xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision 7dd65feb)
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 
36 #include <rdma/ib_cache.h>
37 #include <rdma/ib_pack.h>
38 
39 #include <linux/mlx4/qp.h>
40 
41 #include "mlx4_ib.h"
42 #include "user.h"
43 
44 enum {
45 	MLX4_IB_ACK_REQ_FREQ	= 8,
46 };
47 
48 enum {
49 	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
50 	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f
51 };
52 
53 enum {
54 	/*
55 	 * Largest possible UD header: send with GRH and immediate data.
56 	 */
57 	MLX4_IB_UD_HEADER_SIZE		= 72,
58 	MLX4_IB_LSO_HEADER_SPARE	= 128,
59 };
60 
61 struct mlx4_ib_sqp {
62 	struct mlx4_ib_qp	qp;
63 	int			pkey_index;
64 	u32			qkey;
65 	u32			send_psn;
66 	struct ib_ud_header	ud_header;
67 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
68 };
69 
70 enum {
71 	MLX4_IB_MIN_SQ_STRIDE	= 6,
72 	MLX4_IB_CACHE_LINE_SIZE	= 64,
73 };
74 
75 static const __be32 mlx4_ib_opcode[] = {
76 	[IB_WR_SEND]			= cpu_to_be32(MLX4_OPCODE_SEND),
77 	[IB_WR_LSO]			= cpu_to_be32(MLX4_OPCODE_LSO),
78 	[IB_WR_SEND_WITH_IMM]		= cpu_to_be32(MLX4_OPCODE_SEND_IMM),
79 	[IB_WR_RDMA_WRITE]		= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
80 	[IB_WR_RDMA_WRITE_WITH_IMM]	= cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
81 	[IB_WR_RDMA_READ]		= cpu_to_be32(MLX4_OPCODE_RDMA_READ),
82 	[IB_WR_ATOMIC_CMP_AND_SWP]	= cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
83 	[IB_WR_ATOMIC_FETCH_AND_ADD]	= cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
84 	[IB_WR_SEND_WITH_INV]		= cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
85 	[IB_WR_LOCAL_INV]		= cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
86 	[IB_WR_FAST_REG_MR]		= cpu_to_be32(MLX4_OPCODE_FMR),
87 };
88 
89 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
90 {
91 	return container_of(mqp, struct mlx4_ib_sqp, qp);
92 }
93 
94 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
95 {
96 	return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
97 		qp->mqp.qpn <= dev->dev->caps.sqp_start + 3;
98 }
99 
100 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
101 {
102 	return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
103 		qp->mqp.qpn <= dev->dev->caps.sqp_start + 1;
104 }
105 
106 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
107 {
108 	return mlx4_buf_offset(&qp->buf, offset);
109 }
110 
111 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
112 {
113 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
114 }
115 
116 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
117 {
118 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
119 }
120 
121 /*
122  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
123  * first four bytes of every 64 byte chunk with
124  *     0x7FFFFFF | (invalid_ownership_value << 31).
125  *
126  * When the max work request size is less than or equal to the WQE
127  * basic block size, as an optimization, we can stamp all WQEs with
128  * 0xffffffff, and skip the very first chunk of each WQE.
129  */
130 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
131 {
132 	__be32 *wqe;
133 	int i;
134 	int s;
135 	int ind;
136 	void *buf;
137 	__be32 stamp;
138 	struct mlx4_wqe_ctrl_seg *ctrl;
139 
140 	if (qp->sq_max_wqes_per_wr > 1) {
141 		s = roundup(size, 1U << qp->sq.wqe_shift);
142 		for (i = 0; i < s; i += 64) {
143 			ind = (i >> qp->sq.wqe_shift) + n;
144 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
145 						       cpu_to_be32(0xffffffff);
146 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
147 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
148 			*wqe = stamp;
149 		}
150 	} else {
151 		ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
152 		s = (ctrl->fence_size & 0x3f) << 4;
153 		for (i = 64; i < s; i += 64) {
154 			wqe = buf + i;
155 			*wqe = cpu_to_be32(0xffffffff);
156 		}
157 	}
158 }
159 
160 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
161 {
162 	struct mlx4_wqe_ctrl_seg *ctrl;
163 	struct mlx4_wqe_inline_seg *inl;
164 	void *wqe;
165 	int s;
166 
167 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
168 	s = sizeof(struct mlx4_wqe_ctrl_seg);
169 
170 	if (qp->ibqp.qp_type == IB_QPT_UD) {
171 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
172 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
173 		memset(dgram, 0, sizeof *dgram);
174 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
175 		s += sizeof(struct mlx4_wqe_datagram_seg);
176 	}
177 
178 	/* Pad the remainder of the WQE with an inline data segment. */
179 	if (size > s) {
180 		inl = wqe + s;
181 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
182 	}
183 	ctrl->srcrb_flags = 0;
184 	ctrl->fence_size = size / 16;
185 	/*
186 	 * Make sure descriptor is fully written before setting ownership bit
187 	 * (because HW can start executing as soon as we do).
188 	 */
189 	wmb();
190 
191 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
192 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
193 
194 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
195 }
196 
197 /* Post NOP WQE to prevent wrap-around in the middle of WR */
198 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
199 {
200 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
201 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
202 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
203 		ind += s;
204 	}
205 	return ind;
206 }
207 
208 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
209 {
210 	struct ib_event event;
211 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
212 
213 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
214 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
215 
216 	if (ibqp->event_handler) {
217 		event.device     = ibqp->device;
218 		event.element.qp = ibqp;
219 		switch (type) {
220 		case MLX4_EVENT_TYPE_PATH_MIG:
221 			event.event = IB_EVENT_PATH_MIG;
222 			break;
223 		case MLX4_EVENT_TYPE_COMM_EST:
224 			event.event = IB_EVENT_COMM_EST;
225 			break;
226 		case MLX4_EVENT_TYPE_SQ_DRAINED:
227 			event.event = IB_EVENT_SQ_DRAINED;
228 			break;
229 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
230 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
231 			break;
232 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
233 			event.event = IB_EVENT_QP_FATAL;
234 			break;
235 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
236 			event.event = IB_EVENT_PATH_MIG_ERR;
237 			break;
238 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
239 			event.event = IB_EVENT_QP_REQ_ERR;
240 			break;
241 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
242 			event.event = IB_EVENT_QP_ACCESS_ERR;
243 			break;
244 		default:
245 			printk(KERN_WARNING "mlx4_ib: Unexpected event type %d "
246 			       "on QP %06x\n", type, qp->qpn);
247 			return;
248 		}
249 
250 		ibqp->event_handler(&event, ibqp->qp_context);
251 	}
252 }
253 
254 static int send_wqe_overhead(enum ib_qp_type type, u32 flags)
255 {
256 	/*
257 	 * UD WQEs must have a datagram segment.
258 	 * RC and UC WQEs might have a remote address segment.
259 	 * MLX WQEs need two extra inline data segments (for the UD
260 	 * header and space for the ICRC).
