xref: /openbmc/linux/drivers/infiniband/hw/mlx4/qp.c (revision f42b3800)
1 /*
2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/log2.h>
34 
35 #include <rdma/ib_cache.h>
36 #include <rdma/ib_pack.h>
37 
38 #include <linux/mlx4/qp.h>
39 
40 #include "mlx4_ib.h"
41 #include "user.h"
42 
43 enum {
44 	MLX4_IB_ACK_REQ_FREQ	= 8,
45 };
46 
47 enum {
48 	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
49 	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f
50 };
51 
52 enum {
53 	/*
54 	 * Largest possible UD header: send with GRH and immediate data.
55 	 */
56 	MLX4_IB_UD_HEADER_SIZE		= 72
57 };
58 
59 struct mlx4_ib_sqp {
60 	struct mlx4_ib_qp	qp;
61 	int			pkey_index;
62 	u32			qkey;
63 	u32			send_psn;
64 	struct ib_ud_header	ud_header;
65 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
66 };
67 
68 enum {
69 	MLX4_IB_MIN_SQ_STRIDE = 6
70 };
71 
72 static const __be32 mlx4_ib_opcode[] = {
73 	[IB_WR_SEND]			= __constant_cpu_to_be32(MLX4_OPCODE_SEND),
74 	[IB_WR_LSO]			= __constant_cpu_to_be32(MLX4_OPCODE_LSO),
75 	[IB_WR_SEND_WITH_IMM]		= __constant_cpu_to_be32(MLX4_OPCODE_SEND_IMM),
76 	[IB_WR_RDMA_WRITE]		= __constant_cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
77 	[IB_WR_RDMA_WRITE_WITH_IMM]	= __constant_cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
78 	[IB_WR_RDMA_READ]		= __constant_cpu_to_be32(MLX4_OPCODE_RDMA_READ),
79 	[IB_WR_ATOMIC_CMP_AND_SWP]	= __constant_cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
80 	[IB_WR_ATOMIC_FETCH_AND_ADD]	= __constant_cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
81 };
82 
83 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
84 {
85 	return container_of(mqp, struct mlx4_ib_sqp, qp);
86 }
87 
88 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
89 {
90 	return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
91 		qp->mqp.qpn <= dev->dev->caps.sqp_start + 3;
92 }
93 
94 static int is_qp0(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 + 1;
98 }
99 
100 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
101 {
102 	return mlx4_buf_offset(&qp->buf, offset);
103 }
104 
105 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
106 {
107 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
108 }
109 
110 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
111 {
112 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
113 }
114 
115 /*
116  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
117  * first four bytes of every 64 byte chunk with
118  *     0x7FFFFFF | (invalid_ownership_value << 31).
119  *
120  * When the max work request size is less than or equal to the WQE
121  * basic block size, as an optimization, we can stamp all WQEs with
122  * 0xffffffff, and skip the very first chunk of each WQE.
123  */
124 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
125 {
126 	__be32 *wqe;
127 	int i;
128 	int s;
129 	int ind;
130 	void *buf;
131 	__be32 stamp;
132 
133 	s = roundup(size, 1U << qp->sq.wqe_shift);
134 	if (qp->sq_max_wqes_per_wr > 1) {
135 		for (i = 0; i < s; i += 64) {
136 			ind = (i >> qp->sq.wqe_shift) + n;
137 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
138 						       cpu_to_be32(0xffffffff);
139 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
140 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
141 			*wqe = stamp;
142 		}
143 	} else {
144 		buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
145 		for (i = 64; i < s; i += 64) {
146 			wqe = buf + i;
147 			*wqe = cpu_to_be32(0xffffffff);
148 		}
149 	}
150 }
151 
152 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
153 {
154 	struct mlx4_wqe_ctrl_seg *ctrl;
155 	struct mlx4_wqe_inline_seg *inl;
156 	void *wqe;
157 	int s;
158 
159 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
160 	s = sizeof(struct mlx4_wqe_ctrl_seg);
161 
162 	if (qp->ibqp.qp_type == IB_QPT_UD) {
163 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
164 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
165 		memset(dgram, 0, sizeof *dgram);
166 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
167 		s += sizeof(struct mlx4_wqe_datagram_seg);
168 	}
169 
170 	/* Pad the remainder of the WQE with an inline data segment. */
171 	if (size > s) {
172 		inl = wqe + s;
173 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
174 	}
175 	ctrl->srcrb_flags = 0;
176 	ctrl->fence_size = size / 16;
177 	/*
178 	 * Make sure descriptor is fully written before setting ownership bit
179 	 * (because HW can start executing as soon as we do).
180 	 */
181 	wmb();
182 
183 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
184 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
185 
186 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
187 }
188 
189 /* Post NOP WQE to prevent wrap-around in the middle of WR */
190 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
191 {
192 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
193 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
194 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
195 		ind += s;
196 	}
197 	return ind;
198 }
199 
200 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
201 {
202 	struct ib_event event;
203 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
204 
205 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
206 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
207 
208 	if (ibqp->event_handler) {
209 		event.device     = ibqp->device;
210 		event.element.qp = ibqp;
211 		switch (type) {
212 		case MLX4_EVENT_TYPE_PATH_MIG:
213 			event.event = IB_EVENT_PATH_MIG;
214 			break;
215 		case MLX4_EVENT_TYPE_COMM_EST:
216 			event.event = IB_EVENT_COMM_EST;
217 			break;
218 		case MLX4_EVENT_TYPE_SQ_DRAINED:
219 			event.event = IB_EVENT_SQ_DRAINED;
220 			break;
221 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
222 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
223 			break;
224 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
225 			event.event = IB_EVENT_QP_FATAL;
226 			break;
227 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
228 			event.event = IB_EVENT_PATH_MIG_ERR;
229 			break;
230 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
231 			event.event = IB_EVENT_QP_REQ_ERR;
232 			break;
233 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
234 			event.event = IB_EVENT_QP_ACCESS_ERR;
235 			break;
236 		default:
237 			printk(KERN_WARNING "mlx4_ib: Unexpected event type %d "
238 			       "on QP %06x\n", type, qp->qpn);
239 			return;
240 		}
241 
242 		ibqp->event_handler(&event, ibqp->qp_context);
243 	}
244 }
245 
246 static int send_wqe_overhead(enum ib_qp_type type, u32 flags)
247 {
248 	/*
249 	 * UD WQEs must have a datagram segment.
250 	 * RC and UC WQEs might have a remote address segment.
251 	 * MLX WQEs need two extra inline data segments (for the UD
252 	 * header and space for the ICRC).
