xref: /linux/drivers/infiniband/hw/mlx4/qp.c (revision 367b8112fe2ea5c39a7bb4d263dcdd9b612fae18)
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 };
59 
60 struct mlx4_ib_sqp {
61 	struct mlx4_ib_qp	qp;
62 	int			pkey_index;
63 	u32			qkey;
64 	u32			send_psn;
65 	struct ib_ud_header	ud_header;
66 	u8			header_buf[MLX4_IB_UD_HEADER_SIZE];
67 };
68 
69 enum {
70 	MLX4_IB_MIN_SQ_STRIDE = 6
71 };
72 
73 static const __be32 mlx4_ib_opcode[] = {
74 	[IB_WR_SEND]			= __constant_cpu_to_be32(MLX4_OPCODE_SEND),
75 	[IB_WR_LSO]			= __constant_cpu_to_be32(MLX4_OPCODE_LSO),
76 	[IB_WR_SEND_WITH_IMM]		= __constant_cpu_to_be32(MLX4_OPCODE_SEND_IMM),
77 	[IB_WR_RDMA_WRITE]		= __constant_cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
78 	[IB_WR_RDMA_WRITE_WITH_IMM]	= __constant_cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
79 	[IB_WR_RDMA_READ]		= __constant_cpu_to_be32(MLX4_OPCODE_RDMA_READ),
80 	[IB_WR_ATOMIC_CMP_AND_SWP]	= __constant_cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
81 	[IB_WR_ATOMIC_FETCH_AND_ADD]	= __constant_cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
82 	[IB_WR_SEND_WITH_INV]		= __constant_cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
83 	[IB_WR_LOCAL_INV]		= __constant_cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
84 	[IB_WR_FAST_REG_MR]		= __constant_cpu_to_be32(MLX4_OPCODE_FMR),
85 };
86 
87 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
88 {
89 	return container_of(mqp, struct mlx4_ib_sqp, qp);
90 }
91 
92 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
93 {
94 	return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
95 		qp->mqp.qpn <= dev->dev->caps.sqp_start + 3;
96 }
97 
98 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
99 {
100 	return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
101 		qp->mqp.qpn <= dev->dev->caps.sqp_start + 1;
102 }
103 
104 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
105 {
106 	return mlx4_buf_offset(&qp->buf, offset);
107 }
108 
109 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
110 {
111 	return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
112 }
113 
114 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
115 {
116 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
117 }
118 
119 /*
120  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
121  * first four bytes of every 64 byte chunk with
122  *     0x7FFFFFF | (invalid_ownership_value << 31).
123  *
124  * When the max work request size is less than or equal to the WQE
125  * basic block size, as an optimization, we can stamp all WQEs with
126  * 0xffffffff, and skip the very first chunk of each WQE.
127  */
128 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
129 {
130 	__be32 *wqe;
131 	int i;
132 	int s;
133 	int ind;
134 	void *buf;
135 	__be32 stamp;
136 	struct mlx4_wqe_ctrl_seg *ctrl;
137 
138 	if (qp->sq_max_wqes_per_wr > 1) {
139 		s = roundup(size, 1U << qp->sq.wqe_shift);
140 		for (i = 0; i < s; i += 64) {
141 			ind = (i >> qp->sq.wqe_shift) + n;
142 			stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
143 						       cpu_to_be32(0xffffffff);
144 			buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
145 			wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
146 			*wqe = stamp;
147 		}
148 	} else {
149 		ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
150 		s = (ctrl->fence_size & 0x3f) << 4;
151 		for (i = 64; i < s; i += 64) {
152 			wqe = buf + i;
153 			*wqe = cpu_to_be32(0xffffffff);
154 		}
155 	}
156 }
157 
158 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
159 {
160 	struct mlx4_wqe_ctrl_seg *ctrl;
161 	struct mlx4_wqe_inline_seg *inl;
162 	void *wqe;
163 	int s;
164 
165 	ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
166 	s = sizeof(struct mlx4_wqe_ctrl_seg);
167 
168 	if (qp->ibqp.qp_type == IB_QPT_UD) {
169 		struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
170 		struct mlx4_av *av = (struct mlx4_av *)dgram->av;
171 		memset(dgram, 0, sizeof *dgram);
172 		av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
173 		s += sizeof(struct mlx4_wqe_datagram_seg);
174 	}
175 
176 	/* Pad the remainder of the WQE with an inline data segment. */
177 	if (size > s) {
178 		inl = wqe + s;
179 		inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
180 	}
181 	ctrl->srcrb_flags = 0;
182 	ctrl->fence_size = size / 16;
183 	/*
184 	 * Make sure descriptor is fully written before setting ownership bit
185 	 * (because HW can start executing as soon as we do).