261 	 */
262 	switch (type) {
263 	case IB_QPT_UD:
264 		return sizeof (struct mlx4_wqe_ctrl_seg) +
265 			sizeof (struct mlx4_wqe_datagram_seg) +
266 			((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
267 	case IB_QPT_UC:
268 		return sizeof (struct mlx4_wqe_ctrl_seg) +
269 			sizeof (struct mlx4_wqe_raddr_seg);
270 	case IB_QPT_RC:
271 		return sizeof (struct mlx4_wqe_ctrl_seg) +
272 			sizeof (struct mlx4_wqe_atomic_seg) +
273 			sizeof (struct mlx4_wqe_raddr_seg);
274 	case IB_QPT_SMI:
275 	case IB_QPT_GSI:
276 		return sizeof (struct mlx4_wqe_ctrl_seg) +
277 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
278 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
279 					   MLX4_INLINE_ALIGN) *
280 			      sizeof (struct mlx4_wqe_inline_seg),
281 			      sizeof (struct mlx4_wqe_data_seg)) +
282 			ALIGN(4 +
283 			      sizeof (struct mlx4_wqe_inline_seg),
284 			      sizeof (struct mlx4_wqe_data_seg));
285 	default:
286 		return sizeof (struct mlx4_wqe_ctrl_seg);
287 	}
288 }
289 
290 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
291 		       int is_user, int has_srq, struct mlx4_ib_qp *qp)
292 {
293 	/* Sanity check RQ size before proceeding */
294 	if (cap->max_recv_wr  > dev->dev->caps.max_wqes  ||
295 	    cap->max_recv_sge > dev->dev->caps.max_rq_sg)
296 		return -EINVAL;
297 
298 	if (has_srq) {
299 		/* QPs attached to an SRQ should have no RQ */
300 		if (cap->max_recv_wr)
301 			return -EINVAL;
302 
303 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
304 	} else {
305 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
306 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
307 			return -EINVAL;
308 
309 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
310 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
311 		qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
312 	}
313 
314 	cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
315 	cap->max_recv_sge = qp->rq.max_gs;
316 
317 	return 0;
318 }
319 
320 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
321 			      enum ib_qp_type type, struct mlx4_ib_qp *qp)
322 {
323 	int s;
324 
325 	/* Sanity check SQ size before proceeding */
326 	if (cap->max_send_wr	 > dev->dev->caps.max_wqes  ||
327 	    cap->max_send_sge	 > dev->dev->caps.max_sq_sg ||
328 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
329 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
330 		return -EINVAL;
331 
332 	/*
333 	 * For MLX transport we need 2 extra S/G entries:
334 	 * one for the header and one for the checksum at the end
335 	 */
336 	if ((type == IB_QPT_SMI || type == IB_QPT_GSI) &&
337 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
338 		return -EINVAL;
339 
340 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
341 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
342 		send_wqe_overhead(type, qp->flags);
343 
344 	if (s > dev->dev->caps.max_sq_desc_sz)
345 		return -EINVAL;
346 
347 	/*
348 	 * Hermon supports shrinking WQEs, such that a single work
349 	 * request can include multiple units of 1 << wqe_shift.  This
350 	 * way, work requests can differ in size, and do not have to
351 	 * be a power of 2 in size, saving memory and speeding up send
352 	 * WR posting.  Unfortunately, if we do this then the
353 	 * wqe_index field in CQEs can't be used to look up the WR ID
354 	 * anymore, so we do this only if selective signaling is off.
355 	 *
356 	 * Further, on 32-bit platforms, we can't use vmap() to make
357 	 * the QP buffer virtually contiguous.  Thus we have to use
358 	 * constant-sized WRs to make sure a WR is always fully within
359 	 * a single page-sized chunk.
360 	 *
361 	 * Finally, we use NOP work requests to pad the end of the
362 	 * work queue, to avoid wrap-around in the middle of WR.  We
363 	 * set NEC bit to avoid getting completions with error for
364 	 * these NOP WRs, but since NEC is only supported starting
365 	 * with firmware 2.2.232, we use constant-sized WRs for older
366 	 * firmware.
367 	 *
368 	 * And, since MLX QPs only support SEND, we use constant-sized
369 	 * WRs in this case.
370 	 *
371 	 * We look for the smallest value of wqe_shift such that the
372 	 * resulting number of wqes does not exceed device
373 	 * capabilities.
374 	 *
375 	 * We set WQE size to at least 64 bytes, this way stamping
376 	 * invalidates each WQE.
377 	 */
378 	if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
379 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
380 	    type != IB_QPT_SMI && type != IB_QPT_GSI)
381 		qp->sq.wqe_shift = ilog2(64);
382 	else
383 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
384 
385 	for (;;) {
386 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
387 
388 		/*
389 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
390 		 * allow HW to prefetch.
391 		 */
392 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
393 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
394 						    qp->sq_max_wqes_per_wr +
395 						    qp->sq_spare_wqes);
396 
397 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
398 			break;
399 
400 		if (qp->sq_max_wqes_per_wr <= 1)
401 			return -EINVAL;
402 
403 		++qp->sq.wqe_shift;
404 	}
405 
406 	qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
407 			     (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
408 			 send_wqe_overhead(type, qp->flags)) /
409 		sizeof (struct mlx4_wqe_data_seg);
410 
411 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
412 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
413 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
414 		qp->rq.offset = 0;
415 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
416 	} else {
417 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
418 		qp->sq.offset = 0;
419 	}
420 
421 	cap->max_send_wr  = qp->sq.max_post =
422 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
423 	cap->max_send_sge = min(qp->sq.max_gs,
424 				min(dev->dev->caps.max_sq_sg,
425 				    dev->dev->caps.max_rq_sg));
426 	/* We don't support inline sends for kernel QPs (yet) */
427 	cap->max_inline_data = 0;
428 
429 	return 0;
430 }
431 
432 static int set_user_sq_size(struct mlx4_ib_dev *dev,
433 			    struct mlx4_ib_qp *qp,
434 			    struct mlx4_ib_create_qp *ucmd)
435 {
436 	/* Sanity check SQ size before proceeding */
437 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
438 	    ucmd->log_sq_stride >
439 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
440 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
441 		return -EINVAL;
442 
443 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
444 	qp->sq.wqe_shift = ucmd->log_sq_stride;
445 
446 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
447 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
448 
449 	return 0;
450 }
451 
452 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
453 			    struct ib_qp_init_attr *init_attr,
454 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp *qp)
455 {
456 	int qpn;
457 	int err;
458 
459 	mutex_init(&qp->mutex);
460 	spin_lock_init(&qp->sq.lock);
461 	spin_lock_init(&qp->rq.lock);
462 
463 	qp->state	 = IB_QPS_RESET;
464 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
465 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
466 
467 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, !!init_attr->srq, qp);
468 	if (err)
469 		goto err;
470 
471 	if (pd->uobject) {
472 		struct mlx4_ib_create_qp ucmd;
473 
474 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
475 			err = -EFAULT;
476 			goto err;
477 		}
478 
479 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
480 
481 		err = set_user_sq_size(dev, qp, &ucmd);
482 		if (err)
483 			goto err;
484 
485 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
486 				       qp->buf_size, 0, 0);
487 		if (IS_ERR(qp->umem)) {
488 			err = PTR_ERR(qp->umem);
489 			goto err;
490 		}
491 
492 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
493 				    ilog2(qp->umem->page_size), &qp->mtt);
494 		if (err)
495 			goto err_buf;
496 
497 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
498 		if (err)
499 			goto err_mtt;
500 
501 		if (!init_attr->srq) {
502 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
503 						  ucmd.db_addr, &qp->db);
504 			if (err)
505 				goto err_mtt;
506 		}
507 	} else {
508 		qp->sq_no_prefetch = 0;
509 
510 		if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
511 			qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
512 
513 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
514 			qp->flags |= MLX4_IB_QP_LSO;
515 
516 		err = set_kernel_sq_size(dev, &init_attr->cap, init_attr->qp_type, qp);
517 		if (err)
518 			goto err;
519 
520 		if (!init_attr->srq) {
521 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
522 			if (err)
523 				goto err;
524 
525 			*qp->db.db = 0;
526 		}
527 
528 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf)) {
529 			err = -ENOMEM;
530 			goto err_db;
531 		}
532 
533 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
534 				    &qp->mtt);
535 		if (err)
536 			goto err_buf;
537 
538 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
539 		if (err)
540 			goto err_mtt;
541 
542 		qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), GFP_KERNEL);
543 		qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), GFP_KERNEL);
544 
545 		if (!qp->sq.wrid || !qp->rq.wrid) {
546 			err = -ENOMEM;
547 			goto err_wrid;
548 		}
549 	}
550 
551 	if (sqpn) {
552 		qpn = sqpn;
553 	} else {
554 		err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn);
555 		if (err)
556 			goto err_wrid;
557 	}
558 
559 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
560 	if (err)
561 		goto err_qpn;
562 
563 	/*
564 	 * Hardware wants QPN written in big-endian order (after
565 	 * shifting) for send doorbell.  Precompute this value to save
566 	 * a little bit when posting sends.