253 	 */
254 	switch (type) {
255 	case IB_QPT_UD:
256 		return sizeof (struct mlx4_wqe_ctrl_seg) +
257 			sizeof (struct mlx4_wqe_datagram_seg) +
258 			((flags & MLX4_IB_QP_LSO) ? 64 : 0);
259 	case IB_QPT_UC:
260 		return sizeof (struct mlx4_wqe_ctrl_seg) +
261 			sizeof (struct mlx4_wqe_raddr_seg);
262 	case IB_QPT_RC:
263 		return sizeof (struct mlx4_wqe_ctrl_seg) +
264 			sizeof (struct mlx4_wqe_atomic_seg) +
265 			sizeof (struct mlx4_wqe_raddr_seg);
266 	case IB_QPT_SMI:
267 	case IB_QPT_GSI:
268 		return sizeof (struct mlx4_wqe_ctrl_seg) +
269 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
270 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
271 					   MLX4_INLINE_ALIGN) *
272 			      sizeof (struct mlx4_wqe_inline_seg),
273 			      sizeof (struct mlx4_wqe_data_seg)) +
274 			ALIGN(4 +
275 			      sizeof (struct mlx4_wqe_inline_seg),
276 			      sizeof (struct mlx4_wqe_data_seg));
277 	default:
278 		return sizeof (struct mlx4_wqe_ctrl_seg);
279 	}
280 }
281 
282 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
283 		       int is_user, int has_srq, struct mlx4_ib_qp *qp)
284 {
285 	/* Sanity check RQ size before proceeding */
286 	if (cap->max_recv_wr  > dev->dev->caps.max_wqes  ||
287 	    cap->max_recv_sge > dev->dev->caps.max_rq_sg)
288 		return -EINVAL;
289 
290 	if (has_srq) {
291 		/* QPs attached to an SRQ should have no RQ */
292 		if (cap->max_recv_wr)
293 			return -EINVAL;
294 
295 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
296 	} else {
297 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
298 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
299 			return -EINVAL;
300 
301 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
302 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
303 		qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
304 	}
305 
306 	cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
307 	cap->max_recv_sge = qp->rq.max_gs;
308 
309 	return 0;
310 }
311 
312 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
313 			      enum ib_qp_type type, struct mlx4_ib_qp *qp)
314 {
315 	int s;
316 
317 	/* Sanity check SQ size before proceeding */
318 	if (cap->max_send_wr	 > dev->dev->caps.max_wqes  ||
319 	    cap->max_send_sge	 > dev->dev->caps.max_sq_sg ||
320 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
321 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
322 		return -EINVAL;
323 
324 	/*
325 	 * For MLX transport we need 2 extra S/G entries:
326 	 * one for the header and one for the checksum at the end
327 	 */
328 	if ((type == IB_QPT_SMI || type == IB_QPT_GSI) &&
329 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
330 		return -EINVAL;
331 
332 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
333 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
334 		send_wqe_overhead(type, qp->flags);
335 
336 	/*
337 	 * Hermon supports shrinking WQEs, such that a single work
338 	 * request can include multiple units of 1 << wqe_shift.  This
339 	 * way, work requests can differ in size, and do not have to
340 	 * be a power of 2 in size, saving memory and speeding up send
341 	 * WR posting.  Unfortunately, if we do this then the
342 	 * wqe_index field in CQEs can't be used to look up the WR ID
343 	 * anymore, so we do this only if selective signaling is off.
344 	 *
345 	 * Further, on 32-bit platforms, we can't use vmap() to make
346 	 * the QP buffer virtually contigious.  Thus we have to use
347 	 * constant-sized WRs to make sure a WR is always fully within
348 	 * a single page-sized chunk.
349 	 *
350 	 * Finally, we use NOP work requests to pad the end of the
351 	 * work queue, to avoid wrap-around in the middle of WR.  We
352 	 * set NEC bit to avoid getting completions with error for
353 	 * these NOP WRs, but since NEC is only supported starting
354 	 * with firmware 2.2.232, we use constant-sized WRs for older
355 	 * firmware.
356 	 *
357 	 * And, since MLX QPs only support SEND, we use constant-sized
358 	 * WRs in this case.
359 	 *
360 	 * We look for the smallest value of wqe_shift such that the
361 	 * resulting number of wqes does not exceed device
362 	 * capabilities.
363 	 *
364 	 * We set WQE size to at least 64 bytes, this way stamping
365 	 * invalidates each WQE.
366 	 */
367 	if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
368 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
369 	    type != IB_QPT_SMI && type != IB_QPT_GSI)
370 		qp->sq.wqe_shift = ilog2(64);
371 	else
372 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
373 
374 	for (;;) {
375 		if (1 << qp->sq.wqe_shift > dev->dev->caps.max_sq_desc_sz)
376 			return -EINVAL;
377 
378 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
379 
380 		/*
381 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
382 		 * allow HW to prefetch.
383 		 */
384 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
385 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
386 						    qp->sq_max_wqes_per_wr +
387 						    qp->sq_spare_wqes);
388 
389 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
390 			break;
391 
392 		if (qp->sq_max_wqes_per_wr <= 1)
393 			return -EINVAL;
394 
395 		++qp->sq.wqe_shift;
396 	}
397 
398 	qp->sq.max_gs = ((qp->sq_max_wqes_per_wr << qp->sq.wqe_shift) -
399 			 send_wqe_overhead(type, qp->flags)) /
400 		sizeof (struct mlx4_wqe_data_seg);
401 
402 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
403 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
404 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
405 		qp->rq.offset = 0;
406 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
407 	} else {
408 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
409 		qp->sq.offset = 0;
410 	}
411 
412 	cap->max_send_wr  = qp->sq.max_post =
413 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
414 	cap->max_send_sge = qp->sq.max_gs;
415 	/* We don't support inline sends for kernel QPs (yet) */
416 	cap->max_inline_data = 0;
417 
418 	return 0;
419 }
420 
421 static int set_user_sq_size(struct mlx4_ib_dev *dev,
422 			    struct mlx4_ib_qp *qp,
423 			    struct mlx4_ib_create_qp *ucmd)
424 {
425 	/* Sanity check SQ size before proceeding */
426 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
427 	    ucmd->log_sq_stride >
428 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
429 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
430 		return -EINVAL;
431 
432 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
433 	qp->sq.wqe_shift = ucmd->log_sq_stride;
434 
435 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
436 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
437 
438 	return 0;
439 }
440 
441 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
442 			    struct ib_qp_init_attr *init_attr,
443 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp *qp)
444 {
445 	int err;
446 
447 	mutex_init(&qp->mutex);
448 	spin_lock_init(&qp->sq.lock);
449 	spin_lock_init(&qp->rq.lock);
450 
451 	qp->state	 = IB_QPS_RESET;
452 	qp->atomic_rd_en = 0;
453 	qp->resp_depth   = 0;
454 
455 	qp->rq.head	    = 0;
456 	qp->rq.tail	    = 0;
457 	qp->sq.head	    = 0;
458 	qp->sq.tail	    = 0;
459 	qp->sq_next_wqe     = 0;
460 
461 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
462 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
463 	else