186 	 */
187 	wmb();
188 
189 	ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
190 		(n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
191 
192 	stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
193 }
194 
195 /* Post NOP WQE to prevent wrap-around in the middle of WR */
196 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
197 {
198 	unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
199 	if (unlikely(s < qp->sq_max_wqes_per_wr)) {
200 		post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
201 		ind += s;
202 	}
203 	return ind;
204 }
205 
206 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
207 {
208 	struct ib_event event;
209 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
210 
211 	if (type == MLX4_EVENT_TYPE_PATH_MIG)
212 		to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
213 
214 	if (ibqp->event_handler) {
215 		event.device     = ibqp->device;
216 		event.element.qp = ibqp;
217 		switch (type) {
218 		case MLX4_EVENT_TYPE_PATH_MIG:
219 			event.event = IB_EVENT_PATH_MIG;
220 			break;
221 		case MLX4_EVENT_TYPE_COMM_EST:
222 			event.event = IB_EVENT_COMM_EST;
223 			break;
224 		case MLX4_EVENT_TYPE_SQ_DRAINED:
225 			event.event = IB_EVENT_SQ_DRAINED;
226 			break;
227 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
228 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
229 			break;
230 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
231 			event.event = IB_EVENT_QP_FATAL;
232 			break;
233 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
234 			event.event = IB_EVENT_PATH_MIG_ERR;
235 			break;
236 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
237 			event.event = IB_EVENT_QP_REQ_ERR;
238 			break;
239 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
240 			event.event = IB_EVENT_QP_ACCESS_ERR;
241 			break;
242 		default:
243 			printk(KERN_WARNING "mlx4_ib: Unexpected event type %d "
244 			       "on QP %06x\n", type, qp->qpn);
245 			return;
246 		}
247 
248 		ibqp->event_handler(&event, ibqp->qp_context);
249 	}
250 }
251 
252 static int send_wqe_overhead(enum ib_qp_type type, u32 flags)
253 {
254 	/*
255 	 * UD WQEs must have a datagram segment.
256 	 * RC and UC WQEs might have a remote address segment.
257 	 * MLX WQEs need two extra inline data segments (for the UD
258 	 * header and space for the ICRC).
259 	 */
260 	switch (type) {
261 	case IB_QPT_UD:
262 		return sizeof (struct mlx4_wqe_ctrl_seg) +
263 			sizeof (struct mlx4_wqe_datagram_seg) +
264 			((flags & MLX4_IB_QP_LSO) ? 64 : 0);
265 	case IB_QPT_UC:
266 		return sizeof (struct mlx4_wqe_ctrl_seg) +
267 			sizeof (struct mlx4_wqe_raddr_seg);
268 	case IB_QPT_RC:
269 		return sizeof (struct mlx4_wqe_ctrl_seg) +
270 			sizeof (struct mlx4_wqe_atomic_seg) +
271 			sizeof (struct mlx4_wqe_raddr_seg);
272 	case IB_QPT_SMI:
273 	case IB_QPT_GSI:
274 		return sizeof (struct mlx4_wqe_ctrl_seg) +
275 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
276 			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
277 					   MLX4_INLINE_ALIGN) *
278 			      sizeof (struct mlx4_wqe_inline_seg),
279 			      sizeof (struct mlx4_wqe_data_seg)) +
280 			ALIGN(4 +
281 			      sizeof (struct mlx4_wqe_inline_seg),
282 			      sizeof (struct mlx4_wqe_data_seg));
283 	default:
284 		return sizeof (struct mlx4_wqe_ctrl_seg);
285 	}
286 }
287 
288 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
289 		       int is_user, int has_srq, struct mlx4_ib_qp *qp)
290 {
291 	/* Sanity check RQ size before proceeding */
292 	if (cap->max_recv_wr  > dev->dev->caps.max_wqes  ||