567 	 */
568 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
569 
570 	qp->mqp.event = mlx4_ib_qp_event;
571 
572 	return 0;
573 
574 err_qpn:
575 	if (!sqpn)
576 		mlx4_qp_release_range(dev->dev, qpn, 1);
577 
578 err_wrid:
579 	if (pd->uobject) {
580 		if (!init_attr->srq)
581 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context),
582 					      &qp->db);
583 	} else {
584 		kfree(qp->sq.wrid);
585 		kfree(qp->rq.wrid);
586 	}
587 
588 err_mtt:
589 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
590 
591 err_buf:
592 	if (pd->uobject)
593 		ib_umem_release(qp->umem);
594 	else
595 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
596 
597 err_db:
598 	if (!pd->uobject && !init_attr->srq)
599 		mlx4_db_free(dev->dev, &qp->db);
600 
601 err:
602 	return err;
603 }
604 
605 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
606 {
607 	switch (state) {
608 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
609 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
610 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
611 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
612 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
613 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
614 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
615 	default:		return -1;
616 	}
617 }
618 
619 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
620 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
621 {
622 	if (send_cq == recv_cq) {
623 		spin_lock_irq(&send_cq->lock);
624 		__acquire(&recv_cq->lock);
625 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
626 		spin_lock_irq(&send_cq->lock);
627 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
628 	} else {
629 		spin_lock_irq(&recv_cq->lock);
630 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
631 	}
632 }
633 
634 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
635 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
636 {
637 	if (send_cq == recv_cq) {
638 		__release(&recv_cq->lock);
639 		spin_unlock_irq(&send_cq->lock);
640 	} else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
641 		spin_unlock(&recv_cq->lock);
642 		spin_unlock_irq(&send_cq->lock);
643 	} else {
644 		spin_unlock(&send_cq->lock);
645 		spin_unlock_irq(&recv_cq->lock);
646 	}
647 }
648 
649 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
650 			      int is_user)
651 {
652 	struct mlx4_ib_cq *send_cq, *recv_cq;
653 
654 	if (qp->state != IB_QPS_RESET)
655 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
656 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
657 			printk(KERN_WARNING "mlx4_ib: modify QP %06x to RESET failed.\n",
658 			       qp->mqp.qpn);
659 
660 	send_cq = to_mcq(qp->ibqp.send_cq);
661 	recv_cq = to_mcq(qp->ibqp.recv_cq);
662 
663 	mlx4_ib_lock_cqs(send_cq, recv_cq);
664 
665 	if (!is_user) {
666 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
667 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
668 		if (send_cq != recv_cq)
669 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
670 	}
671 
672 	mlx4_qp_remove(dev->dev, &qp->mqp);
673 
674 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
675 
676 	mlx4_qp_free(dev->dev, &qp->mqp);
677 
678 	if (!is_sqp(dev, qp))
679 		mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
680 
681 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
682 
683 	if (is_user) {
684 		if (!qp->ibqp.srq)
685 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
686 					      &qp->db);
687 		ib_umem_release(qp->umem);
688 	} else {
689 		kfree(qp->sq.wrid);
690 		kfree(qp->rq.wrid);
691 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
692 		if (!qp->ibqp.srq)
693 			mlx4_db_free(dev->dev, &qp->db);
694 	}
695 }
696 
697 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
698 				struct ib_qp_init_attr *init_attr,
699 				struct ib_udata *udata)
700 {
701 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
702 	struct mlx4_ib_sqp *sqp;
703 	struct mlx4_ib_qp *qp;
704 	int err;
705 
706 	/*
707 	 * We only support LSO and multicast loopback blocking, and
708 	 * only for kernel UD QPs.
