
3317 			const struct ib_reg_wr *wr)
3318 {
3319 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
3320 
3321 	fseg->flags		= convert_access(wr->access);
3322 	fseg->mem_key		= cpu_to_be32(wr->key);
3323 	fseg->buf_list		= cpu_to_be64(mr->page_map);
3324 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
3325 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
3326 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
3327 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
3328 	fseg->reserved[0]	= 0;
3329 	fseg->reserved[1]	= 0;
3330 }
3331 
3332 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
3333 {
3334 	memset(iseg, 0, sizeof(*iseg));
3335 	iseg->mem_key = cpu_to_be32(rkey);
3336 }
3337 
3338 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
3339 					  u64 remote_addr, u32 rkey)
3340 {
3341 	rseg->raddr    = cpu_to_be64(remote_addr);
3342 	rseg->rkey     = cpu_to_be32(rkey);
3343 	rseg->reserved = 0;
3344 }
3345 
3346 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
3347 			   const struct ib_atomic_wr *wr)
3348 {
3349 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
3350 		aseg->swap_add = cpu_to_be64(wr->swap);
3351 		aseg->compare  = cpu_to_be64(wr->compare_add);
3352 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
3353 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3354 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
3355 	} else {
3356 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3357 		aseg->compare  = 0;
3358 	}
3359 
3360 }
3361 
3362 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
3363 				  const struct ib_atomic_wr *wr)
3364 {
3365 	aseg->swap_add		= cpu_to_be64(wr->swap);
3366 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
3367 	aseg->compare		= cpu_to_be64(wr->compare_add);
3368 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
3369 }
3370 
3371 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
3372 			     const struct ib_ud_wr *wr)
3373 {
3374 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
3375 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
3376 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
3377 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
3378 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
3379 }
3380 
3381 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
3382 				    struct mlx4_wqe_datagram_seg *dseg,
3383 				    const struct ib_ud_wr *wr,
3384 				    enum mlx4_ib_qp_type qpt)
3385 {
3386 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
3387 	struct mlx4_av sqp_av = {0};
3388 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
3389 
3390 	/* force loopback */
3391 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
3392 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
3393 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
3394 			cpu_to_be32(0xf0000000);
3395 
3396 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
3397 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
3398 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp1_tunnel);
3399 	else
3400 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp0_tunnel);
3401 	/* Use QKEY from the QP context, which is set by master */
3402 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
3403 }
3404 
3405 static void build_tunnel_header(const struct ib_ud_wr *wr, void *wqe,
3406 				unsigned *mlx_seg_len)
3407 {
3408 	struct mlx4_wqe_inline_seg *inl = wqe;
3409 	struct mlx4_ib_tunnel_header hdr;
3410 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3411 	int spc;
3412 	int i;
3413 
3414 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
3415 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
3416 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
3417 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
3418 	memcpy(hdr.mac, ah->av.eth.mac, 6);
3419 	hdr.vlan = ah->av.eth.vlan;
3420 
3421 	spc = MLX4_INLINE_ALIGN -
3422 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3423 	if (sizeof (hdr) <= spc) {
3424 		memcpy(inl + 1, &hdr, sizeof (hdr));
3425 		wmb();
3426 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
3427 		i = 1;
3428 	} else {
3429 		memcpy(inl + 1, &hdr, spc);
3430 		wmb();
3431 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3432 
3433 		inl = (void *) (inl + 1) + spc;
3434 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
3435 		wmb();
3436 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
3437 		i = 2;
3438 	}
3439 
3440 	*mlx_seg_len =
3441 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
3442 }
3443 
3444 static void set_mlx_icrc_seg(void *dseg)
3445 {
3446 	u32 *t = dseg;
3447 	struct mlx4_wqe_inline_seg *iseg = dseg;
3448 
3449 	t[1] = 0;
3450 
3451 	/*
3452 	 * Need a barrier here before writing the byte_count field to
3453 	 * make sure that all the data is visible before the
3454 	 * byte_count field is set.  Otherwise, if the segment begins
3455 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3456 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3457 	 * stale data, and end up sending the wrong data.
3458 	 */
3459 	wmb();
3460 
3461 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
3462 }
3463 
3464 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3465 {
3466 	dseg->lkey       = cpu_to_be32(sg->lkey);
3467 	dseg->addr       = cpu_to_be64(sg->addr);
3468 
3469 	/*
3470 	 * Need a barrier here before writing the byte_count field to
3471 	 * make sure that all the data is visible before the
3472 	 * byte_count field is set.  Otherwise, if the segment begins
3473 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3474 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3475 	 * stale data, and end up sending the wrong data.