293 	    cap->max_recv_sge > dev->dev->caps.max_rq_sg)
294 		return -EINVAL;
295 
296 	if (has_srq) {
297 		/* QPs attached to an SRQ should have no RQ */
298 		if (cap->max_recv_wr)
299 			return -EINVAL;
300 
301 		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
302 	} else {
303 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
304 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
305 			return -EINVAL;
306 
307 		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
308 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
309 		qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
310 	}
311 
312 	cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
313 	cap->max_recv_sge = qp->rq.max_gs;
314 
315 	return 0;
316 }
317 
318 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
319 			      enum ib_qp_type type, struct mlx4_ib_qp *qp)
320 {
321 	int s;
322 
323 	/* Sanity check SQ size before proceeding */
324 	if (cap->max_send_wr	 > dev->dev->caps.max_wqes  ||
325 	    cap->max_send_sge	 > dev->dev->caps.max_sq_sg ||
326 	    cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
327 	    sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
328 		return -EINVAL;
329 
330 	/*
331 	 * For MLX transport we need 2 extra S/G entries:
332 	 * one for the header and one for the checksum at the end
333 	 */
334 	if ((type == IB_QPT_SMI || type == IB_QPT_GSI) &&
335 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
336 		return -EINVAL;
337 
338 	s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
339 		cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
340 		send_wqe_overhead(type, qp->flags);
341 
342 	if (s > dev->dev->caps.max_sq_desc_sz)
343 		return -EINVAL;
344 
345 	/*
346 	 * Hermon supports shrinking WQEs, such that a single work
347 	 * request can include multiple units of 1 << wqe_shift.  This
348 	 * way, work requests can differ in size, and do not have to
349 	 * be a power of 2 in size, saving memory and speeding up send
350 	 * WR posting.  Unfortunately, if we do this then the
351 	 * wqe_index field in CQEs can't be used to look up the WR ID
352 	 * anymore, so we do this only if selective signaling is off.
353 	 *
354 	 * Further, on 32-bit platforms, we can't use vmap() to make
355 	 * the QP buffer virtually contigious.  Thus we have to use
356 	 * constant-sized WRs to make sure a WR is always fully within
357 	 * a single page-sized chunk.
358 	 *
359 	 * Finally, we use NOP work requests to pad the end of the
360 	 * work queue, to avoid wrap-around in the middle of WR.  We
361 	 * set NEC bit to avoid getting completions with error for
362 	 * these NOP WRs, but since NEC is only supported starting
363 	 * with firmware 2.2.232, we use constant-sized WRs for older
364 	 * firmware.
365 	 *
366 	 * And, since MLX QPs only support SEND, we use constant-sized
367 	 * WRs in this case.
368 	 *
369 	 * We look for the smallest value of wqe_shift such that the
370 	 * resulting number of wqes does not exceed device
371 	 * capabilities.
372 	 *
373 	 * We set WQE size to at least 64 bytes, this way stamping
374 	 * invalidates each WQE.
375 	 */
376 	if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
377 	    qp->sq_signal_bits && BITS_PER_LONG == 64 &&
378 	    type != IB_QPT_SMI && type != IB_QPT_GSI)
379 		qp->sq.wqe_shift = ilog2(64);
380 	else
381 		qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
382 
383 	for (;;) {
384 		qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
385 
386 		/*
387 		 * We need to leave 2 KB + 1 WR of headroom in the SQ to
388 		 * allow HW to prefetch.
