xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision 4f3db074)
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/slab.h>
36 #include <linux/netdevice.h>
37 
38 #include <rdma/ib_cache.h>
39 #include <rdma/ib_pack.h>
40 #include <rdma/ib_addr.h>
41 #include <rdma/ib_mad.h>
42 
43 #include <linux/mlx4/driver.h>
44 #include <linux/mlx4/qp.h>
45 
46 #include "mlx4_ib.h"
47 #include "user.h"
48 
49 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
50 			     struct mlx4_ib_cq *recv_cq);
51 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
52 			       struct mlx4_ib_cq *recv_cq);
53 
54 enum {
55 	MLX4_IB_ACK_REQ_FREQ	= 8,
56 };
57 
58 enum {
59 	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
60 	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
61 	MLX4_IB_LINK_TYPE_IB		= 0,
62 	MLX4_IB_LINK_TYPE_ETH		= 1
63 };
64 
65 enum {
66 	/*
67 	 * Largest possible UD header: send with GRH and immediate
68 	 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
69 	 * tag.  (LRH would only use 8 bytes, so Ethernet is the
70 	 * biggest case)
71 	 */
72 	MLX4_IB_UD_HEADER_SIZE		= 82,
73 	MLX4_IB_LSO_HEADER_SPARE	= 128,
74 };
75 
76 enum {
77 	MLX4_IB_IBOE_ETHERTYPE		= 0x8915
78 };
79 
80 struct mlx4_ib_sqp {
81 	struct mlx4_ib_qp	qp;
82 	int			pkey_index;
83 	u32			qkey;
84 	u32			send_psn;
85 	struct ib_ud_header	ud_header;
86 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
87 };
88 
89 enum {
90 	MLX4_IB_MIN_SQ_STRIDE	= 6,
91 	MLX4_IB_CACHE_LINE_SIZE	= 64,
92 };
93 
94 enum {
95 	MLX4_RAW_QP_MTU		= 7,
96 	MLX4_RAW_QP_MSGMAX	= 31,
97 };
98 
99 #ifndef ETH_ALEN
100 #define ETH_ALEN        6
101 #endif
102 
103 static const __be32 mlx4_ib_opcode[] = {
104 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
105 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
106 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
107 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
108 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
109 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
110 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
111 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
112 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
113 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
114 	[IB_WR_FAST_REG_MR]			= cpu_to_be32(MLX4_OPCODE_FMR),
115 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
116 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
117 	[IB_WR_BIND_MW]				= cpu_to_be32(MLX4_OPCODE_BIND_MW),
118 };
119 
120 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
121 {
122 	return container_of(mqp, struct mlx4_ib_sqp, qp);
123 }
124 
125 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
126 {
127 	if (!mlx4_is_master(dev->dev))
128 		return 0;
129 
130 	return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
131 	       qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
132 		8 * MLX4_MFUNC_MAX;
133 }
134 
135 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
136 {
137 	int proxy_sqp = 0;
138 	int real_sqp = 0;
139 	int i;
140 	/* PPF or Native -- real SQP */
141 	real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
142 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
143 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
144 	if (real_sqp)
145 		return 1;
146 	/* VF or PF -- proxy SQP */
147 	if (mlx4_is_mfunc(dev->dev)) {
148 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
149 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] ||
150 			    qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) {
151 				proxy_sqp = 1;
152 				break;
153 			}
154 		}
155 	}
156 	return proxy_sqp;
157 }
158 
159 /* used for INIT/CLOSE port logic */
160 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
161 {
162 	int proxy_qp0 = 0;
163 	int real_qp0 = 0;
164 	int i;
165 	/* PPF or Native -- real QP0 */
166 	real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
167 		    qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
168 		    qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
169 	if (real_qp0)
170 		return 1;
171 	/* VF or PF -- proxy QP0 */
172 	if (mlx4_is_mfunc(dev->dev)) {
173 		for (i = 0; i < dev->dev->caps.num_ports; i++) {
174 			if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) {
175 				proxy_qp0 = 1;
176 				break;
177 			}
178 		}
179 	}
180 	return proxy_qp0;
181 }
182 
183 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
184 {
185 	return mlx4_buf_offset(&qp->buf, offset);
186 }
187 
188 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
189 {
190 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
191 }
192 
193 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
194 {
195 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
196 }
197 
198 /*
199  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
200  * first four bytes of every 64 byte chunk with
201  *     0x7FFFFFF | (invalid_ownership_value << 31).
202  *
203  * When the max work request size is less than or equal to the WQE
204  * basic block size, as an optimization, we can stamp all WQEs with
205  * 0xffffffff, and skip the very first chunk of each WQE.
206  */
207 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
208 {
209 	__be32 *wqe;
210 	int i;
211 	int s;
212 	int ind;
213 	void *buf;
214 	__be32 stamp;
215 	struct mlx4_wqe_ctrl_seg *ctrl;
216 
217 	if (qp->sq_max_wqes_per_wr > 1) {
218 		s = roundup(size, 1U << qp->sq.wqe_shift);
219 		for (i = 0; i < s; i += 64) {
220 			ind = (i >> qp->sq.wqe_shift) + n;
221 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
222 						       cpu_to_be32(0xffffffff);
223 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
224 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
225 			*wqe = stamp;
226 		}
227 	} else {
228 		ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
229 		s = (ctrl->fence_size & 0x3f) << 4;
230 		for (i = 64; i < s; i += 64) {
231 			wqe = buf + i;
232 			*wqe = cpu_to_be32(0xffffffff);
233 		}
234 	}
235 }
236 
237 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
238 {
239 	struct mlx4_wqe_ctrl_seg *ctrl;
240 	struct mlx4_wqe_inline_seg *inl;
241 	void *wqe;
242 	int s;
243 
244 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
245 	s = sizeof(struct mlx4_wqe_ctrl_seg);
246 
247 	if (qp->ibqp.qp_type == IB_QPT_UD) {
248 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
249 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
250 		memset(dgram, 0, sizeof *dgram);
251 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
252 		s += sizeof(struct mlx4_wqe_datagram_seg);
253 	}
254 
255 	/* Pad the remainder of the WQE with an inline data segment. */
256 	if (size > s) {
257 		inl = wqe + s;
258 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
259 	}
260 	ctrl->srcrb_flags = 0;
261 	ctrl->fence_size = size / 16;
262 	/*
263 	 * Make sure descriptor is fully written before setting ownership bit
264 	 * (because HW can start executing as soon as we do).
265 	 */
266 	wmb();
267 
268 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
269 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
270 
271 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
272 }
273 
274 /* Post NOP WQE to prevent wrap-around in the middle of WR */
275 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
276 {
277 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
278 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
279 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
280 		ind += s;
281 	}
282 	return ind;
283 }
284 
285 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
286 {
287 	struct ib_event event;
288 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
289 
290 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
291 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
292 
293 	if (ibqp->event_handler) {
294 		event.device     = ibqp->device;
295 		event.element.qp = ibqp;
296 		switch (type) {
297 		case MLX4_EVENT_TYPE_PATH_MIG:
298 			event.event = IB_EVENT_PATH_MIG;
299 			break;
300 		case MLX4_EVENT_TYPE_COMM_EST:
301 			event.event = IB_EVENT_COMM_EST;
302 			break;
303 		case MLX4_EVENT_TYPE_SQ_DRAINED:
304 			event.event = IB_EVENT_SQ_DRAINED;
305 			break;
306 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
307 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
308 			break;
309 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
310 			event.event = IB_EVENT_QP_FATAL;
311 			break;
312 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
313 			event.event = IB_EVENT_PATH_MIG_ERR;
314 			break;
315 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
316 			event.event = IB_EVENT_QP_REQ_ERR;
317 			break;
318 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
319 			event.event = IB_EVENT_QP_ACCESS_ERR;
320 			break;
321 		default:
322 			pr_warn("Unexpected event type %d "
323 			       "on QP %06x\n", type, qp->qpn);
324 			return;
325 		}
326 
327 		ibqp->event_handler(&event, ibqp->qp_context);
328 	}
329 }
330 
331 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
332 {
333 	/*
334 	 * UD WQEs must have a datagram segment.
335 	 * RC and UC WQEs might have a remote address segment.
336 	 * MLX WQEs need two extra inline data segments (for the UD
337 	 * header and space for the ICRC).
338 	 */
339 	switch (type) {
340 	case MLX4_IB_QPT_UD:
341 		return sizeof (struct mlx4_wqe_ctrl_seg) +
342 			sizeof (struct mlx4_wqe_datagram_seg) +
343 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
344 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
345 	case MLX4_IB_QPT_PROXY_SMI:
346 	case MLX4_IB_QPT_PROXY_GSI:
347 		return sizeof (struct mlx4_wqe_ctrl_seg) +
348 			sizeof (struct mlx4_wqe_datagram_seg) + 64;
349 	case MLX4_IB_QPT_TUN_SMI_OWNER:
350 	case MLX4_IB_QPT_TUN_GSI:
351 		return sizeof (struct mlx4_wqe_ctrl_seg) +
352 			sizeof (struct mlx4_wqe_datagram_seg);
353 
354 	case MLX4_IB_QPT_UC:
355 		return sizeof (struct mlx4_wqe_ctrl_seg) +
356 			sizeof (struct mlx4_wqe_raddr_seg);
357 	case MLX4_IB_QPT_RC:
358 		return sizeof (struct mlx4_wqe_ctrl_seg) +
359 			sizeof (struct mlx4_wqe_atomic_seg) +
360 			sizeof (struct mlx4_wqe_raddr_seg);
361 	case MLX4_IB_QPT_SMI:
362 	case MLX4_IB_QPT_GSI:
363 		return sizeof (struct mlx4_wqe_ctrl_seg) +
364 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
365 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
366 					   MLX4_INLINE_ALIGN) *
367 			      sizeof (struct mlx4_wqe_inline_seg),
368 			      sizeof (struct mlx4_wqe_data_seg)) +
369 			ALIGN(4 +
370 			      sizeof (struct mlx4_wqe_inline_seg),
371 			      sizeof (struct mlx4_wqe_data_seg));
372 	default:
373 		return sizeof (struct mlx4_wqe_ctrl_seg);
374 	}
375 }
376 
377 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
378 		       int is_user, int has_rq, struct mlx4_ib_qp *qp)
379 {
380 	/* Sanity check RQ size before proceeding */
381 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
382 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
383 		return -EINVAL;
384 
385 	if (!has_rq) {
386 		if (cap->max_recv_wr)
387 			return -EINVAL;
388 
389 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
390 	} else {
391 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
392 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
393 			return -EINVAL;
394 
395 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
396 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
397 		qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
398 	}
399 
400 	/* leave userspace return values as they were, so as not to break ABI */
401 	if (is_user) {
402 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
403 		cap->max_recv_sge = qp->rq.max_gs;
404 	} else {
405 		cap->max_recv_wr  = qp->rq.max_post =
406 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
407 		cap->max_recv_sge = min(qp->rq.max_gs,
408 					min(dev->dev->caps.max_sq_sg,
409 					    dev->dev->caps.max_rq_sg));
410 	}
411 
412 	return 0;
413 }
414 
415 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
416 			      enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
417 {
418 	int s;
419 
420 	/* Sanity check SQ size before proceeding */
421 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
422 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
423 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
424 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
425 		return -EINVAL;
426 
427 	/*
428 	 * For MLX transport we need 2 extra S/G entries:
429 	 * one for the header and one for the checksum at the end
430 	 */
431 	if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
432 	     type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
433 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
434 		return -EINVAL;
435 
436 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
437 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
438 		send_wqe_overhead(type, qp->flags);
439 
440 	if (s > dev->dev->caps.max_sq_desc_sz)
441 		return -EINVAL;
442 
443 	/*
444 	 * Hermon supports shrinking WQEs, such that a single work
445 	 * request can include multiple units of 1 << wqe_shift.  This
446 	 * way, work requests can differ in size, and do not have to
447 	 * be a power of 2 in size, saving memory and speeding up send
448 	 * WR posting.  Unfortunately, if we do this then the
449 	 * wqe_index field in CQEs can't be used to look up the WR ID
450 	 * anymore, so we do this only if selective signaling is off.