3476 	 */
3477 	wmb();
3478 
3479 	dseg->byte_count = cpu_to_be32(sg->length);
3480 }
3481 
3482 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3483 {
3484 	dseg->byte_count = cpu_to_be32(sg->length);
3485 	dseg->lkey       = cpu_to_be32(sg->lkey);
3486 	dseg->addr       = cpu_to_be64(sg->addr);
3487 }
3488 
3489 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe,
3490 			 const struct ib_ud_wr *wr, struct mlx4_ib_qp *qp,
3491 			 unsigned *lso_seg_len, __be32 *lso_hdr_sz, __be32 *blh)
3492 {
3493 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
3494 
3495 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
3496 		*blh = cpu_to_be32(1 << 6);
3497 
3498 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
3499 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
3500 		return -EINVAL;
3501 
3502 	memcpy(wqe->header, wr->header, wr->hlen);
3503 
3504 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
3505 	*lso_seg_len = halign;
3506 	return 0;
3507 }
3508 
3509 static __be32 send_ieth(const struct ib_send_wr *wr)
3510 {
3511 	switch (wr->opcode) {
3512 	case IB_WR_SEND_WITH_IMM:
3513 	case IB_WR_RDMA_WRITE_WITH_IMM:
3514 		return wr->ex.imm_data;
3515 
3516 	case IB_WR_SEND_WITH_INV:
3517 		return cpu_to_be32(wr->ex.invalidate_rkey);
3518 
3519 	default:
3520 		return 0;
3521 	}
3522 }
3523 
3524 static void add_zero_len_inline(void *wqe)
3525 {
3526 	struct mlx4_wqe_inline_seg *inl = wqe;
3527 	memset(wqe, 0, 16);
3528 	inl->byte_count = cpu_to_be32(1 << 31);
3529 }
3530 
3531 static int _mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3532 			      const struct ib_send_wr **bad_wr, bool drain)
3533 {
3534 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3535 	void *wqe;
3536 	struct mlx4_wqe_ctrl_seg *ctrl;
3537 	struct mlx4_wqe_data_seg *dseg;
3538 	unsigned long flags;
3539 	int nreq;
3540 	int err = 0;
3541 	unsigned ind;
3542 	int size;
3543 	unsigned seglen;
3544 	__be32 dummy;
3545 	__be32 *lso_wqe;
3546 	__be32 lso_hdr_sz;
3547 	__be32 blh;
3548 	int i;
3549 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3550 
3551 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
3552 		struct mlx4_ib_sqp *sqp = qp->sqp;
3553 
3554 		if (sqp->roce_v2_gsi) {
3555 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
3556 			enum ib_gid_type gid_type;
3557 			union ib_gid gid;
3558 
3559 			if (!fill_gid_by_hw_index(mdev, qp->port,
3560 					   ah->av.ib.gid_index,
3561 					   &gid, &gid_type))
3562 				qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
3563 						to_mqp(sqp->roce_v2_gsi) : qp;
3564 			else
3565 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
3566 				       ah->av.ib.gid_index);
3567 		}
3568 	}
3569 
3570 	spin_lock_irqsave(&qp->sq.lock, flags);
3571 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3572 	    !drain) {
3573 		err = -EIO;
3574 		*bad_wr = wr;
3575 		nreq = 0;
3576 		goto out;
3577 	}
3578 
3579 	ind = qp->sq_next_wqe;
3580 
3581 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3582 		lso_wqe = &dummy;
3583 		blh = 0;
3584 
3585 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
3586 			err = -ENOMEM;
3587 			*bad_wr = wr;
3588 			goto out;
3589 		}
3590 
3591 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3592 			err = -EINVAL;
3593 			*bad_wr = wr;
3594 			goto out;
3595 		}
3596 
3597 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3598 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3599 
3600 		ctrl->srcrb_flags =
3601 			(wr->send_flags & IB_SEND_SIGNALED ?
3602 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3603 			(wr->send_flags & IB_SEND_SOLICITED ?
3604 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3605 			((wr->send_flags & IB_SEND_IP_CSUM) ?
