xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision 95e9fd10)
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 
42 #include <linux/mlx4/qp.h>
43 
44 #include "mlx4_ib.h"
45 #include "user.h"
46 
47 enum {
48 	MLX4_IB_ACK_REQ_FREQ	= 8,
49 };
50 
51 enum {
52 	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
53 	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
54 	MLX4_IB_LINK_TYPE_IB		= 0,
55 	MLX4_IB_LINK_TYPE_ETH		= 1
56 };
57 
58 enum {
59 	/*
60 	 * Largest possible UD header: send with GRH and immediate
61 	 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
62 	 * tag.  (LRH would only use 8 bytes, so Ethernet is the
63 	 * biggest case)
64 	 */
65 	MLX4_IB_UD_HEADER_SIZE		= 82,
66 	MLX4_IB_LSO_HEADER_SPARE	= 128,
67 };
68 
69 enum {
70 	MLX4_IB_IBOE_ETHERTYPE		= 0x8915
71 };
72 
73 struct mlx4_ib_sqp {
74 	struct mlx4_ib_qp	qp;
75 	int			pkey_index;
76 	u32			qkey;
77 	u32			send_psn;
78 	struct ib_ud_header	ud_header;
79 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
80 };
81 
82 enum {
83 	MLX4_IB_MIN_SQ_STRIDE	= 6,
84 	MLX4_IB_CACHE_LINE_SIZE	= 64,
85 };
86 
87 enum {
88 	MLX4_RAW_QP_MTU		= 7,
89 	MLX4_RAW_QP_MSGMAX	= 31,
90 };
91 
92 static const __be32 mlx4_ib_opcode[] = {
93 	[IB_WR_SEND]				= cpu_to_be32(MLX4_OPCODE_SEND),
94 	[IB_WR_LSO]				= cpu_to_be32(MLX4_OPCODE_LSO),
95 	[IB_WR_SEND_WITH_IMM]			= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
96 	[IB_WR_RDMA_WRITE]			= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
97 	[IB_WR_RDMA_WRITE_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
98 	[IB_WR_RDMA_READ]			= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
99 	[IB_WR_ATOMIC_CMP_AND_SWP]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
100 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
101 	[IB_WR_SEND_WITH_INV]			= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
102 	[IB_WR_LOCAL_INV]			= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
103 	[IB_WR_FAST_REG_MR]			= cpu_to_be32(MLX4_OPCODE_FMR),
104 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
105 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
106 };
107 
108 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
109 {
110 	return container_of(mqp, struct mlx4_ib_sqp, qp);
111 }
112 
113 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
114 {
115 	return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
116 		qp->mqp.qpn <= dev->dev->caps.sqp_start + 3;
117 }
118 
119 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
120 {
121 	return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
122 		qp->mqp.qpn <= dev->dev->caps.sqp_start + 1;
123 }
124 
125 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
126 {
127 	return mlx4_buf_offset(&qp->buf, offset);
128 }
129 
130 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
131 {
132 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
133 }
134 
135 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
136 {
137 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
138 }
139 
140 /*
141  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
142  * first four bytes of every 64 byte chunk with
143  *     0x7FFFFFF | (invalid_ownership_value << 31).
144  *
145  * When the max work request size is less than or equal to the WQE
146  * basic block size, as an optimization, we can stamp all WQEs with
147  * 0xffffffff, and skip the very first chunk of each WQE.
148  */
149 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
150 {
151 	__be32 *wqe;
152 	int i;
153 	int s;
154 	int ind;
155 	void *buf;
156 	__be32 stamp;
157 	struct mlx4_wqe_ctrl_seg *ctrl;
158 
159 	if (qp->sq_max_wqes_per_wr > 1) {
160 		s = roundup(size, 1U << qp->sq.wqe_shift);
161 		for (i = 0; i < s; i += 64) {
162 			ind = (i >> qp->sq.wqe_shift) + n;
163 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
164 						       cpu_to_be32(0xffffffff);
165 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
166 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
167 			*wqe = stamp;
168 		}
169 	} else {
170 		ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
171 		s = (ctrl->fence_size & 0x3f) << 4;
172 		for (i = 64; i < s; i += 64) {
173 			wqe = buf + i;
174 			*wqe = cpu_to_be32(0xffffffff);
175 		}
176 	}
177 }
178 
179 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
180 {
181 	struct mlx4_wqe_ctrl_seg *ctrl;
182 	struct mlx4_wqe_inline_seg *inl;
183 	void *wqe;
184 	int s;
185 
186 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
187 	s = sizeof(struct mlx4_wqe_ctrl_seg);
188 
189 	if (qp->ibqp.qp_type == IB_QPT_UD) {
190 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
191 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
192 		memset(dgram, 0, sizeof *dgram);
193 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
194 		s += sizeof(struct mlx4_wqe_datagram_seg);
195 	}
196 
197 	/* Pad the remainder of the WQE with an inline data segment. */
198 	if (size > s) {
199 		inl = wqe + s;
200 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
201 	}
202 	ctrl->srcrb_flags = 0;
203 	ctrl->fence_size = size / 16;
204 	/*
205 	 * Make sure descriptor is fully written before setting ownership bit
206 	 * (because HW can start executing as soon as we do).
207 	 */
208 	wmb();
209 
210 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
211 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
212 
213 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
214 }
215 
216 /* Post NOP WQE to prevent wrap-around in the middle of WR */
217 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
218 {
219 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
220 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
221 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
222 		ind += s;
223 	}
224 	return ind;
225 }
226 
227 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
228 {
229 	struct ib_event event;
230 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
231 
232 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
233 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
234 
235 	if (ibqp->event_handler) {
236 		event.device     = ibqp->device;
237 		event.element.qp = ibqp;
238 		switch (type) {
239 		case MLX4_EVENT_TYPE_PATH_MIG:
240 			event.event = IB_EVENT_PATH_MIG;
241 			break;
242 		case MLX4_EVENT_TYPE_COMM_EST:
243 			event.event = IB_EVENT_COMM_EST;
244 			break;
245 		case MLX4_EVENT_TYPE_SQ_DRAINED:
246 			event.event = IB_EVENT_SQ_DRAINED;
247 			break;
248 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
249 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
250 			break;
251 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
252 			event.event = IB_EVENT_QP_FATAL;
253 			break;
254 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
255 			event.event = IB_EVENT_PATH_MIG_ERR;
256 			break;
257 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
258 			event.event = IB_EVENT_QP_REQ_ERR;
259 			break;
260 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
261 			event.event = IB_EVENT_QP_ACCESS_ERR;
262 			break;
263 		default:
264 			pr_warn("Unexpected event type %d "
265 			       "on QP %06x\n", type, qp->qpn);
266 			return;
267 		}
268 
269 		ibqp->event_handler(&event, ibqp->qp_context);
270 	}
271 }
272 
273 static int send_wqe_overhead(enum ib_qp_type type, u32 flags)
274 {
275 	/*
276 	 * UD WQEs must have a datagram segment.
277 	 * RC and UC WQEs might have a remote address segment.
278 	 * MLX WQEs need two extra inline data segments (for the UD
279 	 * header and space for the ICRC).