451 	 *
452 	 * Further, on 32-bit platforms, we can't use vmap() to make
453 	 * the QP buffer virtually contiguous.  Thus we have to use
454 	 * constant-sized WRs to make sure a WR is always fully within
455 	 * a single page-sized chunk.
456 	 *
457 	 * Finally, we use NOP work requests to pad the end of the
458 	 * work queue, to avoid wrap-around in the middle of WR.  We
459 	 * set NEC bit to avoid getting completions with error for
460 	 * these NOP WRs, but since NEC is only supported starting
461 	 * with firmware 2.2.232, we use constant-sized WRs for older
462 	 * firmware.
463 	 *
464 	 * And, since MLX QPs only support SEND, we use constant-sized
465 	 * WRs in this case.
466 	 *
467 	 * We look for the smallest value of wqe_shift such that the
468 	 * resulting number of wqes does not exceed device
469 	 * capabilities.
470 	 *
471 	 * We set WQE size to at least 64 bytes, this way stamping
472 	 * invalidates each WQE.
473 	 */
474 	if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
475 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
476 	    type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
477 	    !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
478 		      MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
479 		qp->sq.wqe_shift = ilog2(64);
480 	else
481 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
482 
483 	for (;;) {
484 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
485 
486 		/*
487 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
488 		 * allow HW to prefetch.
489 		 */
490 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
491 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
492 						    qp->sq_max_wqes_per_wr +
493 						    qp->sq_spare_wqes);
494 
495 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
496 			break;
497 
498 		if (qp->sq_max_wqes_per_wr <= 1)
499 			return -EINVAL;
500 
501 		++qp->sq.wqe_shift;
502 	}
503 
504 	qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
505 			     (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
506 			 send_wqe_overhead(type, qp->flags)) /
507 		sizeof (struct mlx4_wqe_data_seg);
508 
509 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
510 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
511 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
512 		qp->rq.offset = 0;
513 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
514 	} else {
515 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
516 		qp->sq.offset = 0;
517 	}
518 
519 	cap->max_send_wr  = qp->sq.max_post =
520 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
521 	cap->max_send_sge = min(qp->sq.max_gs,
522 				min(dev->dev->caps.max_sq_sg,
523 				    dev->dev->caps.max_rq_sg));
524 	/* We don't support inline sends for kernel QPs (yet) */
525 	cap->max_inline_data = 0;
526 
527 	return 0;
528 }
529 
530 static int set_user_sq_size(struct mlx4_ib_dev *dev,
531 			    struct mlx4_ib_qp *qp,
532 			    struct mlx4_ib_create_qp *ucmd)
533 {
534 	/* Sanity check SQ size before proceeding */
535 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
536 	    ucmd->log_sq_stride >
537 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
538 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
539 		return -EINVAL;
540 
541 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
542 	qp->sq.wqe_shift = ucmd->log_sq_stride;
543 
544 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
545 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
546 
547 	return 0;
548 }
549 
550 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
551 {
552 	int i;
553 
554 	qp->sqp_proxy_rcv =
555 		kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
556 			GFP_KERNEL);
557 	if (!qp->sqp_proxy_rcv)
558 		return -ENOMEM;
559 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
560 		qp->sqp_proxy_rcv[i].addr =
561 			kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
562 				GFP_KERNEL);
563 		if (!qp->sqp_proxy_rcv[i].addr)
564 			goto err;
565 		qp->sqp_proxy_rcv[i].map =
566 			ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
567 					  sizeof (struct mlx4_ib_proxy_sqp_hdr),
568 					  DMA_FROM_DEVICE);
569 		if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
570 			kfree(qp->sqp_proxy_rcv[i].addr);
571 			goto err;
572 		}
573 	}
574 	return 0;
575 
576 err:
577 	while (i > 0) {
578 		--i;
579 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
580 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
581 				    DMA_FROM_DEVICE);
582 		kfree(qp->sqp_proxy_rcv[i].addr);
583 	}
584 	kfree(qp->sqp_proxy_rcv);
585 	qp->sqp_proxy_rcv = NULL;
586 	return -ENOMEM;
587 }
588 
589 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
590 {
591 	int i;
592 
593 	for (i = 0; i < qp->rq.wqe_cnt; i++) {
594 		ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
595 				    sizeof (struct mlx4_ib_proxy_sqp_hdr),
596 				    DMA_FROM_DEVICE);
597 		kfree(qp->sqp_proxy_rcv[i].addr);
598 	}
599 	kfree(qp->sqp_proxy_rcv);
600 }
601 
602 static int qp_has_rq(struct ib_qp_init_attr *attr)
603 {
604 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
605 		return 0;
606 
607 	return !attr->srq;
608 }
609 
610 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
611 {
612 	int i;
613 	for (i = 0; i < dev->caps.num_ports; i++) {
614 		if (qpn == dev->caps.qp0_proxy[i])
615 			return !!dev->caps.qp0_qkey[i];
616 	}
617 	return 0;
618 }
619 
620 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
621 			    struct ib_qp_init_attr *init_attr,
622 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp,
623 			    gfp_t gfp)
624 {
625 	int qpn;
626 	int err;
627 	struct mlx4_ib_sqp *sqp;
628 	struct mlx4_ib_qp *qp;
629 	enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
630 	struct mlx4_ib_cq *mcq;
631 	unsigned long flags;
632 
633 	/* When tunneling special qps, we use a plain UD qp */
634 	if (sqpn) {
635 		if (mlx4_is_mfunc(dev->dev) &&
636 		    (!mlx4_is_master(dev->dev) ||
637 		     !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
638 			if (init_attr->qp_type == IB_QPT_GSI)
639 				qp_type = MLX4_IB_QPT_PROXY_GSI;
640 			else {
641 				if (mlx4_is_master(dev->dev) ||
642 				    qp0_enabled_vf(dev->dev, sqpn))
643 					qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
644 				else
645 					qp_type = MLX4_IB_QPT_PROXY_SMI;
646 			}
647 		}
648 		qpn = sqpn;
649 		/* add extra sg entry for tunneling */
650 		init_attr->cap.max_recv_sge++;
651 	} else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
652 		struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
653 			container_of(init_attr,
654 				     struct mlx4_ib_qp_tunnel_init_attr, init_attr);
655 		if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
656 		     tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
657 		    !mlx4_is_master(dev->dev))
658 			return -EINVAL;
659 		if (tnl_init->proxy_qp_type == IB_QPT_GSI)
660 			qp_type = MLX4_IB_QPT_TUN_GSI;
661 		else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
662 			 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
663 					     tnl_init->port))
664 			qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
665 		else
666 			qp_type = MLX4_IB_QPT_TUN_SMI;
667 		/* we are definitely in the PPF here, since we are creating
668 		 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
669 		qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
670 			+ tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
671 		sqpn = qpn;
672 	}
673 
674 	if (!*caller_qp) {
675 		if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
676 		    (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
677 				MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
678 			sqp = kzalloc(sizeof (struct mlx4_ib_sqp), gfp);
679 			if (!sqp)
680 				return -ENOMEM;
681 			qp = &sqp->qp;
682 			qp->pri.vid = 0xFFFF;
683 			qp->alt.vid = 0xFFFF;
684 		} else {
685 			qp = kzalloc(sizeof (struct mlx4_ib_qp), gfp);
686 			if (!qp)
687 				return -ENOMEM;
688 			qp->pri.vid = 0xFFFF;
689 			qp->alt.vid = 0xFFFF;
690 		}
691 	} else
692 		qp = *caller_qp;
693 
694 	qp->mlx4_ib_qp_type = qp_type;
695 
696 	mutex_init(&qp->mutex);
697 	spin_lock_init(&qp->sq.lock);
698 	spin_lock_init(&qp->rq.lock);
699 	INIT_LIST_HEAD(&qp->gid_list);
700 	INIT_LIST_HEAD(&qp->steering_rules);
701 
702 	qp->state	 = IB_QPS_RESET;
703 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
704 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
705 
706 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
707 	if (err)
708 		goto err;
709 
710 	if (pd->uobject) {
711 		struct mlx4_ib_create_qp ucmd;
712 
713 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
714 			err = -EFAULT;
715 			goto err;
716 		}
717 
718 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
719 
720 		err = set_user_sq_size(dev, qp, &ucmd);
721 		if (err)
722 			goto err;
723 
724 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
725 				       qp->buf_size, 0, 0);
726 		if (IS_ERR(qp->umem)) {
727 			err = PTR_ERR(qp->umem);
728 			goto err;
729 		}
730 
731 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
732 				    ilog2(qp->umem->page_size), &qp->mtt);
733 		if (err)
734 			goto err_buf;
735 
736 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
737 		if (err)
738 			goto err_mtt;
739 
740 		if (qp_has_rq(init_attr)) {
741 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
742 						  ucmd.db_addr, &qp->db);
743 			if (err)
744 				goto err_mtt;
745 		}
746 	} else {
747 		qp->sq_no_prefetch = 0;
748 
749 		if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
750 			qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
751 
752 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
753 			qp->flags |= MLX4_IB_QP_LSO;
754 
755 		if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
756 			if (dev->steering_support ==
757 			    MLX4_STEERING_MODE_DEVICE_MANAGED)
758 				qp->flags |= MLX4_IB_QP_NETIF;
759 			else
760 				goto err;
761 		}
762 
763 		err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
764 		if (err)
765 			goto err;
766 
767 		if (qp_has_rq(init_attr)) {
768 			err = mlx4_db_alloc(dev->dev, &qp->db, 0, gfp);
769 			if (err)
770 				goto err;
771 
772 			*qp->db.db = 0;
773 		}
774 
775 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf, gfp)) {
776 			err = -ENOMEM;
777 			goto err_db;
778 		}
779 
780 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
781 				    &qp->mtt);
782 		if (err)
783 			goto err_buf;
784 
785 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf, gfp);
786 		if (err)
787 			goto err_mtt;
788 
789 		qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), gfp);
790 		qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), gfp);
791 		if (!qp->sq.wrid || !qp->rq.wrid) {
792 			err = -ENOMEM;
793 			goto err_wrid;
794 		}
795 	}
796 
797 	if (sqpn) {
798 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
799 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
800 			if (alloc_proxy_bufs(pd->device, qp)) {
801 				err = -ENOMEM;
802 				goto err_wrid;
803 			}
804 		}
805 	} else {
806 		/* Raw packet QPNs may not have bits 6,7 set in their qp_num;
807 		 * otherwise, the WQE BlueFlame setup flow wrongly causes
808 		 * VLAN insertion. */
809 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
810 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
811 						    (init_attr->cap.max_send_wr ?
812 						     MLX4_RESERVE_ETH_BF_QP : 0) |
813 						    (init_attr->cap.max_recv_wr ?
814 						     MLX4_RESERVE_A0_QP : 0));
815 		else
816 			if (qp->flags & MLX4_IB_QP_NETIF)
817 				err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
818 			else
819 				err = mlx4_qp_reserve_range(dev->dev, 1, 1,
820 							    &qpn, 0);
821 		if (err)
822 			goto err_proxy;
823 	}
824 
825 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp, gfp);
826 	if (err)
827 		goto err_qpn;
828 
829 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
830 		qp->mqp.qpn |= (1 << 23);
831 
832 	/*
833 	 * Hardware wants QPN written in big-endian order (after
834 	 * shifting) for send doorbell.  Precompute this value to save
835 	 * a little bit when posting sends.