3606 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3607 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3608 			qp->sq_signal_bits;
3609 
3610 		ctrl->imm = send_ieth(wr);
3611 
3612 		wqe += sizeof *ctrl;
3613 		size = sizeof *ctrl / 16;
3614 
3615 		switch (qp->mlx4_ib_qp_type) {
3616 		case MLX4_IB_QPT_RC:
3617 		case MLX4_IB_QPT_UC:
3618 			switch (wr->opcode) {
3619 			case IB_WR_ATOMIC_CMP_AND_SWP:
3620 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3621 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3622 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3623 					      atomic_wr(wr)->rkey);
3624 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3625 
3626 				set_atomic_seg(wqe, atomic_wr(wr));
3627 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
3628 
3629 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3630 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3631 
3632 				break;
3633 
3634 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3635 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3636 					      atomic_wr(wr)->rkey);
3637 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3638 
3639 				set_masked_atomic_seg(wqe, atomic_wr(wr));
3640 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
3641 
3642 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3643 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3644 
3645 				break;
3646 
3647 			case IB_WR_RDMA_READ:
3648 			case IB_WR_RDMA_WRITE:
3649 			case IB_WR_RDMA_WRITE_WITH_IMM:
3650 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3651 					      rdma_wr(wr)->rkey);
3652 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3653 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3654 				break;
3655 
3656 			case IB_WR_LOCAL_INV:
3657 				ctrl->srcrb_flags |=
3658 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3659 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3660 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3661 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3662 				break;
3663 
3664 			case IB_WR_REG_MR:
3665 				ctrl->srcrb_flags |=
3666 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3667 				set_reg_seg(wqe, reg_wr(wr));
3668 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3669 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3670 				break;
3671 
3672 			default:
3673 				/* No extra segments required for sends */
3674 				break;
3675 			}
3676 			break;
3677 
3678 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3679 			err = build_sriov_qp0_header(qp, ud_wr(wr), ctrl,
3680 						     &seglen);
3681 			if (unlikely(err)) {
3682 				*bad_wr = wr;
3683 				goto out;
3684 			}
3685 			wqe  += seglen;
3686 			size += seglen / 16;
3687 			break;
3688 		case MLX4_IB_QPT_TUN_SMI:
3689 		case MLX4_IB_QPT_TUN_GSI:
3690 			/* this is a UD qp used in MAD responses to slaves. */
3691 			set_datagram_seg(wqe, ud_wr(wr));
3692 			/* set the forced-loopback bit in the data seg av */
3693 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3694 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3695 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3696 			break;
3697 		case MLX4_IB_QPT_UD:
3698 			set_datagram_seg(wqe, ud_wr(wr));
3699 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3700 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3701 
3702 			if (wr->opcode == IB_WR_LSO) {
3703 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3704 						&lso_hdr_sz, &blh);
3705 				if (unlikely(err)) {
3706 					*bad_wr = wr;
3707 					goto out;
3708 				}
3709 				lso_wqe = (__be32 *) wqe;
3710 				wqe  += seglen;
3711 				size += seglen / 16;
3712 			}
3713 			break;
3714 
3715 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3716 			err = build_sriov_qp0_header(qp, ud_wr(wr), ctrl,
3717 						     &seglen);
3718 			if (unlikely(err)) {
3719 				*bad_wr = wr;
3720 				goto out;
3721 			}
3722 			wqe  += seglen;
3723 			size += seglen / 16;
3724 			/* to start tunnel header on a cache-line boundary */
3725 			add_zero_len_inline(wqe);
3726 			wqe += 16;
3727 			size++;
3728 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3729 			wqe  += seglen;
3730 			size += seglen / 16;
3731 			break;
3732 		case MLX4_IB_QPT_PROXY_SMI:
3733 		case MLX4_IB_QPT_PROXY_GSI:
3734 			/* If we are tunneling special qps, this is a UD qp.
3735 			 * In this case we first add a UD segment targeting
3736 			 * the tunnel qp, and then add a header with address
3737 			 * information */
3738 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3739 						ud_wr(wr),
3740 						qp->mlx4_ib_qp_type);
3741 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3742 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3743 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3744 			wqe  += seglen;
3745 			size += seglen / 16;
3746 			break;
3747 
3748 		case MLX4_IB_QPT_SMI:
3749 		case MLX4_IB_QPT_GSI:
3750 			err = build_mlx_header(qp, ud_wr(wr), ctrl, &seglen);
3751 			if (unlikely(err)) {
3752 				*bad_wr = wr;
3753 				goto out;
3754 			}
3755 			wqe  += seglen;
3756 			size += seglen / 16;
3757 			break;
3758 
3759 		default:
3760 			break;
3761 		}
3762 
3763 		/*
3764 		 * Write data segments in reverse order, so as to
3765 		 * overwrite cacheline stamp last within each
3766 		 * cacheline.  This avoids issues with WQE
3767 		 * prefetching.
3768 		 */
3769 
3770 		dseg = wqe;
3771 		dseg += wr->num_sge - 1;
3772 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3773 
3774 		/* Add one more inline data segment for ICRC for MLX sends */
3775 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3776 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3777 			     qp->mlx4_ib_qp_type &
3778 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3779 			set_mlx_icrc_seg(dseg + 1);
3780 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3781 		}
3782 
3783 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3784 			set_data_seg(dseg, wr->sg_list + i);
3785 
3786 		/*
3787 		 * Possibly overwrite stamping in cacheline with LSO
3788 		 * segment only after making sure all data segments
3789 		 * are written.