709 	 */
710 	if (init_attr->create_flags & ~(IB_QP_CREATE_IPOIB_UD_LSO |
711 					IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK))
712 		return ERR_PTR(-EINVAL);
713 
714 	if (init_attr->create_flags &&
715 	    (pd->uobject || init_attr->qp_type != IB_QPT_UD))
716 		return ERR_PTR(-EINVAL);
717 
718 	switch (init_attr->qp_type) {
719 	case IB_QPT_RC:
720 	case IB_QPT_UC:
721 	case IB_QPT_UD:
722 	{
723 		qp = kzalloc(sizeof *qp, GFP_KERNEL);
724 		if (!qp)
725 			return ERR_PTR(-ENOMEM);
726 
727 		err = create_qp_common(dev, pd, init_attr, udata, 0, qp);
728 		if (err) {
729 			kfree(qp);
730 			return ERR_PTR(err);
731 		}
732 
733 		qp->ibqp.qp_num = qp->mqp.qpn;
734 
735 		break;
736 	}
737 	case IB_QPT_SMI:
738 	case IB_QPT_GSI:
739 	{
740 		/* Userspace is not allowed to create special QPs: */
741 		if (pd->uobject)
742 			return ERR_PTR(-EINVAL);
743 
744 		sqp = kzalloc(sizeof *sqp, GFP_KERNEL);
745 		if (!sqp)
746 			return ERR_PTR(-ENOMEM);
747 
748 		qp = &sqp->qp;
749 
750 		err = create_qp_common(dev, pd, init_attr, udata,
751 				       dev->dev->caps.sqp_start +
752 				       (init_attr->qp_type == IB_QPT_SMI ? 0 : 2) +
753 				       init_attr->port_num - 1,
754 				       qp);
755 		if (err) {
756 			kfree(sqp);
757 			return ERR_PTR(err);
758 		}
759 
760 		qp->port	= init_attr->port_num;
761 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
762 
763 		break;
764 	}
765 	default:
766 		/* Don't support raw QPs */
767 		return ERR_PTR(-EINVAL);
768 	}
769 
770 	return &qp->ibqp;
771 }
772 
773 int mlx4_ib_destroy_qp(struct ib_qp *qp)
774 {
775 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
776 	struct mlx4_ib_qp *mqp = to_mqp(qp);
777 
778 	if (is_qp0(dev, mqp))
779 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
780 
781 	destroy_qp_common(dev, mqp, !!qp->pd->uobject);
782 
783 	if (is_sqp(dev, mqp))
784 		kfree(to_msqp(mqp));
785 	else
786 		kfree(mqp);
787 
788 	return 0;
789 }
790 
791 static int to_mlx4_st(enum ib_qp_type type)
792 {
793 	switch (type) {
794 	case IB_QPT_RC:		return MLX4_QP_ST_RC;
795 	case IB_QPT_UC:		return MLX4_QP_ST_UC;
796 	case IB_QPT_UD:		return MLX4_QP_ST_UD;
797 	case IB_QPT_SMI:
798 	case IB_QPT_GSI:	return MLX4_QP_ST_MLX;
799 	default:		return -1;
800 	}
801 }
802 
803 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
804 				   int attr_mask)
805 {
806 	u8 dest_rd_atomic;
807 	u32 access_flags;
808 	u32 hw_access_flags = 0;
809 
810 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
811 		dest_rd_atomic = attr->max_dest_rd_atomic;
812 	else
813 		dest_rd_atomic = qp->resp_depth;
814 
815 	if (attr_mask & IB_QP_ACCESS_FLAGS)
816 		access_flags = attr->qp_access_flags;
817 	else
818 		access_flags = qp->atomic_rd_en;
819 
820 	if (!dest_rd_atomic)
821 		access_flags &= IB_ACCESS_REMOTE_WRITE;
822 
823 	if (access_flags & IB_ACCESS_REMOTE_READ)
824 		hw_access_flags |= MLX4_QP_BIT_RRE;
825 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
826 		hw_access_flags |= MLX4_QP_BIT_RAE;
827 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
828 		hw_access_flags |= MLX4_QP_BIT_RWE;
829 
830 	return cpu_to_be32(hw_access_flags);
831 }
832 
833 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
834 			    int attr_mask)
835 {
836 	if (attr_mask & IB_QP_PKEY_INDEX)
837 		sqp->pkey_index = attr->pkey_index;
838 	if (attr_mask & IB_QP_QKEY)
839 		sqp->qkey = attr->qkey;
840 	if (attr_mask & IB_QP_SQ_PSN)
841 		sqp->send_psn = attr->sq_psn;
842 }
843 
844 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
845 {
846 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
847 }
848 
849 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
850 			 struct mlx4_qp_path *path, u8 port)
851 {
852 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
853 	path->rlid	    = cpu_to_be16(ah->dlid);
854 	if (ah->static_rate) {
855 		path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
856 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
857 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
858 			--path->static_rate;
859 	} else
860 		path->static_rate = 0;
861 	path->counter_index = 0xff;
862 
863 	if (ah->ah_flags & IB_AH_GRH) {
864 		if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
865 			printk(KERN_ERR "sgid_index (%u) too large. max is %d\n",
866 			       ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
867 			return -1;
868 		}
869 
870 		path->grh_mylmc |= 1 << 7;
871 		path->mgid_index = ah->grh.sgid_index;
872 		path->hop_limit  = ah->grh.hop_limit;
873 		path->tclass_flowlabel =
874 			cpu_to_be32((ah->grh.traffic_class << 20) |
875 				    (ah->grh.flow_label));
876 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
877 	}
878 
879 	path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
880 		((port - 1) << 6) | ((ah->sl & 0xf) << 2);
881 
882 	return 0;
883 }
884 
885 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
886 			       const struct ib_qp_attr *attr, int attr_mask,
887 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
888 {
889 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
890 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
891 	struct mlx4_qp_context *context;
892 	enum mlx4_qp_optpar optpar = 0;
893 	int sqd_event;
894 	int err = -EINVAL;
895 
896 	context = kzalloc(sizeof *context, GFP_KERNEL);
897 	if (!context)
898 		return -ENOMEM;
899 
900 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
901 				     (to_mlx4_st(ibqp->qp_type) << 16));
902 
903 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
904 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
905 	else {
906 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
907 		switch (attr->path_mig_state) {
908 		case IB_MIG_MIGRATED:
909 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
910 			break;
911 		case IB_MIG_REARM:
912 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
913 			break;
914 		case IB_MIG_ARMED:
915 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
916 			break;
917 		}
918 	}
919 
920 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
921 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
922 	else if (ibqp->qp_type == IB_QPT_UD) {
923 		if (qp->flags & MLX4_IB_QP_LSO)
924 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
925 					      ilog2(dev->dev->caps.max_gso_sz);
926 		else
927 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
928 	} else if (attr_mask & IB_QP_PATH_MTU) {
929 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
930 			printk(KERN_ERR "path MTU (%u) is invalid\n",
931 			       attr->path_mtu);
932 			goto out;
933 		}
934 		context->mtu_msgmax = (attr->path_mtu << 5) |
935 			ilog2(dev->dev->caps.max_msg_sz);
936 	}
937 
938 	if (qp->rq.wqe_cnt)
939 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
940 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
941 
942 	if (qp->sq.wqe_cnt)
943 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
944 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
945 
946 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
947 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
948 
949 	if (qp->ibqp.uobject)
950 		context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
951 	else
952 		context->usr_page = cpu_to_be32(dev->priv_uar.index);
953 
954 	if (attr_mask & IB_QP_DEST_QPN)
955 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
956 
957 	if (attr_mask & IB_QP_PORT) {
958 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
959 		    !(attr_mask & IB_QP_AV)) {
960 			mlx4_set_sched(&context->pri_path, attr->port_num);
961 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
962 		}
963 	}
964 
965 	if (attr_mask & IB_QP_PKEY_INDEX) {
966 		context->pri_path.pkey_index = attr->pkey_index;
967 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
968 	}
969 
970 	if (attr_mask & IB_QP_AV) {
971 		if (mlx4_set_path(dev, &attr->ah_attr, &context->pri_path,
972 				  attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
973 			goto out;
974 
975 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
976 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
977 	}
978 
979 	if (attr_mask & IB_QP_TIMEOUT) {
980 		context->pri_path.ackto = attr->timeout << 3;
981 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
982 	}
983 
984 	if (attr_mask & IB_QP_ALT_PATH) {
985 		if (attr->alt_port_num == 0 ||
986 		    attr->alt_port_num > dev->dev->caps.num_ports)
987 			goto out;
988 
989 		if (attr->alt_pkey_index >=
990 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
991 			goto out;
992 
993 		if (mlx4_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
994 				  attr->alt_port_num))
995 			goto out;
996 
997 		context->alt_path.pkey_index = attr->alt_pkey_index;
998 		context->alt_path.ackto = attr->alt_timeout << 3;
999 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1000 	}
1001 
1002 	context->pd	    = cpu_to_be32(to_mpd(ibqp->pd)->pdn);
1003 	context->params1    = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1004 
1005 	/* Set "fast registration enabled" for all kernel QPs */
1006 	if (!qp->ibqp.uobject)
1007 		context->params1 |= cpu_to_be32(1 << 11);
1008 
1009 	if (attr_mask & IB_QP_RNR_RETRY) {
1010 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1011 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1012 	}
1013 
1014 	if (attr_mask & IB_QP_RETRY_CNT) {
1015 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1016 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1017 	}
1018 
1019 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1020 		if (attr->max_rd_atomic)
1021 			context->params1 |=
1022 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1023 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1024 	}
1025 
1026 	if (attr_mask & IB_QP_SQ_PSN)
1027 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1028 
1029 	context->cqn_send = cpu_to_be32(to_mcq(ibqp->send_cq)->mcq.cqn);
1030 
1031 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1032 		if (attr->max_dest_rd_atomic)
1033 			context->params2 |=
1034 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1035 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1036 	}
1037 
1038 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1039 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1040 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1041 	}
1042 
1043 	if (ibqp->srq)
1044 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1045 
1046 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1047 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1048 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1049 	}
1050 	if (attr_mask & IB_QP_RQ_PSN)
1051 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1052 
1053 	context->cqn_recv = cpu_to_be32(to_mcq(ibqp->recv_cq)->mcq.cqn);
1054 
1055 	if (attr_mask & IB_QP_QKEY) {
1056 		context->qkey = cpu_to_be32(attr->qkey);
1057 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
1058 	}
1059 
1060 	if (ibqp->srq)
1061 		context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1062 
1063 	if (!ibqp->srq && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1064 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1065 
1066 	if (cur_state == IB_QPS_INIT &&
1067 	    new_state == IB_QPS_RTR  &&
1068 	    (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1069 	     ibqp->qp_type == IB_QPT_UD)) {
1070 		context->pri_path.sched_queue = (qp->port - 1) << 6;
1071 		if (is_qp0(dev, qp))
1072 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1073 		else
1074 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1075 	}
1076 
1077 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1078 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1079 		sqd_event = 1;
1080 	else
1081 		sqd_event = 0;
1082 
1083 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1084 		context->rlkey |= (1 << 4);
1085 
1086 	/*
1087 	 * Before passing a kernel QP to the HW, make sure that the
1088 	 * ownership bits of the send queue are set and the SQ
1089 	 * headroom is stamped so that the hardware doesn't start
1090 	 * processing stale work requests.