836 	 */
837 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
838 
839 	qp->mqp.event = mlx4_ib_qp_event;
840 	if (!*caller_qp)
841 		*caller_qp = qp;
842 
843 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
844 	mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
845 			 to_mcq(init_attr->recv_cq));
846 	/* Maintain device to QPs access, needed for further handling
847 	 * via reset flow
848 	 */
849 	list_add_tail(&qp->qps_list, &dev->qp_list);
850 	/* Maintain CQ to QPs access, needed for further handling
851 	 * via reset flow
852 	 */
853 	mcq = to_mcq(init_attr->send_cq);
854 	list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
855 	mcq = to_mcq(init_attr->recv_cq);
856 	list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
857 	mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
858 			   to_mcq(init_attr->recv_cq));
859 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
860 	return 0;
861 
862 err_qpn:
863 	if (!sqpn) {
864 		if (qp->flags & MLX4_IB_QP_NETIF)
865 			mlx4_ib_steer_qp_free(dev, qpn, 1);
866 		else
867 			mlx4_qp_release_range(dev->dev, qpn, 1);
868 	}
869 err_proxy:
870 	if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
871 		free_proxy_bufs(pd->device, qp);
872 err_wrid:
873 	if (pd->uobject) {
874 		if (qp_has_rq(init_attr))
875 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
876 	} else {
877 		kfree(qp->sq.wrid);
878 		kfree(qp->rq.wrid);
879 	}
880 
881 err_mtt:
882 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
883 
884 err_buf:
885 	if (pd->uobject)
886 		ib_umem_release(qp->umem);
887 	else
888 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
889 
890 err_db:
891 	if (!pd->uobject && qp_has_rq(init_attr))
892 		mlx4_db_free(dev->dev, &qp->db);
893 
894 err:
895 	if (!*caller_qp)
896 		kfree(qp);
897 	return err;
898 }
899 
900 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
901 {
902 	switch (state) {
903 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
904 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
905 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
906 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
907 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
908 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
909 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
910 	default:		return -1;
911 	}
912 }
913 
914 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
915 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
916 {
917 	if (send_cq == recv_cq) {
918 		spin_lock(&send_cq->lock);
919 		__acquire(&recv_cq->lock);
920 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
921 		spin_lock(&send_cq->lock);
922 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
923 	} else {
924 		spin_lock(&recv_cq->lock);
925 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
926 	}
927 }
928 
929 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
930 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
931 {
932 	if (send_cq == recv_cq) {
933 		__release(&recv_cq->lock);
934 		spin_unlock(&send_cq->lock);
935 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
936 		spin_unlock(&recv_cq->lock);
937 		spin_unlock(&send_cq->lock);
938 	} else {
939 		spin_unlock(&send_cq->lock);
940 		spin_unlock(&recv_cq->lock);
941 	}
942 }
943 
944 static void del_gid_entries(struct mlx4_ib_qp *qp)
945 {
946 	struct mlx4_ib_gid_entry *ge, *tmp;
947 
948 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
949 		list_del(&ge->list);
950 		kfree(ge);
951 	}
952 }
953 
954 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
955 {
956 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
957 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
958 	else
959 		return to_mpd(qp->ibqp.pd);
960 }
961 
962 static void get_cqs(struct mlx4_ib_qp *qp,
963 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
964 {
965 	switch (qp->ibqp.qp_type) {
966 	case IB_QPT_XRC_TGT:
967 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
968 		*recv_cq = *send_cq;
969 		break;
970 	case IB_QPT_XRC_INI:
971 		*send_cq = to_mcq(qp->ibqp.send_cq);
972 		*recv_cq = *send_cq;
973 		break;
974 	default:
975 		*send_cq = to_mcq(qp->ibqp.send_cq);
976 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
977 		break;
978 	}
979 }
980 
981 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
982 			      int is_user)
983 {
984 	struct mlx4_ib_cq *send_cq, *recv_cq;
985 	unsigned long flags;
986 
987 	if (qp->state != IB_QPS_RESET) {
988 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
989 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
990 			pr_warn("modify QP %06x to RESET failed.\n",
991 			       qp->mqp.qpn);
992 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
993 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
994 			qp->pri.smac = 0;
995 			qp->pri.smac_port = 0;
996 		}
997 		if (qp->alt.smac) {
998 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
999 			qp->alt.smac = 0;
1000 		}
1001 		if (qp->pri.vid < 0x1000) {
1002 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1003 			qp->pri.vid = 0xFFFF;
1004 			qp->pri.candidate_vid = 0xFFFF;
1005 			qp->pri.update_vid = 0;
1006 		}
1007 		if (qp->alt.vid < 0x1000) {
1008 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1009 			qp->alt.vid = 0xFFFF;
1010 			qp->alt.candidate_vid = 0xFFFF;
1011 			qp->alt.update_vid = 0;
1012 		}
1013 	}
1014 
1015 	get_cqs(qp, &send_cq, &recv_cq);
1016 
1017 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1018 	mlx4_ib_lock_cqs(send_cq, recv_cq);
1019 
1020 	/* del from lists under both locks above to protect reset flow paths */
1021 	list_del(&qp->qps_list);
1022 	list_del(&qp->cq_send_list);
1023 	list_del(&qp->cq_recv_list);
1024 	if (!is_user) {
1025 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1026 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1027 		if (send_cq != recv_cq)
1028 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1029 	}
1030 
1031 	mlx4_qp_remove(dev->dev, &qp->mqp);
1032 
1033 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
1034 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1035 
1036 	mlx4_qp_free(dev->dev, &qp->mqp);
1037 
1038 	if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1039 		if (qp->flags & MLX4_IB_QP_NETIF)
1040 			mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1041 		else
1042 			mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1043 	}
1044 
1045 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1046 
1047 	if (is_user) {
1048 		if (qp->rq.wqe_cnt)
1049 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
1050 					      &qp->db);
1051 		ib_umem_release(qp->umem);
1052 	} else {
1053 		kfree(qp->sq.wrid);
1054 		kfree(qp->rq.wrid);
1055 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1056 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1057 			free_proxy_bufs(&dev->ib_dev, qp);
1058 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1059 		if (qp->rq.wqe_cnt)
1060 			mlx4_db_free(dev->dev, &qp->db);
1061 	}
1062 
1063 	del_gid_entries(qp);
1064 }
1065 
1066 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1067 {
1068 	/* Native or PPF */
1069 	if (!mlx4_is_mfunc(dev->dev) ||
1070 	    (mlx4_is_master(dev->dev) &&
1071 	     attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1072 		return  dev->dev->phys_caps.base_sqpn +
1073 			(attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1074 			attr->port_num - 1;
1075 	}
1076 	/* PF or VF -- creating proxies */
1077 	if (attr->qp_type == IB_QPT_SMI)
1078 		return dev->dev->caps.qp0_proxy[attr->port_num - 1];
1079 	else
1080 		return dev->dev->caps.qp1_proxy[attr->port_num - 1];
1081 }
1082 
1083 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1084 				struct ib_qp_init_attr *init_attr,
1085 				struct ib_udata *udata)
1086 {
1087 	struct mlx4_ib_qp *qp = NULL;
1088 	int err;
1089 	u16 xrcdn = 0;
1090 	gfp_t gfp;
1091 
1092 	gfp = (init_attr->create_flags & MLX4_IB_QP_CREATE_USE_GFP_NOIO) ?
1093 		GFP_NOIO : GFP_KERNEL;
1094 	/*
1095 	 * We only support LSO, vendor flag1, and multicast loopback blocking,
1096 	 * and only for kernel UD QPs.
1097 	 */
1098 	if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1099 					MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1100 					MLX4_IB_SRIOV_TUNNEL_QP |
1101 					MLX4_IB_SRIOV_SQP |
1102 					MLX4_IB_QP_NETIF |
1103 					MLX4_IB_QP_CREATE_USE_GFP_NOIO))
1104 		return ERR_PTR(-EINVAL);
1105 
1106 	if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1107 		if (init_attr->qp_type != IB_QPT_UD)
1108 			return ERR_PTR(-EINVAL);
1109 	}
1110 
1111 	if (init_attr->create_flags &&
1112 	    (udata ||
1113 	     ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP | MLX4_IB_QP_CREATE_USE_GFP_NOIO)) &&
1114 	      init_attr->qp_type != IB_QPT_UD) ||
1115 	     ((init_attr->create_flags & MLX4_IB_SRIOV_SQP) &&
1116 	      init_attr->qp_type > IB_QPT_GSI)))
1117 		return ERR_PTR(-EINVAL);
1118 
1119 	switch (init_attr->qp_type) {
1120 	case IB_QPT_XRC_TGT:
1121 		pd = to_mxrcd(init_attr->xrcd)->pd;
1122 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1123 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1124 		/* fall through */
1125 	case IB_QPT_XRC_INI:
1126 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1127 			return ERR_PTR(-ENOSYS);
1128 		init_attr->recv_cq = init_attr->send_cq;
1129 		/* fall through */
1130 	case IB_QPT_RC:
1131 	case IB_QPT_UC:
1132 	case IB_QPT_RAW_PACKET:
1133 		qp = kzalloc(sizeof *qp, gfp);
1134 		if (!qp)
1135 			return ERR_PTR(-ENOMEM);
1136 		qp->pri.vid = 0xFFFF;
1137 		qp->alt.vid = 0xFFFF;
1138 		/* fall through */
1139 	case IB_QPT_UD:
1140 	{
1141 		err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1142 				       udata, 0, &qp, gfp);
1143 		if (err)
1144 			return ERR_PTR(err);
1145 
1146 		qp->ibqp.qp_num = qp->mqp.qpn;
1147 		qp->xrcdn = xrcdn;
1148 
1149 		break;
1150 	}
1151 	case IB_QPT_SMI:
1152 	case IB_QPT_GSI:
1153 	{
1154 		/* Userspace is not allowed to create special QPs: */
1155 		if (udata)
1156 			return ERR_PTR(-EINVAL);
1157 
1158 		err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1159 				       get_sqp_num(to_mdev(pd->device), init_attr),
1160 				       &qp, gfp);
1161 		if (err)
1162 			return ERR_PTR(err);
1163 
1164 		qp->port	= init_attr->port_num;
1165 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
1166 
1167 		break;
1168 	}
1169 	default:
1170 		/* Don't support raw QPs */
1171 		return ERR_PTR(-EINVAL);
1172 	}
1173 
1174 	return &qp->ibqp;
1175 }
1176 
1177 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1178 {
1179 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
1180 	struct mlx4_ib_qp *mqp = to_mqp(qp);
1181 	struct mlx4_ib_pd *pd;
1182 
1183 	if (is_qp0(dev, mqp))
1184 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
1185 
1186 	if (dev->qp1_proxy[mqp->port - 1] == mqp) {
1187 		mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1188 		dev->qp1_proxy[mqp->port - 1] = NULL;
1189 		mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1190 	}
1191 
1192 	pd = get_pd(mqp);
1193 	destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1194 
1195 	if (is_sqp(dev, mqp))
1196 		kfree(to_msqp(mqp));
1197 	else
1198 		kfree(mqp);
1199 
1200 	return 0;
1201 }
1202 
1203 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1204 {
1205 	switch (type) {
1206 	case MLX4_IB_QPT_RC:		return MLX4_QP_ST_RC;
1207 	case MLX4_IB_QPT_UC:		return MLX4_QP_ST_UC;
1208 	case MLX4_IB_QPT_UD:		return MLX4_QP_ST_UD;
1209 	case MLX4_IB_QPT_XRC_INI:
1210 	case MLX4_IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
1211 	case MLX4_IB_QPT_SMI:
1212 	case MLX4_IB_QPT_GSI:
1213 	case MLX4_IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
1214 
1215 	case MLX4_IB_QPT_PROXY_SMI_OWNER:
1216 	case MLX4_IB_QPT_TUN_SMI_OWNER:	return (mlx4_is_mfunc(dev->dev) ?
1217 						MLX4_QP_ST_MLX : -1);
1218 	case MLX4_IB_QPT_PROXY_SMI:
1219 	case MLX4_IB_QPT_TUN_SMI:
1220 	case MLX4_IB_QPT_PROXY_GSI:
1221 	case MLX4_IB_QPT_TUN_GSI:	return (mlx4_is_mfunc(dev->dev) ?