3790 		 */
3791 		wmb();
3792 		*lso_wqe = lso_hdr_sz;
3793 
3794 		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3795 					     MLX4_WQE_CTRL_FENCE : 0) | size;
3796 
3797 		/*
3798 		 * Make sure descriptor is fully written before
3799 		 * setting ownership bit (because HW can start
3800 		 * executing as soon as we do).
3801 		 */
3802 		wmb();
3803 
3804 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3805 			*bad_wr = wr;
3806 			err = -EINVAL;
3807 			goto out;
3808 		}
3809 
3810 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3811 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3812 
3813 		/*
3814 		 * We can improve latency by not stamping the last
3815 		 * send queue WQE until after ringing the doorbell, so
3816 		 * only stamp here if there are still more WQEs to post.
3817 		 */
3818 		if (wr->next)
3819 			stamp_send_wqe(qp, ind + qp->sq_spare_wqes);
3820 		ind++;
3821 	}
3822 
3823 out:
3824 	if (likely(nreq)) {
3825 		qp->sq.head += nreq;
3826 
3827 		/*
3828 		 * Make sure that descriptors are written before
3829 		 * doorbell record.
3830 		 */
3831 		wmb();
3832 
3833 		writel_relaxed(qp->doorbell_qpn,
3834 			to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3835 
3836 		stamp_send_wqe(qp, ind + qp->sq_spare_wqes - 1);
3837 
3838 		qp->sq_next_wqe = ind;
3839 	}
3840 
3841 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3842 
3843 	return err;
3844 }
3845 
3846 int mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3847 		      const struct ib_send_wr **bad_wr)
3848 {
3849 	return _mlx4_ib_post_send(ibqp, wr, bad_wr, false);
3850 }
3851 
3852 static int _mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3853 			      const struct ib_recv_wr **bad_wr, bool drain)
3854 {
3855 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3856 	struct mlx4_wqe_data_seg *scat;
3857 	unsigned long flags;
3858 	int err = 0;
3859 	int nreq;
3860 	int ind;
3861 	int max_gs;
3862 	int i;
3863 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3864 
3865 	max_gs = qp->rq.max_gs;
3866 	spin_lock_irqsave(&qp->rq.lock, flags);
3867 
3868 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3869 	    !drain) {
3870 		err = -EIO;
3871 		*bad_wr = wr;
3872 		nreq = 0;
3873 		goto out;
3874 	}
3875 
3876 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3877 
3878 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3879 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3880 			err = -ENOMEM;
3881 			*bad_wr = wr;
3882 			goto out;
3883 		}
3884 
3885 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3886 			err = -EINVAL;
3887 			*bad_wr = wr;
3888 			goto out;
3889 		}
3890 
3891 		scat = get_recv_wqe(qp, ind);
3892 
3893 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3894 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3895 			ib_dma_sync_single_for_device(ibqp->device,
3896 						      qp->sqp_proxy_rcv[ind].map,
3897 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3898 						      DMA_FROM_DEVICE);
3899 			scat->byte_count =
3900 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3901 			/* use dma lkey from upper layer entry */
3902 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3903 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3904 			scat++;
3905 			max_gs--;
3906 		}
3907 
3908 		for (i = 0; i < wr->num_sge; ++i)
3909 			__set_data_seg(scat + i, wr->sg_list + i);
3910 
3911 		if (i < max_gs) {
3912 			scat[i].byte_count = 0;
3913 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3914 			scat[i].addr       = 0;
3915 		}
3916 
3917 		qp->rq.wrid[ind] = wr->wr_id;
3918 
3919 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3920 	}
3921 
3922 out:
3923 	if (likely(nreq)) {
3924 		qp->rq.head += nreq;
3925 
3926 		/*
3927 		 * Make sure that descriptors are written before
3928 		 * doorbell record.