280 	 */
281 	switch (type) {
282 	case IB_QPT_UD:
283 		return sizeof (struct mlx4_wqe_ctrl_seg) +
284 			sizeof (struct mlx4_wqe_datagram_seg) +
285 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
286 	case IB_QPT_UC:
287 		return sizeof (struct mlx4_wqe_ctrl_seg) +
288 			sizeof (struct mlx4_wqe_raddr_seg);
289 	case IB_QPT_RC:
290 		return sizeof (struct mlx4_wqe_ctrl_seg) +
291 			sizeof (struct mlx4_wqe_atomic_seg) +
292 			sizeof (struct mlx4_wqe_raddr_seg);
293 	case IB_QPT_SMI:
294 	case IB_QPT_GSI:
295 		return sizeof (struct mlx4_wqe_ctrl_seg) +
296 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
297 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
298 					   MLX4_INLINE_ALIGN) *
299 			      sizeof (struct mlx4_wqe_inline_seg),
300 			      sizeof (struct mlx4_wqe_data_seg)) +
301 			ALIGN(4 +
302 			      sizeof (struct mlx4_wqe_inline_seg),
303 			      sizeof (struct mlx4_wqe_data_seg));
304 	default:
305 		return sizeof (struct mlx4_wqe_ctrl_seg);
306 	}
307 }
308 
309 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
310 		       int is_user, int has_rq, struct mlx4_ib_qp *qp)
311 {
312 	/* Sanity check RQ size before proceeding */
313 	if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
314 	    cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
315 		return -EINVAL;
316 
317 	if (!has_rq) {
318 		if (cap->max_recv_wr)
319 			return -EINVAL;
320 
321 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
322 	} else {
323 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
324 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
325 			return -EINVAL;
326 
327 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
328 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
329 		qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
330 	}
331 
332 	/* leave userspace return values as they were, so as not to break ABI */
333 	if (is_user) {
334 		cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
335 		cap->max_recv_sge = qp->rq.max_gs;
336 	} else {
337 		cap->max_recv_wr  = qp->rq.max_post =
338 			min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
339 		cap->max_recv_sge = min(qp->rq.max_gs,
340 					min(dev->dev->caps.max_sq_sg,
341 					    dev->dev->caps.max_rq_sg));
342 	}
343 
344 	return 0;
345 }
346 
347 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
348 			      enum ib_qp_type type, struct mlx4_ib_qp *qp)
349 {
350 	int s;
351 
352 	/* Sanity check SQ size before proceeding */
353 	if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
354 	    cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
355 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
356 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
357 		return -EINVAL;
358 
359 	/*
360 	 * For MLX transport we need 2 extra S/G entries:
361 	 * one for the header and one for the checksum at the end
362 	 */
363 	if ((type == IB_QPT_SMI || type == IB_QPT_GSI) &&
364 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
365 		return -EINVAL;
366 
367 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
368 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
369 		send_wqe_overhead(type, qp->flags);
370 
371 	if (s > dev->dev->caps.max_sq_desc_sz)
372 		return -EINVAL;
373 
374 	/*
375 	 * Hermon supports shrinking WQEs, such that a single work
376 	 * request can include multiple units of 1 << wqe_shift.  This
377 	 * way, work requests can differ in size, and do not have to
378 	 * be a power of 2 in size, saving memory and speeding up send
379 	 * WR posting.  Unfortunately, if we do this then the
380 	 * wqe_index field in CQEs can't be used to look up the WR ID
381 	 * anymore, so we do this only if selective signaling is off.
382 	 *
383 	 * Further, on 32-bit platforms, we can't use vmap() to make
384 	 * the QP buffer virtually contiguous.  Thus we have to use
385 	 * constant-sized WRs to make sure a WR is always fully within
386 	 * a single page-sized chunk.
387 	 *
388 	 * Finally, we use NOP work requests to pad the end of the
389 	 * work queue, to avoid wrap-around in the middle of WR.  We
390 	 * set NEC bit to avoid getting completions with error for
391 	 * these NOP WRs, but since NEC is only supported starting
392 	 * with firmware 2.2.232, we use constant-sized WRs for older
393 	 * firmware.
394 	 *
395 	 * And, since MLX QPs only support SEND, we use constant-sized
396 	 * WRs in this case.
397 	 *
398 	 * We look for the smallest value of wqe_shift such that the
399 	 * resulting number of wqes does not exceed device
400 	 * capabilities.
401 	 *
402 	 * We set WQE size to at least 64 bytes, this way stamping
403 	 * invalidates each WQE.
404 	 */
405 	if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
406 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
407 	    type != IB_QPT_SMI && type != IB_QPT_GSI)
408 		qp->sq.wqe_shift = ilog2(64);
409 	else
410 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
411 
412 	for (;;) {
413 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
414 
415 		/*
416 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
417 		 * allow HW to prefetch.
418 		 */
419 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
420 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
421 						    qp->sq_max_wqes_per_wr +
422 						    qp->sq_spare_wqes);
423 
424 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
425 			break;
426 
427 		if (qp->sq_max_wqes_per_wr <= 1)
428 			return -EINVAL;
429 
430 		++qp->sq.wqe_shift;
431 	}
432 
433 	qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
434 			     (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
435 			 send_wqe_overhead(type, qp->flags)) /
436 		sizeof (struct mlx4_wqe_data_seg);
437 
438 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
439 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
440 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
441 		qp->rq.offset = 0;
442 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
443 	} else {
444 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
445 		qp->sq.offset = 0;
446 	}
447 
448 	cap->max_send_wr  = qp->sq.max_post =
449 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
450 	cap->max_send_sge = min(qp->sq.max_gs,
451 				min(dev->dev->caps.max_sq_sg,
452 				    dev->dev->caps.max_rq_sg));
453 	/* We don't support inline sends for kernel QPs (yet) */
454 	cap->max_inline_data = 0;
455 
456 	return 0;
457 }
458 
459 static int set_user_sq_size(struct mlx4_ib_dev *dev,
460 			    struct mlx4_ib_qp *qp,
461 			    struct mlx4_ib_create_qp *ucmd)
462 {
463 	/* Sanity check SQ size before proceeding */
464 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
465 	    ucmd->log_sq_stride >
466 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
467 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
468 		return -EINVAL;
469 
470 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
471 	qp->sq.wqe_shift = ucmd->log_sq_stride;
472 
473 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
474 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
475 
476 	return 0;
477 }
478 
479 static int qp_has_rq(struct ib_qp_init_attr *attr)
480 {
481 	if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
482 		return 0;
483 
484 	return !attr->srq;
485 }
486 
487 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
488 			    struct ib_qp_init_attr *init_attr,
489 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp *qp)
490 {
491 	int qpn;
492 	int err;
493 
494 	mutex_init(&qp->mutex);
495 	spin_lock_init(&qp->sq.lock);
496 	spin_lock_init(&qp->rq.lock);
497 	INIT_LIST_HEAD(&qp->gid_list);
498 	INIT_LIST_HEAD(&qp->steering_rules);
499 
500 	qp->state	 = IB_QPS_RESET;
501 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
502 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
503 
504 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
505 	if (err)
506 		goto err;
507 
508 	if (pd->uobject) {
509 		struct mlx4_ib_create_qp ucmd;
510 
511 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
512 			err = -EFAULT;
513 			goto err;
514 		}
515 
516 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
517 
518 		err = set_user_sq_size(dev, qp, &ucmd);
519 		if (err)
520 			goto err;
521 
522 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
523 				       qp->buf_size, 0, 0);
524 		if (IS_ERR(qp->umem)) {
525 			err = PTR_ERR(qp->umem);
526 			goto err;
527 		}
528 
529 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
530 				    ilog2(qp->umem->page_size), &qp->mtt);
531 		if (err)
532 			goto err_buf;
533 
534 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
535 		if (err)
536 			goto err_mtt;
537 
538 		if (qp_has_rq(init_attr)) {
539 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
540 						  ucmd.db_addr, &qp->db);
541 			if (err)
542 				goto err_mtt;
543 		}
544 	} else {
545 		qp->sq_no_prefetch = 0;
546 
547 		if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
548 			qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
549 
550 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
551 			qp->flags |= MLX4_IB_QP_LSO;
552 
553 		err = set_kernel_sq_size(dev, &init_attr->cap, init_attr->qp_type, qp);
554 		if (err)
555 			goto err;
556 
557 		if (qp_has_rq(init_attr)) {
558 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
559 			if (err)
560 				goto err;
561 
562 			*qp->db.db = 0;
563 		}
564 
565 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf)) {
566 			err = -ENOMEM;
567 			goto err_db;
568 		}
569 
570 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
571 				    &qp->mtt);
572 		if (err)
573 			goto err_buf;
574 
575 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
576 		if (err)
577 			goto err_mtt;
578 
579 		qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), GFP_KERNEL);
580 		qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), GFP_KERNEL);
581 
582 		if (!qp->sq.wrid || !qp->rq.wrid) {
583 			err = -ENOMEM;
584 			goto err_wrid;
585 		}
586 	}
587 
588 	if (sqpn) {
589 		qpn = sqpn;
590 	} else {
591 		/* Raw packet QPNs must be aligned to 8 bits. If not, the WQE
592 		 * BlueFlame setup flow wrongly causes VLAN insertion. */
593 		if (init_attr->qp_type == IB_QPT_RAW_PACKET)
594 			err = mlx4_qp_reserve_range(dev->dev, 1, 1 << 8, &qpn);
595 		else
596 			err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn);
597 		if (err)
598 			goto err_wrid;
599 	}
600 
601 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
602 	if (err)
603 		goto err_qpn;
604 
605 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
606 		qp->mqp.qpn |= (1 << 23);
607 
608 	/*
609 	 * Hardware wants QPN written in big-endian order (after
610 	 * shifting) for send doorbell.  Precompute this value to save
611 	 * a little bit when posting sends.