1091 	 */
1092 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1093 		struct mlx4_wqe_ctrl_seg *ctrl;
1094 		int i;
1095 
1096 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1097 			ctrl = get_send_wqe(qp, i);
1098 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
1099 			if (qp->sq_max_wqes_per_wr == 1)
1100 				ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1101 
1102 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1103 		}
1104 	}
1105 
1106 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1107 			     to_mlx4_state(new_state), context, optpar,
1108 			     sqd_event, &qp->mqp);
1109 	if (err)
1110 		goto out;
1111 
1112 	qp->state = new_state;
1113 
1114 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1115 		qp->atomic_rd_en = attr->qp_access_flags;
1116 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1117 		qp->resp_depth = attr->max_dest_rd_atomic;
1118 	if (attr_mask & IB_QP_PORT)
1119 		qp->port = attr->port_num;
1120 	if (attr_mask & IB_QP_ALT_PATH)
1121 		qp->alt_port = attr->alt_port_num;
1122 
1123 	if (is_sqp(dev, qp))
1124 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1125 
1126 	/*
1127 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
1128 	 * QP0 to RESET or ERROR, bring the link back down.
1129 	 */
1130 	if (is_qp0(dev, qp)) {
1131 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1132 			if (mlx4_INIT_PORT(dev->dev, qp->port))
1133 				printk(KERN_WARNING "INIT_PORT failed for port %d\n",
1134 				       qp->port);
1135 
1136 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1137 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1138 			mlx4_CLOSE_PORT(dev->dev, qp->port);
1139 	}
1140 
1141 	/*
1142 	 * If we moved a kernel QP to RESET, clean up all old CQ
1143 	 * entries and reinitialize the QP.
1144 	 */
1145 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1146 		mlx4_ib_cq_clean(to_mcq(ibqp->recv_cq), qp->mqp.qpn,
1147 				 ibqp->srq ? to_msrq(ibqp->srq): NULL);
1148 		if (ibqp->send_cq != ibqp->recv_cq)
1149 			mlx4_ib_cq_clean(to_mcq(ibqp->send_cq), qp->mqp.qpn, NULL);
1150 
1151 		qp->rq.head = 0;
1152 		qp->rq.tail = 0;
1153 		qp->sq.head = 0;
1154 		qp->sq.tail = 0;
1155 		qp->sq_next_wqe = 0;
1156 		if (!ibqp->srq)
1157 			*qp->db.db  = 0;
1158 	}
1159 
1160 out:
1161 	kfree(context);
1162 	return err;
1163 }
1164 
1165 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1166 		      int attr_mask, struct ib_udata *udata)
1167 {
1168 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1169 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1170 	enum ib_qp_state cur_state, new_state;
1171 	int err = -EINVAL;
1172 
1173 	mutex_lock(&qp->mutex);
1174 
1175 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1176 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1177 
1178 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask))
1179 		goto out;
1180 
1181 	if ((attr_mask & IB_QP_PORT) &&
1182 	    (attr->port_num == 0 || attr->port_num > dev->dev->caps.num_ports)) {
1183 		goto out;
1184 	}
1185 
1186 	if (attr_mask & IB_QP_PKEY_INDEX) {
1187 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1188 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p])
1189 			goto out;
1190 	}
1191 
1192 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1193 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
1194 		goto out;
1195 	}
1196 
1197 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1198 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
1199 		goto out;
1200 	}
1201 
1202 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1203 		err = 0;
1204 		goto out;
1205 	}
1206 
1207 	err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1208 
1209 out:
1210 	mutex_unlock(&qp->mutex);
1211 	return err;
1212 }
1213 
1214 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
1215 			    void *wqe, unsigned *mlx_seg_len)
1216 {
1217 	struct ib_device *ib_dev = &to_mdev(sqp->qp.ibqp.device)->ib_dev;
1218 	struct mlx4_wqe_mlx_seg *mlx = wqe;
1219 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
1220 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
1221 	u16 pkey;
1222 	int send_size;
1223 	int header_size;
1224 	int spc;
1225 	int i;
1226 
1227 	send_size = 0;
1228 	for (i = 0; i < wr->num_sge; ++i)
1229 		send_size += wr->sg_list[i].length;
1230 
1231 	ib_ud_header_init(send_size, mlx4_ib_ah_grh_present(ah), &sqp->ud_header);
1232 
1233 	sqp->ud_header.lrh.service_level   =
1234 		be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 28;
1235 	sqp->ud_header.lrh.destination_lid = ah->av.dlid;
1236 	sqp->ud_header.lrh.source_lid      = cpu_to_be16(ah->av.g_slid & 0x7f);
1237 	if (mlx4_ib_ah_grh_present(ah)) {
1238 		sqp->ud_header.grh.traffic_class =
1239 			(be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 20) & 0xff;
1240 		sqp->ud_header.grh.flow_label    =
1241 			ah->av.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
1242 		sqp->ud_header.grh.hop_limit     = ah->av.hop_limit;
1243 		ib_get_cached_gid(ib_dev, be32_to_cpu(ah->av.port_pd) >> 24,
1244 				  ah->av.gid_index, &sqp->ud_header.grh.source_gid);
1245 		memcpy(sqp->ud_header.grh.destination_gid.raw,
1246 		       ah->av.dgid, 16);
1247 	}
1248 
1249 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1250 	mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
1251 				  (sqp->ud_header.lrh.destination_lid ==
1252 				   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
1253 				  (sqp->ud_header.lrh.service_level << 8));
1254 	mlx->rlid   = sqp->ud_header.lrh.destination_lid;
1255 
1256 	switch (wr->opcode) {
1257 	case IB_WR_SEND:
1258 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
1259 		sqp->ud_header.immediate_present = 0;
1260 		break;
1261 	case IB_WR_SEND_WITH_IMM:
1262 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1263 		sqp->ud_header.immediate_present = 1;
1264 		sqp->ud_header.immediate_data    = wr->ex.imm_data;
1265 		break;
1266 	default:
1267 		return -EINVAL;
1268 	}
1269 
1270 	sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1271 	if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1272 		sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1273 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1274 	if (!sqp->qp.ibqp.qp_num)
1275 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
1276 	else
1277 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
1278 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1279 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1280 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1281 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1282 					       sqp->qkey : wr->wr.ud.remote_qkey);
1283 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1284 
1285 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
1286 
1287 	if (0) {
1288 		printk(KERN_ERR "built UD header of size %d:\n", header_size);
1289 		for (i = 0; i < header_size / 4; ++i) {
1290 			if (i % 8 == 0)
1291 				printk("  [%02x] ", i * 4);
1292 			printk(" %08x",
1293 			       be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
1294 			if ((i + 1) % 8 == 0)
1295 				printk("\n");
1296 		}
1297 		printk("\n");
1298 	}
1299 
1300 	/*
1301 	 * Inline data segments may not cross a 64 byte boundary.  If
1302 	 * our UD header is bigger than the space available up to the
1303 	 * next 64 byte boundary in the WQE, use two inline data
1304 	 * segments to hold the UD header.