464 		qp->sq_signal_bits = 0;
465 
466 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, !!init_attr->srq, qp);
467 	if (err)
468 		goto err;
469 
470 	if (pd->uobject) {
471 		struct mlx4_ib_create_qp ucmd;
472 
473 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
474 			err = -EFAULT;
475 			goto err;
476 		}
477 
478 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
479 
480 		err = set_user_sq_size(dev, qp, &ucmd);
481 		if (err)
482 			goto err;
483 
484 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
485 				       qp->buf_size, 0);
486 		if (IS_ERR(qp->umem)) {
487 			err = PTR_ERR(qp->umem);
488 			goto err;
489 		}
490 
491 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
492 				    ilog2(qp->umem->page_size), &qp->mtt);
493 		if (err)
494 			goto err_buf;
495 
496 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
497 		if (err)
498 			goto err_mtt;
499 
500 		if (!init_attr->srq) {
501 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
502 						  ucmd.db_addr, &qp->db);
503 			if (err)
504 				goto err_mtt;
505 		}
506 	} else {
507 		qp->sq_no_prefetch = 0;
508 
509 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
510 			qp->flags |= MLX4_IB_QP_LSO;
511 
512 		err = set_kernel_sq_size(dev, &init_attr->cap, init_attr->qp_type, qp);
513 		if (err)
514 			goto err;
515 
516 		if (!init_attr->srq) {
517 			err = mlx4_ib_db_alloc(dev, &qp->db, 0);
518 			if (err)
519 				goto err;
520 
521 			*qp->db.db = 0;
522 		}
523 
524 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf)) {
525 			err = -ENOMEM;
526 			goto err_db;
527 		}
528 
529 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
530 				    &qp->mtt);
531 		if (err)
532 			goto err_buf;
533 
534 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
535 		if (err)
536 			goto err_mtt;
537 
538 		qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), GFP_KERNEL);
539 		qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), GFP_KERNEL);
540 
541 		if (!qp->sq.wrid || !qp->rq.wrid) {
542 			err = -ENOMEM;
543 			goto err_wrid;
544 		}
545 	}
546 
547 	err = mlx4_qp_alloc(dev->dev, sqpn, &qp->mqp);
548 	if (err)
549 		goto err_wrid;
550 
551 	/*
552 	 * Hardware wants QPN written in big-endian order (after
553 	 * shifting) for send doorbell.  Precompute this value to save
554 	 * a little bit when posting sends.
555 	 */
556 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
557 
558 	qp->mqp.event = mlx4_ib_qp_event;
559 
560 	return 0;
561 
562 err_wrid:
563 	if (pd->uobject) {
564 		if (!init_attr->srq)
565 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context),
566 					      &qp->db);
567 	} else {
568 		kfree(qp->sq.wrid);
569 		kfree(qp->rq.wrid);
570 	}
571 
572 err_mtt:
573 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
574 
575 err_buf:
576 	if (pd->uobject)
577 		ib_umem_release(qp->umem);
578 	else
579 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
580 
581 err_db:
582 	if (!pd->uobject && !init_attr->srq)
583 		mlx4_ib_db_free(dev, &qp->db);
584 
585 err:
586 	return err;
587 }
588 
589 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
590 {
591 	switch (state) {
592 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
593 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
594 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
595 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
596 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
597 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
598 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
599 	default:		return -1;
600 	}
601 }
602 
603 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
604 {
605 	if (send_cq == recv_cq)
606 		spin_lock_irq(&send_cq->lock);
607 	else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
608 		spin_lock_irq(&send_cq->lock);
609 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
610 	} else {
611 		spin_lock_irq(&recv_cq->lock);
612 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
613 	}
614 }
615 
616 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
617 {
618 	if (send_cq == recv_cq)
619 		spin_unlock_irq(&send_cq->lock);
620 	else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
621 		spin_unlock(&recv_cq->lock);
622 		spin_unlock_irq(&send_cq->lock);
623 	} else {
624 		spin_unlock(&send_cq->lock);
625 		spin_unlock_irq(&recv_cq->lock);
626 	}
627 }
628 
629 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
630 			      int is_user)
631 {
632 	struct mlx4_ib_cq *send_cq, *recv_cq;
633 
634 	if (qp->state != IB_QPS_RESET)
635 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
636 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
637 			printk(KERN_WARNING "mlx4_ib: modify QP %06x to RESET failed.\n",
638 			       qp->mqp.qpn);
639 
640 	send_cq = to_mcq(qp->ibqp.send_cq);
641 	recv_cq = to_mcq(qp->ibqp.recv_cq);
642 
643 	mlx4_ib_lock_cqs(send_cq, recv_cq);
644 
645 	if (!is_user) {
646 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
647 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
648 		if (send_cq != recv_cq)
649 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
650 	}
651 
652 	mlx4_qp_remove(dev->dev, &qp->mqp);
653 
654 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
655 
656 	mlx4_qp_free(dev->dev, &qp->mqp);
657 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
658 
659 	if (is_user) {
660 		if (!qp->ibqp.srq)
661 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
662 					      &qp->db);
663 		ib_umem_release(qp->umem);
664 	} else {
665 		kfree(qp->sq.wrid);
666 		kfree(qp->rq.wrid);
667 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
668 		if (!qp->ibqp.srq)
669 			mlx4_ib_db_free(dev, &qp->db);
670 	}
671 }
672 
673 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
674 				struct ib_qp_init_attr *init_attr,
675 				struct ib_udata *udata)
676 {
677 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
678 	struct mlx4_ib_sqp *sqp;
679 	struct mlx4_ib_qp *qp;
680 	int err;
681 
682 	/* We only support LSO, and only for kernel UD QPs. */
683 	if (init_attr->create_flags & ~IB_QP_CREATE_IPOIB_UD_LSO)
684 		return ERR_PTR(-EINVAL);
685 	if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO &&
686 	    (pd->uobject || init_attr->qp_type != IB_QPT_UD))
687 		return ERR_PTR(-EINVAL);
688 
689 	switch (init_attr->qp_type) {
690 	case IB_QPT_RC:
691 	case IB_QPT_UC:
692 	case IB_QPT_UD:
693 	{
694 		qp = kmalloc(sizeof *qp, GFP_KERNEL);
695 		if (!qp)
696 			return ERR_PTR(-ENOMEM);
697 
698 		err = create_qp_common(dev, pd, init_attr, udata, 0, qp);
699 		if (err) {
700 			kfree(qp);
701 			return ERR_PTR(err);
702 		}
703 
704 		qp->ibqp.qp_num = qp->mqp.qpn;
705 
706 		break;
707 	}
708 	case IB_QPT_SMI:
709 	case IB_QPT_GSI:
710 	{
711 		/* Userspace is not allowed to create special QPs: */
712 		if (pd->uobject)
713 			return ERR_PTR(-EINVAL);
714 
715 		sqp = kmalloc(sizeof *sqp, GFP_KERNEL);
716 		if (!sqp)
717 			return ERR_PTR(-ENOMEM);
718 
719 		qp = &sqp->qp;
720 
721 		err = create_qp_common(dev, pd, init_attr, udata,
722 				       dev->dev->caps.sqp_start +
723 				       (init_attr->qp_type == IB_QPT_SMI ? 0 : 2) +
724 				       init_attr->port_num - 1,
725 				       qp);
726 		if (err) {
727 			kfree(sqp);
728 			return ERR_PTR(err);
729 		}
730 