1222 						MLX4_QP_ST_UD : -1);
1223 	default:			return -1;
1224 	}
1225 }
1226 
1227 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1228 				   int attr_mask)
1229 {
1230 	u8 dest_rd_atomic;
1231 	u32 access_flags;
1232 	u32 hw_access_flags = 0;
1233 
1234 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1235 		dest_rd_atomic = attr->max_dest_rd_atomic;
1236 	else
1237 		dest_rd_atomic = qp->resp_depth;
1238 
1239 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1240 		access_flags = attr->qp_access_flags;
1241 	else
1242 		access_flags = qp->atomic_rd_en;
1243 
1244 	if (!dest_rd_atomic)
1245 		access_flags &= IB_ACCESS_REMOTE_WRITE;
1246 
1247 	if (access_flags & IB_ACCESS_REMOTE_READ)
1248 		hw_access_flags |= MLX4_QP_BIT_RRE;
1249 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1250 		hw_access_flags |= MLX4_QP_BIT_RAE;
1251 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
1252 		hw_access_flags |= MLX4_QP_BIT_RWE;
1253 
1254 	return cpu_to_be32(hw_access_flags);
1255 }
1256 
1257 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1258 			    int attr_mask)
1259 {
1260 	if (attr_mask & IB_QP_PKEY_INDEX)
1261 		sqp->pkey_index = attr->pkey_index;
1262 	if (attr_mask & IB_QP_QKEY)
1263 		sqp->qkey = attr->qkey;
1264 	if (attr_mask & IB_QP_SQ_PSN)
1265 		sqp->send_psn = attr->sq_psn;
1266 }
1267 
1268 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1269 {
1270 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1271 }
1272 
1273 static int _mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
1274 			  u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1275 			  struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1276 {
1277 	int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
1278 		IB_LINK_LAYER_ETHERNET;
1279 	int vidx;
1280 	int smac_index;
1281 	int err;
1282 
1283 
1284 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
1285 	path->rlid	    = cpu_to_be16(ah->dlid);
1286 	if (ah->static_rate) {
1287 		path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
1288 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1289 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1290 			--path->static_rate;
1291 	} else
1292 		path->static_rate = 0;
1293 
1294 	if (ah->ah_flags & IB_AH_GRH) {
1295 		if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
1296 			pr_err("sgid_index (%u) too large. max is %d\n",
1297 			       ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1298 			return -1;
1299 		}
1300 
1301 		path->grh_mylmc |= 1 << 7;
1302 		path->mgid_index = ah->grh.sgid_index;
1303 		path->hop_limit  = ah->grh.hop_limit;
1304 		path->tclass_flowlabel =
1305 			cpu_to_be32((ah->grh.traffic_class << 20) |
1306 				    (ah->grh.flow_label));
1307 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
1308 	}
1309 
1310 	if (is_eth) {
1311 		if (!(ah->ah_flags & IB_AH_GRH))
1312 			return -1;
1313 
1314 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1315 			((port - 1) << 6) | ((ah->sl & 7) << 3);
1316 
1317 		path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1318 		if (vlan_tag < 0x1000) {
1319 			if (smac_info->vid < 0x1000) {
1320 				/* both valid vlan ids */
1321 				if (smac_info->vid != vlan_tag) {
1322 					/* different VIDs.  unreg old and reg new */
1323 					err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1324 					if (err)
1325 						return err;
1326 					smac_info->candidate_vid = vlan_tag;
1327 					smac_info->candidate_vlan_index = vidx;
1328 					smac_info->candidate_vlan_port = port;
1329 					smac_info->update_vid = 1;
1330 					path->vlan_index = vidx;
1331 				} else {
1332 					path->vlan_index = smac_info->vlan_index;
1333 				}
1334 			} else {
1335 				/* no current vlan tag in qp */
1336 				err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1337 				if (err)
1338 					return err;
1339 				smac_info->candidate_vid = vlan_tag;
1340 				smac_info->candidate_vlan_index = vidx;
1341 				smac_info->candidate_vlan_port = port;
1342 				smac_info->update_vid = 1;
1343 				path->vlan_index = vidx;
1344 			}
1345 			path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1346 			path->fl = 1 << 6;
1347 		} else {
1348 			/* have current vlan tag. unregister it at modify-qp success */
1349 			if (smac_info->vid < 0x1000) {
1350 				smac_info->candidate_vid = 0xFFFF;
1351 				smac_info->update_vid = 1;
1352 			}
1353 		}
1354 
1355 		/* get smac_index for RoCE use.
1356 		 * If no smac was yet assigned, register one.
1357 		 * If one was already assigned, but the new mac differs,
1358 		 * unregister the old one and register the new one.
1359 		*/
1360 		if ((!smac_info->smac && !smac_info->smac_port) ||
1361 		    smac_info->smac != smac) {
1362 			/* register candidate now, unreg if needed, after success */
1363 			smac_index = mlx4_register_mac(dev->dev, port, smac);
1364 			if (smac_index >= 0) {
1365 				smac_info->candidate_smac_index = smac_index;
1366 				smac_info->candidate_smac = smac;
1367 				smac_info->candidate_smac_port = port;
1368 			} else {
1369 				return -EINVAL;
1370 			}
1371 		} else {
1372 			smac_index = smac_info->smac_index;
1373 		}
1374 
1375 		memcpy(path->dmac, ah->dmac, 6);
1376 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
1377 		/* put MAC table smac index for IBoE */
1378 		path->grh_mylmc = (u8) (smac_index) | 0x80;
1379 	} else {
1380 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1381 			((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1382 	}
1383 
1384 	return 0;
1385 }
1386 
1387 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1388 			 enum ib_qp_attr_mask qp_attr_mask,
1389 			 struct mlx4_ib_qp *mqp,
1390 			 struct mlx4_qp_path *path, u8 port)
1391 {
1392 	return _mlx4_set_path(dev, &qp->ah_attr,
1393 			      mlx4_mac_to_u64((u8 *)qp->smac),
1394 			      (qp_attr_mask & IB_QP_VID) ? qp->vlan_id : 0xffff,
1395 			      path, &mqp->pri, port);
1396 }
1397 
1398 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1399 			     const struct ib_qp_attr *qp,
1400 			     enum ib_qp_attr_mask qp_attr_mask,
1401 			     struct mlx4_ib_qp *mqp,
1402 			     struct mlx4_qp_path *path, u8 port)
1403 {
1404 	return _mlx4_set_path(dev, &qp->alt_ah_attr,
1405 			      mlx4_mac_to_u64((u8 *)qp->alt_smac),
1406 			      (qp_attr_mask & IB_QP_ALT_VID) ?
1407 			      qp->alt_vlan_id : 0xffff,
1408 			      path, &mqp->alt, port);
1409 }
1410 
1411 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1412 {
1413 	struct mlx4_ib_gid_entry *ge, *tmp;
1414 
1415 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1416 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1417 			ge->added = 1;
1418 			ge->port = qp->port;
1419 		}
1420 	}
1421 }
1422 
1423 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp, u8 *smac,
1424 				    struct mlx4_qp_context *context)
1425 {
1426 	u64 u64_mac;
1427 	int smac_index;
1428 
1429 	u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1430 
1431 	context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1432 	if (!qp->pri.smac && !qp->pri.smac_port) {
1433 		smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1434 		if (smac_index >= 0) {
1435 			qp->pri.candidate_smac_index = smac_index;
1436 			qp->pri.candidate_smac = u64_mac;
1437 			qp->pri.candidate_smac_port = qp->port;
1438 			context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1439 		} else {
1440 			return -ENOENT;
1441 		}
1442 	}
1443 	return 0;
1444 }
1445 
1446 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1447 			       const struct ib_qp_attr *attr, int attr_mask,
1448 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1449 {
1450 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1451 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1452 	struct mlx4_ib_pd *pd;
1453 	struct mlx4_ib_cq *send_cq, *recv_cq;
1454 	struct mlx4_qp_context *context;
1455 	enum mlx4_qp_optpar optpar = 0;
1456 	int sqd_event;
1457 	int steer_qp = 0;
1458 	int err = -EINVAL;
1459 
1460 	/* APM is not supported under RoCE */
1461 	if (attr_mask & IB_QP_ALT_PATH &&
1462 	    rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1463 	    IB_LINK_LAYER_ETHERNET)
1464 		return -ENOTSUPP;
1465 
1466 	context = kzalloc(sizeof *context, GFP_KERNEL);
1467 	if (!context)
1468 		return -ENOMEM;
1469 
1470 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1471 				     (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1472 
1473 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1474 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1475 	else {
1476 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
1477 		switch (attr->path_mig_state) {
1478 		case IB_MIG_MIGRATED:
1479 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1480 			break;
1481 		case IB_MIG_REARM:
1482 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1483 			break;
1484 		case IB_MIG_ARMED:
1485 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1486 			break;
1487 		}
1488 	}
1489 
1490 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1491 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1492 	else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1493 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1494 	else if (ibqp->qp_type == IB_QPT_UD) {
1495 		if (qp->flags & MLX4_IB_QP_LSO)
1496 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
1497 					      ilog2(dev->dev->caps.max_gso_sz);
1498 		else
1499 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1500 	} else if (attr_mask & IB_QP_PATH_MTU) {
1501 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1502 			pr_err("path MTU (%u) is invalid\n",
1503 			       attr->path_mtu);
1504 			goto out;
1505 		}
1506 		context->mtu_msgmax = (attr->path_mtu << 5) |
1507 			ilog2(dev->dev->caps.max_msg_sz);
1508 	}
1509 
1510 	if (qp->rq.wqe_cnt)
1511 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1512 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
1513 
1514 	if (qp->sq.wqe_cnt)
1515 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1516 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
1517 
1518 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1519 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1520 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1521 		if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1522 			context->param3 |= cpu_to_be32(1 << 30);
1523 	}
1524 
1525 	if (qp->ibqp.uobject)
1526 		context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
1527 	else
1528 		context->usr_page = cpu_to_be32(dev->priv_uar.index);
1529 
1530 	if (attr_mask & IB_QP_DEST_QPN)
1531 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1532 
1533 	if (attr_mask & IB_QP_PORT) {
1534 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1535 		    !(attr_mask & IB_QP_AV)) {
1536 			mlx4_set_sched(&context->pri_path, attr->port_num);
1537 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1538 		}
1539 	}
1540 
1541 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1542 		if (dev->counters[qp->port - 1] != -1) {
1543 			context->pri_path.counter_index =
1544 						dev->counters[qp->port - 1];
1545 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1546 		} else
1547 			context->pri_path.counter_index = 0xff;
1548 
1549 		if (qp->flags & MLX4_IB_QP_NETIF) {
1550 			mlx4_ib_steer_qp_reg(dev, qp, 1);
1551 			steer_qp = 1;
1552 		}
1553 	}
1554 
1555 	if (attr_mask & IB_QP_PKEY_INDEX) {
1556 		if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1557 			context->pri_path.disable_pkey_check = 0x40;
1558 		context->pri_path.pkey_index = attr->pkey_index;
1559 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1560 	}
1561 
1562 	if (attr_mask & IB_QP_AV) {
1563 		if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
1564 				  attr_mask & IB_QP_PORT ?
