xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision 52cdded0)
1 /*
2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/log2.h>
35 #include <linux/etherdevice.h>
36 #include <net/ip.h>
37 #include <linux/slab.h>
38 #include <linux/netdevice.h>
39 
40 #include <rdma/ib_cache.h>
41 #include <rdma/ib_pack.h>
42 #include <rdma/ib_addr.h>
43 #include <rdma/ib_mad.h>
44 #include <rdma/uverbs_ioctl.h>
45 
46 #include <linux/mlx4/driver.h>
47 #include <linux/mlx4/qp.h>
48 
49 #include "mlx4_ib.h"
50 #include <rdma/mlx4-abi.h>
51 
52 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
53 			     struct mlx4_ib_cq *recv_cq);
54 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
55 			       struct mlx4_ib_cq *recv_cq);
56 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state,
57 			      struct ib_udata *udata);
58 
59 enum {
60 	MLX4_IB_ACK_REQ_FREQ	= 8,
61 };
62 
63 enum {
64 	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
65 	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
66 	MLX4_IB_LINK_TYPE_IB		= 0,
67 	MLX4_IB_LINK_TYPE_ETH		= 1
68 };
69 
70 enum {
71 	/*
72 	 * Largest possible UD header: send with GRH and immediate
73 	 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
74 	 * tag.  (LRH would only use 8 bytes, so Ethernet is the
75 	 * biggest case)
76 	 */
77 	MLX4_IB_UD_HEADER_SIZE		= 82,
78 	MLX4_IB_LSO_HEADER_SPARE	= 128,
79 };
80 
81 struct mlx4_ib_sqp {
82 	struct mlx4_ib_qp	qp;
83 	int			pkey_index;
84 	u32			qkey;
85 	u32			send_psn;
86 	struct ib_ud_header	ud_header;
87 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
88 	struct ib_qp		*roce_v2_gsi;
89 };
90 
91 enum {
92 	MLX4_IB_MIN_SQ_STRIDE	= 6,
93 	MLX4_IB_CACHE_LINE_SIZE	= 64,
94 };
95 
96 enum {
97 	MLX4_RAW_QP_MTU		= 7,
98 	MLX4_RAW_QP_MSGMAX	= 31,
99 };
100 
101 #ifndef ETH_ALEN
102 #define ETH_ALEN        6
103 #endif
104 
105 static const __be32 mlx4_ib_opcode[] = {
106 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
107 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
108 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
109 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
110 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
111 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
112 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
113 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
114 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
115 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
116 	[IB_WR_REG_MR]				= cpu_to_be32(MLX4_OPCODE_FMR),
117 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
118 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
119 };
120 
121 enum mlx4_ib_source_type {
122 	MLX4_IB_QP_SRC	= 0,
123 	MLX4_IB_RWQ_SRC	= 1,
124 };
125 
126 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
127 {
128 	return container_of(mqp, struct mlx4_ib_sqp, qp);
129 }
130 
131 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
132 {
133 	if (!mlx4_is_master(dev->dev))
134 		return 0;
135 
136 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
137 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
138 		8 * MLX4_MFUNC_MAX;
139 }
140 
141 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
142 {
143 	int proxy_sqp = 0;
144 	int real_sqp = 0;
145 	int i;
146 	/* PPF or Native -- real SQP */
147 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
148 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
149 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
150 	if (real_sqp)
151 		return 1;
152 	/* VF or PF -- proxy SQP */
153 	if (mlx4_is_mfunc(dev->dev)) {
154 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
155 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy ||
156 			    qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp1_proxy) {
157 				proxy_sqp = 1;
158 				break;
159 			}
160 		}
161 	}
162 	if (proxy_sqp)
163 		return 1;
164 
165 	return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
166 }
167 
168 /* used for INIT/CLOSE port logic */
169 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
170 {
171 	int proxy_qp0 = 0;
172 	int real_qp0 = 0;
173 	int i;
174 	/* PPF or Native -- real QP0 */
175 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
176 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
177 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
178 	if (real_qp0)
179 		return 1;
180 	/* VF or PF -- proxy QP0 */
181 	if (mlx4_is_mfunc(dev->dev)) {
182 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
183 			if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy) {
184 				proxy_qp0 = 1;
185 				break;
186 			}
187 		}
188 	}
189 	return proxy_qp0;
190 }
191 
192 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
193 {
194 	return mlx4_buf_offset(&qp->buf, offset);
195 }
196 
197 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
198 {
199 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
200 }
201 
202 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
203 {
204 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
205 }
206 
207 /*
208  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
209  * first four bytes of every 64 byte chunk with 0xffffffff, except for
210  * the very first chunk of the WQE.
211  */
212 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n)
213 {
214 	__be32 *wqe;
215 	int i;
216 	int s;
217 	void *buf;
218 	struct mlx4_wqe_ctrl_seg *ctrl;
219 
220 	buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
221 	ctrl = (struct mlx4_wqe_ctrl_seg *)buf;
222 	s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4;
223 	for (i = 64; i < s; i += 64) {
224 		wqe = buf + i;
225 		*wqe = cpu_to_be32(0xffffffff);
226 	}
227 }
228 
229 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
230 {
231 	struct ib_event event;
232 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
233 
234 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
235 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
236 
237 	if (ibqp->event_handler) {
238 		event.device     = ibqp->device;
239 		event.element.qp = ibqp;
240 		switch (type) {
241 		case MLX4_EVENT_TYPE_PATH_MIG:
242 			event.event = IB_EVENT_PATH_MIG;
243 			break;
244 		case MLX4_EVENT_TYPE_COMM_EST:
245 			event.event = IB_EVENT_COMM_EST;
246 			break;
247 		case MLX4_EVENT_TYPE_SQ_DRAINED:
248 			event.event = IB_EVENT_SQ_DRAINED;
249 			break;
250 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
251 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
252 			break;
253 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
254 			event.event = IB_EVENT_QP_FATAL;
255 			break;
256 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
257 			event.event = IB_EVENT_PATH_MIG_ERR;
258 			break;
259 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
260 			event.event = IB_EVENT_QP_REQ_ERR;
261 			break;
262 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
263 			event.event = IB_EVENT_QP_ACCESS_ERR;
264 			break;
265 		default:
266 			pr_warn("Unexpected event type %d "
267 			       "on QP %06x\n", type, qp->qpn);
268 			return;
269 		}
270 
271 		ibqp->event_handler(&event, ibqp->qp_context);
272 	}
273 }
274 
275 static void mlx4_ib_wq_event(struct mlx4_qp *qp, enum mlx4_event type)
276 {
277 	pr_warn_ratelimited("Unexpected event type %d on WQ 0x%06x. Events are not supported for WQs\n",
278 			    type, qp->qpn);
279 }
280 
281 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
282 {
283 	/*
284 	 * UD WQEs must have a datagram segment.
285 	 * RC and UC WQEs might have a remote address segment.
286 	 * MLX WQEs need two extra inline data segments (for the UD
287 	 * header and space for the ICRC).
288 	 */
289 	switch (type) {
290 	case MLX4_IB_QPT_UD:
291 		return sizeof (struct mlx4_wqe_ctrl_seg) +
292 			sizeof (struct mlx4_wqe_datagram_seg) +
293 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
294 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
295 	case MLX4_IB_QPT_PROXY_SMI:
296 	case MLX4_IB_QPT_PROXY_GSI:
297 		return sizeof (struct mlx4_wqe_ctrl_seg) +
298 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
299 	case MLX4_IB_QPT_TUN_SMI_OWNER:
300 	case MLX4_IB_QPT_TUN_GSI:
301 		return sizeof (struct mlx4_wqe_ctrl_seg) +
302 			sizeof (struct mlx4_wqe_datagram_seg);
303 
304 	case MLX4_IB_QPT_UC:
305 		return sizeof (struct mlx4_wqe_ctrl_seg) +
306 			sizeof (struct mlx4_wqe_raddr_seg);
307 	case MLX4_IB_QPT_RC:
308 		return sizeof (struct mlx4_wqe_ctrl_seg) +
309 			sizeof (struct mlx4_wqe_masked_atomic_seg) +
310 			sizeof (struct mlx4_wqe_raddr_seg);
311 	case MLX4_IB_QPT_SMI:
312 	case MLX4_IB_QPT_GSI:
313 		return sizeof (struct mlx4_wqe_ctrl_seg) +
314 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
315 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
316 					   MLX4_INLINE_ALIGN) *
317 			      sizeof (struct mlx4_wqe_inline_seg),
318 			      sizeof (struct mlx4_wqe_data_seg)) +
319 			ALIGN(4 +
320 			      sizeof (struct mlx4_wqe_inline_seg),
321 			      sizeof (struct mlx4_wqe_data_seg));
322 	default:
323 		return sizeof (struct mlx4_wqe_ctrl_seg);
324 	}
325 }
326 
327 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
328 		       bool is_user, bool has_rq, struct mlx4_ib_qp *qp,
329 		       u32 inl_recv_sz)
330 {
331 	/* Sanity check RQ size before proceeding */
332 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
333 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
334 		return -EINVAL;
335 
336 	if (!has_rq) {
337 		if (cap->max_recv_wr || inl_recv_sz)
338 			return -EINVAL;
339 
340 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
341 	} else {
342 		u32 max_inl_recv_sz = dev->dev->caps.max_rq_sg *
343 			sizeof(struct mlx4_wqe_data_seg);
344 		u32 wqe_size;
345 
346 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
347 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge ||
348 				inl_recv_sz > max_inl_recv_sz))
349 			return -EINVAL;
350 
351 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
352 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
353 		wqe_size = qp->rq.max_gs * sizeof(struct mlx4_wqe_data_seg);
354 		qp->rq.wqe_shift = ilog2(max_t(u32, wqe_size, inl_recv_sz));
355 	}
356 
357 	/* leave userspace return values as they were, so as not to break ABI */
358 	if (is_user) {
359 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
360 		cap->max_recv_sge = qp->rq.max_gs;
361 	} else {
362 		cap->max_recv_wr  = qp->rq.max_post =
363 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
364 		cap->max_recv_sge = min(qp->rq.max_gs,
365 					min(dev->dev->caps.max_sq_sg,
366 					    dev->dev->caps.max_rq_sg));
367 	}
368 
369 	return 0;
370 }
371 
372 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
373 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
374 {
375 	int s;
376 
377 	/* Sanity check SQ size before proceeding */
378 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
379 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
380 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
381 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
382 		return -EINVAL;
383 
384 	/*
385 	 * For MLX transport we need 2 extra S/G entries:
386 	 * one for the header and one for the checksum at the end
387 	 */
388 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
389 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
390 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
391 		return -EINVAL;
392 
393 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
394 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
395 		send_wqe_overhead(type, qp->flags);
396 
397 	if (s > dev->dev->caps.max_sq_desc_sz)
398 		return -EINVAL;
399 
400 	qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
401 
402 	/*
403 	 * We need to leave 2 KB + 1 WR of headroom in the SQ to
404 	 * allow HW to prefetch.
405 	 */
406 	qp->sq_spare_wqes = MLX4_IB_SQ_HEADROOM(qp->sq.wqe_shift);
407 	qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr +
408 					    qp->sq_spare_wqes);
409 
410 	qp->sq.max_gs =
411 		(min(dev->dev->caps.max_sq_desc_sz,
412 		     (1 << qp->sq.wqe_shift)) -
413 		 send_wqe_overhead(type, qp->flags)) /
414 		sizeof (struct mlx4_wqe_data_seg);
415 
416 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
417 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
418 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
419 		qp->rq.offset = 0;
420 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
421 	} else {
422 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
423 		qp->sq.offset = 0;
424 	}
425 
426 	cap->max_send_wr  = qp->sq.max_post =
427 		qp->sq.wqe_cnt - qp->sq_spare_wqes;
428 	cap->max_send_sge = min(qp->sq.max_gs,
429 				min(dev->dev->caps.max_sq_sg,
430 				    dev->dev->caps.max_rq_sg));
431 	/* We don't support inline sends for kernel QPs (yet) */
432 	cap->max_inline_data = 0;
433 
434 	return 0;
435 }
436 
437 static int set_user_sq_size(struct mlx4_ib_dev *dev,
438 			    struct mlx4_ib_qp *qp,
439 			    struct mlx4_ib_create_qp *ucmd)
440 {
441 	/* Sanity check SQ size before proceeding */
442 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
443 	    ucmd->log_sq_stride >
444 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
445 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
446 		return -EINVAL;
447 
448 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
449 	qp->sq.wqe_shift = ucmd->log_sq_stride;
450 
451 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
452 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
453 
454 	return 0;
455 }
456 
457 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
458 {
459 	int i;
460 
461 	qp->sqp_proxy_rcv =
462 		kmalloc_array(qp->rq.wqe_cnt, sizeof(struct mlx4_ib_buf),
463 			      GFP_KERNEL);
464 	if (!qp->sqp_proxy_rcv)
465 		return -ENOMEM;
466 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
467 		qp->sqp_proxy_rcv[i].addr =
468 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
469 				GFP_KERNEL);
470 		if (!qp->sqp_proxy_rcv[i].addr)
471 			goto err;
472 		qp->sqp_proxy_rcv[i].map =
473 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
474 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
475 					  DMA_FROM_DEVICE);
476 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
477 			kfree(qp->sqp_proxy_rcv[i].addr);
478 			goto err;
479 		}
480 	}
481 	return 0;
482 
483 err:
484 	while (i > 0) {
485 		--i;
486 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
487 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
488 				    DMA_FROM_DEVICE);
489 		kfree(qp->sqp_proxy_rcv[i].addr);
490 	}
491 	kfree(qp->sqp_proxy_rcv);
492 	qp->sqp_proxy_rcv = NULL;
493 	return -ENOMEM;
494 }
495 
496 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
497 {
498 	int i;
499 
500 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
501 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
502 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
503 				    DMA_FROM_DEVICE);
504 		kfree(qp->sqp_proxy_rcv[i].addr);
505 	}
506 	kfree(qp->sqp_proxy_rcv);
507 }
508 
509 static bool qp_has_rq(struct ib_qp_init_attr *attr)
510 {
511 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
512 		return false;
513 
514 	return !attr->srq;
515 }
516 
517 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
518 {
519 	int i;
520 	for (i = 0; i < dev->caps.num_ports; i++) {
521 		if (qpn == dev->caps.spec_qps[i].qp0_proxy)
522 			return !!dev->caps.spec_qps[i].qp0_qkey;
523 	}
524 	return 0;
525 }
526 
527 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
528 				    struct mlx4_ib_qp *qp)
529 {
530 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
531 	mlx4_counter_free(dev->dev, qp->counter_index->index);
532 	list_del(&qp->counter_index->list);
533 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
534 
535 	kfree(qp->counter_index);
536 	qp->counter_index = NULL;
537 }
538 
539 static int set_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_rss *rss_ctx,
540 		      struct ib_qp_init_attr *init_attr,
541 		      struct mlx4_ib_create_qp_rss *ucmd)
542 {
543 	rss_ctx->base_qpn_tbl_sz = init_attr->rwq_ind_tbl->ind_tbl[0]->wq_num |
544 		(init_attr->rwq_ind_tbl->log_ind_tbl_size << 24);
545 
546 	if ((ucmd->rx_hash_function == MLX4_IB_RX_HASH_FUNC_TOEPLITZ) &&
547 	    (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS_TOP)) {
548 		memcpy(rss_ctx->rss_key, ucmd->rx_hash_key,
549 		       MLX4_EN_RSS_KEY_SIZE);
550 	} else {
551 		pr_debug("RX Hash function is not supported\n");
552 		return (-EOPNOTSUPP);
553 	}
554 
555 	if (ucmd->rx_hash_fields_mask & ~(MLX4_IB_RX_HASH_SRC_IPV4	|
556 					  MLX4_IB_RX_HASH_DST_IPV4	|
557 					  MLX4_IB_RX_HASH_SRC_IPV6	|
558 					  MLX4_IB_RX_HASH_DST_IPV6	|
559 					  MLX4_IB_RX_HASH_SRC_PORT_TCP	|
560 					  MLX4_IB_RX_HASH_DST_PORT_TCP	|
561 					  MLX4_IB_RX_HASH_SRC_PORT_UDP	|
562 					  MLX4_IB_RX_HASH_DST_PORT_UDP  |
563 					  MLX4_IB_RX_HASH_INNER)) {
564 		pr_debug("RX Hash fields_mask has unsupported mask (0x%llx)\n",
565 			 ucmd->rx_hash_fields_mask);
566 		return (-EOPNOTSUPP);
567 	}
568 
569 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) &&
570 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
571 		rss_ctx->flags = MLX4_RSS_IPV4;
572 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) ||
573 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
574 		pr_debug("RX Hash fields_mask is not supported - both IPv4 SRC and DST must be set\n");
575 		return (-EOPNOTSUPP);
576 	}
577 
578 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) &&
579 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
580 		rss_ctx->flags |= MLX4_RSS_IPV6;
581 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) ||
582 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
583 		pr_debug("RX Hash fields_mask is not supported - both IPv6 SRC and DST must be set\n");
584 		return (-EOPNOTSUPP);
585 	}
586 
587 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) &&
588 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
589 		if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS)) {
590 			pr_debug("RX Hash fields_mask for UDP is not supported\n");
591 			return (-EOPNOTSUPP);
592 		}
593 
594 		if (rss_ctx->flags & MLX4_RSS_IPV4)
595 			rss_ctx->flags |= MLX4_RSS_UDP_IPV4;
596 		if (rss_ctx->flags & MLX4_RSS_IPV6)
597 			rss_ctx->flags |= MLX4_RSS_UDP_IPV6;
598 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
599 			pr_debug("RX Hash fields_mask is not supported - UDP must be set with IPv4 or IPv6\n");
600 			return (-EOPNOTSUPP);
601 		}
602 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) ||
603 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
604 		pr_debug("RX Hash fields_mask is not supported - both UDP SRC and DST must be set\n");
605 		return (-EOPNOTSUPP);
606 	}
607 
608 	if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) &&
609 	    (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
610 		if (rss_ctx->flags & MLX4_RSS_IPV4)
611 			rss_ctx->flags |= MLX4_RSS_TCP_IPV4;
612 		if (rss_ctx->flags & MLX4_RSS_IPV6)
613 			rss_ctx->flags |= MLX4_RSS_TCP_IPV6;
614 		if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
615 			pr_debug("RX Hash fields_mask is not supported - TCP must be set with IPv4 or IPv6\n");
616 			return (-EOPNOTSUPP);
617 		}
618 	} else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) ||
619 		   (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
620 		pr_debug("RX Hash fields_mask is not supported - both TCP SRC and DST must be set\n");
621 		return (-EOPNOTSUPP);
622 	}
623 
624 	if (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_INNER) {
625 		if (dev->dev->caps.tunnel_offload_mode ==
626 		    MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
627 			/*
628 			 * Hash according to inner headers if exist, otherwise
629 			 * according to outer headers.