612 	 */
613 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
614 
615 	qp->mqp.event = mlx4_ib_qp_event;
616 
617 	return 0;
618 
619 err_qpn:
620 	if (!sqpn)
621 		mlx4_qp_release_range(dev->dev, qpn, 1);
622 
623 err_wrid:
624 	if (pd->uobject) {
625 		if (qp_has_rq(init_attr))
626 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
627 	} else {
628 		kfree(qp->sq.wrid);
629 		kfree(qp->rq.wrid);
630 	}
631 
632 err_mtt:
633 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
634 
635 err_buf:
636 	if (pd->uobject)
637 		ib_umem_release(qp->umem);
638 	else
639 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
640 
641 err_db:
642 	if (!pd->uobject && qp_has_rq(init_attr))
643 		mlx4_db_free(dev->dev, &qp->db);
644 
645 err:
646 	return err;
647 }
648 
649 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
650 {
651 	switch (state) {
652 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
653 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
654 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
655 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
656 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
657 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
658 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
659 	default:		return -1;
660 	}
661 }
662 
663 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
664 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
665 {
666 	if (send_cq == recv_cq) {
667 		spin_lock_irq(&send_cq->lock);
668 		__acquire(&recv_cq->lock);
669 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
670 		spin_lock_irq(&send_cq->lock);
671 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
672 	} else {
673 		spin_lock_irq(&recv_cq->lock);
674 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
675 	}
676 }
677 
678 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
679 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
680 {
681 	if (send_cq == recv_cq) {
682 		__release(&recv_cq->lock);
683 		spin_unlock_irq(&send_cq->lock);
684 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
685 		spin_unlock(&recv_cq->lock);
686 		spin_unlock_irq(&send_cq->lock);
687 	} else {
688 		spin_unlock(&send_cq->lock);
689 		spin_unlock_irq(&recv_cq->lock);
690 	}
691 }
692 
693 static void del_gid_entries(struct mlx4_ib_qp *qp)
694 {
695 	struct mlx4_ib_gid_entry *ge, *tmp;
696 
697 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
698 		list_del(&ge->list);
699 		kfree(ge);
700 	}
701 }
702 
703 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
704 {
705 	if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
706 		return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
707 	else
708 		return to_mpd(qp->ibqp.pd);
709 }
710 
711 static void get_cqs(struct mlx4_ib_qp *qp,
712 		    struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
713 {
714 	switch (qp->ibqp.qp_type) {
715 	case IB_QPT_XRC_TGT:
716 		*send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
717 		*recv_cq = *send_cq;
718 		break;
719 	case IB_QPT_XRC_INI:
720 		*send_cq = to_mcq(qp->ibqp.send_cq);
721 		*recv_cq = *send_cq;
722 		break;
723 	default:
724 		*send_cq = to_mcq(qp->ibqp.send_cq);
725 		*recv_cq = to_mcq(qp->ibqp.recv_cq);
726 		break;
727 	}
728 }
729 
730 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
731 			      int is_user)
732 {
733 	struct mlx4_ib_cq *send_cq, *recv_cq;
734 
735 	if (qp->state != IB_QPS_RESET)
736 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
737 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
738 			pr_warn("modify QP %06x to RESET failed.\n",
739 			       qp->mqp.qpn);
740 
741 	get_cqs(qp, &send_cq, &recv_cq);
742 
743 	mlx4_ib_lock_cqs(send_cq, recv_cq);
744 
745 	if (!is_user) {
746 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
747 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
748 		if (send_cq != recv_cq)
749 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
750 	}
751 
752 	mlx4_qp_remove(dev->dev, &qp->mqp);
753 
754 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
755 
756 	mlx4_qp_free(dev->dev, &qp->mqp);
757 
758 	if (!is_sqp(dev, qp))
759 		mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
760 
761 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
762 
763 	if (is_user) {
764 		if (qp->rq.wqe_cnt)
765 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
766 					      &qp->db);
767 		ib_umem_release(qp->umem);
768 	} else {
769 		kfree(qp->sq.wrid);
770 		kfree(qp->rq.wrid);
771 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
772 		if (qp->rq.wqe_cnt)
773 			mlx4_db_free(dev->dev, &qp->db);
774 	}
775 
776 	del_gid_entries(qp);
777 }
778 
779 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
780 				struct ib_qp_init_attr *init_attr,
781 				struct ib_udata *udata)
782 {
783 	struct mlx4_ib_sqp *sqp;
784 	struct mlx4_ib_qp *qp;
785 	int err;
786 	u16 xrcdn = 0;
787 
788 	/*
789 	 * We only support LSO and multicast loopback blocking, and
790 	 * only for kernel UD QPs.
791 	 */
792 	if (init_attr->create_flags & ~(IB_QP_CREATE_IPOIB_UD_LSO |
793 					IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK))
794 		return ERR_PTR(-EINVAL);
795 
796 	if (init_attr->create_flags &&
797 	    (udata || init_attr->qp_type != IB_QPT_UD))
798 		return ERR_PTR(-EINVAL);
799 
800 	switch (init_attr->qp_type) {
801 	case IB_QPT_XRC_TGT:
802 		pd = to_mxrcd(init_attr->xrcd)->pd;
803 		xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
804 		init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
805 		/* fall through */
806 	case IB_QPT_XRC_INI:
807 		if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
808 			return ERR_PTR(-ENOSYS);
809 		init_attr->recv_cq = init_attr->send_cq;
810 		/* fall through */
811 	case IB_QPT_RC:
812 	case IB_QPT_UC:
813 	case IB_QPT_UD:
814 	case IB_QPT_RAW_PACKET:
815 	{
816 		qp = kzalloc(sizeof *qp, GFP_KERNEL);
817 		if (!qp)
818 			return ERR_PTR(-ENOMEM);
819 
820 		err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata, 0, qp);
821 		if (err) {
822 			kfree(qp);
823 			return ERR_PTR(err);
824 		}
825 
826 		qp->ibqp.qp_num = qp->mqp.qpn;
827 		qp->xrcdn = xrcdn;
828 
829 		break;
830 	}
831 	case IB_QPT_SMI:
832 	case IB_QPT_GSI:
833 	{
834 		/* Userspace is not allowed to create special QPs: */
835 		if (udata)
836 			return ERR_PTR(-EINVAL);
837 
838 		sqp = kzalloc(sizeof *sqp, GFP_KERNEL);
839 		if (!sqp)
840 			return ERR_PTR(-ENOMEM);
841 
842 		qp = &sqp->qp;
843 
844 		err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
845 				       to_mdev(pd->device)->dev->caps.sqp_start +
846 				       (init_attr->qp_type == IB_QPT_SMI ? 0 : 2) +
847 				       init_attr->port_num - 1,
848 				       qp);
849 		if (err) {
850 			kfree(sqp);
851 			return ERR_PTR(err);
852 		}
853 
854 		qp->port	= init_attr->port_num;
855 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
856 
857 		break;
858 	}
859 	default:
860 		/* Don't support raw QPs */
861 		return ERR_PTR(-EINVAL);
862 	}
863 
864 	return &qp->ibqp;
865 }
866 
867 int mlx4_ib_destroy_qp(struct ib_qp *qp)
868 {
869 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
870 	struct mlx4_ib_qp *mqp = to_mqp(qp);
871 	struct mlx4_ib_pd *pd;
872 
873 	if (is_qp0(dev, mqp))
874 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
875 
876 	pd = get_pd(mqp);
877 	destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
878 
879 	if (is_sqp(dev, mqp))
880 		kfree(to_msqp(mqp));
881 	else
882 		kfree(mqp);
883 
884 	return 0;
885 }
886 
887 static int to_mlx4_st(enum ib_qp_type type)
888 {
889 	switch (type) {
890 	case IB_QPT_RC:		return MLX4_QP_ST_RC;
891 	case IB_QPT_UC:		return MLX4_QP_ST_UC;
892 	case IB_QPT_UD:		return MLX4_QP_ST_UD;
893 	case IB_QPT_XRC_INI:
894 	case IB_QPT_XRC_TGT:	return MLX4_QP_ST_XRC;
895 	case IB_QPT_SMI:
896 	case IB_QPT_GSI:
897 	case IB_QPT_RAW_PACKET:	return MLX4_QP_ST_MLX;
898 	default:		return -1;
899 	}
900 }
901 
902 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
903 				   int attr_mask)
904 {
905 	u8 dest_rd_atomic;
906 	u32 access_flags;
907 	u32 hw_access_flags = 0;
908 
909 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
910 		dest_rd_atomic = attr->max_dest_rd_atomic;
911 	else
912 		dest_rd_atomic = qp->resp_depth;
913 
914 	if (attr_mask & IB_QP_ACCESS_FLAGS)
915 		access_flags = attr->qp_access_flags;
916 	else
917 		access_flags = qp->atomic_rd_en;
918 
919 	if (!dest_rd_atomic)
920 		access_flags &= IB_ACCESS_REMOTE_WRITE;
921 
922 	if (access_flags & IB_ACCESS_REMOTE_READ)
923 		hw_access_flags |= MLX4_QP_BIT_RRE;
924 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
925 		hw_access_flags |= MLX4_QP_BIT_RAE;
926 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
927 		hw_access_flags |= MLX4_QP_BIT_RWE;