731 		qp->port	= init_attr->port_num;
732 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
733 
734 		break;
735 	}
736 	default:
737 		/* Don't support raw QPs */
738 		return ERR_PTR(-EINVAL);
739 	}
740 
741 	return &qp->ibqp;
742 }
743 
744 int mlx4_ib_destroy_qp(struct ib_qp *qp)
745 {
746 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
747 	struct mlx4_ib_qp *mqp = to_mqp(qp);
748 
749 	if (is_qp0(dev, mqp))
750 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
751 
752 	destroy_qp_common(dev, mqp, !!qp->pd->uobject);
753 
754 	if (is_sqp(dev, mqp))
755 		kfree(to_msqp(mqp));
756 	else
757 		kfree(mqp);
758 
759 	return 0;
760 }
761 
762 static int to_mlx4_st(enum ib_qp_type type)
763 {
764 	switch (type) {
765 	case IB_QPT_RC:		return MLX4_QP_ST_RC;
766 	case IB_QPT_UC:		return MLX4_QP_ST_UC;
767 	case IB_QPT_UD:		return MLX4_QP_ST_UD;
768 	case IB_QPT_SMI:
769 	case IB_QPT_GSI:	return MLX4_QP_ST_MLX;
770 	default:		return -1;
771 	}
772 }
773 
774 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
775 				   int attr_mask)
776 {
777 	u8 dest_rd_atomic;
778 	u32 access_flags;
779 	u32 hw_access_flags = 0;
780 
781 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
782 		dest_rd_atomic = attr->max_dest_rd_atomic;
783 	else
784 		dest_rd_atomic = qp->resp_depth;
785 
786 	if (attr_mask & IB_QP_ACCESS_FLAGS)
787 		access_flags = attr->qp_access_flags;
788 	else
789 		access_flags = qp->atomic_rd_en;
790 
791 	if (!dest_rd_atomic)
792 		access_flags &= IB_ACCESS_REMOTE_WRITE;
793 
794 	if (access_flags & IB_ACCESS_REMOTE_READ)
795 		hw_access_flags |= MLX4_QP_BIT_RRE;
796 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
797 		hw_access_flags |= MLX4_QP_BIT_RAE;
798 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
799 		hw_access_flags |= MLX4_QP_BIT_RWE;
800 
801 	return cpu_to_be32(hw_access_flags);
802 }
803 
804 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
805 			    int attr_mask)
806 {
807 	if (attr_mask & IB_QP_PKEY_INDEX)
808 		sqp->pkey_index = attr->pkey_index;
809 	if (attr_mask & IB_QP_QKEY)
810 		sqp->qkey = attr->qkey;
811 	if (attr_mask & IB_QP_SQ_PSN)
812 		sqp->send_psn = attr->sq_psn;
813 }
814 
815 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
816 {
817 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
818 }
819 
820 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
821 			 struct mlx4_qp_path *path, u8 port)
822 {
823 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
824 	path->rlid	    = cpu_to_be16(ah->dlid);
825 	if (ah->static_rate) {
826 		path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
827 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
828 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
829 			--path->static_rate;
830 	} else
831 		path->static_rate = 0;
832 	path->counter_index = 0xff;
833 
834 	if (ah->ah_flags & IB_AH_GRH) {
835 		if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
836 			printk(KERN_ERR "sgid_index (%u) too large. max is %d\n",
837 			       ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
838 			return -1;
839 		}
840 
841 		path->grh_mylmc |= 1 << 7;
842 		path->mgid_index = ah->grh.sgid_index;
843 		path->hop_limit  = ah->grh.hop_limit;
844 		path->tclass_flowlabel =
845 			cpu_to_be32((ah->grh.traffic_class << 20) |
846 				    (ah->grh.flow_label));
847 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
848 	}
849 
850 	path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
851 		((port - 1) << 6) | ((ah->sl & 0xf) << 2);
852 
853 	return 0;
854 }
855 
856 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
857 			       const struct ib_qp_attr *attr, int attr_mask,
858 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
859 {
860 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
861 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
862 	struct mlx4_qp_context *context;
863 	enum mlx4_qp_optpar optpar = 0;
864 	int sqd_event;
865 	int err = -EINVAL;
866 
867 	context = kzalloc(sizeof *context, GFP_KERNEL);
868 	if (!context)
869 		return -ENOMEM;
870 
871 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
872 				     (to_mlx4_st(ibqp->qp_type) << 16));
873 	context->flags     |= cpu_to_be32(1 << 8); /* DE? */
874 
875 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
876 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
877 	else {
878 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
879 		switch (attr->path_mig_state) {
880 		case IB_MIG_MIGRATED:
881 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
882 			break;
883 		case IB_MIG_REARM:
884 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
885 			break;
886 		case IB_MIG_ARMED:
887 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
888 			break;
889 		}
890 	}
891 
892 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
893 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
894 	else if (ibqp->qp_type == IB_QPT_UD) {
895 		if (qp->flags & MLX4_IB_QP_LSO)
896 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
897 					      ilog2(dev->dev->caps.max_gso_sz);
898 		else
899 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
900 	} else if (attr_mask & IB_QP_PATH_MTU) {
901 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
902 			printk(KERN_ERR "path MTU (%u) is invalid\n",
903 			       attr->path_mtu);
904 			goto out;
905 		}
906 		context->mtu_msgmax = (attr->path_mtu << 5) | 31;
907 	}
908 
909 	if (qp->rq.wqe_cnt)
910 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
911 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
912 
913 	if (qp->sq.wqe_cnt)
914 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
915 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
916 
917 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
918 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
919 
920 	if (qp->ibqp.uobject)
921 		context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
922 	else
923 		context->usr_page = cpu_to_be32(dev->priv_uar.index);
924 
925 	if (attr_mask & IB_QP_DEST_QPN)
926 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
927 
928 	if (attr_mask & IB_QP_PORT) {
929 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
930 		    !(attr_mask & IB_QP_AV)) {
931 			mlx4_set_sched(&context->pri_path, attr->port_num);
932 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
933 		}
934 	}
935 
936 	if (attr_mask & IB_QP_PKEY_INDEX) {
937 		context->pri_path.pkey_index = attr->pkey_index;
938 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
939 	}
940 
941 	if (attr_mask & IB_QP_AV) {
942 		if (mlx4_set_path(dev, &attr->ah_attr, &context->pri_path,
943 				  attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
944 			goto out;
945 
946 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
947 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
948 	}
949 
950 	if (attr_mask & IB_QP_TIMEOUT) {
951 		context->pri_path.ackto = attr->timeout << 3;
952 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
953 	}
954 
955 	if (attr_mask & IB_QP_ALT_PATH) {
956 		if (attr->alt_port_num == 0 ||
957 		    attr->alt_port_num > dev->dev->caps.num_ports)
958 			goto out;
959 
960 		if (attr->alt_pkey_index >=
961 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
962 			goto out;
963 
964 		if (mlx4_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