630 			 */
631 			rss_ctx->flags |= MLX4_RSS_BY_INNER_HEADERS_IPONLY;
632 		} else {
633 			pr_debug("RSS Hash for inner headers isn't supported\n");
634 			return (-EOPNOTSUPP);
635 		}
636 	}
637 
638 	return 0;
639 }
640 
641 static int create_qp_rss(struct mlx4_ib_dev *dev,
642 			 struct ib_qp_init_attr *init_attr,
643 			 struct mlx4_ib_create_qp_rss *ucmd,
644 			 struct mlx4_ib_qp *qp)
645 {
646 	int qpn;
647 	int err;
648 
649 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
650 
651 	err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 0, qp->mqp.usage);
652 	if (err)
653 		return err;
654 
655 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
656 	if (err)
657 		goto err_qpn;
658 
659 	mutex_init(&qp->mutex);
660 
661 	INIT_LIST_HEAD(&qp->gid_list);
662 	INIT_LIST_HEAD(&qp->steering_rules);
663 
664 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
665 	qp->state = IB_QPS_RESET;
666 
667 	/* Set dummy send resources to be compatible with HV and PRM */
668 	qp->sq_no_prefetch = 1;
669 	qp->sq.wqe_cnt = 1;
670 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
671 	qp->buf_size = qp->sq.wqe_cnt << MLX4_IB_MIN_SQ_STRIDE;
672 	qp->mtt = (to_mqp(
673 		   (struct ib_qp *)init_attr->rwq_ind_tbl->ind_tbl[0]))->mtt;
674 
675 	qp->rss_ctx = kzalloc(sizeof(*qp->rss_ctx), GFP_KERNEL);
676 	if (!qp->rss_ctx) {
677 		err = -ENOMEM;
678 		goto err_qp_alloc;
679 	}
680 
681 	err = set_qp_rss(dev, qp->rss_ctx, init_attr, ucmd);
682 	if (err)
683 		goto err;
684 
685 	return 0;
686 
687 err:
688 	kfree(qp->rss_ctx);
689 
690 err_qp_alloc:
691 	mlx4_qp_remove(dev->dev, &qp->mqp);
692 	mlx4_qp_free(dev->dev, &qp->mqp);
693 
694 err_qpn:
695 	mlx4_qp_release_range(dev->dev, qpn, 1);
696 	return err;
697 }
698 
699 static struct ib_qp *_mlx4_ib_create_qp_rss(struct ib_pd *pd,
700 					    struct ib_qp_init_attr *init_attr,
701 					    struct ib_udata *udata)
702 {
703 	struct mlx4_ib_qp *qp;
704 	struct mlx4_ib_create_qp_rss ucmd = {};
705 	size_t required_cmd_sz;
706 	int err;
707 
708 	if (!udata) {
709 		pr_debug("RSS QP with NULL udata\n");
710 		return ERR_PTR(-EINVAL);
711 	}
712 
713 	if (udata->outlen)
714 		return ERR_PTR(-EOPNOTSUPP);
715 
716 	required_cmd_sz = offsetof(typeof(ucmd), reserved1) +
717 					sizeof(ucmd.reserved1);
718 	if (udata->inlen < required_cmd_sz) {
719 		pr_debug("invalid inlen\n");
720 		return ERR_PTR(-EINVAL);
721 	}
722 
723 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
724 		pr_debug("copy failed\n");
725 		return ERR_PTR(-EFAULT);
726 	}
727 
728 	if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)))
729 		return ERR_PTR(-EOPNOTSUPP);
730 
731 	if (ucmd.comp_mask || ucmd.reserved1)
732 		return ERR_PTR(-EOPNOTSUPP);
733 
734 	if (udata->inlen > sizeof(ucmd) &&
735 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
736 				 udata->inlen - sizeof(ucmd))) {
737 		pr_debug("inlen is not supported\n");
738 		return ERR_PTR(-EOPNOTSUPP);
739 	}
740 
741 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
742 		pr_debug("RSS QP with unsupported QP type %d\n",
743 			 init_attr->qp_type);
744 		return ERR_PTR(-EOPNOTSUPP);
745 	}
746 
747 	if (init_attr->create_flags) {
748 		pr_debug("RSS QP doesn't support create flags\n");
749 		return ERR_PTR(-EOPNOTSUPP);
750 	}
751 
752 	if (init_attr->send_cq || init_attr->cap.max_send_wr) {
753 		pr_debug("RSS QP with unsupported send attributes\n");
754 		return ERR_PTR(-EOPNOTSUPP);
755 	}
756 
757 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
758 	if (!qp)
759 		return ERR_PTR(-ENOMEM);
760 
761 	qp->pri.vid = 0xFFFF;
762 	qp->alt.vid = 0xFFFF;
763 
764 	err = create_qp_rss(to_mdev(pd->device), init_attr, &ucmd, qp);
765 	if (err) {
766 		kfree(qp);
767 		return ERR_PTR(err);
768 	}
769 
770 	qp->ibqp.qp_num = qp->mqp.qpn;
771 
772 	return &qp->ibqp;
773 }
774 
775 /*
776  * This function allocates a WQN from a range which is consecutive and aligned
777  * to its size. In case the range is full, then it creates a new range and
778  * allocates WQN from it. The new range will be used for following allocations.
779  */
780 static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context,
781 			     struct mlx4_ib_qp *qp, int range_size, int *wqn)
782 {
783 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
784 	struct mlx4_wqn_range *range;
785 	int err = 0;
786 
787 	mutex_lock(&context->wqn_ranges_mutex);
788 
789 	range = list_first_entry_or_null(&context->wqn_ranges_list,
790 					 struct mlx4_wqn_range, list);
791 
792 	if (!range || (range->refcount == range->size) || range->dirty) {
793 		range = kzalloc(sizeof(*range), GFP_KERNEL);
794 		if (!range) {
795 			err = -ENOMEM;
796 			goto out;
797 		}
798 
799 		err = mlx4_qp_reserve_range(dev->dev, range_size,
800 					    range_size, &range->base_wqn, 0,
801 					    qp->mqp.usage);
802 		if (err) {
803 			kfree(range);
804 			goto out;
805 		}
806 
807 		range->size = range_size;
808 		list_add(&range->list, &context->wqn_ranges_list);
809 	} else if (range_size != 1) {
810 		/*
811 		 * Requesting a new range (>1) when last range is still open, is
812 		 * not valid.
813 		 */
814 		err = -EINVAL;
815 		goto out;
816 	}
817 
818 	qp->wqn_range = range;
819 
820 	*wqn = range->base_wqn + range->refcount;
821 
822 	range->refcount++;
823 
824 out:
825 	mutex_unlock(&context->wqn_ranges_mutex);
826 
827 	return err;
828 }
829 
830 static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context,
831 				struct mlx4_ib_qp *qp, bool dirty_release)
832 {
833 	struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
834 	struct mlx4_wqn_range *range;
835 
836 	mutex_lock(&context->wqn_ranges_mutex);
837 
838 	range = qp->wqn_range;
839 
840 	range->refcount--;
841 	if (!range->refcount) {
842 		mlx4_qp_release_range(dev->dev, range->base_wqn,
843 				      range->size);
844 		list_del(&range->list);
845 		kfree(range);
846 	} else if (dirty_release) {
847 	/*
848 	 * A range which one of its WQNs is destroyed, won't be able to be
849 	 * reused for further WQN allocations.
850 	 * The next created WQ will allocate a new range.
851 	 */
852 		range->dirty = true;
853 	}
854 
855 	mutex_unlock(&context->wqn_ranges_mutex);
856 }
857 
858 static int create_rq(struct ib_pd *pd, struct ib_qp_init_attr *init_attr,
859 		     struct ib_udata *udata, struct mlx4_ib_qp *qp)
860 {
861 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
862 	int qpn;
863 	int err;
864 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
865 		udata, struct mlx4_ib_ucontext, ibucontext);
866 	struct mlx4_ib_cq *mcq;
867 	unsigned long flags;
868 	int range_size;
869 	struct mlx4_ib_create_wq wq;
870 	size_t copy_len;
871 	int shift;
872 	int n;
873 
874 	qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
875 
876 	mutex_init(&qp->mutex);
877 	spin_lock_init(&qp->sq.lock);
878 	spin_lock_init(&qp->rq.lock);
879 	INIT_LIST_HEAD(&qp->gid_list);
880 	INIT_LIST_HEAD(&qp->steering_rules);
881 
882 	qp->state = IB_QPS_RESET;
883 
884 	copy_len = min(sizeof(struct mlx4_ib_create_wq), udata->inlen);
885 
886 	if (ib_copy_from_udata(&wq, udata, copy_len)) {
887 		err = -EFAULT;
888 		goto err;
889 	}
890 
891 	if (wq.comp_mask || wq.reserved[0] || wq.reserved[1] ||
892 	    wq.reserved[2]) {
893 		pr_debug("user command isn't supported\n");
894 		err = -EOPNOTSUPP;
895 		goto err;
896 	}
897 
898 	if (wq.log_range_size > ilog2(dev->dev->caps.max_rss_tbl_sz)) {
899 		pr_debug("WQN range size must be equal or smaller than %d\n",
900 			 dev->dev->caps.max_rss_tbl_sz);
901 		err = -EOPNOTSUPP;
902 		goto err;
903 	}
904 	range_size = 1 << wq.log_range_size;
905 
906 	if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS)
907 		qp->flags |= MLX4_IB_QP_SCATTER_FCS;
908 
909 	err = set_rq_size(dev, &init_attr->cap, true, true, qp, qp->inl_recv_sz);
910 	if (err)
911 		goto err;
912 
913 	qp->sq_no_prefetch = 1;
914 	qp->sq.wqe_cnt = 1;
915 	qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
916 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
917 		       (qp->sq.wqe_cnt << qp->sq.wqe_shift);
918 
919 	qp->umem = ib_umem_get(pd->device, wq.buf_addr, qp->buf_size, 0);
920 	if (IS_ERR(qp->umem)) {
921 		err = PTR_ERR(qp->umem);
922 		goto err;
923 	}
924 
925 	n = ib_umem_page_count(qp->umem);
926 	shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
927 	err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
928 
929 	if (err)
930 		goto err_buf;
931 
932 	err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
933 	if (err)
934 		goto err_mtt;
935 
936 	err = mlx4_ib_db_map_user(udata, wq.db_addr, &qp->db);
937 	if (err)
938 		goto err_mtt;
939 	qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
940 
941 	err = mlx4_ib_alloc_wqn(context, qp, range_size, &qpn);
942 	if (err)
943 		goto err_wrid;
944 
945 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
946 	if (err)
947 		goto err_qpn;
948 
949 	/*
950 	 * Hardware wants QPN written in big-endian order (after
951 	 * shifting) for send doorbell.  Precompute this value to save
952 	 * a little bit when posting sends.
953 	 */
954 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
955 
956 	qp->mqp.event = mlx4_ib_wq_event;
957 
958 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
959 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
960 			 to_mcq(init_attr->recv_cq));
961 	/* Maintain device to QPs access, needed for further handling
962 	 * via reset flow
963 	 */
964 	list_add_tail(&qp->qps_list, &dev->qp_list);
965 	/* Maintain CQ to QPs access, needed for further handling
966 	 * via reset flow
967 	 */
968 	mcq = to_mcq(init_attr->send_cq);
969 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
970 	mcq = to_mcq(init_attr->recv_cq);
971 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
972 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
973 			   to_mcq(init_attr->recv_cq));
974 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
975 	return 0;
976 
977 err_qpn:
978 	mlx4_ib_release_wqn(context, qp, 0);
979 err_wrid:
980 	mlx4_ib_db_unmap_user(context, &qp->db);
981 
982 err_mtt:
983 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
984 err_buf:
985 	ib_umem_release(qp->umem);
986 err:
987 	return err;
988 }
989 
990 static int create_qp_common(struct ib_pd *pd, struct ib_qp_init_attr *init_attr,
991 			    struct ib_udata *udata, int sqpn,
992 			    struct mlx4_ib_qp **caller_qp)
993 {
994 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
995 	int qpn;
996 	int err;
997 	struct mlx4_ib_sqp *sqp = NULL;
998 	struct mlx4_ib_qp *qp;
999 	struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context(
1000 		udata, struct mlx4_ib_ucontext, ibucontext);
1001 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
1002 	struct mlx4_ib_cq *mcq;
1003 	unsigned long flags;
1004 
1005 	/* When tunneling special qps, we use a plain UD qp */
1006 	if (sqpn) {
1007 		if (mlx4_is_mfunc(dev->dev) &&
1008 		    (!mlx4_is_master(dev->dev) ||
1009 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
1010 			if (init_attr->qp_type == IB_QPT_GSI)
1011 				qp_type = MLX4_IB_QPT_PROXY_GSI;
1012 			else {
1013 				if (mlx4_is_master(dev->dev) ||
1014 				    qp0_enabled_vf(dev->dev, sqpn))
1015 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
1016 				else
1017 					qp_type = MLX4_IB_QPT_PROXY_SMI;
1018 			}
1019 		}
1020 		qpn = sqpn;
1021 		/* add extra sg entry for tunneling */
1022 		init_attr->cap.max_recv_sge++;
1023 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
1024 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
1025 			container_of(init_attr,
1026 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
1027 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
1028 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
1029 		    !mlx4_is_master(dev->dev))
1030 			return -EINVAL;
1031 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
1032 			qp_type = MLX4_IB_QPT_TUN_GSI;
1033 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
1034 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
1035 					     tnl_init->port))
1036 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
1037 		else
1038 			qp_type = MLX4_IB_QPT_TUN_SMI;
1039 		/* we are definitely in the PPF here, since we are creating
1040 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
1041 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
1042 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
1043 		sqpn = qpn;
1044 	}
1045 
1046 	if (!*caller_qp) {
1047 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
1048 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
1049 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
1050 			sqp = kzalloc(sizeof(struct mlx4_ib_sqp), GFP_KERNEL);
1051 			if (!sqp)
1052 				return -ENOMEM;
1053 			qp = &sqp->qp;
1054 		} else {
1055 			qp = kzalloc(sizeof(struct mlx4_ib_qp), GFP_KERNEL);
1056 			if (!qp)
1057 				return -ENOMEM;
1058 		}
1059 		qp->pri.vid = 0xFFFF;
1060 		qp->alt.vid = 0xFFFF;
1061 	} else
1062 		qp = *caller_qp;
1063 
1064 	qp->mlx4_ib_qp_type = qp_type;
1065 
1066 	mutex_init(&qp->mutex);
1067 	spin_lock_init(&qp->sq.lock);
1068 	spin_lock_init(&qp->rq.lock);
1069 	INIT_LIST_HEAD(&qp->gid_list);
1070 	INIT_LIST_HEAD(&qp->steering_rules);
1071 
1072 	qp->state = IB_QPS_RESET;
1073 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1074 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1075 
1076 	if (udata) {
1077 		struct mlx4_ib_create_qp ucmd;
1078 		size_t copy_len;
1079 		int shift;
1080 		int n;
1081 
1082 		copy_len = sizeof(struct mlx4_ib_create_qp);
1083 
1084 		if (ib_copy_from_udata(&ucmd, udata, copy_len)) {
1085 			err = -EFAULT;
1086 			goto err;
1087 		}
1088 
1089 		qp->inl_recv_sz = ucmd.inl_recv_sz;
1090 
1091 		if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1092 			if (!(dev->dev->caps.flags &
1093 			      MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
1094 				pr_debug("scatter FCS is unsupported\n");
1095 				err = -EOPNOTSUPP;
1096 				goto err;
1097 			}
1098 
1099 			qp->flags |= MLX4_IB_QP_SCATTER_FCS;
1100 		}
1101 
1102 		err = set_rq_size(dev, &init_attr->cap, udata,
1103 				  qp_has_rq(init_attr), qp, qp->inl_recv_sz);
1104 		if (err)
1105 			goto err;
1106 
1107 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
1108 
1109 		err = set_user_sq_size(dev, qp, &ucmd);
1110 		if (err)
1111 			goto err;
1112 
1113 		qp->umem =
1114 			ib_umem_get(pd->device, ucmd.buf_addr, qp->buf_size, 0);
1115 		if (IS_ERR(qp->umem)) {
1116 			err = PTR_ERR(qp->umem);
1117 			goto err;
1118 		}
1119 
1120 		n = ib_umem_page_count(qp->umem);
1121 		shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
1122 		err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
1123 
1124 		if (err)
1125 			goto err_buf;
1126 
1127 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
1128 		if (err)
1129 			goto err_mtt;
1130 
1131 		if (qp_has_rq(init_attr)) {
1132 			err = mlx4_ib_db_map_user(udata, ucmd.db_addr, &qp->db);
1133 			if (err)
1134 				goto err_mtt;
1135 		}
1136 		qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
1137 	} else {
1138 		err = set_rq_size(dev, &init_attr->cap, udata,
1139 				  qp_has_rq(init_attr), qp, 0);
1140 		if (err)
1141 			goto err;
1142 
1143 		qp->sq_no_prefetch = 0;
1144 
1145 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
1146 			qp->flags |= MLX4_IB_QP_LSO;
1147 
1148 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1149 			if (dev->steering_support ==
1150 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
1151 				qp->flags |= MLX4_IB_QP_NETIF;
1152 			else
1153 				goto err;
1154 		}
1155 
1156 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
1157 		if (err)
1158 			goto err;
1159 
1160 		if (qp_has_rq(init_attr)) {
1161 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
1162 			if (err)
1163 				goto err;
1164 
1165 			*qp->db.db = 0;
1166 		}
1167 
1168 		if (mlx4_buf_alloc(dev->dev, qp->buf_size,  PAGE_SIZE * 2,
1169 				   &qp->buf)) {
1170 			err = -ENOMEM;
1171 			goto err_db;
1172 		}
1173 
1174 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
1175 				    &qp->mtt);
1176 		if (err)
1177 			goto err_buf;
1178 
1179 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
1180 		if (err)
1181 			goto err_mtt;
1182 
1183 		qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1184 					     sizeof(u64), GFP_KERNEL);
1185 		qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1186 					     sizeof(u64), GFP_KERNEL);
1187 		if (!qp->sq.wrid || !qp->rq.wrid) {
1188 			err = -ENOMEM;
1189 			goto err_wrid;
1190 		}
1191 		qp->mqp.usage = MLX4_RES_USAGE_DRIVER;
1192 	}
1193 
1194 	if (sqpn) {
1195 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1196 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
1197 			if (alloc_proxy_bufs(pd->device, qp)) {
1198 				err = -ENOMEM;
1199 				goto err_wrid;
1200 			}
1201 		}
1202 	} else {
1203 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
1204 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
1205 		 * VLAN insertion. */
1206 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
1207 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
1208 						    (init_attr->cap.max_send_wr ?