1565 				  attr->port_num : qp->port))
1566 			goto out;
1567 
1568 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1569 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
1570 	}
1571 
1572 	if (attr_mask & IB_QP_TIMEOUT) {
1573 		context->pri_path.ackto |= attr->timeout << 3;
1574 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1575 	}
1576 
1577 	if (attr_mask & IB_QP_ALT_PATH) {
1578 		if (attr->alt_port_num == 0 ||
1579 		    attr->alt_port_num > dev->dev->caps.num_ports)
1580 			goto out;
1581 
1582 		if (attr->alt_pkey_index >=
1583 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
1584 			goto out;
1585 
1586 		if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
1587 				      &context->alt_path,
1588 				      attr->alt_port_num))
1589 			goto out;
1590 
1591 		context->alt_path.pkey_index = attr->alt_pkey_index;
1592 		context->alt_path.ackto = attr->alt_timeout << 3;
1593 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1594 	}
1595 
1596 	pd = get_pd(qp);
1597 	get_cqs(qp, &send_cq, &recv_cq);
1598 	context->pd       = cpu_to_be32(pd->pdn);
1599 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1600 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1601 	context->params1  = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1602 
1603 	/* Set "fast registration enabled" for all kernel QPs */
1604 	if (!qp->ibqp.uobject)
1605 		context->params1 |= cpu_to_be32(1 << 11);
1606 
1607 	if (attr_mask & IB_QP_RNR_RETRY) {
1608 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1609 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1610 	}
1611 
1612 	if (attr_mask & IB_QP_RETRY_CNT) {
1613 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1614 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1615 	}
1616 
1617 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1618 		if (attr->max_rd_atomic)
1619 			context->params1 |=
1620 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1621 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1622 	}
1623 
1624 	if (attr_mask & IB_QP_SQ_PSN)
1625 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1626 
1627 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1628 		if (attr->max_dest_rd_atomic)
1629 			context->params2 |=
1630 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1631 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1632 	}
1633 
1634 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1635 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1636 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1637 	}
1638 
1639 	if (ibqp->srq)
1640 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1641 
1642 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1643 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1644 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1645 	}
1646 	if (attr_mask & IB_QP_RQ_PSN)
1647 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1648 
1649 	/* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1650 	if (attr_mask & IB_QP_QKEY) {
1651 		if (qp->mlx4_ib_qp_type &
1652 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1653 			context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1654 		else {
1655 			if (mlx4_is_mfunc(dev->dev) &&
1656 			    !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1657 			    (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1658 			    MLX4_RESERVED_QKEY_BASE) {
1659 				pr_err("Cannot use reserved QKEY"
1660 				       " 0x%x (range 0xffff0000..0xffffffff"
1661 				       " is reserved)\n", attr->qkey);
1662 				err = -EINVAL;
1663 				goto out;
1664 			}
1665 			context->qkey = cpu_to_be32(attr->qkey);
1666 		}
1667 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
1668 	}
1669 
1670 	if (ibqp->srq)
1671 		context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1672 
1673 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1674 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1675 
1676 	if (cur_state == IB_QPS_INIT &&
1677 	    new_state == IB_QPS_RTR  &&
1678 	    (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1679 	     ibqp->qp_type == IB_QPT_UD ||
1680 	     ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1681 		context->pri_path.sched_queue = (qp->port - 1) << 6;
1682 		if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1683 		    qp->mlx4_ib_qp_type &
1684 		    (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1685 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1686 			if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1687 				context->pri_path.fl = 0x80;
1688 		} else {
1689 			if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1690 				context->pri_path.fl = 0x80;
1691 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1692 		}
1693 		if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1694 		    IB_LINK_LAYER_ETHERNET) {
1695 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
1696 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
1697 				context->pri_path.feup = 1 << 7; /* don't fsm */
1698 			/* handle smac_index */
1699 			if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
1700 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
1701 			    qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
1702 				err = handle_eth_ud_smac_index(dev, qp, (u8 *)attr->smac, context);
1703 				if (err) {
1704 					err = -EINVAL;
1705 					goto out;
1706 				}
1707 				if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1708 					dev->qp1_proxy[qp->port - 1] = qp;
1709 			}
1710 		}
1711 	}
1712 
1713 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1714 		context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
1715 					MLX4_IB_LINK_TYPE_ETH;
1716 		if (dev->dev->caps.tunnel_offload_mode ==  MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1717 			/* set QP to receive both tunneled & non-tunneled packets */
1718 			if (!(context->flags & cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET)))
1719 				context->srqn = cpu_to_be32(7 << 28);
1720 		}
1721 	}
1722 
1723 	if (ibqp->qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
1724 		int is_eth = rdma_port_get_link_layer(
1725 				&dev->ib_dev, qp->port) ==
1726 				IB_LINK_LAYER_ETHERNET;
1727 		if (is_eth) {
1728 			context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
1729 			optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
1730 		}
1731 	}
1732 
1733 
1734 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1735 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1736 		sqd_event = 1;
1737 	else
1738 		sqd_event = 0;
1739 
1740 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1741 		context->rlkey |= (1 << 4);
1742 
1743 	/*
1744 	 * Before passing a kernel QP to the HW, make sure that the
1745 	 * ownership bits of the send queue are set and the SQ
1746 	 * headroom is stamped so that the hardware doesn't start
1747 	 * processing stale work requests.
1748 	 */
1749 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1750 		struct mlx4_wqe_ctrl_seg *ctrl;
1751 		int i;
1752 
1753 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1754 			ctrl = get_send_wqe(qp, i);
1755 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
1756 			if (qp->sq_max_wqes_per_wr == 1)
1757 				ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1758 
1759 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1760 		}
1761 	}
1762 
1763 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1764 			     to_mlx4_state(new_state), context, optpar,
1765 			     sqd_event, &qp->mqp);
1766 	if (err)
1767 		goto out;
1768 
1769 	qp->state = new_state;
1770 
1771 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1772 		qp->atomic_rd_en = attr->qp_access_flags;
1773 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1774 		qp->resp_depth = attr->max_dest_rd_atomic;
1775 	if (attr_mask & IB_QP_PORT) {
1776 		qp->port = attr->port_num;
1777 		update_mcg_macs(dev, qp);
1778 	}
1779 	if (attr_mask & IB_QP_ALT_PATH)
1780 		qp->alt_port = attr->alt_port_num;
1781 
1782 	if (is_sqp(dev, qp))
1783 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1784 
1785 	/*
1786 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
1787 	 * QP0 to RESET or ERROR, bring the link back down.
1788 	 */
1789 	if (is_qp0(dev, qp)) {
1790 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1791 			if (mlx4_INIT_PORT(dev->dev, qp->port))
1792 				pr_warn("INIT_PORT failed for port %d\n",
1793 				       qp->port);
1794 
1795 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1796 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1797 			mlx4_CLOSE_PORT(dev->dev, qp->port);
1798 	}
1799 
1800 	/*
1801 	 * If we moved a kernel QP to RESET, clean up all old CQ
1802 	 * entries and reinitialize the QP.
1803 	 */
1804 	if (new_state == IB_QPS_RESET) {
1805 		if (!ibqp->uobject) {
1806 			mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1807 					 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
1808 			if (send_cq != recv_cq)
1809 				mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1810 
1811 			qp->rq.head = 0;
1812 			qp->rq.tail = 0;
1813 			qp->sq.head = 0;
1814 			qp->sq.tail = 0;
1815 			qp->sq_next_wqe = 0;
1816 			if (qp->rq.wqe_cnt)
1817 				*qp->db.db  = 0;
1818 
1819 			if (qp->flags & MLX4_IB_QP_NETIF)
1820 				mlx4_ib_steer_qp_reg(dev, qp, 0);
1821 		}
1822 		if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1823 			mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1824 			qp->pri.smac = 0;
1825 			qp->pri.smac_port = 0;
1826 		}
1827 		if (qp->alt.smac) {
1828 			mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1829 			qp->alt.smac = 0;
1830 		}
1831 		if (qp->pri.vid < 0x1000) {
1832 			mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1833 			qp->pri.vid = 0xFFFF;
1834 			qp->pri.candidate_vid = 0xFFFF;
1835 			qp->pri.update_vid = 0;
1836 		}
1837 
1838 		if (qp->alt.vid < 0x1000) {
1839 			mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1840 			qp->alt.vid = 0xFFFF;
1841 			qp->alt.candidate_vid = 0xFFFF;
1842 			qp->alt.update_vid = 0;
1843 		}
1844 	}
1845 out:
1846 	if (err && steer_qp)
1847 		mlx4_ib_steer_qp_reg(dev, qp, 0);
1848 	kfree(context);
1849 	if (qp->pri.candidate_smac ||
1850 	    (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
1851 		if (err) {
1852 			mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
1853 		} else {
1854 			if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
1855 				mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1856 			qp->pri.smac = qp->pri.candidate_smac;
1857 			qp->pri.smac_index = qp->pri.candidate_smac_index;
1858 			qp->pri.smac_port = qp->pri.candidate_smac_port;
1859 		}
1860 		qp->pri.candidate_smac = 0;
1861 		qp->pri.candidate_smac_index = 0;
1862 		qp->pri.candidate_smac_port = 0;
1863 	}
1864 	if (qp->alt.candidate_smac) {
1865 		if (err) {
1866 			mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
1867 		} else {
1868 			if (qp->alt.smac)
1869 				mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1870 			qp->alt.smac = qp->alt.candidate_smac;
1871 			qp->alt.smac_index = qp->alt.candidate_smac_index;
1872 			qp->alt.smac_port = qp->alt.candidate_smac_port;
1873 		}
1874 		qp->alt.candidate_smac = 0;
1875 		qp->alt.candidate_smac_index = 0;
1876 		qp->alt.candidate_smac_port = 0;
1877 	}
1878 
1879 	if (qp->pri.update_vid) {
1880 		if (err) {
1881 			if (qp->pri.candidate_vid < 0x1000)
1882 				mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
1883 						     qp->pri.candidate_vid);
1884 		} else {
1885 			if (qp->pri.vid < 0x1000)
1886 				mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
1887 						     qp->pri.vid);
1888 			qp->pri.vid = qp->pri.candidate_vid;
1889 			qp->pri.vlan_port = qp->pri.candidate_vlan_port;
1890 			qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
1891 		}
1892 		qp->pri.candidate_vid = 0xFFFF;
1893 		qp->pri.update_vid = 0;
1894 	}
1895 
1896 	if (qp->alt.update_vid) {
1897 		if (err) {
1898 			if (qp->alt.candidate_vid < 0x1000)
1899 				mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
1900 						     qp->alt.candidate_vid);
1901 		} else {
1902 			if (qp->alt.vid < 0x1000)
1903 				mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
1904 						     qp->alt.vid);
1905 			qp->alt.vid = qp->alt.candidate_vid;
1906 			qp->alt.vlan_port = qp->alt.candidate_vlan_port;
1907 			qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
1908 		}
1909 		qp->alt.candidate_vid = 0xFFFF;
1910 		qp->alt.update_vid = 0;
1911 	}
1912 
1913 	return err;
1914 }
1915 
1916 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1917 		      int attr_mask, struct ib_udata *udata)
1918 {
1919 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1920 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1921 	enum ib_qp_state cur_state, new_state;
1922 	int err = -EINVAL;
1923 	int ll;
1924 	mutex_lock(&qp->mutex);
1925 
1926 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1927 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1928 
1929 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1930 		ll = IB_LINK_LAYER_UNSPECIFIED;
1931 	} else {
1932 		int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1933 		ll = rdma_port_get_link_layer(&dev->ib_dev, port);
1934 	}
1935 
1936 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1937 				attr_mask, ll)) {
1938 		pr_debug("qpn 0x%x: invalid attribute mask specified "
1939 			 "for transition %d to %d. qp_type %d,"
1940 			 " attr_mask 0x%x\n",
1941 			 ibqp->qp_num, cur_state, new_state,
1942 			 ibqp->qp_type, attr_mask);
1943 		goto out;
1944 	}
1945 
1946 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
1947 		if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
1948 			if ((ibqp->qp_type == IB_QPT_RC) ||
1949 			    (ibqp->qp_type == IB_QPT_UD) ||
1950 			    (ibqp->qp_type == IB_QPT_UC) ||
1951 			    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
1952 			    (ibqp->qp_type == IB_QPT_XRC_INI)) {
1953 				attr->port_num = mlx4_ib_bond_next_port(dev);
1954 			}
1955 		} else {
1956 			/* no sense in changing port_num
1957 			 * when ports are bonded */
1958 			attr_mask &= ~IB_QP_PORT;
1959 		}
1960 	}
1961 
1962 	if ((attr_mask & IB_QP_PORT) &&
1963 	    (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
1964 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
1965 			 "for transition %d to %d. qp_type %d\n",
1966 			 ibqp->qp_num, attr->port_num, cur_state,
1967 			 new_state, ibqp->qp_type);
1968 		goto out;
1969 	}
1970 
1971 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
1972 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
1973 	     IB_LINK_LAYER_ETHERNET))
1974 		goto out;
1975 
1976 	if (attr_mask & IB_QP_PKEY_INDEX) {
1977 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1978 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
1979 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
1980 				 "for transition %d to %d. qp_type %d\n",
1981 				 ibqp->qp_num, attr->pkey_index, cur_state,
1982 				 new_state, ibqp->qp_type);
1983 			goto out;
1984 		}
1985 	}
1986 
1987 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1988 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
1989 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
1990 			 "Transition %d to %d. qp_type %d\n",
1991 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
1992 			 new_state, ibqp->qp_type);
1993 		goto out;
1994 	}
1995 
1996 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1997 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
1998 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
1999 			 "Transition %d to %d. qp_type %d\n",
2000 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2001 			 new_state, ibqp->qp_type);
2002 		goto out;
2003 	}
2004 
2005 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2006 		err = 0;
2007 		goto out;
2008 	}
2009 
2010 	err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2011 
2012 	if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2013 		attr->port_num = 1;
2014 
2015 out:
2016 	mutex_unlock(&qp->mutex);
2017 	return err;
2018 }
2019 
2020 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2021 {
2022 	int i;
2023 	for (i = 0; i < dev->caps.num_ports; i++) {
2024 		if (qpn == dev->caps.qp0_proxy[i] ||
2025 		    qpn == dev->caps.qp0_tunnel[i]) {
2026 			*qkey = dev->caps.qp0_qkey[i];
2027 			return 0;
2028 		}
2029 	}
2030 	return -EINVAL;
2031 }
2032 
2033 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2034 				  struct ib_send_wr *wr,
2035 				  void *wqe, unsigned *mlx_seg_len)
2036 {
2037 	struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2038 	struct ib_device *ib_dev = &mdev->ib_dev;
2039 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2040 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2041 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2042 	u16 pkey;
2043 	u32 qkey;
2044 	int send_size;
2045 	int header_size;
2046 	int spc;
2047 	int i;
2048 
2049 	if (wr->opcode != IB_WR_SEND)
2050 		return -EINVAL;
2051 
2052 	send_size = 0;
2053 
2054 	for (i = 0; i < wr->num_sge; ++i)
2055 		send_size += wr->sg_list[i].length;
2056 
2057 	/* for proxy-qp0 sends, need to add in size of tunnel header */
2058 	/* for tunnel-qp0 sends, tunnel header is already in s/g list */
2059 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2060 		send_size += sizeof (struct mlx4_ib_tunnel_header);
2061 
2062 	ib_ud_header_init(send_size, 1, 0, 0, 0, 0, &sqp->ud_header);
2063 
2064 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2065 		sqp->ud_header.lrh.service_level =
2066 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2067 		sqp->ud_header.lrh.destination_lid =
2068 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2069 		sqp->ud_header.lrh.source_lid =
2070 			cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2071 	}
2072 
2073 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2074 
2075 	/* force loopback */
2076 	mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2077 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
2078 
2079 	sqp->ud_header.lrh.virtual_lane    = 0;
2080 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
2081 	ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2082 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2083 	if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2084 		sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2085 	else
2086 		sqp->ud_header.bth.destination_qpn =
2087 			cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
2088 
2089 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2090 	if (mlx4_is_master(mdev->dev)) {
2091 		if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2092 			return -EINVAL;
2093 	} else {
2094 		if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2095 			return -EINVAL;
2096 	}
2097 	sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2098 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2099 
2100 	sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2101 	sqp->ud_header.immediate_present = 0;
2102 
2103 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2104 
2105 	/*
2106 	 * Inline data segments may not cross a 64 byte boundary.  If
2107 	 * our UD header is bigger than the space available up to the
2108 	 * next 64 byte boundary in the WQE, use two inline data
2109 	 * segments to hold the UD header.