965 				  attr->alt_port_num))
966 			goto out;
967 
968 		context->alt_path.pkey_index = attr->alt_pkey_index;
969 		context->alt_path.ackto = attr->alt_timeout << 3;
970 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
971 	}
972 
973 	context->pd	    = cpu_to_be32(to_mpd(ibqp->pd)->pdn);
974 	context->params1    = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
975 
976 	if (attr_mask & IB_QP_RNR_RETRY) {
977 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
978 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
979 	}
980 
981 	if (attr_mask & IB_QP_RETRY_CNT) {
982 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
983 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
984 	}
985 
986 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
987 		if (attr->max_rd_atomic)
988 			context->params1 |=
989 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
990 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
991 	}
992 
993 	if (attr_mask & IB_QP_SQ_PSN)
994 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
995 
996 	context->cqn_send = cpu_to_be32(to_mcq(ibqp->send_cq)->mcq.cqn);
997 
998 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
999 		if (attr->max_dest_rd_atomic)
1000 			context->params2 |=
1001 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1002 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1003 	}
1004 
1005 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1006 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1007 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1008 	}
1009 
1010 	if (ibqp->srq)
1011 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1012 
1013 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1014 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1015 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1016 	}
1017 	if (attr_mask & IB_QP_RQ_PSN)
1018 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1019 
1020 	context->cqn_recv = cpu_to_be32(to_mcq(ibqp->recv_cq)->mcq.cqn);
1021 
1022 	if (attr_mask & IB_QP_QKEY) {
1023 		context->qkey = cpu_to_be32(attr->qkey);
1024 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
1025 	}
1026 
1027 	if (ibqp->srq)
1028 		context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1029 
1030 	if (!ibqp->srq && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1031 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1032 
1033 	if (cur_state == IB_QPS_INIT &&
1034 	    new_state == IB_QPS_RTR  &&
1035 	    (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1036 	     ibqp->qp_type == IB_QPT_UD)) {
1037 		context->pri_path.sched_queue = (qp->port - 1) << 6;
1038 		if (is_qp0(dev, qp))
1039 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1040 		else
1041 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1042 	}
1043 
1044 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1045 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1046 		sqd_event = 1;
1047 	else
1048 		sqd_event = 0;
1049 
1050 	/*
1051 	 * Before passing a kernel QP to the HW, make sure that the
1052 	 * ownership bits of the send queue are set and the SQ
1053 	 * headroom is stamped so that the hardware doesn't start
1054 	 * processing stale work requests.
1055 	 */
1056 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1057 		struct mlx4_wqe_ctrl_seg *ctrl;
1058 		int i;
1059 
1060 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1061 			ctrl = get_send_wqe(qp, i);
1062 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
1063 
1064 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1065 		}
1066 	}
1067 
1068 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1069 			     to_mlx4_state(new_state), context, optpar,
1070 			     sqd_event, &qp->mqp);
1071 	if (err)
1072 		goto out;
1073 
1074 	qp->state = new_state;
1075 
1076 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1077 		qp->atomic_rd_en = attr->qp_access_flags;
1078 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1079 		qp->resp_depth = attr->max_dest_rd_atomic;
1080 	if (attr_mask & IB_QP_PORT)
1081 		qp->port = attr->port_num;
1082 	if (attr_mask & IB_QP_ALT_PATH)
1083 		qp->alt_port = attr->alt_port_num;
1084 
1085 	if (is_sqp(dev, qp))
1086 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1087 
1088 	/*
1089 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
1090 	 * QP0 to RESET or ERROR, bring the link back down.
1091 	 */
1092 	if (is_qp0(dev, qp)) {
1093 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1094 			if (mlx4_INIT_PORT(dev->dev, qp->port))
1095 				printk(KERN_WARNING "INIT_PORT failed for port %d\n",
1096 				       qp->port);
1097 
1098 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1099 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1100 			mlx4_CLOSE_PORT(dev->dev, qp->port);
1101 	}
1102 
1103 	/*
1104 	 * If we moved a kernel QP to RESET, clean up all old CQ
1105 	 * entries and reinitialize the QP.
1106 	 */
1107 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1108 		mlx4_ib_cq_clean(to_mcq(ibqp->recv_cq), qp->mqp.qpn,
1109 				 ibqp->srq ? to_msrq(ibqp->srq): NULL);
1110 		if (ibqp->send_cq != ibqp->recv_cq)
1111 			mlx4_ib_cq_clean(to_mcq(ibqp->send_cq), qp->mqp.qpn, NULL);
1112 
1113 		qp->rq.head = 0;
1114 		qp->rq.tail = 0;
1115 		qp->sq.head = 0;
1116 		qp->sq.tail = 0;
1117 		qp->sq_next_wqe = 0;
1118 		if (!ibqp->srq)
1119 			*qp->db.db  = 0;
1120 	}
1121 
1122 out:
1123 	kfree(context);
1124 	return err;
1125 }
1126 
1127 static const struct ib_qp_attr mlx4_ib_qp_attr = { .port_num = 1 };
1128 static const int mlx4_ib_qp_attr_mask_table[IB_QPT_UD + 1] = {
1129 		[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1130 				IB_QP_PORT			|
1131 				IB_QP_QKEY),
1132 		[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
1133 				IB_QP_PORT			|
1134 				IB_QP_ACCESS_FLAGS),
1135 		[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
1136 				IB_QP_PORT			|
1137 				IB_QP_ACCESS_FLAGS),
1138 		[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1139 				IB_QP_QKEY),
1140 		[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1141 				IB_QP_QKEY),
1142 };
1143 
1144 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1145 		      int attr_mask, struct ib_udata *udata)
1146 {
1147 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1148 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1149 	enum ib_qp_state cur_state, new_state;
1150 	int err = -EINVAL;
1151 
1152 	mutex_lock(&qp->mutex);
1153 
1154 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1155 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1156 
1157 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask))
1158 		goto out;
1159 
1160 	if ((attr_mask & IB_QP_PORT) &&
1161 	    (attr->port_num == 0 || attr->port_num > dev->dev->caps.num_ports)) {
1162 		goto out;
1163 	}
1164 
1165 	if (attr_mask & IB_QP_PKEY_INDEX) {
1166 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1167 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p])
1168 			goto out;
1169 	}
1170 
1171 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1172 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
1173 		goto out;
1174 	}
1175 
1176 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1177 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