1209 						     MLX4_RESERVE_ETH_BF_QP : 0) |
1210 						    (init_attr->cap.max_recv_wr ?
1211 						     MLX4_RESERVE_A0_QP : 0),
1212 						    qp->mqp.usage);
1213 		else
1214 			if (qp->flags & MLX4_IB_QP_NETIF)
1215 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
1216 			else
1217 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
1218 							    &qpn, 0, qp->mqp.usage);
1219 		if (err)
1220 			goto err_proxy;
1221 	}
1222 
1223 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1224 		qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1225 
1226 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
1227 	if (err)
1228 		goto err_qpn;
1229 
1230 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1231 		qp->mqp.qpn |= (1 << 23);
1232 
1233 	/*
1234 	 * Hardware wants QPN written in big-endian order (after
1235 	 * shifting) for send doorbell.  Precompute this value to save
1236 	 * a little bit when posting sends.
1237 	 */
1238 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1239 
1240 	qp->mqp.event = mlx4_ib_qp_event;
1241 
1242 	if (!*caller_qp)
1243 		*caller_qp = qp;
1244 
1245 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1246 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
1247 			 to_mcq(init_attr->recv_cq));
1248 	/* Maintain device to QPs access, needed for further handling
1249 	 * via reset flow
1250 	 */
1251 	list_add_tail(&qp->qps_list, &dev->qp_list);
1252 	/* Maintain CQ to QPs access, needed for further handling
1253 	 * via reset flow
1254 	 */
1255 	mcq = to_mcq(init_attr->send_cq);
1256 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
1257 	mcq = to_mcq(init_attr->recv_cq);
1258 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
1259 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
1260 			   to_mcq(init_attr->recv_cq));
1261 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1262 	return 0;
1263 
1264 err_qpn:
1265 	if (!sqpn) {
1266 		if (qp->flags & MLX4_IB_QP_NETIF)
1267 			mlx4_ib_steer_qp_free(dev, qpn, 1);
1268 		else
1269 			mlx4_qp_release_range(dev->dev, qpn, 1);
1270 	}
1271 err_proxy:
1272 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1273 		free_proxy_bufs(pd->device, qp);
1274 err_wrid:
1275 	if (udata) {
1276 		if (qp_has_rq(init_attr))
1277 			mlx4_ib_db_unmap_user(context, &qp->db);
1278 	} else {
1279 		kvfree(qp->sq.wrid);
1280 		kvfree(qp->rq.wrid);
1281 	}
1282 
1283 err_mtt:
1284 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1285 
1286 err_buf:
1287 	if (!qp->umem)
1288 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1289 	ib_umem_release(qp->umem);
1290 
1291 err_db:
1292 	if (!udata && qp_has_rq(init_attr))
1293 		mlx4_db_free(dev->dev, &qp->db);
1294 
1295 err:
1296 	if (!sqp && !*caller_qp)
1297 		kfree(qp);
1298 	kfree(sqp);
1299 
1300 	return err;
1301 }
1302 
1303 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
1304 {
1305 	switch (state) {
1306 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
1307 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
1308 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
1309 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
1310 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
1311 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
1312 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
1313 	default:		return -1;
1314 	}
1315 }
1316 
1317 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1318 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1319 {
1320 	if (send_cq == recv_cq) {
1321 		spin_lock(&send_cq->lock);
1322 		__acquire(&recv_cq->lock);
1323 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1324 		spin_lock(&send_cq->lock);
1325 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1326 	} else {
1327 		spin_lock(&recv_cq->lock);
1328 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1329 	}
1330 }
1331 
1332 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1333 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1334 {
1335 	if (send_cq == recv_cq) {
1336 		__release(&recv_cq->lock);
1337 		spin_unlock(&send_cq->lock);
1338 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1339 		spin_unlock(&recv_cq->lock);
1340 		spin_unlock(&send_cq->lock);
1341 	} else {
1342 		spin_unlock(&send_cq->lock);
1343 		spin_unlock(&recv_cq->lock);
1344 	}
1345 }
1346 
1347 static void del_gid_entries(struct mlx4_ib_qp *qp)
1348 {
1349 	struct mlx4_ib_gid_entry *ge, *tmp;
1350 
1351 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1352 		list_del(&ge->list);
1353 		kfree(ge);
1354 	}
1355 }
1356 
1357 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
1358 {
1359 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
1360 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
1361 	else
1362 		return to_mpd(qp->ibqp.pd);
1363 }
1364 
1365 static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src,
1366 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1367 {
1368 	switch (qp->ibqp.qp_type) {
1369 	case IB_QPT_XRC_TGT:
1370 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1371 		*recv_cq = *send_cq;
1372 		break;
1373 	case IB_QPT_XRC_INI:
1374 		*send_cq = to_mcq(qp->ibqp.send_cq);
1375 		*recv_cq = *send_cq;
1376 		break;
1377 	default:
1378 		*recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) :
1379 						     to_mcq(qp->ibwq.cq);
1380 		*send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) :
1381 						     *recv_cq;
1382 		break;
1383 	}
1384 }
1385 
1386 static void destroy_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1387 {
1388 	if (qp->state != IB_QPS_RESET) {
1389 		int i;
1390 
1391 		for (i = 0; i < (1 << qp->ibqp.rwq_ind_tbl->log_ind_tbl_size);
1392 		     i++) {
1393 			struct ib_wq *ibwq = qp->ibqp.rwq_ind_tbl->ind_tbl[i];
1394 			struct mlx4_ib_qp *wq =	to_mqp((struct ib_qp *)ibwq);
1395 
1396 			mutex_lock(&wq->mutex);
1397 
1398 			wq->rss_usecnt--;
1399 
1400 			mutex_unlock(&wq->mutex);
1401 		}
1402 
1403 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1404 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1405 			pr_warn("modify QP %06x to RESET failed.\n",
1406 				qp->mqp.qpn);
1407 	}
1408 
1409 	mlx4_qp_remove(dev->dev, &qp->mqp);
1410 	mlx4_qp_free(dev->dev, &qp->mqp);
1411 	mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1412 	del_gid_entries(qp);
1413 	kfree(qp->rss_ctx);
1414 }
1415 
1416 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1417 			      enum mlx4_ib_source_type src,
1418 			      struct ib_udata *udata)
1419 {
1420 	struct mlx4_ib_cq *send_cq, *recv_cq;
1421 	unsigned long flags;
1422 
1423 	if (qp->state != IB_QPS_RESET) {
1424 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1425 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1426 			pr_warn("modify QP %06x to RESET failed.\n",
1427 			       qp->mqp.qpn);
1428 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1429 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1430 			qp->pri.smac = 0;
1431 			qp->pri.smac_port = 0;
1432 		}
1433 		if (qp->alt.smac) {
1434 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1435 			qp->alt.smac = 0;
1436 		}
1437 		if (qp->pri.vid < 0x1000) {
1438 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1439 			qp->pri.vid = 0xFFFF;
1440 			qp->pri.candidate_vid = 0xFFFF;
1441 			qp->pri.update_vid = 0;
1442 		}
1443 		if (qp->alt.vid < 0x1000) {
1444 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1445 			qp->alt.vid = 0xFFFF;
1446 			qp->alt.candidate_vid = 0xFFFF;
1447 			qp->alt.update_vid = 0;
1448 		}
1449 	}
1450 
1451 	get_cqs(qp, src, &send_cq, &recv_cq);
1452 
1453 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1454 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1455 
1456 	/* del from lists under both locks above to protect reset flow paths */
1457 	list_del(&qp->qps_list);
1458 	list_del(&qp->cq_send_list);
1459 	list_del(&qp->cq_recv_list);
1460 	if (!udata) {
1461 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1462 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1463 		if (send_cq != recv_cq)
1464 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1465 	}
1466 
1467 	mlx4_qp_remove(dev->dev, &qp->mqp);
1468 
1469 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1470 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1471 
1472 	mlx4_qp_free(dev->dev, &qp->mqp);
1473 
1474 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1475 		if (qp->flags & MLX4_IB_QP_NETIF)
1476 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1477 		else if (src == MLX4_IB_RWQ_SRC)
1478 			mlx4_ib_release_wqn(
1479 				rdma_udata_to_drv_context(
1480 					udata,
1481 					struct mlx4_ib_ucontext,
1482 					ibucontext),
1483 				qp, 1);
1484 		else
1485 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1486 	}
1487 
1488 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1489 
1490 	if (udata) {
1491 		if (qp->rq.wqe_cnt) {
1492 			struct mlx4_ib_ucontext *mcontext =
1493 				rdma_udata_to_drv_context(
1494 					udata,
1495 					struct mlx4_ib_ucontext,
1496 					ibucontext);
1497 
1498 			mlx4_ib_db_unmap_user(mcontext, &qp->db);
1499 		}
1500 	} else {
1501 		kvfree(qp->sq.wrid);
1502 		kvfree(qp->rq.wrid);
1503 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1504 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1505 			free_proxy_bufs(&dev->ib_dev, qp);
1506 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1507 		if (qp->rq.wqe_cnt)
1508 			mlx4_db_free(dev->dev, &qp->db);
1509 	}
1510 	ib_umem_release(qp->umem);
1511 
1512 	del_gid_entries(qp);
1513 }
1514 
1515 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1516 {
1517 	/* Native or PPF */
1518 	if (!mlx4_is_mfunc(dev->dev) ||
1519 	    (mlx4_is_master(dev->dev) &&
1520 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1521 		return  dev->dev->phys_caps.base_sqpn +
1522 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1523 			attr->port_num - 1;
1524 	}
1525 	/* PF or VF -- creating proxies */
1526 	if (attr->qp_type == IB_QPT_SMI)
1527 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp0_proxy;
1528 	else
1529 		return dev->dev->caps.spec_qps[attr->port_num - 1].qp1_proxy;
1530 }
1531 
1532 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1533 					struct ib_qp_init_attr *init_attr,
1534 					struct ib_udata *udata)
1535 {
1536 	struct mlx4_ib_qp *qp = NULL;
1537 	int err;
1538 	int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1539 	u16 xrcdn = 0;
1540 
1541 	if (init_attr->rwq_ind_tbl)
1542 		return _mlx4_ib_create_qp_rss(pd, init_attr, udata);
1543 
1544 	/*
1545 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1546 	 * and only for kernel UD QPs.