1305 	 */
1306 	spc = MLX4_INLINE_ALIGN -
1307 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
1308 	if (header_size <= spc) {
1309 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
1310 		memcpy(inl + 1, sqp->header_buf, header_size);
1311 		i = 1;
1312 	} else {
1313 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
1314 		memcpy(inl + 1, sqp->header_buf, spc);
1315 
1316 		inl = (void *) (inl + 1) + spc;
1317 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
1318 		/*
1319 		 * Need a barrier here to make sure all the data is
1320 		 * visible before the byte_count field is set.
1321 		 * Otherwise the HCA prefetcher could grab the 64-byte
1322 		 * chunk with this inline segment and get a valid (!=
1323 		 * 0xffffffff) byte count but stale data, and end up
1324 		 * generating a packet with bad headers.
1325 		 *
1326 		 * The first inline segment's byte_count field doesn't
1327 		 * need a barrier, because it comes after a
1328 		 * control/MLX segment and therefore is at an offset
1329 		 * of 16 mod 64.
1330 		 */
1331 		wmb();
1332 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
1333 		i = 2;
1334 	}
1335 
1336 	*mlx_seg_len =
1337 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
1338 	return 0;
1339 }
1340 
1341 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1342 {
1343 	unsigned cur;
1344 	struct mlx4_ib_cq *cq;
1345 
1346 	cur = wq->head - wq->tail;
1347 	if (likely(cur + nreq < wq->max_post))
1348 		return 0;
1349 
1350 	cq = to_mcq(ib_cq);
1351 	spin_lock(&cq->lock);
1352 	cur = wq->head - wq->tail;
1353 	spin_unlock(&cq->lock);
1354 
1355 	return cur + nreq >= wq->max_post;
1356 }
1357 
1358 static __be32 convert_access(int acc)
1359 {
1360 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? cpu_to_be32(MLX4_WQE_FMR_PERM_ATOMIC)       : 0) |
1361 	       (acc & IB_ACCESS_REMOTE_WRITE  ? cpu_to_be32(MLX4_WQE_FMR_PERM_REMOTE_WRITE) : 0) |
1362 	       (acc & IB_ACCESS_REMOTE_READ   ? cpu_to_be32(MLX4_WQE_FMR_PERM_REMOTE_READ)  : 0) |
1363 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
1364 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
1365 }
1366 
1367 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr)
1368 {
1369 	struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
1370 	int i;
1371 
1372 	for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i)
1373 		mfrpl->mapped_page_list[i] =
1374 			cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] |
1375 				    MLX4_MTT_FLAG_PRESENT);
1376 
1377 	fseg->flags		= convert_access(wr->wr.fast_reg.access_flags);
1378 	fseg->mem_key		= cpu_to_be32(wr->wr.fast_reg.rkey);
1379 	fseg->buf_list		= cpu_to_be64(mfrpl->map);
1380 	fseg->start_addr	= cpu_to_be64(wr->wr.fast_reg.iova_start);
1381 	fseg->reg_len		= cpu_to_be64(wr->wr.fast_reg.length);
1382 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
1383 	fseg->page_size		= cpu_to_be32(wr->wr.fast_reg.page_shift);
1384 	fseg->reserved[0]	= 0;
1385 	fseg->reserved[1]	= 0;
1386 }
1387 
1388 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
1389 {
1390 	iseg->flags	= 0;
1391 	iseg->mem_key	= cpu_to_be32(rkey);
1392 	iseg->guest_id	= 0;
1393 	iseg->pa	= 0;
1394 }
1395 
1396 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
1397 					  u64 remote_addr, u32 rkey)
1398 {
1399 	rseg->raddr    = cpu_to_be64(remote_addr);
1400 	rseg->rkey     = cpu_to_be32(rkey);
1401 	rseg->reserved = 0;
1402 }
1403 
1404 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
1405 {
1406 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1407 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
1408 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
1409 	} else {
1410 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
1411 		aseg->compare  = 0;
1412 	}
1413 
1414 }
1415 
1416 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
1417 			     struct ib_send_wr *wr)
1418 {
1419 	memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
1420 	dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1421 	dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1422 }
1423 
1424 static void set_mlx_icrc_seg(void *dseg)
1425 {
1426 	u32 *t = dseg;
1427 	struct mlx4_wqe_inline_seg *iseg = dseg;
1428 
1429 	t[1] = 0;
1430 
1431 	/*
1432 	 * Need a barrier here before writing the byte_count field to
1433 	 * make sure that all the data is visible before the
1434 	 * byte_count field is set.  Otherwise, if the segment begins
1435 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
1436 	 * chunk and get a valid (!= * 0xffffffff) byte count but
1437 	 * stale data, and end up sending the wrong data.
1438 	 */
1439 	wmb();
1440 
1441 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
1442 }
1443 
1444 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
1445 {
1446 	dseg->lkey       = cpu_to_be32(sg->lkey);
1447 	dseg->addr       = cpu_to_be64(sg->addr);
1448 
1449 	/*
1450 	 * Need a barrier here before writing the byte_count field to
1451 	 * make sure that all the data is visible before the
1452 	 * byte_count field is set.  Otherwise, if the segment begins
1453 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
1454 	 * chunk and get a valid (!= * 0xffffffff) byte count but
1455 	 * stale data, and end up sending the wrong data.