1178 		goto out;
1179 	}
1180 
1181 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1182 		err = 0;
1183 		goto out;
1184 	}
1185 
1186 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_ERR) {
1187 		err = __mlx4_ib_modify_qp(ibqp, &mlx4_ib_qp_attr,
1188 					  mlx4_ib_qp_attr_mask_table[ibqp->qp_type],
1189 					  IB_QPS_RESET, IB_QPS_INIT);
1190 		if (err)
1191 			goto out;
1192 		cur_state = IB_QPS_INIT;
1193 	}
1194 
1195 	err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1196 
1197 out:
1198 	mutex_unlock(&qp->mutex);
1199 	return err;
1200 }
1201 
1202 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
1203 			    void *wqe, unsigned *mlx_seg_len)
1204 {
1205 	struct ib_device *ib_dev = &to_mdev(sqp->qp.ibqp.device)->ib_dev;
1206 	struct mlx4_wqe_mlx_seg *mlx = wqe;
1207 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
1208 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
1209 	u16 pkey;
1210 	int send_size;
1211 	int header_size;
1212 	int spc;
1213 	int i;
1214 
1215 	send_size = 0;
1216 	for (i = 0; i < wr->num_sge; ++i)
1217 		send_size += wr->sg_list[i].length;
1218 
1219 	ib_ud_header_init(send_size, mlx4_ib_ah_grh_present(ah), &sqp->ud_header);
1220 
1221 	sqp->ud_header.lrh.service_level   =
1222 		be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 28;
1223 	sqp->ud_header.lrh.destination_lid = ah->av.dlid;
1224 	sqp->ud_header.lrh.source_lid      = cpu_to_be16(ah->av.g_slid & 0x7f);
1225 	if (mlx4_ib_ah_grh_present(ah)) {
1226 		sqp->ud_header.grh.traffic_class =
1227 			(be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 20) & 0xff;
1228 		sqp->ud_header.grh.flow_label    =
1229 			ah->av.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
1230 		sqp->ud_header.grh.hop_limit     = ah->av.hop_limit;
1231 		ib_get_cached_gid(ib_dev, be32_to_cpu(ah->av.port_pd) >> 24,
1232 				  ah->av.gid_index, &sqp->ud_header.grh.source_gid);
1233 		memcpy(sqp->ud_header.grh.destination_gid.raw,
1234 		       ah->av.dgid, 16);
1235 	}
1236 
1237 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1238 	mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
1239 				  (sqp->ud_header.lrh.destination_lid ==
1240 				   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
1241 				  (sqp->ud_header.lrh.service_level << 8));
1242 	mlx->rlid   = sqp->ud_header.lrh.destination_lid;
1243 
1244 	switch (wr->opcode) {
1245 	case IB_WR_SEND:
1246 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
1247 		sqp->ud_header.immediate_present = 0;
1248 		break;
1249 	case IB_WR_SEND_WITH_IMM:
1250 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1251 		sqp->ud_header.immediate_present = 1;
1252 		sqp->ud_header.immediate_data    = wr->ex.imm_data;
1253 		break;
1254 	default:
1255 		return -EINVAL;
1256 	}
1257 
1258 	sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1259 	if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1260 		sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1261 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1262 	if (!sqp->qp.ibqp.qp_num)
1263 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
1264 	else
1265 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
1266 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1267 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1268 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1269 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1270 					       sqp->qkey : wr->wr.ud.remote_qkey);
1271 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1272 
1273 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
1274 
1275 	if (0) {
1276 		printk(KERN_ERR "built UD header of size %d:\n", header_size);
1277 		for (i = 0; i < header_size / 4; ++i) {
1278 			if (i % 8 == 0)
1279 				printk("  [%02x] ", i * 4);
1280 			printk(" %08x",
1281 			       be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
1282 			if ((i + 1) % 8 == 0)
1283 				printk("\n");
1284 		}
1285 		printk("\n");
1286 	}
1287 
1288 	/*
1289 	 * Inline data segments may not cross a 64 byte boundary.  If
1290 	 * our UD header is bigger than the space available up to the
1291 	 * next 64 byte boundary in the WQE, use two inline data
1292 	 * segments to hold the UD header.
1293 	 */
1294 	spc = MLX4_INLINE_ALIGN -
1295 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
1296 	if (header_size <= spc) {
1297 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
1298 		memcpy(inl + 1, sqp->header_buf, header_size);
1299 		i = 1;
1300 	} else {
1301 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
1302 		memcpy(inl + 1, sqp->header_buf, spc);
1303 
1304 		inl = (void *) (inl + 1) + spc;
1305 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
1306 		/*
1307 		 * Need a barrier here to make sure all the data is
1308 		 * visible before the byte_count field is set.
1309 		 * Otherwise the HCA prefetcher could grab the 64-byte
1310 		 * chunk with this inline segment and get a valid (!=
1311 		 * 0xffffffff) byte count but stale data, and end up
1312 		 * generating a packet with bad headers.
1313 		 *
1314 		 * The first inline segment's byte_count field doesn't
1315 		 * need a barrier, because it comes after a
1316 		 * control/MLX segment and therefore is at an offset
1317 		 * of 16 mod 64.
1318 		 */
1319 		wmb();
1320 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
1321 		i = 2;
1322 	}
1323 
1324 	*mlx_seg_len =
1325 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
1326 	return 0;
1327 }
1328 
1329 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1330 {
1331 	unsigned cur;
1332 	struct mlx4_ib_cq *cq;
1333 
1334 	cur = wq->head - wq->tail;
1335 	if (likely(cur + nreq < wq->max_post))
1336 		return 0;
1337 
1338 	cq = to_mcq(ib_cq);
1339 	spin_lock(&cq->lock);
1340 	cur = wq->head - wq->tail;
1341 	spin_unlock(&cq->lock);
1342 
1343 	return cur + nreq >= wq->max_post;
1344 }
1345 
1346 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
1347 					  u64 remote_addr, u32 rkey)
1348 {
1349 	rseg->raddr    = cpu_to_be64(remote_addr);
1350 	rseg->rkey     = cpu_to_be32(rkey);
1351 	rseg->reserved = 0;
1352 }
1353 
1354 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
1355 {
1356 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1357 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
1358 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
1359 	} else {
1360 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
1361 		aseg->compare  = 0;
1362 	}
1363 
1364 }
1365 
1366 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
1367 			     struct ib_send_wr *wr)
1368 {
1369 	memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
1370 	dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1371 	dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1372 }
1373 
1374 static void set_mlx_icrc_seg(void *dseg)
1375 {
1376 	u32 *t = dseg;
1377 	struct mlx4_wqe_inline_seg *iseg = dseg;
1378 
1379 	t[1] = 0;
1380 
1381 	/*
1382 	 * Need a barrier here before writing the byte_count field to
1383 	 * make sure that all the data is visible before the
1384 	 * byte_count field is set.  Otherwise, if the segment begins
1385 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
1386 	 * chunk and get a valid (!= * 0xffffffff) byte count but
1387 	 * stale data, and end up sending the wrong data.