1547 	 */
1548 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1549 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1550 					MLX4_IB_SRIOV_TUNNEL_QP |
1551 					MLX4_IB_SRIOV_SQP |
1552 					MLX4_IB_QP_NETIF |
1553 					MLX4_IB_QP_CREATE_ROCE_V2_GSI))
1554 		return ERR_PTR(-EINVAL);
1555 
1556 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1557 		if (init_attr->qp_type != IB_QPT_UD)
1558 			return ERR_PTR(-EINVAL);
1559 	}
1560 
1561 	if (init_attr->create_flags) {
1562 		if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1563 			return ERR_PTR(-EINVAL);
1564 
1565 		if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1566 						 MLX4_IB_QP_CREATE_ROCE_V2_GSI  |
1567 						 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1568 		     init_attr->qp_type != IB_QPT_UD) ||
1569 		    (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1570 		     init_attr->qp_type > IB_QPT_GSI) ||
1571 		    (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1572 		     init_attr->qp_type != IB_QPT_GSI))
1573 			return ERR_PTR(-EINVAL);
1574 	}
1575 
1576 	switch (init_attr->qp_type) {
1577 	case IB_QPT_XRC_TGT:
1578 		pd = to_mxrcd(init_attr->xrcd)->pd;
1579 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1580 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1581 		/* fall through */
1582 	case IB_QPT_XRC_INI:
1583 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1584 			return ERR_PTR(-ENOSYS);
1585 		init_attr->recv_cq = init_attr->send_cq;
1586 		/* fall through */
1587 	case IB_QPT_RC:
1588 	case IB_QPT_UC:
1589 	case IB_QPT_RAW_PACKET:
1590 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1591 		if (!qp)
1592 			return ERR_PTR(-ENOMEM);
1593 		qp->pri.vid = 0xFFFF;
1594 		qp->alt.vid = 0xFFFF;
1595 		/* fall through */
1596 	case IB_QPT_UD:
1597 	{
1598 		err = create_qp_common(pd, init_attr, udata, 0, &qp);
1599 		if (err) {
1600 			kfree(qp);
1601 			return ERR_PTR(err);
1602 		}
1603 
1604 		qp->ibqp.qp_num = qp->mqp.qpn;
1605 		qp->xrcdn = xrcdn;
1606 
1607 		break;
1608 	}
1609 	case IB_QPT_SMI:
1610 	case IB_QPT_GSI:
1611 	{
1612 		int sqpn;
1613 
1614 		/* Userspace is not allowed to create special QPs: */
1615 		if (udata)
1616 			return ERR_PTR(-EINVAL);
1617 		if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1618 			int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev,
1619 							1, 1, &sqpn, 0,
1620 							MLX4_RES_USAGE_DRIVER);
1621 
1622 			if (res)
1623 				return ERR_PTR(res);
1624 		} else {
1625 			sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1626 		}
1627 
1628 		err = create_qp_common(pd, init_attr, udata, sqpn, &qp);
1629 		if (err)
1630 			return ERR_PTR(err);
1631 
1632 		qp->port	= init_attr->port_num;
1633 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1634 			init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1635 		break;
1636 	}
1637 	default:
1638 		/* Don't support raw QPs */
1639 		return ERR_PTR(-EOPNOTSUPP);
1640 	}
1641 
1642 	return &qp->ibqp;
1643 }
1644 
1645 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1646 				struct ib_qp_init_attr *init_attr,
1647 				struct ib_udata *udata) {
1648 	struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1649 	struct ib_qp *ibqp;
1650 	struct mlx4_ib_dev *dev = to_mdev(device);
1651 
1652 	ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1653 
1654 	if (!IS_ERR(ibqp) &&
1655 	    (init_attr->qp_type == IB_QPT_GSI) &&
1656 	    !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1657 		struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1658 		int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1659 
1660 		if (is_eth &&
1661 		    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1662 			init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1663 			sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1664 
1665 			if (IS_ERR(sqp->roce_v2_gsi)) {
1666 				pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1667 				sqp->roce_v2_gsi = NULL;
1668 			} else {
1669 				sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1670 				sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1671 			}
1672 
1673 			init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1674 		}
1675 	}
1676 	return ibqp;
1677 }
1678 
1679 static int _mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
1680 {
1681 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1682 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1683 
1684 	if (is_qp0(dev, mqp))
1685 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1686 
1687 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1688 	    dev->qp1_proxy[mqp->port - 1] == mqp) {
1689 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1690 		dev->qp1_proxy[mqp->port - 1] = NULL;
1691 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1692 	}
1693 
1694 	if (mqp->counter_index)
1695 		mlx4_ib_free_qp_counter(dev, mqp);
1696 
1697 	if (qp->rwq_ind_tbl) {
1698 		destroy_qp_rss(dev, mqp);
1699 	} else {
1700 		destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, udata);
1701 	}
1702 
1703 	if (is_sqp(dev, mqp))
1704 		kfree(to_msqp(mqp));
1705 	else
1706 		kfree(mqp);
1707 
1708 	return 0;
1709 }
1710 
1711 int mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
1712 {
1713 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1714 
1715 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1716 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1717 
1718 		if (sqp->roce_v2_gsi)
1719 			ib_destroy_qp(sqp->roce_v2_gsi);
1720 	}
1721 
1722 	return _mlx4_ib_destroy_qp(qp, udata);
1723 }
1724 
1725 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1726 {
1727 	switch (type) {
1728 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1729 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1730 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1731 	case MLX4_IB_QPT_XRC_INI:
1732 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1733 	case MLX4_IB_QPT_SMI:
1734 	case MLX4_IB_QPT_GSI:
1735 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1736 
1737 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1738 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1739 						MLX4_QP_ST_MLX : -1);
1740 	case MLX4_IB_QPT_PROXY_SMI:
1741 	case MLX4_IB_QPT_TUN_SMI:
1742 	case MLX4_IB_QPT_PROXY_GSI:
1743 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1744 						MLX4_QP_ST_UD : -1);
1745 	default:			return -1;
1746 	}
1747 }
1748 
1749 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1750 				   int attr_mask)
1751 {
1752 	u8 dest_rd_atomic;
1753 	u32 access_flags;
1754 	u32 hw_access_flags = 0;
1755 
1756 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1757 		dest_rd_atomic = attr->max_dest_rd_atomic;
1758 	else
1759 		dest_rd_atomic = qp->resp_depth;
1760 
1761 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1762 		access_flags = attr->qp_access_flags;
1763 	else
1764 		access_flags = qp->atomic_rd_en;
1765 
1766 	if (!dest_rd_atomic)
1767 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1768 
1769 	if (access_flags & IB_ACCESS_REMOTE_READ)
1770 		hw_access_flags |= MLX4_QP_BIT_RRE;
1771 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1772 		hw_access_flags |= MLX4_QP_BIT_RAE;
1773 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1774 		hw_access_flags |= MLX4_QP_BIT_RWE;
1775 
1776 	return cpu_to_be32(hw_access_flags);
1777 }
1778 
1779 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1780 			    int attr_mask)
1781 {
1782 	if (attr_mask & IB_QP_PKEY_INDEX)
1783 		sqp->pkey_index = attr->pkey_index;
1784 	if (attr_mask & IB_QP_QKEY)
1785 		sqp->qkey = attr->qkey;
1786 	if (attr_mask & IB_QP_SQ_PSN)
1787 		sqp->send_psn = attr->sq_psn;
1788 }
1789 
1790 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1791 {
1792 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1793 }
1794 
1795 static int _mlx4_set_path(struct mlx4_ib_dev *dev,
1796 			  const struct rdma_ah_attr *ah,
1797 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1798 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1799 {
1800 	int vidx;
1801 	int smac_index;
1802 	int err;
1803 
1804 	path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
1805 	path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
1806 	if (rdma_ah_get_static_rate(ah)) {
1807 		path->static_rate = rdma_ah_get_static_rate(ah) +
1808 				    MLX4_STAT_RATE_OFFSET;
1809 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1810 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1811 			--path->static_rate;
1812 	} else
1813 		path->static_rate = 0;
1814 
1815 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
1816 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
1817 		int real_sgid_index =
1818 			mlx4_ib_gid_index_to_real_index(dev, grh->sgid_attr);
1819 
1820 		if (real_sgid_index < 0)
1821 			return real_sgid_index;
1822 		if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1823 			pr_err("sgid_index (%u) too large. max is %d\n",
1824 			       real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1825 			return -1;
1826 		}
1827 
1828 		path->grh_mylmc |= 1 << 7;
1829 		path->mgid_index = real_sgid_index;
1830 		path->hop_limit  = grh->hop_limit;
1831 		path->tclass_flowlabel =
1832 			cpu_to_be32((grh->traffic_class << 20) |
1833 				    (grh->flow_label));
1834 		memcpy(path->rgid, grh->dgid.raw, 16);
1835 	}
1836 
1837 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
1838 		if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
1839 			return -1;
1840 
1841 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1842 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
1843 
1844 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1845 		if (vlan_tag < 0x1000) {
1846 			if (smac_info->vid < 0x1000) {
1847 				/* both valid vlan ids */
1848 				if (smac_info->vid != vlan_tag) {
1849 					/* different VIDs.  unreg old and reg new */
1850 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1851 					if (err)
1852 						return err;
1853 					smac_info->candidate_vid = vlan_tag;
1854 					smac_info->candidate_vlan_index = vidx;
1855 					smac_info->candidate_vlan_port = port;
1856 					smac_info->update_vid = 1;
1857 					path->vlan_index = vidx;
1858 				} else {
1859 					path->vlan_index = smac_info->vlan_index;
1860 				}
1861 			} else {
1862 				/* no current vlan tag in qp */
1863 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1864 				if (err)
1865 					return err;
1866 				smac_info->candidate_vid = vlan_tag;
1867 				smac_info->candidate_vlan_index = vidx;
1868 				smac_info->candidate_vlan_port = port;
1869 				smac_info->update_vid = 1;
1870 				path->vlan_index = vidx;
1871 			}
1872 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1873 			path->fl = 1 << 6;
1874 		} else {
1875 			/* have current vlan tag. unregister it at modify-qp success */
1876 			if (smac_info->vid < 0x1000) {
1877 				smac_info->candidate_vid = 0xFFFF;
1878 				smac_info->update_vid = 1;
1879 			}
1880 		}
1881 
1882 		/* get smac_index for RoCE use.
1883 		 * If no smac was yet assigned, register one.
1884 		 * If one was already assigned, but the new mac differs,
1885 		 * unregister the old one and register the new one.
1886 		*/
1887 		if ((!smac_info->smac && !smac_info->smac_port) ||
1888 		    smac_info->smac != smac) {
1889 			/* register candidate now, unreg if needed, after success */
1890 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1891 			if (smac_index >= 0) {
1892 				smac_info->candidate_smac_index = smac_index;
1893 				smac_info->candidate_smac = smac;
1894 				smac_info->candidate_smac_port = port;
1895 			} else {
1896 				return -EINVAL;
1897 			}
1898 		} else {
1899 			smac_index = smac_info->smac_index;
1900 		}
1901 		memcpy(path->dmac, ah->roce.dmac, 6);
1902 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1903 		/* put MAC table smac index for IBoE */
1904 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1905 	} else {
1906 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1907 			((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
1908 	}
1909 
1910 	return 0;
1911 }
1912 
1913 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1914 			 enum ib_qp_attr_mask qp_attr_mask,
1915 			 struct mlx4_ib_qp *mqp,
1916 			 struct mlx4_qp_path *path, u8 port,
1917 			 u16 vlan_id, u8 *smac)
1918 {
1919 	return _mlx4_set_path(dev, &qp->ah_attr,
1920 			      mlx4_mac_to_u64(smac),
1921 			      vlan_id,
1922 			      path, &mqp->pri, port);
1923 }
1924 
1925 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1926 			     const struct ib_qp_attr *qp,
1927 			     enum ib_qp_attr_mask qp_attr_mask,
1928 			     struct mlx4_ib_qp *mqp,
1929 			     struct mlx4_qp_path *path, u8 port)
1930 {
1931 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1932 			      0,
1933 			      0xffff,
1934 			      path, &mqp->alt, port);
1935 }
1936 
1937 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1938 {
1939 	struct mlx4_ib_gid_entry *ge, *tmp;
1940 
1941 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1942 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1943 			ge->added = 1;
1944 			ge->port = qp->port;
1945 		}
1946 	}
1947 }
1948 
1949 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1950 				    struct mlx4_ib_qp *qp,
1951 				    struct mlx4_qp_context *context)
1952 {
1953 	u64 u64_mac;
1954 	int smac_index;
1955 
1956 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1957 
1958 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1959 	if (!qp->pri.smac && !qp->pri.smac_port) {
1960 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1961 		if (smac_index >= 0) {
1962 			qp->pri.candidate_smac_index = smac_index;
1963 			qp->pri.candidate_smac = u64_mac;
1964 			qp->pri.candidate_smac_port = qp->port;
1965 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1966 		} else {
1967 			return -ENOENT;
1968 		}
1969 	}
1970 	return 0;
1971 }
1972 
1973 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1974 {
1975 	struct counter_index *new_counter_index;
1976 	int err;
1977 	u32 tmp_idx;
1978 
1979 	if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1980 	    IB_LINK_LAYER_ETHERNET ||
1981 	    !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1982 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1983 		return 0;
1984 
1985 	err = mlx4_counter_alloc(dev->dev, &tmp_idx, MLX4_RES_USAGE_DRIVER);
1986 	if (err)
1987 		return err;
1988 
1989 	new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1990 	if (!new_counter_index) {
1991 		mlx4_counter_free(dev->dev, tmp_idx);
1992 		return -ENOMEM;
1993 	}
1994 
1995 	new_counter_index->index = tmp_idx;
1996 	new_counter_index->allocated = 1;
1997 	qp->counter_index = new_counter_index;
1998 
1999 	mutex_lock(&dev->counters_table[qp->port - 1].mutex);
2000 	list_add_tail(&new_counter_index->list,
2001 		      &dev->counters_table[qp->port - 1].counters_list);
2002 	mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
2003 
2004 	return 0;
2005 }
2006 
2007 enum {
2008 	MLX4_QPC_ROCE_MODE_1 = 0,
2009 	MLX4_QPC_ROCE_MODE_2 = 2,
2010 	MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
2011 };
2012 
2013 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
2014 {
2015 	switch (gid_type) {
2016 	case IB_GID_TYPE_ROCE:
2017 		return MLX4_QPC_ROCE_MODE_1;
2018 	case IB_GID_TYPE_ROCE_UDP_ENCAP:
2019 		return MLX4_QPC_ROCE_MODE_2;
2020 	default:
2021 		return MLX4_QPC_ROCE_MODE_UNDEFINED;
2022 	}
2023 }
2024 
2025 /*
2026  * Go over all RSS QP's childes (WQs) and apply their HW state according to
2027  * their logic state if the RSS QP is the first RSS QP associated for the WQ.
2028  */
2029 static int bringup_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, u8 port_num,
2030 			    struct ib_udata *udata)
2031 {
2032 	int err = 0;
2033 	int i;
2034 
2035 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2036 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2037 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2038 
2039 		mutex_lock(&wq->mutex);
2040 
2041 		/* Mlx4_ib restrictions:
2042 		 * WQ's is associated to a port according to the RSS QP it is
2043 		 * associates to.
2044 		 * In case the WQ is associated to a different port by another
2045 		 * RSS QP, return a failure.
2046 		 */
2047 		if ((wq->rss_usecnt > 0) && (wq->port != port_num)) {
2048 			err = -EINVAL;
2049 			mutex_unlock(&wq->mutex);
2050 			break;
2051 		}
2052 		wq->port = port_num;
2053 		if ((wq->rss_usecnt == 0) && (ibwq->state == IB_WQS_RDY)) {
2054 			err = _mlx4_ib_modify_wq(ibwq, IB_WQS_RDY, udata);
2055 			if (err) {
2056 				mutex_unlock(&wq->mutex);
2057 				break;
2058 			}
2059 		}
2060 		wq->rss_usecnt++;
2061 
2062 		mutex_unlock(&wq->mutex);
2063 	}
2064 
2065 	if (i && err) {
2066 		int j;
2067 
2068 		for (j = (i - 1); j >= 0; j--) {
2069 			struct ib_wq *ibwq = ind_tbl->ind_tbl[j];
2070 			struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2071 
2072 			mutex_lock(&wq->mutex);
2073 
2074 			if ((wq->rss_usecnt == 1) &&
2075 			    (ibwq->state == IB_WQS_RDY))
2076 				if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET,
2077 						       udata))
2078 					pr_warn("failed to reverse WQN=0x%06x\n",
2079 						ibwq->wq_num);
2080 			wq->rss_usecnt--;
2081 
2082 			mutex_unlock(&wq->mutex);
2083 		}
2084 	}
2085 
2086 	return err;
2087 }
2088 
2089 static void bring_down_rss_rwqs(struct ib_rwq_ind_table *ind_tbl,
2090 				struct ib_udata *udata)
2091 {
2092 	int i;
2093 
2094 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2095 		struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2096 		struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2097 