389 		 */
390 		qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
391 		qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
392 						    qp->sq_max_wqes_per_wr +
393 						    qp->sq_spare_wqes);
394 
395 		if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
396 			break;
397 
398 		if (qp->sq_max_wqes_per_wr <= 1)
399 			return -EINVAL;
400 
401 		++qp->sq.wqe_shift;
402 	}
403 
404 	qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
405 			     (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
406 			 send_wqe_overhead(type, qp->flags)) /
407 		sizeof (struct mlx4_wqe_data_seg);
408 
409 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
410 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
411 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
412 		qp->rq.offset = 0;
413 		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
414 	} else {
415 		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
416 		qp->sq.offset = 0;
417 	}
418 
419 	cap->max_send_wr  = qp->sq.max_post =
420 		(qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
421 	cap->max_send_sge = min(qp->sq.max_gs,
422 				min(dev->dev->caps.max_sq_sg,
423 				    dev->dev->caps.max_rq_sg));
424 	/* We don't support inline sends for kernel QPs (yet) */
425 	cap->max_inline_data = 0;
426 
427 	return 0;
428 }
429 
430 static int set_user_sq_size(struct mlx4_ib_dev *dev,
431 			    struct mlx4_ib_qp *qp,
432 			    struct mlx4_ib_create_qp *ucmd)
433 {
434 	/* Sanity check SQ size before proceeding */
435 	if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes	 ||
436 	    ucmd->log_sq_stride >
437 		ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
438 	    ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
439 		return -EINVAL;
440 
441 	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
442 	qp->sq.wqe_shift = ucmd->log_sq_stride;
443 
444 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
445 		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
446 
447 	return 0;
448 }
449 
450 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
451 			    struct ib_qp_init_attr *init_attr,
452 			    struct ib_udata *udata, int sqpn, struct mlx4_ib_qp *qp)
453 {
454 	int qpn;
455 	int err;
456 
457 	mutex_init(&qp->mutex);
458 	spin_lock_init(&qp->sq.lock);
459 	spin_lock_init(&qp->rq.lock);
460 
461 	qp->state	 = IB_QPS_RESET;
462 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
463 		qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
464 
465 	err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, !!init_attr->srq, qp);
466 	if (err)
467 		goto err;
468 
469 	if (pd->uobject) {
470 		struct mlx4_ib_create_qp ucmd;
471 
472 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
473 			err = -EFAULT;
474 			goto err;
475 		}
476 
477 		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
478 
479 		err = set_user_sq_size(dev, qp, &ucmd);
480 		if (err)
481 			goto err;
482 
483 		qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
484 				       qp->buf_size, 0, 0);
485 		if (IS_ERR(qp->umem)) {
486 			err = PTR_ERR(qp->umem);
487 			goto err;
488 		}
489 
490 		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
491 				    ilog2(qp->umem->page_size), &qp->mtt);
492 		if (err)
493 			goto err_buf;
494 
495 		err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
496 		if (err)
497 			goto err_mtt;
498 
499 		if (!init_attr->srq) {
500 			err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
501 						  ucmd.db_addr, &qp->db);
502 			if (err)
503 				goto err_mtt;
504 		}
505 	} else {
506 		qp->sq_no_prefetch = 0;
507 
508 		if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
509 			qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
510 
511 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
512 			qp->flags |= MLX4_IB_QP_LSO;
513 
514 		err = set_kernel_sq_size(dev, &init_attr->cap, init_attr->qp_type, qp);
515 		if (err)
516 			goto err;
517 
518 		if (!init_attr->srq) {
519 			err = mlx4_db_alloc(dev->dev, &qp->db, 0);
520 			if (err)
521 				goto err;
522 
523 			*qp->db.db = 0;
524 		}
525 
526 		if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf)) {
527 			err = -ENOMEM;
528 			goto err_db;
529 		}
530 
531 		err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
532 				    &qp->mtt);
533 		if (err)
534 			goto err_buf;
535 
536 		err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
537 		if (err)
538 			goto err_mtt;
539 
540 		qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), GFP_KERNEL);
541 		qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), GFP_KERNEL);
542 
543 		if (!qp->sq.wrid || !qp->rq.wrid) {
544 			err = -ENOMEM;
545 			goto err_wrid;
546 		}
547 	}
548 
549 	if (sqpn) {
550 		qpn = sqpn;
551 	} else {
552 		err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn);
553 		if (err)
554 			goto err_wrid;
555 	}
556 
557 	err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
558 	if (err)
559 		goto err_qpn;
560 
561 	/*
562 	 * Hardware wants QPN written in big-endian order (after
563 	 * shifting) for send doorbell.  Precompute this value to save
564 	 * a little bit when posting sends.