1388 	 */
1389 	wmb();
1390 
1391 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
1392 }
1393 
1394 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
1395 {
1396 	dseg->lkey       = cpu_to_be32(sg->lkey);
1397 	dseg->addr       = cpu_to_be64(sg->addr);
1398 
1399 	/*
1400 	 * Need a barrier here before writing the byte_count field to
1401 	 * make sure that all the data is visible before the
1402 	 * byte_count field is set.  Otherwise, if the segment begins
1403 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
1404 	 * chunk and get a valid (!= * 0xffffffff) byte count but
1405 	 * stale data, and end up sending the wrong data.
1406 	 */
1407 	wmb();
1408 
1409 	dseg->byte_count = cpu_to_be32(sg->length);
1410 }
1411 
1412 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
1413 {
1414 	dseg->byte_count = cpu_to_be32(sg->length);
1415 	dseg->lkey       = cpu_to_be32(sg->lkey);
1416 	dseg->addr       = cpu_to_be64(sg->addr);
1417 }
1418 
1419 static int build_lso_seg(struct mlx4_lso_seg *wqe, struct ib_send_wr *wr,
1420 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len)
1421 {
1422 	unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
1423 
1424 	/*
1425 	 * This is a temporary limitation and will be removed in
1426 	 * a forthcoming FW release:
1427 	 */
1428 	if (unlikely(halign > 64))
1429 		return -EINVAL;
1430 
1431 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
1432 		     wr->num_sge > qp->sq.max_gs - (halign >> 4)))
1433 		return -EINVAL;
1434 
1435 	memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
1436 
1437 	/* make sure LSO header is written before overwriting stamping */
1438 	wmb();
1439 
1440 	wqe->mss_hdr_size = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 |
1441 					wr->wr.ud.hlen);
1442 
1443 	*lso_seg_len = halign;
1444 	return 0;
1445 }
1446 
1447 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1448 		      struct ib_send_wr **bad_wr)
1449 {
1450 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1451 	void *wqe;
1452 	struct mlx4_wqe_ctrl_seg *ctrl;
1453 	struct mlx4_wqe_data_seg *dseg;
1454 	unsigned long flags;
1455 	int nreq;
1456 	int err = 0;
1457 	unsigned ind;
1458 	int uninitialized_var(stamp);
1459 	int uninitialized_var(size);
1460 	unsigned seglen;
1461 	int i;
1462 
1463 	spin_lock_irqsave(&qp->sq.lock, flags);
1464 
1465 	ind = qp->sq_next_wqe;
1466 
1467 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1468 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1469 			err = -ENOMEM;
1470 			*bad_wr = wr;
1471 			goto out;
1472 		}
1473 
1474 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
1475 			err = -EINVAL;
1476 			*bad_wr = wr;
1477 			goto out;
1478 		}
1479 
1480 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
1481 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
1482 
1483 		ctrl->srcrb_flags =
1484 			(wr->send_flags & IB_SEND_SIGNALED ?
1485 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
1486 			(wr->send_flags & IB_SEND_SOLICITED ?
1487 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
1488 			((wr->send_flags & IB_SEND_IP_CSUM) ?
1489 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
1490 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
1491 			qp->sq_signal_bits;
1492 
1493 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1494 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1495 			ctrl->imm = wr->ex.imm_data;
1496 		else
1497 			ctrl->imm = 0;
1498 
1499 		wqe += sizeof *ctrl;
1500 		size = sizeof *ctrl / 16;
1501 
1502 		switch (ibqp->qp_type) {
1503 		case IB_QPT_RC:
1504 		case IB_QPT_UC:
1505 			switch (wr->opcode) {
1506 			case IB_WR_ATOMIC_CMP_AND_SWP:
1507 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1508 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
1509 					      wr->wr.atomic.rkey);
1510 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
1511 
1512 				set_atomic_seg(wqe, wr);
1513 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
1514 
1515 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
1516 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
1517 
1518 				break;
1519 
1520 			case IB_WR_RDMA_READ:
1521 			case IB_WR_RDMA_WRITE:
1522 			case IB_WR_RDMA_WRITE_WITH_IMM:
1523 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1524 					      wr->wr.rdma.rkey);
1525 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
1526 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
1527 				break;
1528 
1529 			default:
1530 				/* No extra segments required for sends */
1531 				break;
1532 			}
1533 			break;
1534 
1535 		case IB_QPT_UD:
1536 			set_datagram_seg(wqe, wr);
1537 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
1538 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
1539 
1540 			if (wr->opcode == IB_WR_LSO) {
1541 				err = build_lso_seg(wqe, wr, qp, &seglen);
1542 				if (unlikely(err)) {
1543 					*bad_wr = wr;
1544 					goto out;
1545 				}
1546 				wqe  += seglen;
1547 				size += seglen / 16;
1548 			}
1549 			break;
1550 
1551 		case IB_QPT_SMI:
1552 		case IB_QPT_GSI:
1553 			err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
1554 			if (unlikely(err)) {
1555 				*bad_wr = wr;
1556 				goto out;
1557 			}
1558 			wqe  += seglen;
1559 			size += seglen / 16;
1560 			break;
1561 
1562 		default:
1563 			break;
1564 		}
1565 
1566 		/*
1567 		 * Write data segments in reverse order, so as to
1568 		 * overwrite cacheline stamp last within each
1569 		 * cacheline.  This avoids issues with WQE
1570 		 * prefetching.
1571 		 */
1572 
1573 		dseg = wqe;
1574 		dseg += wr->num_sge - 1;
1575 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
1576 
1577 		/* Add one more inline data segment for ICRC for MLX sends */
1578 		if (unlikely(qp->ibqp.qp_type == IB_QPT_SMI ||
1579 			     qp->ibqp.qp_type == IB_QPT_GSI)) {
1580 			set_mlx_icrc_seg(dseg + 1);
1581 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
1582 		}
1583 
1584 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
1585 			set_data_seg(dseg, wr->sg_list + i);
1586 
1587 		ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
1588 				    MLX4_WQE_CTRL_FENCE : 0) | size;
1589 
1590 		/*
1591 		 * Make sure descriptor is fully written before
1592 		 * setting ownership bit (because HW can start
1593 		 * executing as soon as we do).
1594 		 */
1595 		wmb();
1596 
1597 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
1598 			err = -EINVAL;
1599 			goto out;
1600 		}
1601 
1602 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
1603 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
1604 
1605 		stamp = ind + qp->sq_spare_wqes;
1606 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
1607 
1608 		/*
1609 		 * We can improve latency by not stamping the last
1610 		 * send queue WQE until after ringing the doorbell, so
1611 		 * only stamp here if there are still more WQEs to post.
1612 		 *
1613 		 * Same optimization applies to padding with NOP wqe
1614 		 * in case of WQE shrinking (used to prevent wrap-around
1615 		 * in the middle of WR).