2098 		mutex_lock(&wq->mutex);
2099 
2100 		if ((wq->rss_usecnt == 1) && (ibwq->state == IB_WQS_RDY))
2101 			if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET, udata))
2102 				pr_warn("failed to reverse WQN=%x\n",
2103 					ibwq->wq_num);
2104 		wq->rss_usecnt--;
2105 
2106 		mutex_unlock(&wq->mutex);
2107 	}
2108 }
2109 
2110 static void fill_qp_rss_context(struct mlx4_qp_context *context,
2111 				struct mlx4_ib_qp *qp)
2112 {
2113 	struct mlx4_rss_context *rss_context;
2114 
2115 	rss_context = (void *)context + offsetof(struct mlx4_qp_context,
2116 			pri_path) + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
2117 
2118 	rss_context->base_qpn = cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz);
2119 	rss_context->default_qpn =
2120 		cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz & 0xffffff);
2121 	if (qp->rss_ctx->flags & (MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6))
2122 		rss_context->base_qpn_udp = rss_context->default_qpn;
2123 	rss_context->flags = qp->rss_ctx->flags;
2124 	/* Currently support just toeplitz */
2125 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
2126 
2127 	memcpy(rss_context->rss_key, qp->rss_ctx->rss_key,
2128 	       MLX4_EN_RSS_KEY_SIZE);
2129 }
2130 
2131 static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type,
2132 			       const struct ib_qp_attr *attr, int attr_mask,
2133 			       enum ib_qp_state cur_state,
2134 			       enum ib_qp_state new_state,
2135 			       struct ib_udata *udata)
2136 {
2137 	struct ib_srq  *ibsrq;
2138 	const struct ib_gid_attr *gid_attr = NULL;
2139 	struct ib_rwq_ind_table *rwq_ind_tbl;
2140 	enum ib_qp_type qp_type;
2141 	struct mlx4_ib_dev *dev;
2142 	struct mlx4_ib_qp *qp;
2143 	struct mlx4_ib_pd *pd;
2144 	struct mlx4_ib_cq *send_cq, *recv_cq;
2145 	struct mlx4_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2146 		udata, struct mlx4_ib_ucontext, ibucontext);
2147 	struct mlx4_qp_context *context;
2148 	enum mlx4_qp_optpar optpar = 0;
2149 	int sqd_event;
2150 	int steer_qp = 0;
2151 	int err = -EINVAL;
2152 	int counter_index;
2153 
2154 	if (src_type == MLX4_IB_RWQ_SRC) {
2155 		struct ib_wq *ibwq;
2156 
2157 		ibwq	    = (struct ib_wq *)src;
2158 		ibsrq	    = NULL;
2159 		rwq_ind_tbl = NULL;
2160 		qp_type     = IB_QPT_RAW_PACKET;
2161 		qp	    = to_mqp((struct ib_qp *)ibwq);
2162 		dev	    = to_mdev(ibwq->device);
2163 		pd	    = to_mpd(ibwq->pd);
2164 	} else {
2165 		struct ib_qp *ibqp;
2166 
2167 		ibqp	    = (struct ib_qp *)src;
2168 		ibsrq	    = ibqp->srq;
2169 		rwq_ind_tbl = ibqp->rwq_ind_tbl;
2170 		qp_type     = ibqp->qp_type;
2171 		qp	    = to_mqp(ibqp);
2172 		dev	    = to_mdev(ibqp->device);
2173 		pd	    = get_pd(qp);
2174 	}
2175 
2176 	/* APM is not supported under RoCE */
2177 	if (attr_mask & IB_QP_ALT_PATH &&
2178 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2179 	    IB_LINK_LAYER_ETHERNET)
2180 		return -ENOTSUPP;
2181 
2182 	context = kzalloc(sizeof *context, GFP_KERNEL);
2183 	if (!context)
2184 		return -ENOMEM;
2185 
2186 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
2187 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
2188 
2189 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
2190 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2191 	else {
2192 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
2193 		switch (attr->path_mig_state) {
2194 		case IB_MIG_MIGRATED:
2195 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2196 			break;
2197 		case IB_MIG_REARM:
2198 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
2199 			break;
2200 		case IB_MIG_ARMED:
2201 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
2202 			break;
2203 		}
2204 	}
2205 
2206 	if (qp->inl_recv_sz)
2207 		context->param3 |= cpu_to_be32(1 << 25);
2208 
2209 	if (qp->flags & MLX4_IB_QP_SCATTER_FCS)
2210 		context->param3 |= cpu_to_be32(1 << 29);
2211 
2212 	if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI)
2213 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
2214 	else if (qp_type == IB_QPT_RAW_PACKET)
2215 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
2216 	else if (qp_type == IB_QPT_UD) {
2217 		if (qp->flags & MLX4_IB_QP_LSO)
2218 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
2219 					      ilog2(dev->dev->caps.max_gso_sz);
2220 		else
2221 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 13;
2222 	} else if (attr_mask & IB_QP_PATH_MTU) {
2223 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
2224 			pr_err("path MTU (%u) is invalid\n",
2225 			       attr->path_mtu);
2226 			goto out;
2227 		}
2228 		context->mtu_msgmax = (attr->path_mtu << 5) |
2229 			ilog2(dev->dev->caps.max_msg_sz);
2230 	}
2231 
2232 	if (!rwq_ind_tbl) { /* PRM RSS receive side should be left zeros */
2233 		if (qp->rq.wqe_cnt)
2234 			context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
2235 		context->rq_size_stride |= qp->rq.wqe_shift - 4;
2236 	}
2237 
2238 	if (qp->sq.wqe_cnt)
2239 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
2240 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
2241 
2242 	if (new_state == IB_QPS_RESET && qp->counter_index)
2243 		mlx4_ib_free_qp_counter(dev, qp);
2244 
2245 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2246 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
2247 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
2248 		if (qp_type == IB_QPT_RAW_PACKET)
2249 			context->param3 |= cpu_to_be32(1 << 30);
2250 	}
2251 
2252 	if (ucontext)
2253 		context->usr_page = cpu_to_be32(
2254 			mlx4_to_hw_uar_index(dev->dev, ucontext->uar.index));
2255 	else
2256 		context->usr_page = cpu_to_be32(
2257 			mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
2258 
2259 	if (attr_mask & IB_QP_DEST_QPN)
2260 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
2261 
2262 	if (attr_mask & IB_QP_PORT) {
2263 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
2264 		    !(attr_mask & IB_QP_AV)) {
2265 			mlx4_set_sched(&context->pri_path, attr->port_num);
2266 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
2267 		}
2268 	}
2269 
2270 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
2271 		err = create_qp_lb_counter(dev, qp);
2272 		if (err)
2273 			goto out;
2274 
2275 		counter_index =
2276 			dev->counters_table[qp->port - 1].default_counter;
2277 		if (qp->counter_index)
2278 			counter_index = qp->counter_index->index;
2279 
2280 		if (counter_index != -1) {
2281 			context->pri_path.counter_index = counter_index;
2282 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
2283 			if (qp->counter_index) {
2284 				context->pri_path.fl |=
2285 					MLX4_FL_ETH_SRC_CHECK_MC_LB;
2286 				context->pri_path.vlan_control |=
2287 					MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
2288 			}
2289 		} else
2290 			context->pri_path.counter_index =
2291 				MLX4_SINK_COUNTER_INDEX(dev->dev);
2292 
2293 		if (qp->flags & MLX4_IB_QP_NETIF) {
2294 			mlx4_ib_steer_qp_reg(dev, qp, 1);
2295 			steer_qp = 1;
2296 		}
2297 
2298 		if (qp_type == IB_QPT_GSI) {
2299 			enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
2300 				IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
2301 			u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
2302 
2303 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2304 		}
2305 	}
2306 
2307 	if (attr_mask & IB_QP_PKEY_INDEX) {
2308 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2309 			context->pri_path.disable_pkey_check = 0x40;
2310 		context->pri_path.pkey_index = attr->pkey_index;
2311 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
2312 	}
2313 
2314 	if (attr_mask & IB_QP_AV) {
2315 		u8 port_num = mlx4_is_bonded(dev->dev) ? 1 :
2316 			attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2317 		u16 vlan = 0xffff;
2318 		u8 smac[ETH_ALEN];
2319 		int is_eth =
2320 			rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
2321 			rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
2322 
2323 		if (is_eth) {
2324 			gid_attr = attr->ah_attr.grh.sgid_attr;
2325 			err = rdma_read_gid_l2_fields(gid_attr, &vlan,
2326 						      &smac[0]);
2327 			if (err)
2328 				goto out;
2329 		}
2330 
2331 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
2332 				  port_num, vlan, smac))
2333 			goto out;
2334 
2335 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
2336 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
2337 
2338 		if (is_eth &&
2339 		    (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
2340 			u8 qpc_roce_mode = gid_type_to_qpc(gid_attr->gid_type);
2341 
2342 			if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
2343 				err = -EINVAL;
2344 				goto out;
2345 			}
2346 			context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2347 		}
2348 
2349 	}
2350 
2351 	if (attr_mask & IB_QP_TIMEOUT) {
2352 		context->pri_path.ackto |= attr->timeout << 3;
2353 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
2354 	}
2355 
2356 	if (attr_mask & IB_QP_ALT_PATH) {
2357 		if (attr->alt_port_num == 0 ||
2358 		    attr->alt_port_num > dev->dev->caps.num_ports)
2359 			goto out;
2360 
2361 		if (attr->alt_pkey_index >=
2362 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
2363 			goto out;
2364 
2365 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
2366 				      &context->alt_path,
2367 				      attr->alt_port_num))
2368 			goto out;
2369 
2370 		context->alt_path.pkey_index = attr->alt_pkey_index;
2371 		context->alt_path.ackto = attr->alt_timeout << 3;
2372 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
2373 	}
2374 
2375 	context->pd = cpu_to_be32(pd->pdn);
2376 
2377 	if (!rwq_ind_tbl) {
2378 		context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
2379 		get_cqs(qp, src_type, &send_cq, &recv_cq);
2380 	} else { /* Set dummy CQs to be compatible with HV and PRM */
2381 		send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq);
2382 		recv_cq = send_cq;
2383 	}
2384 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
2385 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
2386 
2387 	/* Set "fast registration enabled" for all kernel QPs */
2388 	if (!ucontext)
2389 		context->params1 |= cpu_to_be32(1 << 11);
2390 
2391 	if (attr_mask & IB_QP_RNR_RETRY) {
2392 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2393 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
2394 	}
2395 
2396 	if (attr_mask & IB_QP_RETRY_CNT) {
2397 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2398 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
2399 	}
2400 
2401 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2402 		if (attr->max_rd_atomic)
2403 			context->params1 |=
2404 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2405 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
2406 	}
2407 
2408 	if (attr_mask & IB_QP_SQ_PSN)
2409 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
2410 
2411 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2412 		if (attr->max_dest_rd_atomic)
2413 			context->params2 |=
2414 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2415 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
2416 	}
2417 
2418 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
2419 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
2420 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
2421 	}
2422 
2423 	if (ibsrq)
2424 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
2425 
2426 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
2427 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2428 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
2429 	}
2430 	if (attr_mask & IB_QP_RQ_PSN)
2431 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2432 
2433 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
2434 	if (attr_mask & IB_QP_QKEY) {
2435 		if (qp->mlx4_ib_qp_type &
2436 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
2437 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2438 		else {
2439 			if (mlx4_is_mfunc(dev->dev) &&
2440 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
2441 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
2442 			    MLX4_RESERVED_QKEY_BASE) {
2443 				pr_err("Cannot use reserved QKEY"
2444 				       " 0x%x (range 0xffff0000..0xffffffff"
2445 				       " is reserved)\n", attr->qkey);
2446 				err = -EINVAL;
2447 				goto out;
2448 			}
2449 			context->qkey = cpu_to_be32(attr->qkey);
2450 		}
2451 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
2452 	}
2453 
2454 	if (ibsrq)
2455 		context->srqn = cpu_to_be32(1 << 24 |
2456 					    to_msrq(ibsrq)->msrq.srqn);
2457 
2458 	if (qp->rq.wqe_cnt &&
2459 	    cur_state == IB_QPS_RESET &&
2460 	    new_state == IB_QPS_INIT)
2461 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
2462 
2463 	if (cur_state == IB_QPS_INIT &&
2464 	    new_state == IB_QPS_RTR  &&
2465 	    (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI ||
2466 	     qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) {
2467 		context->pri_path.sched_queue = (qp->port - 1) << 6;
2468 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2469 		    qp->mlx4_ib_qp_type &
2470 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
2471 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
2472 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
2473 				context->pri_path.fl = 0x80;
2474 		} else {
2475 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2476 				context->pri_path.fl = 0x80;
2477 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
2478 		}
2479 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2480 		    IB_LINK_LAYER_ETHERNET) {
2481 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
2482 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
2483 				context->pri_path.feup = 1 << 7; /* don't fsm */
2484 			/* handle smac_index */
2485 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
2486 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
2487 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
2488 				err = handle_eth_ud_smac_index(dev, qp, context);
2489 				if (err) {
2490 					err = -EINVAL;
2491 					goto out;
2492 				}
2493 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
2494 					dev->qp1_proxy[qp->port - 1] = qp;
2495 			}
2496 		}
2497 	}
2498 
2499 	if (qp_type == IB_QPT_RAW_PACKET) {
2500 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
2501 					MLX4_IB_LINK_TYPE_ETH;
2502 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
2503 			/* set QP to receive both tunneled & non-tunneled packets */
2504 			if (!rwq_ind_tbl)
2505 				context->srqn = cpu_to_be32(7 << 28);
2506 		}
2507 	}
2508 
2509 	if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
2510 		int is_eth = rdma_port_get_link_layer(
2511 				&dev->ib_dev, qp->port) ==
2512 				IB_LINK_LAYER_ETHERNET;
2513 		if (is_eth) {
2514 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
2515 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
2516 		}
2517 	}
2518 
2519 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
2520 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2521 		sqd_event = 1;
2522 	else
2523 		sqd_event = 0;
2524 
2525 	if (!ucontext &&
2526 	    cur_state == IB_QPS_RESET &&
2527 	    new_state == IB_QPS_INIT)
2528 		context->rlkey_roce_mode |= (1 << 4);
2529 
2530 	/*
2531 	 * Before passing a kernel QP to the HW, make sure that the
2532 	 * ownership bits of the send queue are set and the SQ
2533 	 * headroom is stamped so that the hardware doesn't start
2534 	 * processing stale work requests.
2535 	 */
2536 	if (!ucontext &&
2537 	    cur_state == IB_QPS_RESET &&
2538 	    new_state == IB_QPS_INIT) {
2539 		struct mlx4_wqe_ctrl_seg *ctrl;
2540 		int i;
2541 
2542 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
2543 			ctrl = get_send_wqe(qp, i);
2544 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
2545 			ctrl->qpn_vlan.fence_size =
2546 				1 << (qp->sq.wqe_shift - 4);
2547 			stamp_send_wqe(qp, i);
2548 		}
2549 	}
2550 
2551 	if (rwq_ind_tbl	&&
2552 	    cur_state == IB_QPS_RESET &&
2553 	    new_state == IB_QPS_INIT) {
2554 		fill_qp_rss_context(context, qp);
2555 		context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET);
2556 	}
2557 
2558 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2559 			     to_mlx4_state(new_state), context, optpar,
2560 			     sqd_event, &qp->mqp);
2561 	if (err)
2562 		goto out;
2563 
2564 	qp->state = new_state;
2565 
2566 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2567 		qp->atomic_rd_en = attr->qp_access_flags;
2568 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2569 		qp->resp_depth = attr->max_dest_rd_atomic;
2570 	if (attr_mask & IB_QP_PORT) {
2571 		qp->port = attr->port_num;
2572 		update_mcg_macs(dev, qp);
2573 	}
2574 	if (attr_mask & IB_QP_ALT_PATH)
2575 		qp->alt_port = attr->alt_port_num;
2576 
2577 	if (is_sqp(dev, qp))
2578 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2579 
2580 	/*
2581 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
2582 	 * QP0 to RESET or ERROR, bring the link back down.
2583 	 */
2584 	if (is_qp0(dev, qp)) {
2585 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2586 			if (mlx4_INIT_PORT(dev->dev, qp->port))
2587 				pr_warn("INIT_PORT failed for port %d\n",
2588 				       qp->port);
2589 
2590 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2591 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2592 			mlx4_CLOSE_PORT(dev->dev, qp->port);
2593 	}
2594 
2595 	/*
2596 	 * If we moved a kernel QP to RESET, clean up all old CQ
2597 	 * entries and reinitialize the QP.