565 	 */
566 	qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
567 
568 	qp->mqp.event = mlx4_ib_qp_event;
569 
570 	return 0;
571 
572 err_qpn:
573 	if (!sqpn)
574 		mlx4_qp_release_range(dev->dev, qpn, 1);
575 
576 err_wrid:
577 	if (pd->uobject) {
578 		if (!init_attr->srq)
579 			mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context),
580 					      &qp->db);
581 	} else {
582 		kfree(qp->sq.wrid);
583 		kfree(qp->rq.wrid);
584 	}
585 
586 err_mtt:
587 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
588 
589 err_buf:
590 	if (pd->uobject)
591 		ib_umem_release(qp->umem);
592 	else
593 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
594 
595 err_db:
596 	if (!pd->uobject && !init_attr->srq)
597 		mlx4_db_free(dev->dev, &qp->db);
598 
599 err:
600 	return err;
601 }
602 
603 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
604 {
605 	switch (state) {
606 	case IB_QPS_RESET:	return MLX4_QP_STATE_RST;
607 	case IB_QPS_INIT:	return MLX4_QP_STATE_INIT;
608 	case IB_QPS_RTR:	return MLX4_QP_STATE_RTR;
609 	case IB_QPS_RTS:	return MLX4_QP_STATE_RTS;
610 	case IB_QPS_SQD:	return MLX4_QP_STATE_SQD;
611 	case IB_QPS_SQE:	return MLX4_QP_STATE_SQER;
612 	case IB_QPS_ERR:	return MLX4_QP_STATE_ERR;
613 	default:		return -1;
614 	}
615 }
616 
617 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
618 {
619 	if (send_cq == recv_cq)
620 		spin_lock_irq(&send_cq->lock);
621 	else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
622 		spin_lock_irq(&send_cq->lock);
623 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
624 	} else {
625 		spin_lock_irq(&recv_cq->lock);
626 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
627 	}
628 }
629 
630 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
631 {
632 	if (send_cq == recv_cq)
633 		spin_unlock_irq(&send_cq->lock);
634 	else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
635 		spin_unlock(&recv_cq->lock);
636 		spin_unlock_irq(&send_cq->lock);
637 	} else {
638 		spin_unlock(&send_cq->lock);
639 		spin_unlock_irq(&recv_cq->lock);
640 	}
641 }
642 
643 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
644 			      int is_user)
645 {
646 	struct mlx4_ib_cq *send_cq, *recv_cq;
647 
648 	if (qp->state != IB_QPS_RESET)
649 		if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
650 				   MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
651 			printk(KERN_WARNING "mlx4_ib: modify QP %06x to RESET failed.\n",
652 			       qp->mqp.qpn);
653 
654 	send_cq = to_mcq(qp->ibqp.send_cq);
655 	recv_cq = to_mcq(qp->ibqp.recv_cq);
656 
657 	mlx4_ib_lock_cqs(send_cq, recv_cq);
658 
659 	if (!is_user) {
660 		__mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
661 				 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
662 		if (send_cq != recv_cq)
663 			__mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
664 	}
665 
666 	mlx4_qp_remove(dev->dev, &qp->mqp);
667 
668 	mlx4_ib_unlock_cqs(send_cq, recv_cq);
669 
670 	mlx4_qp_free(dev->dev, &qp->mqp);
671 
672 	if (!is_sqp(dev, qp))
673 		mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
674 
675 	mlx4_mtt_cleanup(dev->dev, &qp->mtt);
676 
677 	if (is_user) {
678 		if (!qp->ibqp.srq)
679 			mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
680 					      &qp->db);
681 		ib_umem_release(qp->umem);
682 	} else {
683 		kfree(qp->sq.wrid);
684 		kfree(qp->rq.wrid);
685 		mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
686 		if (!qp->ibqp.srq)
687 			mlx4_db_free(dev->dev, &qp->db);
688 	}
689 }
690 
691 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
692 				struct ib_qp_init_attr *init_attr,
693 				struct ib_udata *udata)
694 {
695 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
696 	struct mlx4_ib_sqp *sqp;
697 	struct mlx4_ib_qp *qp;
698 	int err;
699 
700 	/*
701 	 * We only support LSO and multicast loopback blocking, and
702 	 * only for kernel UD QPs.