1616 		 */
1617 		if (wr->next) {
1618 			stamp_send_wqe(qp, stamp, size * 16);
1619 			ind = pad_wraparound(qp, ind);
1620 		}
1621 
1622 	}
1623 
1624 out:
1625 	if (likely(nreq)) {
1626 		qp->sq.head += nreq;
1627 
1628 		/*
1629 		 * Make sure that descriptors are written before
1630 		 * doorbell record.
1631 		 */
1632 		wmb();
1633 
1634 		writel(qp->doorbell_qpn,
1635 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
1636 
1637 		/*
1638 		 * Make sure doorbells don't leak out of SQ spinlock
1639 		 * and reach the HCA out of order.
1640 		 */
1641 		mmiowb();
1642 
1643 		stamp_send_wqe(qp, stamp, size * 16);
1644 
1645 		ind = pad_wraparound(qp, ind);
1646 		qp->sq_next_wqe = ind;
1647 	}
1648 
1649 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1650 
1651 	return err;
1652 }
1653 
1654 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1655 		      struct ib_recv_wr **bad_wr)
1656 {
1657 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1658 	struct mlx4_wqe_data_seg *scat;
1659 	unsigned long flags;
1660 	int err = 0;
1661 	int nreq;
1662 	int ind;
1663 	int i;
1664 
1665 	spin_lock_irqsave(&qp->rq.lock, flags);
1666 
1667 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
1668 
1669 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1670 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.send_cq)) {
1671 			err = -ENOMEM;
1672 			*bad_wr = wr;
1673 			goto out;
1674 		}
1675 
1676 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1677 			err = -EINVAL;
1678 			*bad_wr = wr;
1679 			goto out;
1680 		}
1681 
1682 		scat = get_recv_wqe(qp, ind);
1683 
1684 		for (i = 0; i < wr->num_sge; ++i)
1685 			__set_data_seg(scat + i, wr->sg_list + i);
1686 
1687 		if (i < qp->rq.max_gs) {
1688 			scat[i].byte_count = 0;
1689 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
1690 			scat[i].addr       = 0;
1691 		}
1692 
1693 		qp->rq.wrid[ind] = wr->wr_id;
1694 
1695 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
1696 	}
1697 
1698 out:
1699 	if (likely(nreq)) {
1700 		qp->rq.head += nreq;
1701 
1702 		/*
1703 		 * Make sure that descriptors are written before
1704 		 * doorbell record.
1705 		 */
1706 		wmb();
1707 
1708 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
1709 	}
1710 
1711 	spin_unlock_irqrestore(&qp->rq.lock, flags);
1712 
1713 	return err;
1714 }
1715 
1716 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
1717 {
1718 	switch (mlx4_state) {
1719 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
1720 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
1721 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
1722 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
1723 	case MLX4_QP_STATE_SQ_DRAINING:
1724 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
1725 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
1726 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
1727 	default:		     return -1;
1728 	}
1729 }
1730 
1731 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
1732 {
1733 	switch (mlx4_mig_state) {
1734 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
1735 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
1736 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
1737 	default: return -1;
1738 	}
1739 }
1740 
1741 static int to_ib_qp_access_flags(int mlx4_flags)
1742 {
1743 	int ib_flags = 0;
1744 
1745 	if (mlx4_flags & MLX4_QP_BIT_RRE)
1746 		ib_flags |= IB_ACCESS_REMOTE_READ;
1747 	if (mlx4_flags & MLX4_QP_BIT_RWE)
1748 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
1749 	if (mlx4_flags & MLX4_QP_BIT_RAE)
1750 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
1751 
1752 	return ib_flags;
1753 }
1754 
1755 static void to_ib_ah_attr(struct mlx4_dev *dev, struct ib_ah_attr *ib_ah_attr,
1756 				struct mlx4_qp_path *path)
1757 {
1758 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
1759 	ib_ah_attr->port_num	  = path->sched_queue & 0x40 ? 2 : 1;
1760 
1761 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
1762 		return;
1763 
1764 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
1765 	ib_ah_attr->sl		  = (path->sched_queue >> 2) & 0xf;
1766 	ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
1767 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
1768 	ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
1769 	if (ib_ah_attr->ah_flags) {
1770 		ib_ah_attr->grh.sgid_index = path->mgid_index;
1771 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
1772 		ib_ah_attr->grh.traffic_class =
1773 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
1774 		ib_ah_attr->grh.flow_label =
1775 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
1776 		memcpy(ib_ah_attr->grh.dgid.raw,
1777 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
1778 	}
1779 }
1780 
1781 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
1782 		     struct ib_qp_init_attr *qp_init_attr)
1783 {
1784 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1785 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1786 	struct mlx4_qp_context context;
1787 	int mlx4_state;
1788 	int err = 0;
1789 
1790 	mutex_lock(&qp->mutex);
1791 
1792 	if (qp->state == IB_QPS_RESET) {
1793 		qp_attr->qp_state = IB_QPS_RESET;
1794 		goto done;
1795 	}
1796 
1797 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
1798 	if (err) {
1799 		err = -EINVAL;
1800 		goto out;
1801 	}
1802 
1803 	mlx4_state = be32_to_cpu(context.flags) >> 28;
1804 
1805 	qp->state		     = to_ib_qp_state(mlx4_state);
1806 	qp_attr->qp_state	     = qp->state;
1807 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
1808 	qp_attr->path_mig_state	     =
1809 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
1810 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
1811 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
1812 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
1813 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
1814 	qp_attr->qp_access_flags     =
1815 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
1816 
1817 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
1818 		to_ib_ah_attr(dev->dev, &qp_attr->ah_attr, &context.pri_path);
1819 		to_ib_ah_attr(dev->dev, &qp_attr->alt_ah_attr, &context.alt_path);
1820 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
1821 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
1822 	}
1823 
1824 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
1825 	if (qp_attr->qp_state == IB_QPS_INIT)
1826 		qp_attr->port_num = qp->port;
1827 	else
1828 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
1829 
1830 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
1831 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
1832 
1833 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
1834 
1835 	qp_attr->max_dest_rd_atomic =
1836 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
1837 	qp_attr->min_rnr_timer	    =
1838 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
1839 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
1840 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
1841 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
1842 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
1843 
1844 done:
1845 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
1846 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
1847 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
1848 
1849 	if (!ibqp->uobject) {
1850 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
1851 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
1852 	} else {
1853 		qp_attr->cap.max_send_wr  = 0;
1854 		qp_attr->cap.max_send_sge = 0;
1855 	}
1856 
1857 	/*
1858 	 * We don't support inline sends for kernel QPs (yet), and we
1859 	 * don't know what userspace's value should be.
1860 	 */
1861 	qp_attr->cap.max_inline_data = 0;
1862 
1863 	qp_init_attr->cap	     = qp_attr->cap;
1864 
1865 out:
1866 	mutex_unlock(&qp->mutex);
1867 	return err;
1868 }
1869 
1870