2598 	 */
2599 	if (new_state == IB_QPS_RESET) {
2600 		if (!ucontext) {
2601 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2602 					 ibsrq ? to_msrq(ibsrq) : NULL);
2603 			if (send_cq != recv_cq)
2604 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2605 
2606 			qp->rq.head = 0;
2607 			qp->rq.tail = 0;
2608 			qp->sq.head = 0;
2609 			qp->sq.tail = 0;
2610 			qp->sq_next_wqe = 0;
2611 			if (qp->rq.wqe_cnt)
2612 				*qp->db.db  = 0;
2613 
2614 			if (qp->flags & MLX4_IB_QP_NETIF)
2615 				mlx4_ib_steer_qp_reg(dev, qp, 0);
2616 		}
2617 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2618 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2619 			qp->pri.smac = 0;
2620 			qp->pri.smac_port = 0;
2621 		}
2622 		if (qp->alt.smac) {
2623 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2624 			qp->alt.smac = 0;
2625 		}
2626 		if (qp->pri.vid < 0x1000) {
2627 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2628 			qp->pri.vid = 0xFFFF;
2629 			qp->pri.candidate_vid = 0xFFFF;
2630 			qp->pri.update_vid = 0;
2631 		}
2632 
2633 		if (qp->alt.vid < 0x1000) {
2634 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2635 			qp->alt.vid = 0xFFFF;
2636 			qp->alt.candidate_vid = 0xFFFF;
2637 			qp->alt.update_vid = 0;
2638 		}
2639 	}
2640 out:
2641 	if (err && qp->counter_index)
2642 		mlx4_ib_free_qp_counter(dev, qp);
2643 	if (err && steer_qp)
2644 		mlx4_ib_steer_qp_reg(dev, qp, 0);
2645 	kfree(context);
2646 	if (qp->pri.candidate_smac ||
2647 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2648 		if (err) {
2649 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2650 		} else {
2651 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2652 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2653 			qp->pri.smac = qp->pri.candidate_smac;
2654 			qp->pri.smac_index = qp->pri.candidate_smac_index;
2655 			qp->pri.smac_port = qp->pri.candidate_smac_port;
2656 		}
2657 		qp->pri.candidate_smac = 0;
2658 		qp->pri.candidate_smac_index = 0;
2659 		qp->pri.candidate_smac_port = 0;
2660 	}
2661 	if (qp->alt.candidate_smac) {
2662 		if (err) {
2663 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2664 		} else {
2665 			if (qp->alt.smac)
2666 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2667 			qp->alt.smac = qp->alt.candidate_smac;
2668 			qp->alt.smac_index = qp->alt.candidate_smac_index;
2669 			qp->alt.smac_port = qp->alt.candidate_smac_port;
2670 		}
2671 		qp->alt.candidate_smac = 0;
2672 		qp->alt.candidate_smac_index = 0;
2673 		qp->alt.candidate_smac_port = 0;
2674 	}
2675 
2676 	if (qp->pri.update_vid) {
2677 		if (err) {
2678 			if (qp->pri.candidate_vid < 0x1000)
2679 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2680 						     qp->pri.candidate_vid);
2681 		} else {
2682 			if (qp->pri.vid < 0x1000)
2683 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2684 						     qp->pri.vid);
2685 			qp->pri.vid = qp->pri.candidate_vid;
2686 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2687 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
2688 		}
2689 		qp->pri.candidate_vid = 0xFFFF;
2690 		qp->pri.update_vid = 0;
2691 	}
2692 
2693 	if (qp->alt.update_vid) {
2694 		if (err) {
2695 			if (qp->alt.candidate_vid < 0x1000)
2696 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2697 						     qp->alt.candidate_vid);
2698 		} else {
2699 			if (qp->alt.vid < 0x1000)
2700 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2701 						     qp->alt.vid);
2702 			qp->alt.vid = qp->alt.candidate_vid;
2703 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2704 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
2705 		}
2706 		qp->alt.candidate_vid = 0xFFFF;
2707 		qp->alt.update_vid = 0;
2708 	}
2709 
2710 	return err;
2711 }
2712 
2713 enum {
2714 	MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK = (IB_QP_STATE	|
2715 					      IB_QP_PORT),
2716 };
2717 
2718 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2719 			      int attr_mask, struct ib_udata *udata)
2720 {
2721 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2722 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2723 	enum ib_qp_state cur_state, new_state;
2724 	int err = -EINVAL;
2725 	mutex_lock(&qp->mutex);
2726 
2727 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2728 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2729 
2730 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2731 				attr_mask)) {
2732 		pr_debug("qpn 0x%x: invalid attribute mask specified "
2733 			 "for transition %d to %d. qp_type %d,"
2734 			 " attr_mask 0x%x\n",
2735 			 ibqp->qp_num, cur_state, new_state,
2736 			 ibqp->qp_type, attr_mask);
2737 		goto out;
2738 	}
2739 
2740 	if (ibqp->rwq_ind_tbl) {
2741 		if (!(((cur_state == IB_QPS_RESET) &&
2742 		       (new_state == IB_QPS_INIT)) ||
2743 		      ((cur_state == IB_QPS_INIT)  &&
2744 		       (new_state == IB_QPS_RTR)))) {
2745 			pr_debug("qpn 0x%x: RSS QP unsupported transition %d to %d\n",
2746 				 ibqp->qp_num, cur_state, new_state);
2747 
2748 			err = -EOPNOTSUPP;
2749 			goto out;
2750 		}
2751 
2752 		if (attr_mask & ~MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK) {
2753 			pr_debug("qpn 0x%x: RSS QP unsupported attribute mask 0x%x for transition %d to %d\n",
2754 				 ibqp->qp_num, attr_mask, cur_state, new_state);
2755 
2756 			err = -EOPNOTSUPP;
2757 			goto out;
2758 		}
2759 	}
2760 
2761 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2762 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2763 			if ((ibqp->qp_type == IB_QPT_RC) ||
2764 			    (ibqp->qp_type == IB_QPT_UD) ||
2765 			    (ibqp->qp_type == IB_QPT_UC) ||
2766 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2767 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
2768 				attr->port_num = mlx4_ib_bond_next_port(dev);
2769 			}
2770 		} else {
2771 			/* no sense in changing port_num
2772 			 * when ports are bonded */
2773 			attr_mask &= ~IB_QP_PORT;
2774 		}
2775 	}
2776 
2777 	if ((attr_mask & IB_QP_PORT) &&
2778 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2779 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
2780 			 "for transition %d to %d. qp_type %d\n",
2781 			 ibqp->qp_num, attr->port_num, cur_state,
2782 			 new_state, ibqp->qp_type);
2783 		goto out;
2784 	}
2785 
2786 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2787 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2788 	     IB_LINK_LAYER_ETHERNET))
2789 		goto out;
2790 
2791 	if (attr_mask & IB_QP_PKEY_INDEX) {
2792 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2793 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2794 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2795 				 "for transition %d to %d. qp_type %d\n",
2796 				 ibqp->qp_num, attr->pkey_index, cur_state,
2797 				 new_state, ibqp->qp_type);
2798 			goto out;
2799 		}
2800 	}
2801 
2802 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2803 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2804 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2805 			 "Transition %d to %d. qp_type %d\n",
2806 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2807 			 new_state, ibqp->qp_type);
2808 		goto out;
2809 	}
2810 
2811 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2812 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2813 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2814 			 "Transition %d to %d. qp_type %d\n",
2815 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2816 			 new_state, ibqp->qp_type);
2817 		goto out;
2818 	}
2819 
2820 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2821 		err = 0;
2822 		goto out;
2823 	}
2824 
2825 	if (ibqp->rwq_ind_tbl && (new_state == IB_QPS_INIT)) {
2826 		err = bringup_rss_rwqs(ibqp->rwq_ind_tbl, attr->port_num,
2827 				       udata);
2828 		if (err)
2829 			goto out;
2830 	}
2831 
2832 	err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask,
2833 				  cur_state, new_state, udata);
2834 
2835 	if (ibqp->rwq_ind_tbl && err)
2836 		bring_down_rss_rwqs(ibqp->rwq_ind_tbl, udata);
2837 
2838 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2839 		attr->port_num = 1;
2840 
2841 out:
2842 	mutex_unlock(&qp->mutex);
2843 	return err;
2844 }
2845 
2846 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2847 		      int attr_mask, struct ib_udata *udata)
2848 {
2849 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2850 	int ret;
2851 
2852 	ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2853 
2854 	if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2855 		struct mlx4_ib_sqp *sqp = to_msqp(mqp);
2856 		int err = 0;
2857 
2858 		if (sqp->roce_v2_gsi)
2859 			err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2860 		if (err)
2861 			pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2862 			       err);
2863 	}
2864 	return ret;
2865 }
2866 
2867 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2868 {
2869 	int i;
2870 	for (i = 0; i < dev->caps.num_ports; i++) {
2871 		if (qpn == dev->caps.spec_qps[i].qp0_proxy ||
2872 		    qpn == dev->caps.spec_qps[i].qp0_tunnel) {
2873 			*qkey = dev->caps.spec_qps[i].qp0_qkey;
2874 			return 0;
2875 		}
2876 	}
2877 	return -EINVAL;
2878 }
2879 
2880 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2881 				  const struct ib_ud_wr *wr,
2882 				  void *wqe, unsigned *mlx_seg_len)
2883 {
2884 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2885 	struct ib_device *ib_dev = &mdev->ib_dev;
2886 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2887 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2888 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
2889 	u16 pkey;
2890 	u32 qkey;
2891 	int send_size;
2892 	int header_size;
2893 	int spc;
2894 	int err;
2895 	int i;
2896 
2897 	if (wr->wr.opcode != IB_WR_SEND)
2898 		return -EINVAL;
2899 
2900 	send_size = 0;
2901 
2902 	for (i = 0; i < wr->wr.num_sge; ++i)
2903 		send_size += wr->wr.sg_list[i].length;
2904 
2905 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2906 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2907 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2908 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2909 
2910 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2911 
2912 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2913 		sqp->ud_header.lrh.service_level =
2914 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2915 		sqp->ud_header.lrh.destination_lid =
2916 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2917 		sqp->ud_header.lrh.source_lid =
2918 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2919 	}
2920 
2921 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2922 
2923 	/* force loopback */
2924 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2925 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2926 
2927 	sqp->ud_header.lrh.virtual_lane    = 0;
2928 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2929 	err = ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2930 	if (err)
2931 		return err;
2932 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2933 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2934 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2935 	else
2936 		sqp->ud_header.bth.destination_qpn =
2937 			cpu_to_be32(mdev->dev->caps.spec_qps[sqp->qp.port - 1].qp0_tunnel);
2938 
2939 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2940 	if (mlx4_is_master(mdev->dev)) {
2941 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2942 			return -EINVAL;
2943 	} else {
2944 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2945 			return -EINVAL;
2946 	}
2947 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2948 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2949 
2950 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2951 	sqp->ud_header.immediate_present = 0;
2952 
2953 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2954 
2955 	/*
2956 	 * Inline data segments may not cross a 64 byte boundary.  If
2957 	 * our UD header is bigger than the space available up to the
2958 	 * next 64 byte boundary in the WQE, use two inline data
2959 	 * segments to hold the UD header.
2960 	 */
2961 	spc = MLX4_INLINE_ALIGN -
2962 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2963 	if (header_size <= spc) {
2964 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2965 		memcpy(inl + 1, sqp->header_buf, header_size);
2966 		i = 1;
2967 	} else {
2968 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2969 		memcpy(inl + 1, sqp->header_buf, spc);
2970 
2971 		inl = (void *) (inl + 1) + spc;
2972 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2973 		/*
2974 		 * Need a barrier here to make sure all the data is
2975 		 * visible before the byte_count field is set.
2976 		 * Otherwise the HCA prefetcher could grab the 64-byte
2977 		 * chunk with this inline segment and get a valid (!=
2978 		 * 0xffffffff) byte count but stale data, and end up
2979 		 * generating a packet with bad headers.
2980 		 *
2981 		 * The first inline segment's byte_count field doesn't
2982 		 * need a barrier, because it comes after a
2983 		 * control/MLX segment and therefore is at an offset
2984 		 * of 16 mod 64.
2985 		 */
2986 		wmb();
2987 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2988 		i = 2;
2989 	}
2990 
2991 	*mlx_seg_len =
2992 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2993 	return 0;
2994 }
2995 
2996 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
2997 {
2998 	union sl2vl_tbl_to_u64 tmp_vltab;
2999 	u8 vl;
3000 
3001 	if (sl > 15)
3002 		return 0xf;
3003 	tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
3004 	vl = tmp_vltab.sl8[sl >> 1];
3005 	if (sl & 1)
3006 		vl &= 0x0f;
3007 	else
3008 		vl >>= 4;
3009 	return vl;
3010 }
3011 
3012 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
3013 				int index, union ib_gid *gid,
3014 				enum ib_gid_type *gid_type)
3015 {
3016 	struct mlx4_ib_iboe *iboe = &ibdev->iboe;
3017 	struct mlx4_port_gid_table *port_gid_table;
3018 	unsigned long flags;
3019 
3020 	port_gid_table = &iboe->gids[port_num - 1];
3021 	spin_lock_irqsave(&iboe->lock, flags);
3022 	memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
3023 	*gid_type = port_gid_table->gids[index].gid_type;
3024 	spin_unlock_irqrestore(&iboe->lock, flags);
3025 	if (rdma_is_zero_gid(gid))
3026 		return -ENOENT;
3027 
3028 	return 0;
3029 }
3030 
3031 #define MLX4_ROCEV2_QP1_SPORT 0xC000
3032 static int build_mlx_header(struct mlx4_ib_sqp *sqp, const struct ib_ud_wr *wr,
3033 			    void *wqe, unsigned *mlx_seg_len)
3034 {
3035 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
3036 	struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
3037 	struct mlx4_wqe_mlx_seg *mlx = wqe;
3038 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
3039 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
3040 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3041 	union ib_gid sgid;
3042 	u16 pkey;
3043 	int send_size;
3044 	int header_size;
3045 	int spc;
3046 	int i;
3047 	int err = 0;
3048 	u16 vlan = 0xffff;
3049 	bool is_eth;
3050 	bool is_vlan = false;
3051 	bool is_grh;
3052 	bool is_udp = false;
3053 	int ip_version = 0;
3054 
3055 	send_size = 0;
3056 	for (i = 0; i < wr->wr.num_sge; ++i)
3057 		send_size += wr->wr.sg_list[i].length;
3058 
3059 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
3060 	is_grh = mlx4_ib_ah_grh_present(ah);
3061 	if (is_eth) {
3062 		enum ib_gid_type gid_type;
3063 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3064 			/* When multi-function is enabled, the ib_core gid
3065 			 * indexes don't necessarily match the hw ones, so
3066 			 * we must use our own cache */
3067 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
3068 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
3069 							   ah->av.ib.gid_index, &sgid.raw[0]);
3070 			if (err)
3071 				return err;
3072 		} else  {
3073 			err = fill_gid_by_hw_index(ibdev, sqp->qp.port,
3074 					    ah->av.ib.gid_index,
3075 					    &sgid, &gid_type);
3076 			if (!err) {
3077 				is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
3078 				if (is_udp) {
3079 					if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
3080 						ip_version = 4;
3081 					else
3082 						ip_version = 6;
3083 					is_grh = false;
3084 				}
3085 			} else {
3086 				return err;
3087 			}
3088 		}
3089 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
3090 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
3091 			is_vlan = true;
3092 		}
3093 	}
3094 	err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
3095 			  ip_version, is_udp, 0, &sqp->ud_header);
3096 	if (err)
3097 		return err;
3098 
3099 	if (!is_eth) {
3100 		sqp->ud_header.lrh.service_level =
3101 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
3102 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
3103 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
3104 	}
3105 
3106 	if (is_grh || (ip_version == 6)) {
3107 		sqp->ud_header.grh.traffic_class =
3108 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3109 		sqp->ud_header.grh.flow_label    =
3110 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
3111 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
3112 		if (is_eth) {
3113 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
3114 		} else {
3115 			if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3116 				/* When multi-function is enabled, the ib_core gid
3117 				 * indexes don't necessarily match the hw ones, so
3118 				 * we must use our own cache
3119 				 */
3120 				sqp->ud_header.grh.source_gid.global.subnet_prefix =
3121 					cpu_to_be64(atomic64_read(&(to_mdev(ib_dev)->sriov.
3122 								    demux[sqp->qp.port - 1].
3123 								    subnet_prefix)));
3124 				sqp->ud_header.grh.source_gid.global.interface_id =
3125 					to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
3126 						       guid_cache[ah->av.ib.gid_index];
3127 			} else {
3128 				sqp->ud_header.grh.source_gid =
3129 					ah->ibah.sgid_attr->gid;
3130 			}
3131 		}
3132 		memcpy(sqp->ud_header.grh.destination_gid.raw,
3133 		       ah->av.ib.dgid, 16);
3134 	}
3135 
3136 	if (ip_version == 4) {
3137 		sqp->ud_header.ip4.tos =
3138 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3139 		sqp->ud_header.ip4.id = 0;
3140 		sqp->ud_header.ip4.frag_off = htons(IP_DF);
3141 		sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
3142 
3143 		memcpy(&sqp->ud_header.ip4.saddr,
3144 		       sgid.raw + 12, 4);
3145 		memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
3146 		sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
3147 	}
3148 
3149 	if (is_udp) {
3150 		sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
3151 		sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
3152 		sqp->ud_header.udp.csum = 0;
3153 	}
3154 
3155 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
3156 
3157 	if (!is_eth) {
3158 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
3159 					  (sqp->ud_header.lrh.destination_lid ==
3160 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
3161 					  (sqp->ud_header.lrh.service_level << 8));
3162 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
3163 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
3164 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
3165 	}
3166 
3167 	switch (wr->wr.opcode) {
3168 	case IB_WR_SEND:
3169 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
3170 		sqp->ud_header.immediate_present = 0;
3171 		break;
3172 	case IB_WR_SEND_WITH_IMM:
3173 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
3174 		sqp->ud_header.immediate_present = 1;
3175 		sqp->ud_header.immediate_data    = wr->wr.ex.imm_data;
3176 		break;
3177 	default:
3178 		return -EINVAL;
3179 	}
3180 
3181 	if (is_eth) {
3182 		struct in6_addr in6;
3183 		u16 ether_type;
3184 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
3185 
3186 		ether_type = (!is_udp) ? ETH_P_IBOE:
3187 			(ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
3188 
3189 		mlx->sched_prio = cpu_to_be16(pcp);
3190 
3191 		ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
3192 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
3193 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
3194 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
3195 		memcpy(&in6, sgid.raw, sizeof(in6));
3196 
3197 
3198 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
3199 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
3200 		if (!is_vlan) {
3201 			sqp->ud_header.eth.type = cpu_to_be16(ether_type);
3202 		} else {
3203 			sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
3204 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
3205 		}
3206 	} else {
3207 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 :
3208 							sl_to_vl(to_mdev(ib_dev),
3209 								 sqp->ud_header.lrh.service_level,
3210 								 sqp->qp.port);
3211 		if (sqp->qp.ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
3212 			return -EINVAL;
3213 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
3214 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
3215 	}
3216 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
3217 	if (!sqp->qp.ibqp.qp_num)
3218 		err = ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index,
3219 					 &pkey);
3220 	else
3221 		err = ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index,
3222 					 &pkey);
3223 	if (err)
3224 		return err;
3225 
3226 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
3227 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
3228 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
3229 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
3230 					       sqp->qkey : wr->remote_qkey);
3231 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
3232 
3233 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3234 
3235 	if (0) {
3236 		pr_err("built UD header of size %d:\n", header_size);
3237 		for (i = 0; i < header_size / 4; ++i) {
3238 			if (i % 8 == 0)
3239 				pr_err("  [%02x] ", i * 4);
3240 			pr_cont(" %08x",
3241 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
3242 			if ((i + 1) % 8 == 0)
3243 				pr_cont("\n");
3244 		}
3245 		pr_err("\n");
3246 	}
3247 
3248 	/*
3249 	 * Inline data segments may not cross a 64 byte boundary.  If
3250 	 * our UD header is bigger than the space available up to the
3251 	 * next 64 byte boundary in the WQE, use two inline data
3252 	 * segments to hold the UD header.
3253 	 */
3254 	spc = MLX4_INLINE_ALIGN -
3255 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3256 	if (header_size <= spc) {
3257 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3258 		memcpy(inl + 1, sqp->header_buf, header_size);
3259 		i = 1;
3260 	} else {
3261 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3262 		memcpy(inl + 1, sqp->header_buf, spc);
3263 
3264 		inl = (void *) (inl + 1) + spc;
3265 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3266 		/*
3267 		 * Need a barrier here to make sure all the data is
3268 		 * visible before the byte_count field is set.
3269 		 * Otherwise the HCA prefetcher could grab the 64-byte
3270 		 * chunk with this inline segment and get a valid (!=
3271 		 * 0xffffffff) byte count but stale data, and end up
3272 		 * generating a packet with bad headers.
3273 		 *
3274 		 * The first inline segment's byte_count field doesn't
3275 		 * need a barrier, because it comes after a
3276 		 * control/MLX segment and therefore is at an offset
3277 		 * of 16 mod 64.
3278 		 */
3279 		wmb();
3280 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3281 		i = 2;
3282 	}
3283 
3284 	*mlx_seg_len =
3285 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3286 	return 0;
3287 }
3288 
3289 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3290 {
3291 	unsigned cur;
3292 	struct mlx4_ib_cq *cq;
3293 
3294 	cur = wq->head - wq->tail;
3295 	if (likely(cur + nreq < wq->max_post))
3296 		return 0;
3297 
3298 	cq = to_mcq(ib_cq);
3299 	spin_lock(&cq->lock);
3300 	cur = wq->head - wq->tail;
3301 	spin_unlock(&cq->lock);
3302 
3303 	return cur + nreq >= wq->max_post;
3304 }
3305 
3306 static __be32 convert_access(int acc)
3307 {
3308 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
3309 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
3310 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
3311 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
3312 	       (acc & IB_ACCESS_REMOTE_READ   ?