2110 	 */
2111 	spc = MLX4_INLINE_ALIGN -
2112 	      ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2113 	if (header_size <= spc) {
2114 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2115 		memcpy(inl + 1, sqp->header_buf, header_size);
2116 		i = 1;
2117 	} else {
2118 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2119 		memcpy(inl + 1, sqp->header_buf, spc);
2120 
2121 		inl = (void *) (inl + 1) + spc;
2122 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2123 		/*
2124 		 * Need a barrier here to make sure all the data is
2125 		 * visible before the byte_count field is set.
2126 		 * Otherwise the HCA prefetcher could grab the 64-byte
2127 		 * chunk with this inline segment and get a valid (!=
2128 		 * 0xffffffff) byte count but stale data, and end up
2129 		 * generating a packet with bad headers.
2130 		 *
2131 		 * The first inline segment's byte_count field doesn't
2132 		 * need a barrier, because it comes after a
2133 		 * control/MLX segment and therefore is at an offset
2134 		 * of 16 mod 64.
2135 		 */
2136 		wmb();
2137 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2138 		i = 2;
2139 	}
2140 
2141 	*mlx_seg_len =
2142 	ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2143 	return 0;
2144 }
2145 
2146 static void mlx4_u64_to_smac(u8 *dst_mac, u64 src_mac)
2147 {
2148 	int i;
2149 
2150 	for (i = ETH_ALEN; i; i--) {
2151 		dst_mac[i - 1] = src_mac & 0xff;
2152 		src_mac >>= 8;
2153 	}
2154 }
2155 
2156 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
2157 			    void *wqe, unsigned *mlx_seg_len)
2158 {
2159 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
2160 	struct mlx4_wqe_mlx_seg *mlx = wqe;
2161 	struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2162 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2163 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2164 	union ib_gid sgid;
2165 	u16 pkey;
2166 	int send_size;
2167 	int header_size;
2168 	int spc;
2169 	int i;
2170 	int err = 0;
2171 	u16 vlan = 0xffff;
2172 	bool is_eth;
2173 	bool is_vlan = false;
2174 	bool is_grh;
2175 
2176 	send_size = 0;
2177 	for (i = 0; i < wr->num_sge; ++i)
2178 		send_size += wr->sg_list[i].length;
2179 
2180 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2181 	is_grh = mlx4_ib_ah_grh_present(ah);
2182 	if (is_eth) {
2183 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2184 			/* When multi-function is enabled, the ib_core gid
2185 			 * indexes don't necessarily match the hw ones, so
2186 			 * we must use our own cache */
2187 			err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2188 							   be32_to_cpu(ah->av.ib.port_pd) >> 24,
2189 							   ah->av.ib.gid_index, &sgid.raw[0]);
2190 			if (err)
2191 				return err;
2192 		} else  {
2193 			err = ib_get_cached_gid(ib_dev,
2194 						be32_to_cpu(ah->av.ib.port_pd) >> 24,
2195 						ah->av.ib.gid_index, &sgid);
2196 			if (err)
2197 				return err;
2198 		}
2199 
2200 		if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
2201 			vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
2202 			is_vlan = 1;
2203 		}
2204 	}
2205 	ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh, 0, &sqp->ud_header);
2206 
2207 	if (!is_eth) {
2208 		sqp->ud_header.lrh.service_level =
2209 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2210 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
2211 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2212 	}
2213 
2214 	if (is_grh) {
2215 		sqp->ud_header.grh.traffic_class =
2216 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2217 		sqp->ud_header.grh.flow_label    =
2218 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
2219 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
2220 		if (is_eth)
2221 			memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
2222 		else {
2223 		if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2224 			/* When multi-function is enabled, the ib_core gid
2225 			 * indexes don't necessarily match the hw ones, so
2226 			 * we must use our own cache */
2227 			sqp->ud_header.grh.source_gid.global.subnet_prefix =
2228 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2229 						       subnet_prefix;
2230 			sqp->ud_header.grh.source_gid.global.interface_id =
2231 				to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2232 					       guid_cache[ah->av.ib.gid_index];
2233 		} else
2234 			ib_get_cached_gid(ib_dev,
2235 					  be32_to_cpu(ah->av.ib.port_pd) >> 24,
2236 					  ah->av.ib.gid_index,
2237 					  &sqp->ud_header.grh.source_gid);
2238 		}
2239 		memcpy(sqp->ud_header.grh.destination_gid.raw,
2240 		       ah->av.ib.dgid, 16);
2241 	}
2242 
2243 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2244 
2245 	if (!is_eth) {
2246 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
2247 					  (sqp->ud_header.lrh.destination_lid ==
2248 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
2249 					  (sqp->ud_header.lrh.service_level << 8));
2250 		if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
2251 			mlx->flags |= cpu_to_be32(0x1); /* force loopback */
2252 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
2253 	}
2254 
2255 	switch (wr->opcode) {
2256 	case IB_WR_SEND:
2257 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
2258 		sqp->ud_header.immediate_present = 0;
2259 		break;
2260 	case IB_WR_SEND_WITH_IMM:
2261 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2262 		sqp->ud_header.immediate_present = 1;
2263 		sqp->ud_header.immediate_data    = wr->ex.imm_data;
2264 		break;
2265 	default:
2266 		return -EINVAL;
2267 	}
2268 
2269 	if (is_eth) {
2270 		struct in6_addr in6;
2271 
2272 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
2273 
2274 		mlx->sched_prio = cpu_to_be16(pcp);
2275 
2276 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
2277 		/* FIXME: cache smac value? */
2278 		memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
2279 		memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
2280 		memcpy(&in6, sgid.raw, sizeof(in6));
2281 
2282 		if (!mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2283 			u64 mac = atomic64_read(&to_mdev(ib_dev)->iboe.mac[sqp->qp.port - 1]);
2284 			u8 smac[ETH_ALEN];
2285 
2286 			mlx4_u64_to_smac(smac, mac);
2287 			memcpy(sqp->ud_header.eth.smac_h, smac, ETH_ALEN);
2288 		} else {
2289 			/* use the src mac of the tunnel */
2290 			memcpy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac, ETH_ALEN);
2291 		}
2292 
2293 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
2294 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
2295 		if (!is_vlan) {
2296 			sqp->ud_header.eth.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2297 		} else {
2298 			sqp->ud_header.vlan.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2299 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
2300 		}
2301 	} else {
2302 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
2303 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
2304 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
2305 	}
2306 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
2307 	if (!sqp->qp.ibqp.qp_num)
2308 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
2309 	else
2310 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
2311 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2312 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2313 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2314 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
2315 					       sqp->qkey : wr->wr.ud.remote_qkey);
2316 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
2317 
2318 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2319 
2320 	if (0) {
2321 		pr_err("built UD header of size %d:\n", header_size);
2322 		for (i = 0; i < header_size / 4; ++i) {
2323 			if (i % 8 == 0)
2324 				pr_err("  [%02x] ", i * 4);
2325 			pr_cont(" %08x",
2326 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
2327 			if ((i + 1) % 8 == 0)
2328 				pr_cont("\n");
2329 		}
2330 		pr_err("\n");
2331 	}
2332 
2333 	/*
2334 	 * Inline data segments may not cross a 64 byte boundary.  If
2335 	 * our UD header is bigger than the space available up to the
2336 	 * next 64 byte boundary in the WQE, use two inline data
2337 	 * segments to hold the UD header.
2338 	 */
2339 	spc = MLX4_INLINE_ALIGN -
2340 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2341 	if (header_size <= spc) {
2342 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2343 		memcpy(inl + 1, sqp->header_buf, header_size);
2344 		i = 1;
2345 	} else {
2346 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2347 		memcpy(inl + 1, sqp->header_buf, spc);
2348 
2349 		inl = (void *) (inl + 1) + spc;
2350 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2351 		/*
2352 		 * Need a barrier here to make sure all the data is
2353 		 * visible before the byte_count field is set.
2354 		 * Otherwise the HCA prefetcher could grab the 64-byte
2355 		 * chunk with this inline segment and get a valid (!=
2356 		 * 0xffffffff) byte count but stale data, and end up
2357 		 * generating a packet with bad headers.
2358 		 *
2359 		 * The first inline segment's byte_count field doesn't
2360 		 * need a barrier, because it comes after a
2361 		 * control/MLX segment and therefore is at an offset
2362 		 * of 16 mod 64.
2363 		 */
2364 		wmb();
2365 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2366 		i = 2;
2367 	}
2368 
2369 	*mlx_seg_len =
2370 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2371 	return 0;
2372 }
2373 
2374 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2375 {
2376 	unsigned cur;
2377 	struct mlx4_ib_cq *cq;
2378 
2379 	cur = wq->head - wq->tail;
2380 	if (likely(cur + nreq < wq->max_post))
2381 		return 0;
2382 
2383 	cq = to_mcq(ib_cq);
2384 	spin_lock(&cq->lock);
2385 	cur = wq->head - wq->tail;
2386 	spin_unlock(&cq->lock);
2387 
2388 	return cur + nreq >= wq->max_post;
2389 }
2390 
2391 static __be32 convert_access(int acc)
2392 {
2393 	return (acc & IB_ACCESS_REMOTE_ATOMIC ?
2394 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
2395 	       (acc & IB_ACCESS_REMOTE_WRITE  ?
2396 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
2397 	       (acc & IB_ACCESS_REMOTE_READ   ?