3929 		 */
3930 		wmb();
3931 
3932 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3933 	}
3934 
3935 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3936 
3937 	return err;
3938 }
3939 
3940 int mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3941 		      const struct ib_recv_wr **bad_wr)
3942 {
3943 	return _mlx4_ib_post_recv(ibqp, wr, bad_wr, false);
3944 }
3945 
3946 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3947 {
3948 	switch (mlx4_state) {
3949 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3950 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3951 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3952 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3953 	case MLX4_QP_STATE_SQ_DRAINING:
3954 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3955 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3956 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3957 	default:		     return -1;
3958 	}
3959 }
3960 
3961 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3962 {
3963 	switch (mlx4_mig_state) {
3964 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3965 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3966 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3967 	default: return -1;
3968 	}
3969 }
3970 
3971 static int to_ib_qp_access_flags(int mlx4_flags)
3972 {
3973 	int ib_flags = 0;
3974 
3975 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3976 		ib_flags |= IB_ACCESS_REMOTE_READ;
3977 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3978 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3979 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3980 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3981 
3982 	return ib_flags;
3983 }
3984 
3985 static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
3986 			    struct rdma_ah_attr *ah_attr,
3987 			    struct mlx4_qp_path *path)
3988 {
3989 	struct mlx4_dev *dev = ibdev->dev;
3990 	u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
3991 
3992 	memset(ah_attr, 0, sizeof(*ah_attr));
3993 	if (port_num == 0 || port_num > dev->caps.num_ports)
3994 		return;
3995 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
3996 
3997 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
3998 		rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
3999 			       ((path->sched_queue & 4) << 1));
4000 	else
4001 		rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
4002 	rdma_ah_set_port_num(ah_attr, port_num);
4003 
4004 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
4005 	rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
4006 	rdma_ah_set_static_rate(ah_attr,
4007 				path->static_rate ? path->static_rate - 5 : 0);
4008 	if (path->grh_mylmc & (1 << 7)) {
4009 		rdma_ah_set_grh(ah_attr, NULL,
4010 				be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
4011 				path->mgid_index,
4012 				path->hop_limit,
4013 				(be32_to_cpu(path->tclass_flowlabel)
4014 				 >> 20) & 0xff);
4015 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
4016 	}
4017 }
4018 
4019 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
4020 		     struct ib_qp_init_attr *qp_init_attr)
4021 {
4022 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
4023 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
4024 	struct mlx4_qp_context context;
4025 	int mlx4_state;
4026 	int err = 0;
4027 
4028 	if (ibqp->rwq_ind_tbl)
4029 		return -EOPNOTSUPP;
4030 
4031 	mutex_lock(&qp->mutex);
4032 
4033 	if (qp->state == IB_QPS_RESET) {
4034 		qp_attr->qp_state = IB_QPS_RESET;
4035 		goto done;
4036 	}
4037 
4038 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
4039 	if (err) {
4040 		err = -EINVAL;
4041 		goto out;
4042 	}
4043 
4044 	mlx4_state = be32_to_cpu(context.flags) >> 28;
4045 
4046 	qp->state		     = to_ib_qp_state(mlx4_state);
4047 	qp_attr->qp_state	     = qp->state;
4048 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
4049 	qp_attr->path_mig_state	     =
4050 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
4051 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
4052 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
4053 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
4054 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
4055 	qp_attr->qp_access_flags     =
4056 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
4057 
4058 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC ||
4059 	    qp->ibqp.qp_type == IB_QPT_XRC_INI ||
4060 	    qp->ibqp.qp_type == IB_QPT_XRC_TGT) {
4061 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
4062 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
4063 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
4064 		qp_attr->alt_port_num	=
4065 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
4066 	}
4067 
4068 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
4069 	if (qp_attr->qp_state == IB_QPS_INIT)
4070 		qp_attr->port_num = qp->port;
4071 	else
4072 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
4073 
4074 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4075 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
4076 
4077 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
4078 
4079 	qp_attr->max_dest_rd_atomic =
4080 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
4081 	qp_attr->min_rnr_timer	    =
4082 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
4083 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
4084 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
4085 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
4086 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
4087 
4088 done:
4089 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4090 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4091 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4092 
4093 	if (!ibqp->uobject) {
4094 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
4095 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4096 	} else {
4097 		qp_attr->cap.max_send_wr  = 0;
4098 		qp_attr->cap.max_send_sge = 0;
4099 	}
4100 
4101 	/*
4102 	 * We don't support inline sends for kernel QPs (yet), and we
4103 	 * don't know what userspace's value should be.