3313 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
3314 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
3315 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
3316 }
3317 
3318 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
3319 			const struct ib_reg_wr *wr)
3320 {
3321 	struct mlx4_ib_mr *mr = to_mmr(wr->mr);
3322 
3323 	fseg->flags		= convert_access(wr->access);
3324 	fseg->mem_key		= cpu_to_be32(wr->key);
3325 	fseg->buf_list		= cpu_to_be64(mr->page_map);
3326 	fseg->start_addr	= cpu_to_be64(mr->ibmr.iova);
3327 	fseg->reg_len		= cpu_to_be64(mr->ibmr.length);
3328 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
3329 	fseg->page_size		= cpu_to_be32(ilog2(mr->ibmr.page_size));
3330 	fseg->reserved[0]	= 0;
3331 	fseg->reserved[1]	= 0;
3332 }
3333 
3334 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
3335 {
3336 	memset(iseg, 0, sizeof(*iseg));
3337 	iseg->mem_key = cpu_to_be32(rkey);
3338 }
3339 
3340 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
3341 					  u64 remote_addr, u32 rkey)
3342 {
3343 	rseg->raddr    = cpu_to_be64(remote_addr);
3344 	rseg->rkey     = cpu_to_be32(rkey);
3345 	rseg->reserved = 0;
3346 }
3347 
3348 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
3349 			   const struct ib_atomic_wr *wr)
3350 {
3351 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
3352 		aseg->swap_add = cpu_to_be64(wr->swap);
3353 		aseg->compare  = cpu_to_be64(wr->compare_add);
3354 	} else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
3355 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3356 		aseg->compare  = cpu_to_be64(wr->compare_add_mask);
3357 	} else {
3358 		aseg->swap_add = cpu_to_be64(wr->compare_add);
3359 		aseg->compare  = 0;
3360 	}
3361 
3362 }
3363 
3364 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
3365 				  const struct ib_atomic_wr *wr)
3366 {
3367 	aseg->swap_add		= cpu_to_be64(wr->swap);
3368 	aseg->swap_add_mask	= cpu_to_be64(wr->swap_mask);
3369 	aseg->compare		= cpu_to_be64(wr->compare_add);
3370 	aseg->compare_mask	= cpu_to_be64(wr->compare_add_mask);
3371 }
3372 
3373 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
3374 			     const struct ib_ud_wr *wr)
3375 {
3376 	memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
3377 	dseg->dqpn = cpu_to_be32(wr->remote_qpn);
3378 	dseg->qkey = cpu_to_be32(wr->remote_qkey);
3379 	dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
3380 	memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
3381 }
3382 
3383 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
3384 				    struct mlx4_wqe_datagram_seg *dseg,
3385 				    const struct ib_ud_wr *wr,
3386 				    enum mlx4_ib_qp_type qpt)
3387 {
3388 	union mlx4_ext_av *av = &to_mah(wr->ah)->av;
3389 	struct mlx4_av sqp_av = {0};
3390 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
3391 
3392 	/* force loopback */
3393 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
3394 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
3395 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
3396 			cpu_to_be32(0xf0000000);
3397 
3398 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
3399 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
3400 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp1_tunnel);
3401 	else
3402 		dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp0_tunnel);
3403 	/* Use QKEY from the QP context, which is set by master */
3404 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
3405 }
3406 
3407 static void build_tunnel_header(const struct ib_ud_wr *wr, void *wqe,
3408 				unsigned *mlx_seg_len)
3409 {
3410 	struct mlx4_wqe_inline_seg *inl = wqe;
3411 	struct mlx4_ib_tunnel_header hdr;
3412 	struct mlx4_ib_ah *ah = to_mah(wr->ah);
3413 	int spc;
3414 	int i;
3415 
3416 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
3417 	hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
3418 	hdr.pkey_index = cpu_to_be16(wr->pkey_index);
3419 	hdr.qkey = cpu_to_be32(wr->remote_qkey);
3420 	memcpy(hdr.mac, ah->av.eth.mac, 6);
3421 	hdr.vlan = ah->av.eth.vlan;
3422 
3423 	spc = MLX4_INLINE_ALIGN -
3424 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3425 	if (sizeof (hdr) <= spc) {
3426 		memcpy(inl + 1, &hdr, sizeof (hdr));
3427 		wmb();
3428 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
3429 		i = 1;
3430 	} else {
3431 		memcpy(inl + 1, &hdr, spc);
3432 		wmb();
3433 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
3434 
3435 		inl = (void *) (inl + 1) + spc;
3436 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
3437 		wmb();
3438 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
3439 		i = 2;
3440 	}
3441 
3442 	*mlx_seg_len =
3443 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
3444 }
3445 
3446 static void set_mlx_icrc_seg(void *dseg)
3447 {
3448 	u32 *t = dseg;
3449 	struct mlx4_wqe_inline_seg *iseg = dseg;
3450 
3451 	t[1] = 0;
3452 
3453 	/*
3454 	 * Need a barrier here before writing the byte_count field to
3455 	 * make sure that all the data is visible before the
3456 	 * byte_count field is set.  Otherwise, if the segment begins
3457 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3458 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3459 	 * stale data, and end up sending the wrong data.
3460 	 */
3461 	wmb();
3462 
3463 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
3464 }
3465 
3466 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3467 {
3468 	dseg->lkey       = cpu_to_be32(sg->lkey);
3469 	dseg->addr       = cpu_to_be64(sg->addr);
3470 
3471 	/*
3472 	 * Need a barrier here before writing the byte_count field to
3473 	 * make sure that all the data is visible before the
3474 	 * byte_count field is set.  Otherwise, if the segment begins
3475 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
3476 	 * chunk and get a valid (!= * 0xffffffff) byte count but
3477 	 * stale data, and end up sending the wrong data.
3478 	 */
3479 	wmb();
3480 
3481 	dseg->byte_count = cpu_to_be32(sg->length);
3482 }
3483 
3484 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3485 {
3486 	dseg->byte_count = cpu_to_be32(sg->length);
3487 	dseg->lkey       = cpu_to_be32(sg->lkey);
3488 	dseg->addr       = cpu_to_be64(sg->addr);
3489 }
3490 
3491 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe,
3492 			 const struct ib_ud_wr *wr, struct mlx4_ib_qp *qp,
3493 			 unsigned *lso_seg_len, __be32 *lso_hdr_sz, __be32 *blh)
3494 {
3495 	unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
3496 
3497 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
3498 		*blh = cpu_to_be32(1 << 6);
3499 
3500 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
3501 		     wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
3502 		return -EINVAL;
3503 
3504 	memcpy(wqe->header, wr->header, wr->hlen);
3505 
3506 	*lso_hdr_sz  = cpu_to_be32(wr->mss << 16 | wr->hlen);
3507 	*lso_seg_len = halign;
3508 	return 0;
3509 }
3510 
3511 static __be32 send_ieth(const struct ib_send_wr *wr)
3512 {
3513 	switch (wr->opcode) {
3514 	case IB_WR_SEND_WITH_IMM:
3515 	case IB_WR_RDMA_WRITE_WITH_IMM:
3516 		return wr->ex.imm_data;
3517 
3518 	case IB_WR_SEND_WITH_INV:
3519 		return cpu_to_be32(wr->ex.invalidate_rkey);
3520 
3521 	default:
3522 		return 0;
3523 	}
3524 }
3525 
3526 static void add_zero_len_inline(void *wqe)
3527 {
3528 	struct mlx4_wqe_inline_seg *inl = wqe;
3529 	memset(wqe, 0, 16);
3530 	inl->byte_count = cpu_to_be32(1 << 31);
3531 }
3532 
3533 static int _mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3534 			      const struct ib_send_wr **bad_wr, bool drain)
3535 {
3536 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3537 	void *wqe;
3538 	struct mlx4_wqe_ctrl_seg *ctrl;
3539 	struct mlx4_wqe_data_seg *dseg;
3540 	unsigned long flags;
3541 	int nreq;
3542 	int err = 0;
3543 	unsigned ind;
3544 	int size;
3545 	unsigned seglen;
3546 	__be32 dummy;
3547 	__be32 *lso_wqe;
3548 	__be32 lso_hdr_sz;
3549 	__be32 blh;
3550 	int i;
3551 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3552 
3553 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
3554 		struct mlx4_ib_sqp *sqp = to_msqp(qp);
3555 
3556 		if (sqp->roce_v2_gsi) {
3557 			struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
3558 			enum ib_gid_type gid_type;
3559 			union ib_gid gid;
3560 
3561 			if (!fill_gid_by_hw_index(mdev, sqp->qp.port,
3562 					   ah->av.ib.gid_index,
3563 					   &gid, &gid_type))
3564 				qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
3565 						to_mqp(sqp->roce_v2_gsi) : qp;
3566 			else
3567 				pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
3568 				       ah->av.ib.gid_index);
3569 		}
3570 	}
3571 
3572 	spin_lock_irqsave(&qp->sq.lock, flags);
3573 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3574 	    !drain) {
3575 		err = -EIO;
3576 		*bad_wr = wr;
3577 		nreq = 0;
3578 		goto out;
3579 	}
3580 
3581 	ind = qp->sq_next_wqe;
3582 
3583 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3584 		lso_wqe = &dummy;
3585 		blh = 0;
3586 
3587 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
3588 			err = -ENOMEM;
3589 			*bad_wr = wr;
3590 			goto out;
3591 		}
3592 
3593 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3594 			err = -EINVAL;
3595 			*bad_wr = wr;
3596 			goto out;
3597 		}
3598 
3599 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3600 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3601 
3602 		ctrl->srcrb_flags =
3603 			(wr->send_flags & IB_SEND_SIGNALED ?
3604 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3605 			(wr->send_flags & IB_SEND_SOLICITED ?
3606 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3607 			((wr->send_flags & IB_SEND_IP_CSUM) ?
3608 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3609 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3610 			qp->sq_signal_bits;
3611 
3612 		ctrl->imm = send_ieth(wr);
3613 
3614 		wqe += sizeof *ctrl;
3615 		size = sizeof *ctrl / 16;
3616 
3617 		switch (qp->mlx4_ib_qp_type) {
3618 		case MLX4_IB_QPT_RC:
3619 		case MLX4_IB_QPT_UC:
3620 			switch (wr->opcode) {
3621 			case IB_WR_ATOMIC_CMP_AND_SWP:
3622 			case IB_WR_ATOMIC_FETCH_AND_ADD:
3623 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3624 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3625 					      atomic_wr(wr)->rkey);
3626 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3627 
3628 				set_atomic_seg(wqe, atomic_wr(wr));
3629 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
3630 
3631 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3632 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3633 
3634 				break;
3635 
3636 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3637 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3638 					      atomic_wr(wr)->rkey);
3639 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3640 
3641 				set_masked_atomic_seg(wqe, atomic_wr(wr));
3642 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
3643 
3644 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
3645 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3646 
3647 				break;
3648 
3649 			case IB_WR_RDMA_READ:
3650 			case IB_WR_RDMA_WRITE:
3651 			case IB_WR_RDMA_WRITE_WITH_IMM:
3652 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3653 					      rdma_wr(wr)->rkey);
3654 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
3655 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3656 				break;
3657 
3658 			case IB_WR_LOCAL_INV:
3659 				ctrl->srcrb_flags |=
3660 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3661 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3662 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
3663 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3664 				break;
3665 
3666 			case IB_WR_REG_MR:
3667 				ctrl->srcrb_flags |=
3668 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3669 				set_reg_seg(wqe, reg_wr(wr));
3670 				wqe  += sizeof(struct mlx4_wqe_fmr_seg);
3671 				size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3672 				break;
3673 
3674 			default:
3675 				/* No extra segments required for sends */
3676 				break;
3677 			}
3678 			break;
3679 
3680 		case MLX4_IB_QPT_TUN_SMI_OWNER:
3681 			err =  build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3682 					ctrl, &seglen);
3683 			if (unlikely(err)) {
3684 				*bad_wr = wr;
3685 				goto out;
3686 			}
3687 			wqe  += seglen;
3688 			size += seglen / 16;
3689 			break;
3690 		case MLX4_IB_QPT_TUN_SMI:
3691 		case MLX4_IB_QPT_TUN_GSI:
3692 			/* this is a UD qp used in MAD responses to slaves. */
3693 			set_datagram_seg(wqe, ud_wr(wr));
3694 			/* set the forced-loopback bit in the data seg av */
3695 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
3696 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3697 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3698 			break;
3699 		case MLX4_IB_QPT_UD:
3700 			set_datagram_seg(wqe, ud_wr(wr));
3701 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3702 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3703 
3704 			if (wr->opcode == IB_WR_LSO) {
3705 				err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3706 						&lso_hdr_sz, &blh);
3707 				if (unlikely(err)) {
3708 					*bad_wr = wr;
3709 					goto out;
3710 				}
3711 				lso_wqe = (__be32 *) wqe;
3712 				wqe  += seglen;
3713 				size += seglen / 16;
3714 			}
3715 			break;
3716 
3717 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
3718 			err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3719 					ctrl, &seglen);
3720 			if (unlikely(err)) {
3721 				*bad_wr = wr;
3722 				goto out;
3723 			}
3724 			wqe  += seglen;
3725 			size += seglen / 16;
3726 			/* to start tunnel header on a cache-line boundary */
3727 			add_zero_len_inline(wqe);
3728 			wqe += 16;
3729 			size++;
3730 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3731 			wqe  += seglen;
3732 			size += seglen / 16;
3733 			break;
3734 		case MLX4_IB_QPT_PROXY_SMI:
3735 		case MLX4_IB_QPT_PROXY_GSI:
3736 			/* If we are tunneling special qps, this is a UD qp.
3737 			 * In this case we first add a UD segment targeting
3738 			 * the tunnel qp, and then add a header with address
3739 			 * information */
3740 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3741 						ud_wr(wr),
3742 						qp->mlx4_ib_qp_type);
3743 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
3744 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3745 			build_tunnel_header(ud_wr(wr), wqe, &seglen);
3746 			wqe  += seglen;
3747 			size += seglen / 16;
3748 			break;
3749 
3750 		case MLX4_IB_QPT_SMI:
3751 		case MLX4_IB_QPT_GSI:
3752 			err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
3753 					&seglen);
3754 			if (unlikely(err)) {
3755 				*bad_wr = wr;
3756 				goto out;
3757 			}
3758 			wqe  += seglen;
3759 			size += seglen / 16;
3760 			break;
3761 
3762 		default:
3763 			break;
3764 		}
3765 
3766 		/*
3767 		 * Write data segments in reverse order, so as to
3768 		 * overwrite cacheline stamp last within each
3769 		 * cacheline.  This avoids issues with WQE
3770 		 * prefetching.
3771 		 */
3772 
3773 		dseg = wqe;
3774 		dseg += wr->num_sge - 1;
3775 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3776 
3777 		/* Add one more inline data segment for ICRC for MLX sends */
3778 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3779 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3780 			     qp->mlx4_ib_qp_type &
3781 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3782 			set_mlx_icrc_seg(dseg + 1);
3783 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
3784 		}
3785 
3786 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3787 			set_data_seg(dseg, wr->sg_list + i);
3788 
3789 		/*
3790 		 * Possibly overwrite stamping in cacheline with LSO
3791 		 * segment only after making sure all data segments
3792 		 * are written.
3793 		 */
3794 		wmb();
3795 		*lso_wqe = lso_hdr_sz;
3796 
3797 		ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3798 					     MLX4_WQE_CTRL_FENCE : 0) | size;
3799 
3800 		/*
3801 		 * Make sure descriptor is fully written before
3802 		 * setting ownership bit (because HW can start
3803 		 * executing as soon as we do).
3804 		 */
3805 		wmb();
3806 
3807 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3808 			*bad_wr = wr;
3809 			err = -EINVAL;
3810 			goto out;
3811 		}
3812 
3813 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3814 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3815 
3816 		/*
3817 		 * We can improve latency by not stamping the last
3818 		 * send queue WQE until after ringing the doorbell, so
3819 		 * only stamp here if there are still more WQEs to post.
3820 		 */
3821 		if (wr->next)
3822 			stamp_send_wqe(qp, ind + qp->sq_spare_wqes);
3823 		ind++;
3824 	}
3825 
3826 out:
3827 	if (likely(nreq)) {
3828 		qp->sq.head += nreq;
3829 
3830 		/*
3831 		 * Make sure that descriptors are written before
3832 		 * doorbell record.