928 
929 	return cpu_to_be32(hw_access_flags);
930 }
931 
932 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
933 			    int attr_mask)
934 {
935 	if (attr_mask & IB_QP_PKEY_INDEX)
936 		sqp->pkey_index = attr->pkey_index;
937 	if (attr_mask & IB_QP_QKEY)
938 		sqp->qkey = attr->qkey;
939 	if (attr_mask & IB_QP_SQ_PSN)
940 		sqp->send_psn = attr->sq_psn;
941 }
942 
943 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
944 {
945 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
946 }
947 
948 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
949 			 struct mlx4_qp_path *path, u8 port)
950 {
951 	int err;
952 	int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
953 		IB_LINK_LAYER_ETHERNET;
954 	u8 mac[6];
955 	int is_mcast;
956 	u16 vlan_tag;
957 	int vidx;
958 
959 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
960 	path->rlid	    = cpu_to_be16(ah->dlid);
961 	if (ah->static_rate) {
962 		path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
963 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
964 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
965 			--path->static_rate;
966 	} else
967 		path->static_rate = 0;
968 
969 	if (ah->ah_flags & IB_AH_GRH) {
970 		if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
971 			pr_err("sgid_index (%u) too large. max is %d\n",
972 			       ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
973 			return -1;
974 		}
975 
976 		path->grh_mylmc |= 1 << 7;
977 		path->mgid_index = ah->grh.sgid_index;
978 		path->hop_limit  = ah->grh.hop_limit;
979 		path->tclass_flowlabel =
980 			cpu_to_be32((ah->grh.traffic_class << 20) |
981 				    (ah->grh.flow_label));
982 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
983 	}
984 
985 	if (is_eth) {
986 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
987 			((port - 1) << 6) | ((ah->sl & 7) << 3);
988 
989 		if (!(ah->ah_flags & IB_AH_GRH))
990 			return -1;
991 
992 		err = mlx4_ib_resolve_grh(dev, ah, mac, &is_mcast, port);
993 		if (err)
994 			return err;
995 
996 		memcpy(path->dmac, mac, 6);
997 		path->ackto = MLX4_IB_LINK_TYPE_ETH;
998 		/* use index 0 into MAC table for IBoE */
999 		path->grh_mylmc &= 0x80;
1000 
1001 		vlan_tag = rdma_get_vlan_id(&dev->iboe.gid_table[port - 1][ah->grh.sgid_index]);
1002 		if (vlan_tag < 0x1000) {
1003 			if (mlx4_find_cached_vlan(dev->dev, port, vlan_tag, &vidx))
1004 				return -ENOENT;
1005 
1006 			path->vlan_index = vidx;
1007 			path->fl = 1 << 6;
1008 		}
1009 	} else
1010 		path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1011 			((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1012 
1013 	return 0;
1014 }
1015 
1016 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1017 {
1018 	struct mlx4_ib_gid_entry *ge, *tmp;
1019 
1020 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1021 		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1022 			ge->added = 1;
1023 			ge->port = qp->port;
1024 		}
1025 	}
1026 }
1027 
1028 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1029 			       const struct ib_qp_attr *attr, int attr_mask,
1030 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
1031 {
1032 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1033 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1034 	struct mlx4_ib_pd *pd;
1035 	struct mlx4_ib_cq *send_cq, *recv_cq;
1036 	struct mlx4_qp_context *context;
1037 	enum mlx4_qp_optpar optpar = 0;
1038 	int sqd_event;
1039 	int err = -EINVAL;
1040 
1041 	context = kzalloc(sizeof *context, GFP_KERNEL);
1042 	if (!context)
1043 		return -ENOMEM;
1044 
1045 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1046 				     (to_mlx4_st(ibqp->qp_type) << 16));
1047 
1048 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1049 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1050 	else {
1051 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
1052 		switch (attr->path_mig_state) {
1053 		case IB_MIG_MIGRATED:
1054 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1055 			break;
1056 		case IB_MIG_REARM:
1057 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1058 			break;
1059 		case IB_MIG_ARMED:
1060 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1061 			break;
1062 		}
1063 	}
1064 
1065 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1066 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1067 	else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1068 		context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1069 	else if (ibqp->qp_type == IB_QPT_UD) {
1070 		if (qp->flags & MLX4_IB_QP_LSO)
1071 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
1072 					      ilog2(dev->dev->caps.max_gso_sz);
1073 		else
1074 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1075 	} else if (attr_mask & IB_QP_PATH_MTU) {
1076 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1077 			pr_err("path MTU (%u) is invalid\n",
1078 			       attr->path_mtu);
1079 			goto out;
1080 		}
1081 		context->mtu_msgmax = (attr->path_mtu << 5) |
1082 			ilog2(dev->dev->caps.max_msg_sz);
1083 	}
1084 
1085 	if (qp->rq.wqe_cnt)
1086 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1087 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
1088 
1089 	if (qp->sq.wqe_cnt)
1090 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1091 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
1092 
1093 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1094 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1095 		context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1096 	}
1097 
1098 	if (qp->ibqp.uobject)
1099 		context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
1100 	else
1101 		context->usr_page = cpu_to_be32(dev->priv_uar.index);
1102 
1103 	if (attr_mask & IB_QP_DEST_QPN)
1104 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1105 
1106 	if (attr_mask & IB_QP_PORT) {
1107 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1108 		    !(attr_mask & IB_QP_AV)) {
1109 			mlx4_set_sched(&context->pri_path, attr->port_num);
1110 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1111 		}
1112 	}
1113 
1114 	if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1115 		if (dev->counters[qp->port - 1] != -1) {
1116 			context->pri_path.counter_index =
1117 						dev->counters[qp->port - 1];
1118 			optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1119 		} else
1120 			context->pri_path.counter_index = 0xff;
1121 	}
1122 
1123 	if (attr_mask & IB_QP_PKEY_INDEX) {
1124 		context->pri_path.pkey_index = attr->pkey_index;
1125 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1126 	}
1127 
1128 	if (attr_mask & IB_QP_AV) {
1129 		if (mlx4_set_path(dev, &attr->ah_attr, &context->pri_path,
1130 				  attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
1131 			goto out;
1132 
1133 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1134 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
1135 	}
1136 
1137 	if (attr_mask & IB_QP_TIMEOUT) {
1138 		context->pri_path.ackto |= attr->timeout << 3;
1139 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1140 	}
1141 
1142 	if (attr_mask & IB_QP_ALT_PATH) {
1143 		if (attr->alt_port_num == 0 ||
1144 		    attr->alt_port_num > dev->dev->caps.num_ports)
1145 			goto out;
1146 
1147 		if (attr->alt_pkey_index >=
1148 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
1149 			goto out;
1150 
1151 		if (mlx4_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
1152 				  attr->alt_port_num))
1153 			goto out;
1154 
1155 		context->alt_path.pkey_index = attr->alt_pkey_index;
1156 		context->alt_path.ackto = attr->alt_timeout << 3;
1157 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1158 	}
1159 
1160 	pd = get_pd(qp);
1161 	get_cqs(qp, &send_cq, &recv_cq);
1162 	context->pd       = cpu_to_be32(pd->pdn);
1163 	context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1164 	context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1165 	context->params1  = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1166 
1167 	/* Set "fast registration enabled" for all kernel QPs */
1168 	if (!qp->ibqp.uobject)
1169 		context->params1 |= cpu_to_be32(1 << 11);
1170 
1171 	if (attr_mask & IB_QP_RNR_RETRY) {
1172 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1173 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1174 	}
1175 
1176 	if (attr_mask & IB_QP_RETRY_CNT) {
1177 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1178 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1179 	}
1180 
1181 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1182 		if (attr->max_rd_atomic)
1183 			context->params1 |=
1184 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1185 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1186 	}
1187 