703 	 */
704 	if (init_attr->create_flags & ~(IB_QP_CREATE_IPOIB_UD_LSO |
705 					IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK))
706 		return ERR_PTR(-EINVAL);
707 
708 	if (init_attr->create_flags &&
709 	    (pd->uobject || init_attr->qp_type != IB_QPT_UD))
710 		return ERR_PTR(-EINVAL);
711 
712 	switch (init_attr->qp_type) {
713 	case IB_QPT_RC:
714 	case IB_QPT_UC:
715 	case IB_QPT_UD:
716 	{
717 		qp = kzalloc(sizeof *qp, GFP_KERNEL);
718 		if (!qp)
719 			return ERR_PTR(-ENOMEM);
720 
721 		err = create_qp_common(dev, pd, init_attr, udata, 0, qp);
722 		if (err) {
723 			kfree(qp);
724 			return ERR_PTR(err);
725 		}
726 
727 		qp->ibqp.qp_num = qp->mqp.qpn;
728 
729 		break;
730 	}
731 	case IB_QPT_SMI:
732 	case IB_QPT_GSI:
733 	{
734 		/* Userspace is not allowed to create special QPs: */
735 		if (pd->uobject)
736 			return ERR_PTR(-EINVAL);
737 
738 		sqp = kzalloc(sizeof *sqp, GFP_KERNEL);
739 		if (!sqp)
740 			return ERR_PTR(-ENOMEM);
741 
742 		qp = &sqp->qp;
743 
744 		err = create_qp_common(dev, pd, init_attr, udata,
745 				       dev->dev->caps.sqp_start +
746 				       (init_attr->qp_type == IB_QPT_SMI ? 0 : 2) +
747 				       init_attr->port_num - 1,
748 				       qp);
749 		if (err) {
750 			kfree(sqp);
751 			return ERR_PTR(err);
752 		}
753 
754 		qp->port	= init_attr->port_num;
755 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
756 
757 		break;
758 	}
759 	default:
760 		/* Don't support raw QPs */
761 		return ERR_PTR(-EINVAL);
762 	}
763 
764 	return &qp->ibqp;
765 }
766 
767 int mlx4_ib_destroy_qp(struct ib_qp *qp)
768 {
769 	struct mlx4_ib_dev *dev = to_mdev(qp->device);
770 	struct mlx4_ib_qp *mqp = to_mqp(qp);
771 
772 	if (is_qp0(dev, mqp))
773 		mlx4_CLOSE_PORT(dev->dev, mqp->port);
774 
775 	destroy_qp_common(dev, mqp, !!qp->pd->uobject);
776 
777 	if (is_sqp(dev, mqp))
778 		kfree(to_msqp(mqp));
779 	else
780 		kfree(mqp);
781 
782 	return 0;
783 }
784 
785 static int to_mlx4_st(enum ib_qp_type type)
786 {
787 	switch (type) {
788 	case IB_QPT_RC:		return MLX4_QP_ST_RC;
789 	case IB_QPT_UC:		return MLX4_QP_ST_UC;
790 	case IB_QPT_UD:		return MLX4_QP_ST_UD;
791 	case IB_QPT_SMI:
792 	case IB_QPT_GSI:	return MLX4_QP_ST_MLX;
793 	default:		return -1;
794 	}
795 }
796 
797 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
798 				   int attr_mask)
799 {
800 	u8 dest_rd_atomic;
801 	u32 access_flags;
802 	u32 hw_access_flags = 0;
803 
804 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
805 		dest_rd_atomic = attr->max_dest_rd_atomic;
806 	else
807 		dest_rd_atomic = qp->resp_depth;
808 
809 	if (attr_mask & IB_QP_ACCESS_FLAGS)
810 		access_flags = attr->qp_access_flags;
811 	else
812 		access_flags = qp->atomic_rd_en;
813 
814 	if (!dest_rd_atomic)
815 		access_flags &= IB_ACCESS_REMOTE_WRITE;
816 
817 	if (access_flags & IB_ACCESS_REMOTE_READ)
818 		hw_access_flags |= MLX4_QP_BIT_RRE;
819 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
820 		hw_access_flags |= MLX4_QP_BIT_RAE;
821 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
822 		hw_access_flags |= MLX4_QP_BIT_RWE;
823 
824 	return cpu_to_be32(hw_access_flags);
825 }
826 
827 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
828 			    int attr_mask)
829 {
830 	if (attr_mask & IB_QP_PKEY_INDEX)
831 		sqp->pkey_index = attr->pkey_index;
832 	if (attr_mask & IB_QP_QKEY)
833 		sqp->qkey = attr->qkey;
834 	if (attr_mask & IB_QP_SQ_PSN)
835 		sqp->send_psn = attr->sq_psn;
836 }
837 
838 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
839 {
840 	path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
841 }
842 
843 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
844 			 struct mlx4_qp_path *path, u8 port)
845 {
846 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
847 	path->rlid	    = cpu_to_be16(ah->dlid);
848 	if (ah->static_rate) {
849 		path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
850 		while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
851 		       !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
852 			--path->static_rate;
853 	} else
854 		path->static_rate = 0;
855 	path->counter_index = 0xff;
856 
857 	if (ah->ah_flags & IB_AH_GRH) {
858 		if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
859 			printk(KERN_ERR "sgid_index (%u) too large. max is %d\n",
860 			       ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