4104 	 */
4105 	qp_attr->cap.max_inline_data = 0;
4106 
4107 	qp_init_attr->cap	     = qp_attr->cap;
4108 
4109 	qp_init_attr->create_flags = 0;
4110 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4111 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4112 
4113 	if (qp->flags & MLX4_IB_QP_LSO)
4114 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
4115 
4116 	if (qp->flags & MLX4_IB_QP_NETIF)
4117 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
4118 
4119 	qp_init_attr->sq_sig_type =
4120 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
4121 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4122 
4123 out:
4124 	mutex_unlock(&qp->mutex);
4125 	return err;
4126 }
4127 
4128 struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
4129 				struct ib_wq_init_attr *init_attr,
4130 				struct ib_udata *udata)
4131 {
4132 	struct mlx4_dev *dev = to_mdev(pd->device)->dev;
4133 	struct ib_qp_init_attr ib_qp_init_attr = {};
4134 	struct mlx4_ib_qp *qp;
4135 	struct mlx4_ib_create_wq ucmd;
4136 	int err, required_cmd_sz;
4137 
4138 	if (!udata)
4139 		return ERR_PTR(-EINVAL);
4140 
4141 	required_cmd_sz = offsetof(typeof(ucmd), comp_mask) +
4142 			  sizeof(ucmd.comp_mask);
4143 	if (udata->inlen < required_cmd_sz) {
4144 		pr_debug("invalid inlen\n");
4145 		return ERR_PTR(-EINVAL);
4146 	}
4147 
4148 	if (udata->inlen > sizeof(ucmd) &&
4149 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4150 				 udata->inlen - sizeof(ucmd))) {
4151 		pr_debug("inlen is not supported\n");
4152 		return ERR_PTR(-EOPNOTSUPP);
4153 	}
4154 
4155 	if (udata->outlen)
4156 		return ERR_PTR(-EOPNOTSUPP);
4157 
4158 	if (init_attr->wq_type != IB_WQT_RQ) {
4159 		pr_debug("unsupported wq type %d\n", init_attr->wq_type);
4160 		return ERR_PTR(-EOPNOTSUPP);
4161 	}
4162 
4163 	if (init_attr->create_flags & ~IB_WQ_FLAGS_SCATTER_FCS ||
4164 	    !(dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
4165 		pr_debug("unsupported create_flags %u\n",
4166 			 init_attr->create_flags);
4167 		return ERR_PTR(-EOPNOTSUPP);
4168 	}
4169 
4170 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
4171 	if (!qp)
4172 		return ERR_PTR(-ENOMEM);
4173 
4174 	mutex_init(&qp->mutex);
4175 	qp->pri.vid = 0xFFFF;
4176 	qp->alt.vid = 0xFFFF;
4177 
4178 	ib_qp_init_attr.qp_context = init_attr->wq_context;
4179 	ib_qp_init_attr.qp_type = IB_QPT_RAW_PACKET;
4180 	ib_qp_init_attr.cap.max_recv_wr = init_attr->max_wr;
4181 	ib_qp_init_attr.cap.max_recv_sge = init_attr->max_sge;
4182 	ib_qp_init_attr.recv_cq = init_attr->cq;
4183 	ib_qp_init_attr.send_cq = ib_qp_init_attr.recv_cq; /* Dummy CQ */
4184 
4185 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS)
4186 		ib_qp_init_attr.create_flags |= IB_QP_CREATE_SCATTER_FCS;
4187 
4188 	err = create_rq(pd, &ib_qp_init_attr, udata, qp);
4189 	if (err) {
4190 		kfree(qp);
4191 		return ERR_PTR(err);
4192 	}
4193 
4194 	qp->ibwq.event_handler = init_attr->event_handler;
4195 	qp->ibwq.wq_num = qp->mqp.qpn;
4196 	qp->ibwq.state = IB_WQS_RESET;
4197 
4198 	return &qp->ibwq;
4199 }
4200 
4201 static int ib_wq2qp_state(enum ib_wq_state state)
4202 {
4203 	switch (state) {
4204 	case IB_WQS_RESET:
4205 		return IB_QPS_RESET;
4206 	case IB_WQS_RDY:
4207 		return IB_QPS_RTR;
4208 	default:
4209 		return IB_QPS_ERR;
4210 	}
4211 }
4212 
4213 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state,
4214 			      struct ib_udata *udata)
4215 {
4216 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4217 	enum ib_qp_state qp_cur_state;
4218 	enum ib_qp_state qp_new_state;
4219 	int attr_mask;
4220 	int err;
4221 
4222 	/* ib_qp.state represents the WQ HW state while ib_wq.state represents
4223 	 * the WQ logic state.
4224 	 */
4225 	qp_cur_state = qp->state;
4226 	qp_new_state = ib_wq2qp_state(new_state);
4227 
4228 	if (ib_wq2qp_state(new_state) == qp_cur_state)
4229 		return 0;
4230 
4231 	if (new_state == IB_WQS_RDY) {
4232 		struct ib_qp_attr attr = {};
4233 
4234 		attr.port_num = qp->port;
4235 		attr_mask = IB_QP_PORT;
4236 
4237 		err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, &attr,
4238 					  attr_mask, IB_QPS_RESET, IB_QPS_INIT,
4239 					  udata);
4240 		if (err) {
4241 			pr_debug("WQN=0x%06x failed to apply RST->INIT on the HW QP\n",
4242 				 ibwq->wq_num);
4243 			return err;
4244 		}