1188 	if (attr_mask & IB_QP_SQ_PSN)
1189 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1190 
1191 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1192 		if (attr->max_dest_rd_atomic)
1193 			context->params2 |=
1194 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1195 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1196 	}
1197 
1198 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1199 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1200 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1201 	}
1202 
1203 	if (ibqp->srq)
1204 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1205 
1206 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1207 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1208 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1209 	}
1210 	if (attr_mask & IB_QP_RQ_PSN)
1211 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1212 
1213 	if (attr_mask & IB_QP_QKEY) {
1214 		context->qkey = cpu_to_be32(attr->qkey);
1215 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
1216 	}
1217 
1218 	if (ibqp->srq)
1219 		context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1220 
1221 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1222 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1223 
1224 	if (cur_state == IB_QPS_INIT &&
1225 	    new_state == IB_QPS_RTR  &&
1226 	    (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1227 	     ibqp->qp_type == IB_QPT_UD ||
1228 	     ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1229 		context->pri_path.sched_queue = (qp->port - 1) << 6;
1230 		if (is_qp0(dev, qp))
1231 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1232 		else
1233 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1234 	}
1235 
1236 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1237 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1238 		sqd_event = 1;
1239 	else
1240 		sqd_event = 0;
1241 
1242 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1243 		context->rlkey |= (1 << 4);
1244 
1245 	/*
1246 	 * Before passing a kernel QP to the HW, make sure that the
1247 	 * ownership bits of the send queue are set and the SQ
1248 	 * headroom is stamped so that the hardware doesn't start
1249 	 * processing stale work requests.
1250 	 */
1251 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1252 		struct mlx4_wqe_ctrl_seg *ctrl;
1253 		int i;
1254 
1255 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1256 			ctrl = get_send_wqe(qp, i);
1257 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
1258 			if (qp->sq_max_wqes_per_wr == 1)
1259 				ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1260 
1261 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1262 		}
1263 	}
1264 
1265 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1266 			     to_mlx4_state(new_state), context, optpar,
1267 			     sqd_event, &qp->mqp);
1268 	if (err)
1269 		goto out;
1270 
1271 	qp->state = new_state;
1272 
1273 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1274 		qp->atomic_rd_en = attr->qp_access_flags;
1275 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1276 		qp->resp_depth = attr->max_dest_rd_atomic;
1277 	if (attr_mask & IB_QP_PORT) {
1278 		qp->port = attr->port_num;
1279 		update_mcg_macs(dev, qp);
1280 	}
1281 	if (attr_mask & IB_QP_ALT_PATH)
1282 		qp->alt_port = attr->alt_port_num;
1283 
1284 	if (is_sqp(dev, qp))
1285 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1286 
1287 	/*
1288 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
1289 	 * QP0 to RESET or ERROR, bring the link back down.
1290 	 */
1291 	if (is_qp0(dev, qp)) {
1292 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1293 			if (mlx4_INIT_PORT(dev->dev, qp->port))
1294 				pr_warn("INIT_PORT failed for port %d\n",
1295 				       qp->port);
1296 
1297 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1298 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1299 			mlx4_CLOSE_PORT(dev->dev, qp->port);
1300 	}
1301 
1302 	/*
1303 	 * If we moved a kernel QP to RESET, clean up all old CQ
1304 	 * entries and reinitialize the QP.
1305 	 */
1306 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1307 		mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1308 				 ibqp->srq ? to_msrq(ibqp->srq): NULL);
1309 		if (send_cq != recv_cq)
1310 			mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1311 
1312 		qp->rq.head = 0;
1313 		qp->rq.tail = 0;
1314 		qp->sq.head = 0;
1315 		qp->sq.tail = 0;
1316 		qp->sq_next_wqe = 0;
1317 		if (qp->rq.wqe_cnt)
1318 			*qp->db.db  = 0;
1319 	}
1320 
1321 out:
1322 	kfree(context);
1323 	return err;
1324 }
1325 
1326 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1327 		      int attr_mask, struct ib_udata *udata)
1328 {
1329 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1330 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1331 	enum ib_qp_state cur_state, new_state;
1332 	int err = -EINVAL;
1333 
1334 	mutex_lock(&qp->mutex);
1335 
1336 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1337 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1338 
1339 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
1340 		pr_debug("qpn 0x%x: invalid attribute mask specified "
1341 			 "for transition %d to %d. qp_type %d,"
1342 			 " attr_mask 0x%x\n",
1343 			 ibqp->qp_num, cur_state, new_state,
1344 			 ibqp->qp_type, attr_mask);
1345 		goto out;
1346 	}
1347 
1348 	if ((attr_mask & IB_QP_PORT) &&
1349 	    (attr->port_num == 0 || attr->port_num > dev->dev->caps.num_ports)) {
1350 		pr_debug("qpn 0x%x: invalid port number (%d) specified "
1351 			 "for transition %d to %d. qp_type %d\n",
1352 			 ibqp->qp_num, attr->port_num, cur_state,
1353 			 new_state, ibqp->qp_type);
1354 		goto out;
1355 	}
1356 
1357 	if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
1358 	    (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
1359 	     IB_LINK_LAYER_ETHERNET))
1360 		goto out;
1361 
1362 	if (attr_mask & IB_QP_PKEY_INDEX) {
1363 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1364 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
1365 			pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
1366 				 "for transition %d to %d. qp_type %d\n",
1367 				 ibqp->qp_num, attr->pkey_index, cur_state,
1368 				 new_state, ibqp->qp_type);
1369 			goto out;
1370 		}
1371 	}
1372 
1373 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1374 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
1375 		pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
1376 			 "Transition %d to %d. qp_type %d\n",
1377 			 ibqp->qp_num, attr->max_rd_atomic, cur_state,
1378 			 new_state, ibqp->qp_type);
1379 		goto out;
1380 	}
1381 
1382 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1383 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
1384 		pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
1385 			 "Transition %d to %d. qp_type %d\n",
1386 			 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
1387 			 new_state, ibqp->qp_type);
1388 		goto out;
1389 	}
1390 
1391 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1392 		err = 0;
1393 		goto out;
1394 	}
1395 
1396 	err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1397 
1398 out:
1399 	mutex_unlock(&qp->mutex);
1400 	return err;
1401 }
1402 
1403 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
1404 			    void *wqe, unsigned *mlx_seg_len)
1405 {
1406 	struct ib_device *ib_dev = sqp->qp.ibqp.device;
1407 	struct mlx4_wqe_mlx_seg *mlx = wqe;
1408 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
1409 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
1410 	struct net_device *ndev;
1411 	union ib_gid sgid;
1412 	u16 pkey;
1413 	int send_size;
1414 	int header_size;
1415 	int spc;
1416 	int i;
1417 	int is_eth;
1418 	int is_vlan = 0;
1419 	int is_grh;
1420 	u16 vlan;
1421 
1422 	send_size = 0;
1423 	for (i = 0; i < wr->num_sge; ++i)
1424 		send_size += wr->sg_list[i].length;
1425 
1426 	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
1427 	is_grh = mlx4_ib_ah_grh_present(ah);
1428 	if (is_eth) {
1429 		ib_get_cached_gid(ib_dev, be32_to_cpu(ah->av.ib.port_pd) >> 24,
1430 				  ah->av.ib.gid_index, &sgid);
1431 		vlan = rdma_get_vlan_id(&sgid);
1432 		is_vlan = vlan < 0x1000;
1433 	}
1434 	ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh, 0, &sqp->ud_header);
1435 
1436 	if (!is_eth) {
1437 		sqp->ud_header.lrh.service_level =
1438 			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
1439 		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
1440 		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
1441 	}
1442 
1443 	if (is_grh) {
1444 		sqp->ud_header.grh.traffic_class =
1445 			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
1446 		sqp->ud_header.grh.flow_label    =