861 			return -1;
862 		}
863 
864 		path->grh_mylmc |= 1 << 7;
865 		path->mgid_index = ah->grh.sgid_index;
866 		path->hop_limit  = ah->grh.hop_limit;
867 		path->tclass_flowlabel =
868 			cpu_to_be32((ah->grh.traffic_class << 20) |
869 				    (ah->grh.flow_label));
870 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
871 	}
872 
873 	path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
874 		((port - 1) << 6) | ((ah->sl & 0xf) << 2);
875 
876 	return 0;
877 }
878 
879 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
880 			       const struct ib_qp_attr *attr, int attr_mask,
881 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
882 {
883 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
884 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
885 	struct mlx4_qp_context *context;
886 	enum mlx4_qp_optpar optpar = 0;
887 	int sqd_event;
888 	int err = -EINVAL;
889 
890 	context = kzalloc(sizeof *context, GFP_KERNEL);
891 	if (!context)
892 		return -ENOMEM;
893 
894 	context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
895 				     (to_mlx4_st(ibqp->qp_type) << 16));
896 	context->flags     |= cpu_to_be32(1 << 8); /* DE? */
897 
898 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
899 		context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
900 	else {
901 		optpar |= MLX4_QP_OPTPAR_PM_STATE;
902 		switch (attr->path_mig_state) {
903 		case IB_MIG_MIGRATED:
904 			context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
905 			break;
906 		case IB_MIG_REARM:
907 			context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
908 			break;
909 		case IB_MIG_ARMED:
910 			context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
911 			break;
912 		}
913 	}
914 
915 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
916 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
917 	else if (ibqp->qp_type == IB_QPT_UD) {
918 		if (qp->flags & MLX4_IB_QP_LSO)
919 			context->mtu_msgmax = (IB_MTU_4096 << 5) |
920 					      ilog2(dev->dev->caps.max_gso_sz);
921 		else
922 			context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
923 	} else if (attr_mask & IB_QP_PATH_MTU) {
924 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
925 			printk(KERN_ERR "path MTU (%u) is invalid\n",
926 			       attr->path_mtu);
927 			goto out;
928 		}
929 		context->mtu_msgmax = (attr->path_mtu << 5) |
930 			ilog2(dev->dev->caps.max_msg_sz);
931 	}
932 
933 	if (qp->rq.wqe_cnt)
934 		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
935 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
936 
937 	if (qp->sq.wqe_cnt)
938 		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
939 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
940 
941 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
942 		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
943 
944 	if (qp->ibqp.uobject)
945 		context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
946 	else
947 		context->usr_page = cpu_to_be32(dev->priv_uar.index);
948 
949 	if (attr_mask & IB_QP_DEST_QPN)
950 		context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
951 
952 	if (attr_mask & IB_QP_PORT) {
953 		if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
954 		    !(attr_mask & IB_QP_AV)) {
955 			mlx4_set_sched(&context->pri_path, attr->port_num);
956 			optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
957 		}
958 	}
959 
960 	if (attr_mask & IB_QP_PKEY_INDEX) {
961 		context->pri_path.pkey_index = attr->pkey_index;
962 		optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
963 	}
964 
965 	if (attr_mask & IB_QP_AV) {
966 		if (mlx4_set_path(dev, &attr->ah_attr, &context->pri_path,
967 				  attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
968 			goto out;
969 
970 		optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
971 			   MLX4_QP_OPTPAR_SCHED_QUEUE);
972 	}
973 
974 	if (attr_mask & IB_QP_TIMEOUT) {
975 		context->pri_path.ackto = attr->timeout << 3;
976 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
977 	}
978 
979 	if (attr_mask & IB_QP_ALT_PATH) {
980 		if (attr->alt_port_num == 0 ||
981 		    attr->alt_port_num > dev->dev->caps.num_ports)
982 			goto out;
983 
984 		if (attr->alt_pkey_index >=
985 		    dev->dev->caps.pkey_table_len[attr->alt_port_num])
986 			goto out;
987 