4245 
4246 		qp_cur_state = IB_QPS_INIT;
4247 	}
4248 
4249 	attr_mask = 0;
4250 	err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, attr_mask,
4251 				  qp_cur_state,  qp_new_state, udata);
4252 
4253 	if (err && (qp_cur_state == IB_QPS_INIT)) {
4254 		qp_new_state = IB_QPS_RESET;
4255 		if (__mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL,
4256 					attr_mask, IB_QPS_INIT, IB_QPS_RESET,
4257 					udata)) {
4258 			pr_warn("WQN=0x%06x failed with reverting HW's resources failure\n",
4259 				ibwq->wq_num);
4260 			qp_new_state = IB_QPS_INIT;
4261 		}
4262 	}
4263 
4264 	qp->state = qp_new_state;
4265 
4266 	return err;
4267 }
4268 
4269 int mlx4_ib_modify_wq(struct ib_wq *ibwq, struct ib_wq_attr *wq_attr,
4270 		      u32 wq_attr_mask, struct ib_udata *udata)
4271 {
4272 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4273 	struct mlx4_ib_modify_wq ucmd = {};
4274 	size_t required_cmd_sz;
4275 	enum ib_wq_state cur_state, new_state;
4276 	int err = 0;
4277 
4278 	required_cmd_sz = offsetof(typeof(ucmd), reserved) +
4279 				   sizeof(ucmd.reserved);
4280 	if (udata->inlen < required_cmd_sz)
4281 		return -EINVAL;
4282 
4283 	if (udata->inlen > sizeof(ucmd) &&
4284 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4285 				 udata->inlen - sizeof(ucmd)))
4286 		return -EOPNOTSUPP;
4287 
4288 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4289 		return -EFAULT;
4290 
4291 	if (ucmd.comp_mask || ucmd.reserved)
4292 		return -EOPNOTSUPP;
4293 
4294 	if (wq_attr_mask & IB_WQ_FLAGS)
4295 		return -EOPNOTSUPP;
4296 
4297 	cur_state = wq_attr->curr_wq_state;
4298 	new_state = wq_attr->wq_state;
4299 
4300 	if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR))
4301 		return -EINVAL;
4302 
4303 	if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET))
4304 		return -EINVAL;
4305 
4306 	/* Need to protect against the parent RSS which also may modify WQ
4307 	 * state.
4308 	 */
4309 	mutex_lock(&qp->mutex);
4310 
4311 	/* Can update HW state only if a RSS QP has already associated to this
4312 	 * WQ, so we can apply its port on the WQ.
4313 	 */
4314 	if (qp->rss_usecnt)
4315 		err = _mlx4_ib_modify_wq(ibwq, new_state, udata);
4316 
4317 	if (!err)
4318 		ibwq->state = new_state;
4319 
4320 	mutex_unlock(&qp->mutex);
4321 
4322 	return err;
4323 }
4324 
4325 int mlx4_ib_destroy_wq(struct ib_wq *ibwq, struct ib_udata *udata)
4326 {
4327 	struct mlx4_ib_dev *dev = to_mdev(ibwq->device);
4328 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4329 
4330 	if (qp->counter_index)
4331 		mlx4_ib_free_qp_counter(dev, qp);
4332 
4333 	destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, udata);
4334 
4335 	kfree(qp);
4336 	return 0;
4337 }
4338 
4339 int mlx4_ib_create_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table,
4340 				 struct ib_rwq_ind_table_init_attr *init_attr,
4341 				 struct ib_udata *udata)
4342 {
4343 	struct mlx4_ib_create_rwq_ind_tbl_resp resp = {};
4344 	unsigned int ind_tbl_size = 1 << init_attr->log_ind_tbl_size;
4345 	struct ib_device *device = rwq_ind_table->device;
4346 	unsigned int base_wqn;
4347 	size_t min_resp_len;
4348 	int i, err = 0;
4349 
4350 	if (udata->inlen > 0 &&
4351 	    !ib_is_udata_cleared(udata, 0,
4352 				 udata->inlen))
4353 		return -EOPNOTSUPP;
4354 
4355 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4356 	if (udata->outlen && udata->outlen < min_resp_len)
4357 		return -EINVAL;
4358 
4359 	if (ind_tbl_size >
4360 	    device->attrs.rss_caps.max_rwq_indirection_table_size) {
4361 		pr_debug("log_ind_tbl_size = %d is bigger than supported = %d\n",
4362 			 ind_tbl_size,
4363 			 device->attrs.rss_caps.max_rwq_indirection_table_size);
4364 		return -EINVAL;
4365 	}
4366 
4367 	base_wqn = init_attr->ind_tbl[0]->wq_num;
4368 
4369 	if (base_wqn % ind_tbl_size) {
4370 		pr_debug("WQN=0x%x isn't aligned with indirection table size\n",
4371 			 base_wqn);
4372 		return -EINVAL;
4373 	}
4374 
4375 	for (i = 1; i < ind_tbl_size; i++) {
4376 		if (++base_wqn != init_attr->ind_tbl[i]->wq_num) {
4377 			pr_debug("indirection table's WQNs aren't consecutive\n");
4378 			return -EINVAL;
4379 		}
4380 	}
4381 
4382 	if (udata->outlen) {
4383 		resp.response_length = offsetof(typeof(resp), response_length) +
4384 					sizeof(resp.response_length);