1447 			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
1448 		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
1449 		ib_get_cached_gid(ib_dev, be32_to_cpu(ah->av.ib.port_pd) >> 24,
1450 				  ah->av.ib.gid_index, &sqp->ud_header.grh.source_gid);
1451 		memcpy(sqp->ud_header.grh.destination_gid.raw,
1452 		       ah->av.ib.dgid, 16);
1453 	}
1454 
1455 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1456 
1457 	if (!is_eth) {
1458 		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
1459 					  (sqp->ud_header.lrh.destination_lid ==
1460 					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
1461 					  (sqp->ud_header.lrh.service_level << 8));
1462 		mlx->rlid = sqp->ud_header.lrh.destination_lid;
1463 	}
1464 
1465 	switch (wr->opcode) {
1466 	case IB_WR_SEND:
1467 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
1468 		sqp->ud_header.immediate_present = 0;
1469 		break;
1470 	case IB_WR_SEND_WITH_IMM:
1471 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1472 		sqp->ud_header.immediate_present = 1;
1473 		sqp->ud_header.immediate_data    = wr->ex.imm_data;
1474 		break;
1475 	default:
1476 		return -EINVAL;
1477 	}
1478 
1479 	if (is_eth) {
1480 		u8 *smac;
1481 		u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
1482 
1483 		mlx->sched_prio = cpu_to_be16(pcp);
1484 
1485 		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
1486 		/* FIXME: cache smac value? */
1487 		ndev = to_mdev(sqp->qp.ibqp.device)->iboe.netdevs[sqp->qp.port - 1];
1488 		if (!ndev)
1489 			return -ENODEV;
1490 		smac = ndev->dev_addr;
1491 		memcpy(sqp->ud_header.eth.smac_h, smac, 6);
1492 		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
1493 			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
1494 		if (!is_vlan) {
1495 			sqp->ud_header.eth.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
1496 		} else {
1497 			sqp->ud_header.vlan.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
1498 			sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
1499 		}
1500 	} else {
1501 		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1502 		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1503 			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1504 	}
1505 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1506 	if (!sqp->qp.ibqp.qp_num)
1507 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
1508 	else
1509 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
1510 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1511 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1512 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1513 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1514 					       sqp->qkey : wr->wr.ud.remote_qkey);
1515 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1516 
1517 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
1518 
1519 	if (0) {
1520 		pr_err("built UD header of size %d:\n", header_size);
1521 		for (i = 0; i < header_size / 4; ++i) {
1522 			if (i % 8 == 0)
1523 				pr_err("  [%02x] ", i * 4);
1524 			pr_cont(" %08x",
1525 				be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
1526 			if ((i + 1) % 8 == 0)
1527 				pr_cont("\n");
1528 		}
1529 		pr_err("\n");
1530 	}
1531 
1532 	/*
1533 	 * Inline data segments may not cross a 64 byte boundary.  If
1534 	 * our UD header is bigger than the space available up to the
1535 	 * next 64 byte boundary in the WQE, use two inline data
1536 	 * segments to hold the UD header.
1537 	 */
1538 	spc = MLX4_INLINE_ALIGN -
1539 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
1540 	if (header_size <= spc) {
1541 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
1542 		memcpy(inl + 1, sqp->header_buf, header_size);
1543 		i = 1;
1544 	} else {
1545 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
1546 		memcpy(inl + 1, sqp->header_buf, spc);
1547 
1548 		inl = (void *) (inl + 1) + spc;
1549 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
1550 		/*
1551 		 * Need a barrier here to make sure all the data is
1552 		 * visible before the byte_count field is set.
1553 		 * Otherwise the HCA prefetcher could grab the 64-byte
1554 		 * chunk with this inline segment and get a valid (!=
1555 		 * 0xffffffff) byte count but stale data, and end up
1556 		 * generating a packet with bad headers.
1557 		 *
1558 		 * The first inline segment's byte_count field doesn't
1559 		 * need a barrier, because it comes after a
1560 		 * control/MLX segment and therefore is at an offset
1561 		 * of 16 mod 64.
1562 		 */
1563 		wmb();
1564 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
1565 		i = 2;
1566 	}
1567 
1568 	*mlx_seg_len =
1569 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
1570 	return 0;
1571 }
1572 
1573 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1574 {
1575 	unsigned cur;
1576 	struct mlx4_ib_cq *cq;
1577 
1578 	cur = wq->head - wq->tail;
1579 	if (likely(cur + nreq < wq->max_post))
1580 		return 0;
1581 
1582 	cq = to_mcq(ib_cq);
1583 	spin_lock(&cq->lock);
1584 	cur = wq->head - wq->tail;
1585 	spin_unlock(&cq->lock);
1586 
1587 	return cur + nreq >= wq->max_post;
1588 }
1589 
1590 static __be32 convert_access(int acc)
1591 {
1592 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? cpu_to_be32(MLX4_WQE_FMR_PERM_ATOMIC)       : 0) |
1593 	       (acc & IB_ACCESS_REMOTE_WRITE  ? cpu_to_be32(MLX4_WQE_FMR_PERM_REMOTE_WRITE) : 0) |
1594 	       (acc & IB_ACCESS_REMOTE_READ   ? cpu_to_be32(MLX4_WQE_FMR_PERM_REMOTE_READ)  : 0) |
1595 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
1596 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
1597 }
1598 
1599 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr)
1600 {
1601 	struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
1602 	int i;
1603 
1604 	for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i)
1605 		mfrpl->mapped_page_list[i] =
1606 			cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] |
1607 				    MLX4_MTT_FLAG_PRESENT);
1608 
1609 	fseg->flags		= convert_access(wr->wr.fast_reg.access_flags);
1610 	fseg->mem_key		= cpu_to_be32(wr->wr.fast_reg.rkey);
1611 	fseg->buf_list		= cpu_to_be64(mfrpl->map);
1612 	fseg->start_addr	= cpu_to_be64(wr->wr.fast_reg.iova_start);
1613 	fseg->reg_len		= cpu_to_be64(wr->wr.fast_reg.length);
1614 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
1615 	fseg->page_size		= cpu_to_be32(wr->wr.fast_reg.page_shift);
1616 	fseg->reserved[0]	= 0;
1617 	fseg->reserved[1]	= 0;
1618 }
1619 
1620 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
1621 {
1622 	iseg->flags	= 0;
1623 	iseg->mem_key	= cpu_to_be32(rkey);
1624 	iseg->guest_id	= 0;
1625 	iseg->pa	= 0;
1626 }
1627 
1628 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
1629 					  u64 remote_addr, u32 rkey)
1630 {
1631 	rseg->raddr    = cpu_to_be64(remote_addr);
1632 	rseg->rkey     = cpu_to_be32(rkey);
1633 	rseg->reserved = 0;
1634 }
1635 
1636 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
1637 {
1638 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1639 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
1640 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
1641 	} else if (wr->opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
1642 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
1643 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add_mask);
1644 	} else {
1645 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
1646 		aseg->compare  = 0;
1647 	}
1648 
1649 }
1650 
1651 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
1652 				  struct ib_send_wr *wr)
1653 {
1654 	aseg->swap_add		= cpu_to_be64(wr->wr.atomic.swap);
1655 	aseg->swap_add_mask	= cpu_to_be64(wr->wr.atomic.swap_mask);
1656 	aseg->compare		= cpu_to_be64(wr->wr.atomic.compare_add);
1657 	aseg->compare_mask	= cpu_to_be64(wr->wr.atomic.compare_add_mask);
1658 }
1659 
1660 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
1661 			     struct ib_send_wr *wr)
1662 {
1663 	memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
1664 	dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1665 	dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1666 	dseg->vlan = to_mah(wr->wr.ud.ah)->av.eth.vlan;
1667 	memcpy(dseg->mac, to_mah(wr->wr.ud.ah)->av.eth.mac, 6);
1668 }
1669 
1670 static void set_mlx_icrc_seg(void *dseg)
1671 {
1672 	u32 *t = dseg;
1673 	struct mlx4_wqe_inline_seg *iseg = dseg;
1674 
1675 	t[1] = 0;
1676 
1677 	/*
1678 	 * Need a barrier here before writing the byte_count field to
1679 	 * make sure that all the data is visible before the
1680 	 * byte_count field is set.  Otherwise, if the segment begins
1681 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
1682 	 * chunk and get a valid (!= * 0xffffffff) byte count but
1683 	 * stale data, and end up sending the wrong data.