1456 	 */
1457 	wmb();
1458 
1459 	dseg->byte_count = cpu_to_be32(sg->length);
1460 }
1461 
1462 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
1463 {
1464 	dseg->byte_count = cpu_to_be32(sg->length);
1465 	dseg->lkey       = cpu_to_be32(sg->lkey);
1466 	dseg->addr       = cpu_to_be64(sg->addr);
1467 }
1468 
1469 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr,
1470 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
1471 			 __be32 *lso_hdr_sz, __be32 *blh)
1472 {
1473 	unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
1474 
1475 	if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
1476 		*blh = cpu_to_be32(1 << 6);
1477 
1478 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
1479 		     wr->num_sge > qp->sq.max_gs - (halign >> 4)))
1480 		return -EINVAL;
1481 
1482 	memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
1483 
1484 	*lso_hdr_sz  = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 |
1485 				   wr->wr.ud.hlen);
1486 	*lso_seg_len = halign;
1487 	return 0;
1488 }
1489 
1490 static __be32 send_ieth(struct ib_send_wr *wr)
1491 {
1492 	switch (wr->opcode) {
1493 	case IB_WR_SEND_WITH_IMM:
1494 	case IB_WR_RDMA_WRITE_WITH_IMM:
1495 		return wr->ex.imm_data;
1496 
1497 	case IB_WR_SEND_WITH_INV:
1498 		return cpu_to_be32(wr->ex.invalidate_rkey);
1499 
1500 	default:
1501 		return 0;
1502 	}
1503 }
1504 
1505 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1506 		      struct ib_send_wr **bad_wr)
1507 {
1508 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1509 	void *wqe;
1510 	struct mlx4_wqe_ctrl_seg *ctrl;
1511 	struct mlx4_wqe_data_seg *dseg;
1512 	unsigned long flags;
1513 	int nreq;
1514 	int err = 0;
1515 	unsigned ind;
1516 	int uninitialized_var(stamp);
1517 	int uninitialized_var(size);
1518 	unsigned uninitialized_var(seglen);
1519 	__be32 dummy;
1520 	__be32 *lso_wqe;
1521 	__be32 uninitialized_var(lso_hdr_sz);
1522 	__be32 blh;
1523 	int i;
1524 
1525 	spin_lock_irqsave(&qp->sq.lock, flags);
1526 
1527 	ind = qp->sq_next_wqe;
1528 
1529 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1530 		lso_wqe = &dummy;
1531 		blh = 0;
1532 
1533 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1534 			err = -ENOMEM;
1535 			*bad_wr = wr;
1536 			goto out;
1537 		}
1538 
1539 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
1540 			err = -EINVAL;
1541 			*bad_wr = wr;
1542 			goto out;
1543 		}
1544 
1545 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
1546 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
1547 
1548 		ctrl->srcrb_flags =
1549 			(wr->send_flags & IB_SEND_SIGNALED ?
1550 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
1551 			(wr->send_flags & IB_SEND_SOLICITED ?
1552 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
1553 			((wr->send_flags & IB_SEND_IP_CSUM) ?
1554 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
1555 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
1556 			qp->sq_signal_bits;
1557 
1558 		ctrl->imm = send_ieth(wr);
1559 
1560 		wqe += sizeof *ctrl;
1561 		size = sizeof *ctrl / 16;
1562 
1563 		switch (ibqp->qp_type) {
1564 		case IB_QPT_RC:
1565 		case IB_QPT_UC:
1566 			switch (wr->opcode) {
1567 			case IB_WR_ATOMIC_CMP_AND_SWP:
1568 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1569 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
1570 					      wr->wr.atomic.rkey);
1571 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
1572 
1573 				set_atomic_seg(wqe, wr);
1574 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
1575 
1576 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
1577 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
1578 
1579 				break;
1580 
1581 			case IB_WR_RDMA_READ:
1582 			case IB_WR_RDMA_WRITE:
1583 			case IB_WR_RDMA_WRITE_WITH_IMM:
1584 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1585 					      wr->wr.rdma.rkey);
1586 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
1587 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
1588 				break;
1589 
1590 			case IB_WR_LOCAL_INV:
1591 				ctrl->srcrb_flags |=
1592 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
1593 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
1594 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
1595 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
1596 				break;
1597 
1598 			case IB_WR_FAST_REG_MR:
1599 				ctrl->srcrb_flags |=
1600 					cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
1601 				set_fmr_seg(wqe, wr);
1602 				wqe  += sizeof (struct mlx4_wqe_fmr_seg);
1603 				size += sizeof (struct mlx4_wqe_fmr_seg) / 16;
1604 				break;
1605 
1606 			default:
1607 				/* No extra segments required for sends */
1608 				break;
1609 			}
1610 			break;
1611 
1612 		case IB_QPT_UD:
1613 			set_datagram_seg(wqe, wr);
1614 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
1615 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
1616 
1617 			if (wr->opcode == IB_WR_LSO) {
1618 				err = build_lso_seg(wqe, wr, qp, &seglen, &lso_hdr_sz, &blh);
1619 				if (unlikely(err)) {
1620 					*bad_wr = wr;
1621 					goto out;
1622 				}
1623 				lso_wqe = (__be32 *) wqe;
1624 				wqe  += seglen;
1625 				size += seglen / 16;
1626 			}
1627 			break;
1628 
1629 		case IB_QPT_SMI:
1630 		case IB_QPT_GSI:
1631 			err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
1632 			if (unlikely(err)) {
1633 				*bad_wr = wr;
1634 				goto out;
1635 			}
1636 			wqe  += seglen;
1637 			size += seglen / 16;
1638 			break;
1639 
1640 		default:
1641 			break;
1642 		}
1643 
1644 		/*
1645 		 * Write data segments in reverse order, so as to
1646 		 * overwrite cacheline stamp last within each
1647 		 * cacheline.  This avoids issues with WQE
1648 		 * prefetching.
1649 		 */
1650 
1651 		dseg = wqe;
1652 		dseg += wr->num_sge - 1;
1653 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
1654 
1655 		/* Add one more inline data segment for ICRC for MLX sends */
1656 		if (unlikely(qp->ibqp.qp_type == IB_QPT_SMI ||
1657 			     qp->ibqp.qp_type == IB_QPT_GSI)) {
1658 			set_mlx_icrc_seg(dseg + 1);
1659 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
1660 		}
1661 
1662 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
1663 			set_data_seg(dseg, wr->sg_list + i);
1664 
1665 		/*
1666 		 * Possibly overwrite stamping in cacheline with LSO
1667 		 * segment only after making sure all data segments
1668 		 * are written.
1669 		 */
1670 		wmb();
1671 		*lso_wqe = lso_hdr_sz;
1672 
1673 		ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
1674 				    MLX4_WQE_CTRL_FENCE : 0) | size;
1675 
1676 		/*
1677 		 * Make sure descriptor is fully written before
1678 		 * setting ownership bit (because HW can start
1679 		 * executing as soon as we do).
1680 		 */
1681 		wmb();
1682 
1683 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
1684 			err = -EINVAL;
1685 			goto out;
1686 		}
1687 
1688 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
1689 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
1690 
1691 		stamp = ind + qp->sq_spare_wqes;
1692 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
1693 
1694 		/*
1695 		 * We can improve latency by not stamping the last
1696 		 * send queue WQE until after ringing the doorbell, so
1697 		 * only stamp here if there are still more WQEs to post.