4385 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4386 	}
4387 
4388 	return err;
4389 }
4390 
4391 struct mlx4_ib_drain_cqe {
4392 	struct ib_cqe cqe;
4393 	struct completion done;
4394 };
4395 
4396 static void mlx4_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
4397 {
4398 	struct mlx4_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
4399 						     struct mlx4_ib_drain_cqe,
4400 						     cqe);
4401 
4402 	complete(&cqe->done);
4403 }
4404 
4405 /* This function returns only once the drained WR was completed */
4406 static void handle_drain_completion(struct ib_cq *cq,
4407 				    struct mlx4_ib_drain_cqe *sdrain,
4408 				    struct mlx4_ib_dev *dev)
4409 {
4410 	struct mlx4_dev *mdev = dev->dev;
4411 
4412 	if (cq->poll_ctx == IB_POLL_DIRECT) {
4413 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
4414 			ib_process_cq_direct(cq, -1);
4415 		return;
4416 	}
4417 
4418 	if (mdev->persist->state == MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4419 		struct mlx4_ib_cq *mcq = to_mcq(cq);
4420 		bool triggered = false;
4421 		unsigned long flags;
4422 
4423 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
4424 		/* Make sure that the CQ handler won't run if wasn't run yet */
4425 		if (!mcq->mcq.reset_notify_added)
4426 			mcq->mcq.reset_notify_added = 1;
4427 		else
4428 			triggered = true;
4429 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
4430 
4431 		if (triggered) {
4432 			/* Wait for any scheduled/running task to be ended */
4433 			switch (cq->poll_ctx) {
4434 			case IB_POLL_SOFTIRQ:
4435 				irq_poll_disable(&cq->iop);
4436 				irq_poll_enable(&cq->iop);
4437 				break;
4438 			case IB_POLL_WORKQUEUE:
4439 				cancel_work_sync(&cq->work);
4440 				break;
4441 			default:
4442 				WARN_ON_ONCE(1);
4443 			}
4444 		}
4445 
4446 		/* Run the CQ handler - this makes sure that the drain WR will
4447 		 * be processed if wasn't processed yet.
4448 		 */
4449 		mcq->mcq.comp(&mcq->mcq);
4450 	}
4451 
4452 	wait_for_completion(&sdrain->done);
4453 }
4454 
4455 void mlx4_ib_drain_sq(struct ib_qp *qp)
4456 {
4457 	struct ib_cq *cq = qp->send_cq;
4458 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4459 	struct mlx4_ib_drain_cqe sdrain;
4460 	const struct ib_send_wr *bad_swr;
4461 	struct ib_rdma_wr swr = {
4462 		.wr = {
4463 			.next = NULL,
4464 			{ .wr_cqe	= &sdrain.cqe, },
4465 			.opcode	= IB_WR_RDMA_WRITE,
4466 		},
4467 	};
4468 	int ret;
4469 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
4470 	struct mlx4_dev *mdev = dev->dev;
4471 
4472 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4473 	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4474 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4475 		return;
4476 	}
4477 
4478 	sdrain.cqe.done = mlx4_ib_drain_qp_done;
4479 	init_completion(&sdrain.done);
4480 
4481 	ret = _mlx4_ib_post_send(qp, &swr.wr, &bad_swr, true);
4482 	if (ret) {
4483 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4484 		return;
4485 	}
4486 
4487 	handle_drain_completion(cq, &sdrain, dev);
4488 }
4489 
4490 void mlx4_ib_drain_rq(struct ib_qp *qp)
4491 {
4492 	struct ib_cq *cq = qp->recv_cq;
4493 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4494 	struct mlx4_ib_drain_cqe rdrain;
4495 	struct ib_recv_wr rwr = {};
4496 	const struct ib_recv_wr *bad_rwr;
4497 	int ret;
4498 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
4499 	struct mlx4_dev *mdev = dev->dev;
4500 
4501 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4502 	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4503 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4504 		return;
4505 	}
4506 
4507 	rwr.wr_cqe = &rdrain.cqe;
4508 	rdrain.cqe.done = mlx4_ib_drain_qp_done;
4509 	init_completion(&rdrain.done);
4510 
4511 	ret = _mlx4_ib_post_recv(qp, &rwr, &bad_rwr, true);
4512 	if (ret) {
4513 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4514 		return;
4515 	}
4516 
4517 	handle_drain_completion(cq, &rdrain, dev);
4518 }
4519 
4520 int mlx4_ib_qp_event_init(void)
4521 {
4522 	mlx4_ib_qp_event_wq = alloc_ordered_workqueue("mlx4_ib_qp_event_wq", 0);
4523 	if (!mlx4_ib_qp_event_wq)
4524 		return -ENOMEM;
4525 
4526 	return 0;
4527 }
4528 
4529 void mlx4_ib_qp_event_cleanup(void)
4530 {
4531 	destroy_workqueue(mlx4_ib_qp_event_wq);
4532 }
4533