1684 	 */
1685 	wmb();
1686 
1687 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
1688 }
1689 
1690 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
1691 {
1692 	dseg->lkey       = cpu_to_be32(sg->lkey);
1693 	dseg->addr       = cpu_to_be64(sg->addr);
1694 
1695 	/*
1696 	 * Need a barrier here before writing the byte_count field to
1697 	 * make sure that all the data is visible before the
1698 	 * byte_count field is set.  Otherwise, if the segment begins
1699 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
1700 	 * chunk and get a valid (!= * 0xffffffff) byte count but
1701 	 * stale data, and end up sending the wrong data.
1702 	 */
1703 	wmb();
1704 
1705 	dseg->byte_count = cpu_to_be32(sg->length);
1706 }
1707 
1708 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
1709 {
1710 	dseg->byte_count = cpu_to_be32(sg->length);
1711 	dseg->lkey       = cpu_to_be32(sg->lkey);
1712 	dseg->addr       = cpu_to_be64(sg->addr);
1713 }
1714 
1715 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr,
1716 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
1717 			 __be32 *lso_hdr_sz, __be32 *blh)
1718 {
1719 	unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
1720 
1721 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
1722 		*blh = cpu_to_be32(1 << 6);
1723 
1724 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
1725 		     wr->num_sge > qp->sq.max_gs - (halign >> 4)))
1726 		return -EINVAL;
1727 
1728 	memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
1729 
1730 	*lso_hdr_sz  = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 |
1731 				   wr->wr.ud.hlen);
1732 	*lso_seg_len = halign;
1733 	return 0;
1734 }
1735 
1736 static __be32 send_ieth(struct ib_send_wr *wr)
1737 {
1738 	switch (wr->opcode) {
1739 	case IB_WR_SEND_WITH_IMM:
1740 	case IB_WR_RDMA_WRITE_WITH_IMM:
1741 		return wr->ex.imm_data;
1742 
1743 	case IB_WR_SEND_WITH_INV:
1744 		return cpu_to_be32(wr->ex.invalidate_rkey);
1745 
1746 	default:
1747 		return 0;
1748 	}
1749 }
1750 
1751 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1752 		      struct ib_send_wr **bad_wr)
1753 {
1754 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1755 	void *wqe;
1756 	struct mlx4_wqe_ctrl_seg *ctrl;
1757 	struct mlx4_wqe_data_seg *dseg;
1758 	unsigned long flags;
1759 	int nreq;
1760 	int err = 0;
1761 	unsigned ind;
1762 	int uninitialized_var(stamp);
1763 	int uninitialized_var(size);
1764 	unsigned uninitialized_var(seglen);
1765 	__be32 dummy;
1766 	__be32 *lso_wqe;
1767 	__be32 uninitialized_var(lso_hdr_sz);
1768 	__be32 blh;
1769 	int i;
1770 
1771 	spin_lock_irqsave(&qp->sq.lock, flags);
1772 
1773 	ind = qp->sq_next_wqe;
1774 
1775 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1776 		lso_wqe = &dummy;
1777 		blh = 0;
1778 
1779 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1780 			err = -ENOMEM;
1781 			*bad_wr = wr;
1782 			goto out;
1783 		}
1784 
1785 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
1786 			err = -EINVAL;
1787 			*bad_wr = wr;
1788 			goto out;
1789 		}
1790 
1791 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
1792 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
1793 
1794 		ctrl->srcrb_flags =
1795 			(wr->send_flags & IB_SEND_SIGNALED ?
1796 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
1797 			(wr->send_flags & IB_SEND_SOLICITED ?
1798 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
1799 			((wr->send_flags & IB_SEND_IP_CSUM) ?
1800 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
1801 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
1802 			qp->sq_signal_bits;
1803 
1804 		ctrl->imm = send_ieth(wr);
1805 
1806 		wqe += sizeof *ctrl;
1807 		size = sizeof *ctrl / 16;
1808 
1809 		switch (ibqp->qp_type) {
1810 		case IB_QPT_RC:
1811 		case IB_QPT_UC:
1812 			switch (wr->opcode) {
1813 			case IB_WR_ATOMIC_CMP_AND_SWP:
1814 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1815 			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
1816 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
1817 					      wr->wr.atomic.rkey);
1818 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
1819 
1820 				set_atomic_seg(wqe, wr);
1821 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
1822 
1823 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
1824 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
1825 
1826 				break;
1827 
1828 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
1829 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
1830 					      wr->wr.atomic.rkey);
1831 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
1832 
1833 				set_masked_atomic_seg(wqe, wr);
1834 				wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
1835 
1836 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
1837 					 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
1838 
1839 				break;
1840 
1841 			case IB_WR_RDMA_READ:
1842 			case IB_WR_RDMA_WRITE:
1843 			case IB_WR_RDMA_WRITE_WITH_IMM:
1844 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1845 					      wr->wr.rdma.rkey);
1846 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
1847 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
1848 				break;
1849 
1850 			case IB_WR_LOCAL_INV:
1851 				ctrl->srcrb_flags |=
1852 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
1853 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
1854 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
1855 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
1856 				break;
1857 
1858 			case IB_WR_FAST_REG_MR:
1859 				ctrl->srcrb_flags |=
1860 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
1861 				set_fmr_seg(wqe, wr);
1862 				wqe  += sizeof (struct mlx4_wqe_fmr_seg);
1863 				size += sizeof (struct mlx4_wqe_fmr_seg) / 16;
1864 				break;
1865 
1866 			default:
1867 				/* No extra segments required for sends */
1868 				break;
1869 			}
1870 			break;
1871 
1872 		case IB_QPT_UD:
1873 			set_datagram_seg(wqe, wr);
1874 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
1875 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
1876 
1877 			if (wr->opcode == IB_WR_LSO) {
1878 				err = build_lso_seg(wqe, wr, qp, &seglen, &lso_hdr_sz, &blh);
1879 				if (unlikely(err)) {
1880 					*bad_wr = wr;
1881 					goto out;
1882 				}
1883 				lso_wqe = (__be32 *) wqe;
1884 				wqe  += seglen;
1885 				size += seglen / 16;
1886 			}
1887 			break;
1888 
1889 		case IB_QPT_SMI:
1890 		case IB_QPT_GSI:
1891 			err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
1892 			if (unlikely(err)) {
1893 				*bad_wr = wr;
1894 				goto out;
1895 			}
1896 			wqe  += seglen;
1897 			size += seglen / 16;
1898 			break;
1899 
1900 		default:
1901 			break;
1902 		}
1903 
1904 		/*
1905 		 * Write data segments in reverse order, so as to
1906 		 * overwrite cacheline stamp last within each
1907 		 * cacheline.  This avoids issues with WQE
1908 		 * prefetching.
1909 		 */
1910 
1911 		dseg = wqe;
1912 		dseg += wr->num_sge - 1;
1913 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
1914 
1915 		/* Add one more inline data segment for ICRC for MLX sends */
1916 		if (unlikely(qp->ibqp.qp_type == IB_QPT_SMI ||
1917 			     qp->ibqp.qp_type == IB_QPT_GSI)) {
1918 			set_mlx_icrc_seg(dseg + 1);
1919 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
1920 		}
1921 
1922 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
1923 			set_data_seg(dseg, wr->sg_list + i);
1924 
1925 		/*
1926 		 * Possibly overwrite stamping in cacheline with LSO
1927 		 * segment only after making sure all data segments
1928 		 * are written.