2398 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
2399 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
2400 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
2401 }
2402 
2403 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr)
2404 {
2405 	struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
2406 	int i;
2407 
2408 	for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i)
2409 		mfrpl->mapped_page_list[i] =
2410 			cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] |
2411 				    MLX4_MTT_FLAG_PRESENT);
2412 
2413 	fseg->flags		= convert_access(wr->wr.fast_reg.access_flags);
2414 	fseg->mem_key		= cpu_to_be32(wr->wr.fast_reg.rkey);
2415 	fseg->buf_list		= cpu_to_be64(mfrpl->map);
2416 	fseg->start_addr	= cpu_to_be64(wr->wr.fast_reg.iova_start);
2417 	fseg->reg_len		= cpu_to_be64(wr->wr.fast_reg.length);
2418 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
2419 	fseg->page_size		= cpu_to_be32(wr->wr.fast_reg.page_shift);
2420 	fseg->reserved[0]	= 0;
2421 	fseg->reserved[1]	= 0;
2422 }
2423 
2424 static void set_bind_seg(struct mlx4_wqe_bind_seg *bseg, struct ib_send_wr *wr)
2425 {
2426 	bseg->flags1 =
2427 		convert_access(wr->wr.bind_mw.bind_info.mw_access_flags) &
2428 		cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ  |
2429 			    MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE |
2430 			    MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC);
2431 	bseg->flags2 = 0;
2432 	if (wr->wr.bind_mw.mw->type == IB_MW_TYPE_2)
2433 		bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_TYPE_2);
2434 	if (wr->wr.bind_mw.bind_info.mw_access_flags & IB_ZERO_BASED)
2435 		bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_ZERO_BASED);
2436 	bseg->new_rkey = cpu_to_be32(wr->wr.bind_mw.rkey);
2437 	bseg->lkey = cpu_to_be32(wr->wr.bind_mw.bind_info.mr->lkey);
2438 	bseg->addr = cpu_to_be64(wr->wr.bind_mw.bind_info.addr);
2439 	bseg->length = cpu_to_be64(wr->wr.bind_mw.bind_info.length);
2440 }
2441 
2442 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2443 {
2444 	memset(iseg, 0, sizeof(*iseg));
2445 	iseg->mem_key = cpu_to_be32(rkey);
2446 }
2447 
2448 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2449 					  u64 remote_addr, u32 rkey)
2450 {
2451 	rseg->raddr    = cpu_to_be64(remote_addr);
2452 	rseg->rkey     = cpu_to_be32(rkey);
2453 	rseg->reserved = 0;
2454 }
2455 
2456 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
2457 {
2458 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2459 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
2460 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
2461 	} else if (wr->opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2462 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2463 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add_mask);
2464 	} else {
2465 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2466 		aseg->compare  = 0;
2467 	}
2468 
2469 }
2470 
2471 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2472 				  struct ib_send_wr *wr)
2473 {
2474 	aseg->swap_add		= cpu_to_be64(wr->wr.atomic.swap);
2475 	aseg->swap_add_mask	= cpu_to_be64(wr->wr.atomic.swap_mask);
2476 	aseg->compare		= cpu_to_be64(wr->wr.atomic.compare_add);
2477 	aseg->compare_mask	= cpu_to_be64(wr->wr.atomic.compare_add_mask);
2478 }
2479 
2480 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2481 			     struct ib_send_wr *wr)
2482 {
2483 	memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
2484 	dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2485 	dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2486 	dseg->vlan = to_mah(wr->wr.ud.ah)->av.eth.vlan;
2487 	memcpy(dseg->mac, to_mah(wr->wr.ud.ah)->av.eth.mac, 6);
2488 }
2489 
2490 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2491 				    struct mlx4_wqe_datagram_seg *dseg,
2492 				    struct ib_send_wr *wr,
2493 				    enum mlx4_ib_qp_type qpt)
2494 {
2495 	union mlx4_ext_av *av = &to_mah(wr->wr.ud.ah)->av;
2496 	struct mlx4_av sqp_av = {0};
2497 	int port = *((u8 *) &av->ib.port_pd) & 0x3;
2498 
2499 	/* force loopback */
2500 	sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2501 	sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2502 	sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2503 			cpu_to_be32(0xf0000000);
2504 
2505 	memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2506 	if (qpt == MLX4_IB_QPT_PROXY_GSI)
2507 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2508 	else
2509 		dseg->dqpn = cpu_to_be32(dev->dev->caps.qp0_tunnel[port - 1]);
2510 	/* Use QKEY from the QP context, which is set by master */
2511 	dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2512 }
2513 
2514 static void build_tunnel_header(struct ib_send_wr *wr, void *wqe, unsigned *mlx_seg_len)
2515 {
2516 	struct mlx4_wqe_inline_seg *inl = wqe;
2517 	struct mlx4_ib_tunnel_header hdr;
2518 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2519 	int spc;
2520 	int i;
2521 
2522 	memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2523 	hdr.remote_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2524 	hdr.pkey_index = cpu_to_be16(wr->wr.ud.pkey_index);
2525 	hdr.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2526 	memcpy(hdr.mac, ah->av.eth.mac, 6);
2527 	hdr.vlan = ah->av.eth.vlan;
2528 
2529 	spc = MLX4_INLINE_ALIGN -
2530 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2531 	if (sizeof (hdr) <= spc) {
2532 		memcpy(inl + 1, &hdr, sizeof (hdr));
2533 		wmb();
2534 		inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
2535 		i = 1;
2536 	} else {
2537 		memcpy(inl + 1, &hdr, spc);
2538 		wmb();
2539 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
2540 
2541 		inl = (void *) (inl + 1) + spc;
2542 		memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2543 		wmb();
2544 		inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
2545 		i = 2;
2546 	}
2547 
2548 	*mlx_seg_len =
2549 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2550 }
2551 
2552 static void set_mlx_icrc_seg(void *dseg)
2553 {
2554 	u32 *t = dseg;
2555 	struct mlx4_wqe_inline_seg *iseg = dseg;
2556 
2557 	t[1] = 0;
2558 
2559 	/*
2560 	 * Need a barrier here before writing the byte_count field to
2561 	 * make sure that all the data is visible before the
2562 	 * byte_count field is set.  Otherwise, if the segment begins
2563 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2564 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2565 	 * stale data, and end up sending the wrong data.
2566 	 */
2567 	wmb();
2568 
2569 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2570 }
2571 
2572 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2573 {
2574 	dseg->lkey       = cpu_to_be32(sg->lkey);
2575 	dseg->addr       = cpu_to_be64(sg->addr);
2576 
2577 	/*
2578 	 * Need a barrier here before writing the byte_count field to
2579 	 * make sure that all the data is visible before the
2580 	 * byte_count field is set.  Otherwise, if the segment begins
2581 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
2582 	 * chunk and get a valid (!= * 0xffffffff) byte count but
2583 	 * stale data, and end up sending the wrong data.
2584 	 */
2585 	wmb();
2586 
2587 	dseg->byte_count = cpu_to_be32(sg->length);
2588 }
2589 
2590 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2591 {
2592 	dseg->byte_count = cpu_to_be32(sg->length);
2593 	dseg->lkey       = cpu_to_be32(sg->lkey);
2594 	dseg->addr       = cpu_to_be64(sg->addr);
2595 }
2596 
2597 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr,
2598 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2599 			 __be32 *lso_hdr_sz, __be32 *blh)
2600 {
2601 	unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
2602 
2603 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2604 		*blh = cpu_to_be32(1 << 6);
2605 
2606 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2607 		     wr->num_sge > qp->sq.max_gs - (halign >> 4)))
2608 		return -EINVAL;
2609 
2610 	memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
2611 
2612 	*lso_hdr_sz  = cpu_to_be32(wr->wr.ud.mss << 16 | wr->wr.ud.hlen);
2613 	*lso_seg_len = halign;
2614 	return 0;
2615 }
2616 
2617 static __be32 send_ieth(struct ib_send_wr *wr)
2618 {
2619 	switch (wr->opcode) {
2620 	case IB_WR_SEND_WITH_IMM:
2621 	case IB_WR_RDMA_WRITE_WITH_IMM:
2622 		return wr->ex.imm_data;
2623 
2624 	case IB_WR_SEND_WITH_INV:
2625 		return cpu_to_be32(wr->ex.invalidate_rkey);
2626 
2627 	default:
2628 		return 0;
2629 	}
2630 }
2631 
2632 static void add_zero_len_inline(void *wqe)
2633 {
2634 	struct mlx4_wqe_inline_seg *inl = wqe;
2635 	memset(wqe, 0, 16);
2636 	inl->byte_count = cpu_to_be32(1 << 31);
2637 }
2638 
2639 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2640 		      struct ib_send_wr **bad_wr)
2641 {
2642 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2643 	void *wqe;
2644 	struct mlx4_wqe_ctrl_seg *ctrl;
2645 	struct mlx4_wqe_data_seg *dseg;
2646 	unsigned long flags;
2647 	int nreq;
2648 	int err = 0;
2649 	unsigned ind;
2650 	int uninitialized_var(stamp);
2651 	int uninitialized_var(size);
2652 	unsigned uninitialized_var(seglen);
2653 	__be32 dummy;
2654 	__be32 *lso_wqe;
2655 	__be32 uninitialized_var(lso_hdr_sz);
2656 	__be32 blh;
2657 	int i;
2658 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
2659 
2660 	spin_lock_irqsave(&qp->sq.lock, flags);
2661 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
2662 		err = -EIO;
2663 		*bad_wr = wr;
2664 		nreq = 0;
2665 		goto out;
2666 	}
2667 
2668 	ind = qp->sq_next_wqe;
2669 
2670 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2671 		lso_wqe = &dummy;
2672 		blh = 0;
2673 
2674 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2675 			err = -ENOMEM;
2676 			*bad_wr = wr;
2677 			goto out;
2678 		}
2679 
2680 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
2681 			err = -EINVAL;
2682 			*bad_wr = wr;
2683 			goto out;
2684 		}
2685 
2686 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
2687 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
2688 
2689 		ctrl->srcrb_flags =
2690 			(wr->send_flags & IB_SEND_SIGNALED ?
2691 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
2692 			(wr->send_flags & IB_SEND_SOLICITED ?
2693 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
2694 			((wr->send_flags & IB_SEND_IP_CSUM) ?
2695 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
2696 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
2697 			qp->sq_signal_bits;
2698 
2699 		ctrl->imm = send_ieth(wr);
2700 
2701 		wqe += sizeof *ctrl;
2702 		size = sizeof *ctrl / 16;
2703 
2704 		switch (qp->mlx4_ib_qp_type) {
2705 		case MLX4_IB_QPT_RC:
2706 		case MLX4_IB_QPT_UC:
2707 			switch (wr->opcode) {
2708 			case IB_WR_ATOMIC_CMP_AND_SWP:
2709 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2710 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
2711 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2712 					      wr->wr.atomic.rkey);
2713 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2714 
2715 				set_atomic_seg(wqe, wr);
2716 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
2717 
2718 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2719 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
2720 
2721 				break;
2722 
2723 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2724 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2725 					      wr->wr.atomic.rkey);
2726 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2727 
2728 				set_masked_atomic_seg(wqe, wr);
2729 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
2730 
2731 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
2732 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
2733 
2734 				break;
2735 
2736 			case IB_WR_RDMA_READ:
2737 			case IB_WR_RDMA_WRITE:
2738 			case IB_WR_RDMA_WRITE_WITH_IMM:
2739 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2740 					      wr->wr.rdma.rkey);
2741 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2742 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
2743 				break;
2744 
2745 			case IB_WR_LOCAL_INV:
2746 				ctrl->srcrb_flags |=
2747 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2748 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
2749 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
2750 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
2751 				break;
2752 
2753 			case IB_WR_FAST_REG_MR:
2754 				ctrl->srcrb_flags |=
2755 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2756 				set_fmr_seg(wqe, wr);
2757 				wqe  += sizeof (struct mlx4_wqe_fmr_seg);
2758 				size += sizeof (struct mlx4_wqe_fmr_seg) / 16;
2759 				break;
2760 
2761 			case IB_WR_BIND_MW:
2762 				ctrl->srcrb_flags |=
2763 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2764 				set_bind_seg(wqe, wr);
2765 				wqe  += sizeof(struct mlx4_wqe_bind_seg);
2766 				size += sizeof(struct mlx4_wqe_bind_seg) / 16;
2767 				break;
2768 			default:
2769 				/* No extra segments required for sends */
2770 				break;
2771 			}
2772 			break;
2773 
2774 		case MLX4_IB_QPT_TUN_SMI_OWNER:
2775 			err =  build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2776 			if (unlikely(err)) {
2777 				*bad_wr = wr;
2778 				goto out;
2779 			}
2780 			wqe  += seglen;
2781 			size += seglen / 16;
2782 			break;
2783 		case MLX4_IB_QPT_TUN_SMI:
2784 		case MLX4_IB_QPT_TUN_GSI:
2785 			/* this is a UD qp used in MAD responses to slaves. */
2786 			set_datagram_seg(wqe, wr);
2787 			/* set the forced-loopback bit in the data seg av */
2788 			*(__be32 *) wqe |= cpu_to_be32(0x80000000);
2789 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2790 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2791 			break;
2792 		case MLX4_IB_QPT_UD:
2793 			set_datagram_seg(wqe, wr);
2794 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2795 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2796 
2797 			if (wr->opcode == IB_WR_LSO) {
2798 				err = build_lso_seg(wqe, wr, qp, &seglen, &lso_hdr_sz, &blh);
2799 				if (unlikely(err)) {
2800 					*bad_wr = wr;
2801 					goto out;
2802 				}
2803 				lso_wqe = (__be32 *) wqe;
2804 				wqe  += seglen;
2805 				size += seglen / 16;
2806 			}
2807 			break;
2808 
2809 		case MLX4_IB_QPT_PROXY_SMI_OWNER:
2810 			err = build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2811 			if (unlikely(err)) {
2812 				*bad_wr = wr;
2813 				goto out;
2814 			}
2815 			wqe  += seglen;
2816 			size += seglen / 16;
2817 			/* to start tunnel header on a cache-line boundary */
2818 			add_zero_len_inline(wqe);
2819 			wqe += 16;
2820 			size++;
2821 			build_tunnel_header(wr, wqe, &seglen);
2822 			wqe  += seglen;
2823 			size += seglen / 16;
2824 			break;
2825 		case MLX4_IB_QPT_PROXY_SMI:
2826 		case MLX4_IB_QPT_PROXY_GSI:
2827 			/* If we are tunneling special qps, this is a UD qp.