1698 		 *
1699 		 * Same optimization applies to padding with NOP wqe
1700 		 * in case of WQE shrinking (used to prevent wrap-around
1701 		 * in the middle of WR).
1702 		 */
1703 		if (wr->next) {
1704 			stamp_send_wqe(qp, stamp, size * 16);
1705 			ind = pad_wraparound(qp, ind);
1706 		}
1707 	}
1708 
1709 out:
1710 	if (likely(nreq)) {
1711 		qp->sq.head += nreq;
1712 
1713 		/*
1714 		 * Make sure that descriptors are written before
1715 		 * doorbell record.
1716 		 */
1717 		wmb();
1718 
1719 		writel(qp->doorbell_qpn,
1720 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
1721 
1722 		/*
1723 		 * Make sure doorbells don't leak out of SQ spinlock
1724 		 * and reach the HCA out of order.
1725 		 */
1726 		mmiowb();
1727 
1728 		stamp_send_wqe(qp, stamp, size * 16);
1729 
1730 		ind = pad_wraparound(qp, ind);
1731 		qp->sq_next_wqe = ind;
1732 	}
1733 
1734 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1735 
1736 	return err;
1737 }
1738 
1739 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1740 		      struct ib_recv_wr **bad_wr)
1741 {
1742 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1743 	struct mlx4_wqe_data_seg *scat;
1744 	unsigned long flags;
1745 	int err = 0;
1746 	int nreq;
1747 	int ind;
1748 	int i;
1749 
1750 	spin_lock_irqsave(&qp->rq.lock, flags);
1751 
1752 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
1753 
1754 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1755 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1756 			err = -ENOMEM;
1757 			*bad_wr = wr;
1758 			goto out;
1759 		}
1760 
1761 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1762 			err = -EINVAL;
1763 			*bad_wr = wr;
1764 			goto out;
1765 		}
1766 
1767 		scat = get_recv_wqe(qp, ind);
1768 
1769 		for (i = 0; i < wr->num_sge; ++i)
1770 			__set_data_seg(scat + i, wr->sg_list + i);
1771 
1772 		if (i < qp->rq.max_gs) {
1773 			scat[i].byte_count = 0;
1774 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
1775 			scat[i].addr       = 0;
1776 		}
1777 
1778 		qp->rq.wrid[ind] = wr->wr_id;
1779 
1780 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
1781 	}
1782 
1783 out:
1784 	if (likely(nreq)) {
1785 		qp->rq.head += nreq;
1786 
1787 		/*
1788 		 * Make sure that descriptors are written before
1789 		 * doorbell record.
1790 		 */
1791 		wmb();
1792 
1793 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
1794 	}
1795 
1796 	spin_unlock_irqrestore(&qp->rq.lock, flags);
1797 
1798 	return err;
1799 }
1800 
1801 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
1802 {
1803 	switch (mlx4_state) {
1804 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
1805 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
1806 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
1807 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
1808 	case MLX4_QP_STATE_SQ_DRAINING:
1809 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
1810 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
1811 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
1812 	default:		     return -1;
1813 	}
1814 }
1815 
1816 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
1817 {
1818 	switch (mlx4_mig_state) {
1819 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
1820 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
1821 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
1822 	default: return -1;
1823 	}
1824 }
1825 
1826 static int to_ib_qp_access_flags(int mlx4_flags)
1827 {
1828 	int ib_flags = 0;
1829 
1830 	if (mlx4_flags & MLX4_QP_BIT_RRE)
1831 		ib_flags |= IB_ACCESS_REMOTE_READ;
1832 	if (mlx4_flags & MLX4_QP_BIT_RWE)
1833 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
1834 	if (mlx4_flags & MLX4_QP_BIT_RAE)
1835 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
1836 
1837 	return ib_flags;
1838 }
1839 
1840 static void to_ib_ah_attr(struct mlx4_dev *dev, struct ib_ah_attr *ib_ah_attr,
1841 				struct mlx4_qp_path *path)
1842 {
1843 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
1844 	ib_ah_attr->port_num	  = path->sched_queue & 0x40 ? 2 : 1;
1845 
1846 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
1847 		return;
1848 
1849 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
1850 	ib_ah_attr->sl		  = (path->sched_queue >> 2) & 0xf;
1851 	ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
1852 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
1853 	ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
1854 	if (ib_ah_attr->ah_flags) {
1855 		ib_ah_attr->grh.sgid_index = path->mgid_index;
1856 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
1857 		ib_ah_attr->grh.traffic_class =
1858 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
1859 		ib_ah_attr->grh.flow_label =
1860 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
1861 		memcpy(ib_ah_attr->grh.dgid.raw,
1862 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
1863 	}
1864 }
1865 
1866 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
1867 		     struct ib_qp_init_attr *qp_init_attr)
1868 {
1869 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1870 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1871 	struct mlx4_qp_context context;
1872 	int mlx4_state;
1873 	int err = 0;
1874 
1875 	mutex_lock(&qp->mutex);
1876 
1877 	if (qp->state == IB_QPS_RESET) {
1878 		qp_attr->qp_state = IB_QPS_RESET;
1879 		goto done;
1880 	}
1881 
1882 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
1883 	if (err) {
1884 		err = -EINVAL;
1885 		goto out;
1886 	}
1887 
1888 	mlx4_state = be32_to_cpu(context.flags) >> 28;
1889 
1890 	qp->state		     = to_ib_qp_state(mlx4_state);
1891 	qp_attr->qp_state	     = qp->state;
1892 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
1893 	qp_attr->path_mig_state	     =
1894 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
1895 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
1896 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
1897 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
1898 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
1899 	qp_attr->qp_access_flags     =
1900 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
1901 
1902 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
1903 		to_ib_ah_attr(dev->dev, &qp_attr->ah_attr, &context.pri_path);
1904 		to_ib_ah_attr(dev->dev, &qp_attr->alt_ah_attr, &context.alt_path);
1905 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
1906 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
1907 	}
1908 
1909 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
1910 	if (qp_attr->qp_state == IB_QPS_INIT)
1911 		qp_attr->port_num = qp->port;
1912 	else
1913 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
1914 
1915 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
1916 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
1917 
1918 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
1919 
1920 	qp_attr->max_dest_rd_atomic =
1921 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
1922 	qp_attr->min_rnr_timer	    =
1923 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
1924 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
1925 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
1926 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
1927 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
1928 
1929 done:
1930 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
1931 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
1932 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
1933 
1934 	if (!ibqp->uobject) {
1935 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
1936 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
1937 	} else {
1938 		qp_attr->cap.max_send_wr  = 0;
1939 		qp_attr->cap.max_send_sge = 0;
1940 	}
1941 
1942 	/*
1943 	 * We don't support inline sends for kernel QPs (yet), and we
1944 	 * don't know what userspace's value should be.
1945 	 */
1946 	qp_attr->cap.max_inline_data = 0;
1947 
1948 	qp_init_attr->cap	     = qp_attr->cap;
1949 
1950 	qp_init_attr->create_flags = 0;
1951 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
1952 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
1953 
1954 	if (qp->flags & MLX4_IB_QP_LSO)
1955 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
1956 
1957 out:
1958 	mutex_unlock(&qp->mutex);
1959 	return err;
1960 }
1961 
1962