1929 		 */
1930 		wmb();
1931 		*lso_wqe = lso_hdr_sz;
1932 
1933 		ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
1934 				    MLX4_WQE_CTRL_FENCE : 0) | size;
1935 
1936 		/*
1937 		 * Make sure descriptor is fully written before
1938 		 * setting ownership bit (because HW can start
1939 		 * executing as soon as we do).
1940 		 */
1941 		wmb();
1942 
1943 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
1944 			*bad_wr = wr;
1945 			err = -EINVAL;
1946 			goto out;
1947 		}
1948 
1949 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
1950 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
1951 
1952 		stamp = ind + qp->sq_spare_wqes;
1953 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
1954 
1955 		/*
1956 		 * We can improve latency by not stamping the last
1957 		 * send queue WQE until after ringing the doorbell, so
1958 		 * only stamp here if there are still more WQEs to post.
1959 		 *
1960 		 * Same optimization applies to padding with NOP wqe
1961 		 * in case of WQE shrinking (used to prevent wrap-around
1962 		 * in the middle of WR).
1963 		 */
1964 		if (wr->next) {
1965 			stamp_send_wqe(qp, stamp, size * 16);
1966 			ind = pad_wraparound(qp, ind);
1967 		}
1968 	}
1969 
1970 out:
1971 	if (likely(nreq)) {
1972 		qp->sq.head += nreq;
1973 
1974 		/*
1975 		 * Make sure that descriptors are written before
1976 		 * doorbell record.
1977 		 */
1978 		wmb();
1979 
1980 		writel(qp->doorbell_qpn,
1981 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
1982 
1983 		/*
1984 		 * Make sure doorbells don't leak out of SQ spinlock
1985 		 * and reach the HCA out of order.
1986 		 */
1987 		mmiowb();
1988 
1989 		stamp_send_wqe(qp, stamp, size * 16);
1990 
1991 		ind = pad_wraparound(qp, ind);
1992 		qp->sq_next_wqe = ind;
1993 	}
1994 
1995 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1996 
1997 	return err;
1998 }
1999 
2000 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2001 		      struct ib_recv_wr **bad_wr)
2002 {
2003 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2004 	struct mlx4_wqe_data_seg *scat;
2005 	unsigned long flags;
2006 	int err = 0;
2007 	int nreq;
2008 	int ind;
2009 	int i;
2010 
2011 	spin_lock_irqsave(&qp->rq.lock, flags);
2012 
2013 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
2014 
2015 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2016 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2017 			err = -ENOMEM;
2018 			*bad_wr = wr;
2019 			goto out;
2020 		}
2021 
2022 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2023 			err = -EINVAL;
2024 			*bad_wr = wr;
2025 			goto out;
2026 		}
2027 
2028 		scat = get_recv_wqe(qp, ind);
2029 
2030 		for (i = 0; i < wr->num_sge; ++i)
2031 			__set_data_seg(scat + i, wr->sg_list + i);
2032 
2033 		if (i < qp->rq.max_gs) {
2034 			scat[i].byte_count = 0;
2035 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
2036 			scat[i].addr       = 0;
2037 		}
2038 
2039 		qp->rq.wrid[ind] = wr->wr_id;
2040 
2041 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
2042 	}
2043 
2044 out:
2045 	if (likely(nreq)) {
2046 		qp->rq.head += nreq;
2047 
2048 		/*
2049 		 * Make sure that descriptors are written before
2050 		 * doorbell record.
2051 		 */
2052 		wmb();
2053 
2054 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
2055 	}
2056 
2057 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2058 
2059 	return err;
2060 }
2061 
2062 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
2063 {
2064 	switch (mlx4_state) {
2065 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
2066 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
2067 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
2068 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
2069 	case MLX4_QP_STATE_SQ_DRAINING:
2070 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
2071 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
2072 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
2073 	default:		     return -1;
2074 	}
2075 }
2076 
2077 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
2078 {
2079 	switch (mlx4_mig_state) {
2080 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
2081 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
2082 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
2083 	default: return -1;
2084 	}
2085 }
2086 
2087 static int to_ib_qp_access_flags(int mlx4_flags)
2088 {
2089 	int ib_flags = 0;
2090 
2091 	if (mlx4_flags & MLX4_QP_BIT_RRE)
2092 		ib_flags |= IB_ACCESS_REMOTE_READ;
2093 	if (mlx4_flags & MLX4_QP_BIT_RWE)
2094 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
2095 	if (mlx4_flags & MLX4_QP_BIT_RAE)
2096 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
2097 
2098 	return ib_flags;
2099 }
2100 
2101 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
2102 				struct mlx4_qp_path *path)
2103 {
2104 	struct mlx4_dev *dev = ibdev->dev;
2105 	int is_eth;
2106 
2107 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
2108 	ib_ah_attr->port_num	  = path->sched_queue & 0x40 ? 2 : 1;
2109 
2110 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
2111 		return;
2112 
2113 	is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
2114 		IB_LINK_LAYER_ETHERNET;
2115 	if (is_eth)
2116 		ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
2117 		((path->sched_queue & 4) << 1);
2118 	else
2119 		ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
2120 
2121 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
2122 	ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
2123 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
2124 	ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
2125 	if (ib_ah_attr->ah_flags) {
2126 		ib_ah_attr->grh.sgid_index = path->mgid_index;
2127 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
2128 		ib_ah_attr->grh.traffic_class =
2129 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
2130 		ib_ah_attr->grh.flow_label =
2131 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
2132 		memcpy(ib_ah_attr->grh.dgid.raw,
2133 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
2134 	}
2135 }
2136 
2137 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
2138 		     struct ib_qp_init_attr *qp_init_attr)
2139 {
2140 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2141 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
2142 	struct mlx4_qp_context context;
2143 	int mlx4_state;
2144 	int err = 0;
2145 
2146 	mutex_lock(&qp->mutex);
2147 
2148 	if (qp->state == IB_QPS_RESET) {
2149 		qp_attr->qp_state = IB_QPS_RESET;
2150 		goto done;
2151 	}
2152 
2153 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
2154 	if (err) {
2155 		err = -EINVAL;
2156 		goto out;
2157 	}
2158 
2159 	mlx4_state = be32_to_cpu(context.flags) >> 28;
2160 
2161 	qp->state		     = to_ib_qp_state(mlx4_state);
2162 	qp_attr->qp_state	     = qp->state;
2163 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
2164 	qp_attr->path_mig_state	     =
2165 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
2166 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
2167 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
2168 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
2169 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
2170 	qp_attr->qp_access_flags     =
2171 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
2172 
2173 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
2174 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
2175 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
2176 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
2177 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
2178 	}
2179 
2180 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
2181 	if (qp_attr->qp_state == IB_QPS_INIT)
2182 		qp_attr->port_num = qp->port;
2183 	else
2184 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
2185 
2186 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
2187 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
2188 
2189 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
2190 
2191 	qp_attr->max_dest_rd_atomic =
2192 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
2193 	qp_attr->min_rnr_timer	    =
2194 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
2195 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
2196 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
2197 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
2198 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
2199 
2200 done:
2201 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
2202 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
2203 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
2204 
2205 	if (!ibqp->uobject) {
2206 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
2207 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
2208 	} else {
2209 		qp_attr->cap.max_send_wr  = 0;
2210 		qp_attr->cap.max_send_sge = 0;
2211 	}
2212 
2213 	/*
2214 	 * We don't support inline sends for kernel QPs (yet), and we
2215 	 * don't know what userspace's value should be.
2216 	 */
2217 	qp_attr->cap.max_inline_data = 0;
2218 
2219 	qp_init_attr->cap	     = qp_attr->cap;
2220 
2221 	qp_init_attr->create_flags = 0;
2222 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
2223 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
2224 
2225 	if (qp->flags & MLX4_IB_QP_LSO)
2226 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
2227 
2228 out:
2229 	mutex_unlock(&qp->mutex);
2230 	return err;
2231 }
2232 
2233