2828 			 * In this case we first add a UD segment targeting
2829 			 * the tunnel qp, and then add a header with address
2830 			 * information */
2831 			set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe, wr,
2832 						qp->mlx4_ib_qp_type);
2833 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2834 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2835 			build_tunnel_header(wr, wqe, &seglen);
2836 			wqe  += seglen;
2837 			size += seglen / 16;
2838 			break;
2839 
2840 		case MLX4_IB_QPT_SMI:
2841 		case MLX4_IB_QPT_GSI:
2842 			err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
2843 			if (unlikely(err)) {
2844 				*bad_wr = wr;
2845 				goto out;
2846 			}
2847 			wqe  += seglen;
2848 			size += seglen / 16;
2849 			break;
2850 
2851 		default:
2852 			break;
2853 		}
2854 
2855 		/*
2856 		 * Write data segments in reverse order, so as to
2857 		 * overwrite cacheline stamp last within each
2858 		 * cacheline.  This avoids issues with WQE
2859 		 * prefetching.
2860 		 */
2861 
2862 		dseg = wqe;
2863 		dseg += wr->num_sge - 1;
2864 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
2865 
2866 		/* Add one more inline data segment for ICRC for MLX sends */
2867 		if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2868 			     qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
2869 			     qp->mlx4_ib_qp_type &
2870 			     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
2871 			set_mlx_icrc_seg(dseg + 1);
2872 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
2873 		}
2874 
2875 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
2876 			set_data_seg(dseg, wr->sg_list + i);
2877 
2878 		/*
2879 		 * Possibly overwrite stamping in cacheline with LSO
2880 		 * segment only after making sure all data segments
2881 		 * are written.
2882 		 */
2883 		wmb();
2884 		*lso_wqe = lso_hdr_sz;
2885 
2886 		ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
2887 				    MLX4_WQE_CTRL_FENCE : 0) | size;
2888 
2889 		/*
2890 		 * Make sure descriptor is fully written before
2891 		 * setting ownership bit (because HW can start
2892 		 * executing as soon as we do).
2893 		 */
2894 		wmb();
2895 
2896 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
2897 			*bad_wr = wr;
2898 			err = -EINVAL;
2899 			goto out;
2900 		}
2901 
2902 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
2903 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
2904 
2905 		stamp = ind + qp->sq_spare_wqes;
2906 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
2907 
2908 		/*
2909 		 * We can improve latency by not stamping the last
2910 		 * send queue WQE until after ringing the doorbell, so
2911 		 * only stamp here if there are still more WQEs to post.
2912 		 *
2913 		 * Same optimization applies to padding with NOP wqe
2914 		 * in case of WQE shrinking (used to prevent wrap-around
2915 		 * in the middle of WR).
2916 		 */
2917 		if (wr->next) {
2918 			stamp_send_wqe(qp, stamp, size * 16);
2919 			ind = pad_wraparound(qp, ind);
2920 		}
2921 	}
2922 
2923 out:
2924 	if (likely(nreq)) {
2925 		qp->sq.head += nreq;
2926 
2927 		/*
2928 		 * Make sure that descriptors are written before
2929 		 * doorbell record.
2930 		 */
2931 		wmb();
2932 
2933 		writel(qp->doorbell_qpn,
2934 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
2935 
2936 		/*
2937 		 * Make sure doorbells don't leak out of SQ spinlock
2938 		 * and reach the HCA out of order.
2939 		 */
2940 		mmiowb();
2941 
2942 		stamp_send_wqe(qp, stamp, size * 16);
2943 
2944 		ind = pad_wraparound(qp, ind);
2945 		qp->sq_next_wqe = ind;
2946 	}
2947 
2948 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2949 
2950 	return err;
2951 }
2952 
2953 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2954 		      struct ib_recv_wr **bad_wr)
2955 {
2956 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2957 	struct mlx4_wqe_data_seg *scat;
2958 	unsigned long flags;
2959 	int err = 0;
2960 	int nreq;
2961 	int ind;
2962 	int max_gs;
2963 	int i;
2964 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
2965 
2966 	max_gs = qp->rq.max_gs;
2967 	spin_lock_irqsave(&qp->rq.lock, flags);
2968 
2969 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
2970 		err = -EIO;
2971 		*bad_wr = wr;
2972 		nreq = 0;
2973 		goto out;
2974 	}
2975 
2976 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
2977 
2978 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2979 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2980 			err = -ENOMEM;
2981 			*bad_wr = wr;
2982 			goto out;
2983 		}
2984 
2985 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2986 			err = -EINVAL;
2987 			*bad_wr = wr;
2988 			goto out;
2989 		}
2990 
2991 		scat = get_recv_wqe(qp, ind);
2992 
2993 		if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
2994 		    MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
2995 			ib_dma_sync_single_for_device(ibqp->device,
2996 						      qp->sqp_proxy_rcv[ind].map,
2997 						      sizeof (struct mlx4_ib_proxy_sqp_hdr),
2998 						      DMA_FROM_DEVICE);
2999 			scat->byte_count =
3000 				cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3001 			/* use dma lkey from upper layer entry */
3002 			scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3003 			scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3004 			scat++;
3005 			max_gs--;
3006 		}
3007 
3008 		for (i = 0; i < wr->num_sge; ++i)
3009 			__set_data_seg(scat + i, wr->sg_list + i);
3010 
3011 		if (i < max_gs) {
3012 			scat[i].byte_count = 0;
3013 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
3014 			scat[i].addr       = 0;
3015 		}
3016 
3017 		qp->rq.wrid[ind] = wr->wr_id;
3018 
3019 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3020 	}
3021 
3022 out:
3023 	if (likely(nreq)) {
3024 		qp->rq.head += nreq;
3025 
3026 		/*
3027 		 * Make sure that descriptors are written before
3028 		 * doorbell record.
3029 		 */
3030 		wmb();
3031 
3032 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3033 	}
3034 
3035 	spin_unlock_irqrestore(&qp->rq.lock, flags);
3036 
3037 	return err;
3038 }
3039 
3040 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3041 {
3042 	switch (mlx4_state) {
3043 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
3044 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
3045 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
3046 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
3047 	case MLX4_QP_STATE_SQ_DRAINING:
3048 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
3049 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
3050 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
3051 	default:		     return -1;
3052 	}
3053 }
3054 
3055 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3056 {
3057 	switch (mlx4_mig_state) {
3058 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
3059 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
3060 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
3061 	default: return -1;
3062 	}
3063 }
3064 
3065 static int to_ib_qp_access_flags(int mlx4_flags)
3066 {
3067 	int ib_flags = 0;
3068 
3069 	if (mlx4_flags & MLX4_QP_BIT_RRE)
3070 		ib_flags |= IB_ACCESS_REMOTE_READ;
3071 	if (mlx4_flags & MLX4_QP_BIT_RWE)
3072 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
3073 	if (mlx4_flags & MLX4_QP_BIT_RAE)
3074 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3075 
3076 	return ib_flags;
3077 }
3078 
3079 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
3080 				struct mlx4_qp_path *path)
3081 {
3082 	struct mlx4_dev *dev = ibdev->dev;
3083 	int is_eth;
3084 
3085 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
3086 	ib_ah_attr->port_num	  = path->sched_queue & 0x40 ? 2 : 1;
3087 
3088 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
3089 		return;
3090 
3091 	is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
3092 		IB_LINK_LAYER_ETHERNET;
3093 	if (is_eth)
3094 		ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
3095 		((path->sched_queue & 4) << 1);
3096 	else
3097 		ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
3098 
3099 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
3100 	ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
3101 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
3102 	ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
3103 	if (ib_ah_attr->ah_flags) {
3104 		ib_ah_attr->grh.sgid_index = path->mgid_index;
3105 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
3106 		ib_ah_attr->grh.traffic_class =
3107 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3108 		ib_ah_attr->grh.flow_label =
3109 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3110 		memcpy(ib_ah_attr->grh.dgid.raw,
3111 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
3112 	}
3113 }
3114 
3115 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3116 		     struct ib_qp_init_attr *qp_init_attr)
3117 {
3118 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3119 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
3120 	struct mlx4_qp_context context;
3121 	int mlx4_state;
3122 	int err = 0;
3123 
3124 	mutex_lock(&qp->mutex);
3125 
3126 	if (qp->state == IB_QPS_RESET) {
3127 		qp_attr->qp_state = IB_QPS_RESET;
3128 		goto done;
3129 	}
3130 
3131 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3132 	if (err) {
3133 		err = -EINVAL;
3134 		goto out;
3135 	}
3136 
3137 	mlx4_state = be32_to_cpu(context.flags) >> 28;
3138 
3139 	qp->state		     = to_ib_qp_state(mlx4_state);
3140 	qp_attr->qp_state	     = qp->state;
3141 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
3142 	qp_attr->path_mig_state	     =
3143 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3144 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
3145 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3146 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
3147 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
3148 	qp_attr->qp_access_flags     =
3149 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
3150 
3151 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3152 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3153 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3154 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3155 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
3156 	}
3157 
3158 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3159 	if (qp_attr->qp_state == IB_QPS_INIT)
3160 		qp_attr->port_num = qp->port;
3161 	else
3162 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3163 
3164 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3165 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3166 
3167 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3168 
3169 	qp_attr->max_dest_rd_atomic =
3170 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3171 	qp_attr->min_rnr_timer	    =
3172 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3173 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
3174 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
3175 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
3176 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
3177 
3178 done:
3179 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
3180 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
3181 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
3182 
3183 	if (!ibqp->uobject) {
3184 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
3185 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
3186 	} else {
3187 		qp_attr->cap.max_send_wr  = 0;
3188 		qp_attr->cap.max_send_sge = 0;
3189 	}
3190 
3191 	/*
3192 	 * We don't support inline sends for kernel QPs (yet), and we
3193 	 * don't know what userspace's value should be.
3194 	 */
3195 	qp_attr->cap.max_inline_data = 0;
3196 
3197 	qp_init_attr->cap	     = qp_attr->cap;
3198 
3199 	qp_init_attr->create_flags = 0;
3200 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3201 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3202 
3203 	if (qp->flags & MLX4_IB_QP_LSO)
3204 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
3205 
3206 	if (qp->flags & MLX4_IB_QP_NETIF)
3207 		qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
3208 
3209 	qp_init_attr->sq_sig_type =
3210 		qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
3211 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3212 
3213 out:
3214 	mutex_unlock(&qp->mutex);
3215 	return err;
3216 }
3217 
3218