3833 		 */
3834 		wmb();
3835 
3836 		writel_relaxed(qp->doorbell_qpn,
3837 			to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3838 
3839 		stamp_send_wqe(qp, ind + qp->sq_spare_wqes - 1);
3840 
3841 		qp->sq_next_wqe = ind;
3842 	}
3843 
3844 	spin_unlock_irqrestore(&qp->sq.lock, flags);
3845 
3846 	return err;
3847 }
3848 
3849 int mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3850 		      const struct ib_send_wr **bad_wr)
3851 {
3852 	return _mlx4_ib_post_send(ibqp, wr, bad_wr, false);
3853 }
3854 
3855 static int _mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3856 			      const struct ib_recv_wr **bad_wr, bool drain)
3857 {
3858 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3859 	struct mlx4_wqe_data_seg *scat;
3860 	unsigned long flags;
3861 	int err = 0;
3862 	int nreq;
3863 	int ind;
3864 	int max_gs;
3865 	int i;
3866 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3867 
3868 	max_gs = qp->rq.max_gs;
3869 	spin_lock_irqsave(&qp->rq.lock, flags);
3870 
3871 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3872 	    !drain) {
3873 		err = -EIO;
3874 		*bad_wr = wr;
3875 		nreq = 0;
3876 		goto out;
3877 	}
3878 
3879 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3880 
3881 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
3882 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3883 			err = -ENOMEM;
3884 			*bad_wr = wr;
3885 			goto out;
3886 		}
3887 
3888 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3889 			err = -EINVAL;
3890 			*bad_wr = wr;
3891 			goto out;
3892 		}
3893 
3894 		scat = get_recv_wqe(qp, ind);
3895 
3896 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3897 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3898 			ib_dma_sync_single_for_device(ibqp->device,
3899 						      qp->sqp_proxy_rcv[ind].map,
3900 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
3901 						      DMA_FROM_DEVICE);
3902 			scat->byte_count =
3903 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3904 			/* use dma lkey from upper layer entry */
3905 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3906 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3907 			scat++;
3908 			max_gs--;
3909 		}
3910 
3911 		for (i = 0; i < wr->num_sge; ++i)
3912 			__set_data_seg(scat + i, wr->sg_list + i);
3913 
3914 		if (i < max_gs) {
3915 			scat[i].byte_count = 0;
3916 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3917 			scat[i].addr       = 0;
3918 		}
3919 
3920 		qp->rq.wrid[ind] = wr->wr_id;
3921 
3922 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3923 	}
3924 
3925 out:
3926 	if (likely(nreq)) {
3927 		qp->rq.head += nreq;
3928 
3929 		/*
3930 		 * Make sure that descriptors are written before
3931 		 * doorbell record.
3932 		 */
3933 		wmb();
3934 
3935 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3936 	}
3937 
3938 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3939 
3940 	return err;
3941 }
3942 
3943 int mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3944 		      const struct ib_recv_wr **bad_wr)
3945 {
3946 	return _mlx4_ib_post_recv(ibqp, wr, bad_wr, false);
3947 }
3948 
3949 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3950 {
3951 	switch (mlx4_state) {
3952 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3953 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3954 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3955 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3956 	case MLX4_QP_STATE_SQ_DRAINING:
3957 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3958 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3959 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3960 	default:		     return -1;
3961 	}
3962 }
3963 
3964 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3965 {
3966 	switch (mlx4_mig_state) {
3967 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3968 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3969 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3970 	default: return -1;
3971 	}
3972 }
3973 
3974 static int to_ib_qp_access_flags(int mlx4_flags)
3975 {
3976 	int ib_flags = 0;
3977 
3978 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3979 		ib_flags |= IB_ACCESS_REMOTE_READ;
3980 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3981 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3982 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3983 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3984 
3985 	return ib_flags;
3986 }
3987 
3988 static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
3989 			    struct rdma_ah_attr *ah_attr,
3990 			    struct mlx4_qp_path *path)
3991 {
3992 	struct mlx4_dev *dev = ibdev->dev;
3993 	u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
3994 
3995 	memset(ah_attr, 0, sizeof(*ah_attr));
3996 	if (port_num == 0 || port_num > dev->caps.num_ports)
3997 		return;
3998 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
3999 
4000 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
4001 		rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
4002 			       ((path->sched_queue & 4) << 1));
4003 	else
4004 		rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
4005 	rdma_ah_set_port_num(ah_attr, port_num);
4006 
4007 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
4008 	rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
4009 	rdma_ah_set_static_rate(ah_attr,
4010 				path->static_rate ? path->static_rate - 5 : 0);
4011 	if (path->grh_mylmc & (1 << 7)) {
4012 		rdma_ah_set_grh(ah_attr, NULL,
4013 				be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
4014 				path->mgid_index,
4015 				path->hop_limit,
4016 				(be32_to_cpu(path->tclass_flowlabel)
4017 				 >> 20) & 0xff);
4018 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
4019 	}
4020 }
4021 
4022 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
4023 		     struct ib_qp_init_attr *qp_init_attr)
4024 {
4025 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
4026 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
4027 	struct mlx4_qp_context context;
4028 	int mlx4_state;
4029 	int err = 0;
4030 
4031 	if (ibqp->rwq_ind_tbl)
4032 		return -EOPNOTSUPP;
4033 
4034 	mutex_lock(&qp->mutex);
4035 
4036 	if (qp->state == IB_QPS_RESET) {
4037 		qp_attr->qp_state = IB_QPS_RESET;
4038 		goto done;
4039 	}
4040 
4041 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
4042 	if (err) {
4043 		err = -EINVAL;
4044 		goto out;
4045 	}
4046 
4047 	mlx4_state = be32_to_cpu(context.flags) >> 28;
4048 
4049 	qp->state		     = to_ib_qp_state(mlx4_state);
4050 	qp_attr->qp_state	     = qp->state;
4051 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
4052 	qp_attr->path_mig_state	     =
4053 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
4054 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
4055 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
4056 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
4057 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
4058 	qp_attr->qp_access_flags     =
4059 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
4060 
4061 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4062 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
4063 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
4064 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
4065 		qp_attr->alt_port_num	=
4066 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
4067 	}
4068 
4069 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
4070 	if (qp_attr->qp_state == IB_QPS_INIT)
4071 		qp_attr->port_num = qp->port;
4072 	else
4073 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
4074 
4075 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4076 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
4077 
4078 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
4079 
4080 	qp_attr->max_dest_rd_atomic =
4081 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
4082 	qp_attr->min_rnr_timer	    =
4083 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
4084 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
4085 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
4086 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
4087 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
4088 
4089 done:
4090 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4091 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4092 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4093 
4094 	if (!ibqp->uobject) {
4095 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
4096 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4097 	} else {
4098 		qp_attr->cap.max_send_wr  = 0;
4099 		qp_attr->cap.max_send_sge = 0;
4100 	}
4101 
4102 	/*
4103 	 * We don't support inline sends for kernel QPs (yet), and we
4104 	 * don't know what userspace's value should be.
4105 	 */
4106 	qp_attr->cap.max_inline_data = 0;
4107 
4108 	qp_init_attr->cap	     = qp_attr->cap;
4109 
4110 	qp_init_attr->create_flags = 0;
4111 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4112 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4113 
4114 	if (qp->flags & MLX4_IB_QP_LSO)
4115 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
4116 
4117 	if (qp->flags & MLX4_IB_QP_NETIF)
4118 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
4119 
4120 	qp_init_attr->sq_sig_type =
4121 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
4122 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4123 
4124 out:
4125 	mutex_unlock(&qp->mutex);
4126 	return err;
4127 }
4128 
4129 struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
4130 				struct ib_wq_init_attr *init_attr,
4131 				struct ib_udata *udata)
4132 {
4133 	struct mlx4_dev *dev = to_mdev(pd->device)->dev;
4134 	struct ib_qp_init_attr ib_qp_init_attr = {};
4135 	struct mlx4_ib_qp *qp;
4136 	struct mlx4_ib_create_wq ucmd;
4137 	int err, required_cmd_sz;
4138 
4139 	if (!udata)
4140 		return ERR_PTR(-EINVAL);
4141 
4142 	required_cmd_sz = offsetof(typeof(ucmd), comp_mask) +
4143 			  sizeof(ucmd.comp_mask);
4144 	if (udata->inlen < required_cmd_sz) {
4145 		pr_debug("invalid inlen\n");
4146 		return ERR_PTR(-EINVAL);
4147 	}
4148 
4149 	if (udata->inlen > sizeof(ucmd) &&
4150 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4151 				 udata->inlen - sizeof(ucmd))) {
4152 		pr_debug("inlen is not supported\n");
4153 		return ERR_PTR(-EOPNOTSUPP);
4154 	}
4155 
4156 	if (udata->outlen)
4157 		return ERR_PTR(-EOPNOTSUPP);
4158 
4159 	if (init_attr->wq_type != IB_WQT_RQ) {
4160 		pr_debug("unsupported wq type %d\n", init_attr->wq_type);
4161 		return ERR_PTR(-EOPNOTSUPP);
4162 	}
4163 
4164 	if (init_attr->create_flags & ~IB_WQ_FLAGS_SCATTER_FCS ||
4165 	    !(dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
4166 		pr_debug("unsupported create_flags %u\n",
4167 			 init_attr->create_flags);
4168 		return ERR_PTR(-EOPNOTSUPP);
4169 	}
4170 
4171 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
4172 	if (!qp)
4173 		return ERR_PTR(-ENOMEM);
4174 
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_mask & IB_WQ_CUR_STATE ? wq_attr->curr_wq_state :
4298 						     ibwq->state;
4299 	new_state = wq_attr_mask & IB_WQ_STATE ? wq_attr->wq_state : cur_state;
4300 
4301 	if (cur_state  < IB_WQS_RESET || cur_state  > IB_WQS_ERR ||
4302 	    new_state < IB_WQS_RESET || new_state > IB_WQS_ERR)
4303 		return -EINVAL;
4304 
4305 	if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR))
4306 		return -EINVAL;
4307 
4308 	if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET))
4309 		return -EINVAL;
4310 
4311 	/* Need to protect against the parent RSS which also may modify WQ
4312 	 * state.
4313 	 */
4314 	mutex_lock(&qp->mutex);
4315 
4316 	/* Can update HW state only if a RSS QP has already associated to this
4317 	 * WQ, so we can apply its port on the WQ.
4318 	 */
4319 	if (qp->rss_usecnt)
4320 		err = _mlx4_ib_modify_wq(ibwq, new_state, udata);
4321 
4322 	if (!err)
4323 		ibwq->state = new_state;
4324 
4325 	mutex_unlock(&qp->mutex);
4326 
4327 	return err;
4328 }
4329 
4330 void mlx4_ib_destroy_wq(struct ib_wq *ibwq, struct ib_udata *udata)
4331 {
4332 	struct mlx4_ib_dev *dev = to_mdev(ibwq->device);
4333 	struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4334 
4335 	if (qp->counter_index)
4336 		mlx4_ib_free_qp_counter(dev, qp);
4337 
4338 	destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, udata);
4339 
4340 	kfree(qp);
4341 }
4342 
4343 struct ib_rwq_ind_table
4344 *mlx4_ib_create_rwq_ind_table(struct ib_device *device,
4345 			      struct ib_rwq_ind_table_init_attr *init_attr,
4346 			      struct ib_udata *udata)
4347 {
4348 	struct ib_rwq_ind_table *rwq_ind_table;
4349 	struct mlx4_ib_create_rwq_ind_tbl_resp resp = {};
4350 	unsigned int ind_tbl_size = 1 << init_attr->log_ind_tbl_size;
4351 	unsigned int base_wqn;
4352 	size_t min_resp_len;
4353 	int i;
4354 	int err;
4355 
4356 	if (udata->inlen > 0 &&
4357 	    !ib_is_udata_cleared(udata, 0,
4358 				 udata->inlen))
4359 		return ERR_PTR(-EOPNOTSUPP);
4360 
4361 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4362 	if (udata->outlen && udata->outlen < min_resp_len)
4363 		return ERR_PTR(-EINVAL);
4364 
4365 	if (ind_tbl_size >
4366 	    device->attrs.rss_caps.max_rwq_indirection_table_size) {
4367 		pr_debug("log_ind_tbl_size = %d is bigger than supported = %d\n",
4368 			 ind_tbl_size,
4369 			 device->attrs.rss_caps.max_rwq_indirection_table_size);
4370 		return ERR_PTR(-EINVAL);
4371 	}
4372 
4373 	base_wqn = init_attr->ind_tbl[0]->wq_num;
4374 
4375 	if (base_wqn % ind_tbl_size) {
4376 		pr_debug("WQN=0x%x isn't aligned with indirection table size\n",
4377 			 base_wqn);
4378 		return ERR_PTR(-EINVAL);
4379 	}
4380 
4381 	for (i = 1; i < ind_tbl_size; i++) {
4382 		if (++base_wqn != init_attr->ind_tbl[i]->wq_num) {
4383 			pr_debug("indirection table's WQNs aren't consecutive\n");
4384 			return ERR_PTR(-EINVAL);
4385 		}
4386 	}
4387 
4388 	rwq_ind_table = kzalloc(sizeof(*rwq_ind_table), GFP_KERNEL);
4389 	if (!rwq_ind_table)
4390 		return ERR_PTR(-ENOMEM);
4391 
4392 	if (udata->outlen) {
4393 		resp.response_length = offsetof(typeof(resp), response_length) +
4394 					sizeof(resp.response_length);
4395 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4396 		if (err)
4397 			goto err;
4398 	}
4399 
4400 	return rwq_ind_table;
4401 
4402 err:
4403 	kfree(rwq_ind_table);
4404 	return ERR_PTR(err);
4405 }
4406 
4407 int mlx4_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4408 {
4409 	kfree(ib_rwq_ind_tbl);
4410 	return 0;
4411 }
4412 
4413 struct mlx4_ib_drain_cqe {
4414 	struct ib_cqe cqe;
4415 	struct completion done;
4416 };
4417 
4418 static void mlx4_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
4419 {
4420 	struct mlx4_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
4421 						     struct mlx4_ib_drain_cqe,
4422 						     cqe);
4423 
4424 	complete(&cqe->done);
4425 }
4426 
4427 /* This function returns only once the drained WR was completed */
4428 static void handle_drain_completion(struct ib_cq *cq,
4429 				    struct mlx4_ib_drain_cqe *sdrain,
4430 				    struct mlx4_ib_dev *dev)
4431 {
4432 	struct mlx4_dev *mdev = dev->dev;
4433 
4434 	if (cq->poll_ctx == IB_POLL_DIRECT) {
4435 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
4436 			ib_process_cq_direct(cq, -1);
4437 		return;
4438 	}
4439 
4440 	if (mdev->persist->state == MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4441 		struct mlx4_ib_cq *mcq = to_mcq(cq);
4442 		bool triggered = false;
4443 		unsigned long flags;
4444 
4445 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
4446 		/* Make sure that the CQ handler won't run if wasn't run yet */
4447 		if (!mcq->mcq.reset_notify_added)
4448 			mcq->mcq.reset_notify_added = 1;
4449 		else
4450 			triggered = true;
4451 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
4452 
4453 		if (triggered) {
4454 			/* Wait for any scheduled/running task to be ended */
4455 			switch (cq->poll_ctx) {
4456 			case IB_POLL_SOFTIRQ:
4457 				irq_poll_disable(&cq->iop);
4458 				irq_poll_enable(&cq->iop);
4459 				break;
4460 			case IB_POLL_WORKQUEUE:
4461 				cancel_work_sync(&cq->work);
4462 				break;
4463 			default:
4464 				WARN_ON_ONCE(1);
4465 			}
4466 		}
4467 
4468 		/* Run the CQ handler - this makes sure that the drain WR will
4469 		 * be processed if wasn't processed yet.
4470 		 */
4471 		mcq->mcq.comp(&mcq->mcq);
4472 	}
4473 
4474 	wait_for_completion(&sdrain->done);
4475 }
4476 
4477 void mlx4_ib_drain_sq(struct ib_qp *qp)
4478 {
4479 	struct ib_cq *cq = qp->send_cq;
4480 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4481 	struct mlx4_ib_drain_cqe sdrain;
4482 	const struct ib_send_wr *bad_swr;
4483 	struct ib_rdma_wr swr = {
4484 		.wr = {
4485 			.next = NULL,
4486 			{ .wr_cqe	= &sdrain.cqe, },
4487 			.opcode	= IB_WR_RDMA_WRITE,
4488 		},
4489 	};
4490 	int ret;
4491 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
4492 	struct mlx4_dev *mdev = dev->dev;
4493 
4494 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4495 	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4496 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4497 		return;
4498 	}
4499 
4500 	sdrain.cqe.done = mlx4_ib_drain_qp_done;
4501 	init_completion(&sdrain.done);
4502 
4503 	ret = _mlx4_ib_post_send(qp, &swr.wr, &bad_swr, true);
4504 	if (ret) {
4505 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4506 		return;
4507 	}
4508 
4509 	handle_drain_completion(cq, &sdrain, dev);
4510 }
4511 
4512 void mlx4_ib_drain_rq(struct ib_qp *qp)
4513 {
4514 	struct ib_cq *cq = qp->recv_cq;
4515 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4516 	struct mlx4_ib_drain_cqe rdrain;
4517 	struct ib_recv_wr rwr = {};
4518 	const struct ib_recv_wr *bad_rwr;
4519 	int ret;
4520 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
4521 	struct mlx4_dev *mdev = dev->dev;
4522 
4523 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4524 	if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4525 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4526 		return;
4527 	}
4528 
4529 	rwr.wr_cqe = &rdrain.cqe;
4530 	rdrain.cqe.done = mlx4_ib_drain_qp_done;
4531 	init_completion(&rdrain.done);
4532 
4533 	ret = _mlx4_ib_post_recv(qp, &rwr, &bad_rwr, true);
4534 	if (ret) {
4535 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4536 		return;
4537 	}
4538 
4539 	handle_drain_completion(cq, &rdrain, dev);
4540 }
4541