988 		if (mlx4_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
989 				  attr->alt_port_num))
990 			goto out;
991 
992 		context->alt_path.pkey_index = attr->alt_pkey_index;
993 		context->alt_path.ackto = attr->alt_timeout << 3;
994 		optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
995 	}
996 
997 	context->pd	    = cpu_to_be32(to_mpd(ibqp->pd)->pdn);
998 	context->params1    = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
999 
1000 	/* Set "fast registration enabled" for all kernel QPs */
1001 	if (!qp->ibqp.uobject)
1002 		context->params1 |= cpu_to_be32(1 << 11);
1003 
1004 	if (attr_mask & IB_QP_RNR_RETRY) {
1005 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1006 		optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1007 	}
1008 
1009 	if (attr_mask & IB_QP_RETRY_CNT) {
1010 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1011 		optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1012 	}
1013 
1014 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1015 		if (attr->max_rd_atomic)
1016 			context->params1 |=
1017 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1018 		optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1019 	}
1020 
1021 	if (attr_mask & IB_QP_SQ_PSN)
1022 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
1023 
1024 	context->cqn_send = cpu_to_be32(to_mcq(ibqp->send_cq)->mcq.cqn);
1025 
1026 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1027 		if (attr->max_dest_rd_atomic)
1028 			context->params2 |=
1029 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1030 		optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1031 	}
1032 
1033 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1034 		context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1035 		optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1036 	}
1037 
1038 	if (ibqp->srq)
1039 		context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1040 
1041 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1042 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1043 		optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1044 	}
1045 	if (attr_mask & IB_QP_RQ_PSN)
1046 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1047 
1048 	context->cqn_recv = cpu_to_be32(to_mcq(ibqp->recv_cq)->mcq.cqn);
1049 
1050 	if (attr_mask & IB_QP_QKEY) {
1051 		context->qkey = cpu_to_be32(attr->qkey);
1052 		optpar |= MLX4_QP_OPTPAR_Q_KEY;
1053 	}
1054 
1055 	if (ibqp->srq)
1056 		context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1057 
1058 	if (!ibqp->srq && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1059 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
1060 
1061 	if (cur_state == IB_QPS_INIT &&
1062 	    new_state == IB_QPS_RTR  &&
1063 	    (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1064 	     ibqp->qp_type == IB_QPT_UD)) {
1065 		context->pri_path.sched_queue = (qp->port - 1) << 6;
1066 		if (is_qp0(dev, qp))
1067 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1068 		else
1069 			context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1070 	}
1071 
1072 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
1073 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1074 		sqd_event = 1;
1075 	else
1076 		sqd_event = 0;
1077 
1078 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1079 		context->rlkey |= (1 << 4);
1080 
1081 	/*
1082 	 * Before passing a kernel QP to the HW, make sure that the
1083 	 * ownership bits of the send queue are set and the SQ
1084 	 * headroom is stamped so that the hardware doesn't start
1085 	 * processing stale work requests.
1086 	 */
1087 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1088 		struct mlx4_wqe_ctrl_seg *ctrl;
1089 		int i;
1090 
1091 		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1092 			ctrl = get_send_wqe(qp, i);
1093 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
1094 			if (qp->sq_max_wqes_per_wr == 1)
1095 				ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1096 
1097 			stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1098 		}
1099 	}
1100 
1101 	err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1102 			     to_mlx4_state(new_state), context, optpar,
1103 			     sqd_event, &qp->mqp);
1104 	if (err)
1105 		goto out;
1106 
1107 	qp->state = new_state;
1108 
1109 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1110 		qp->atomic_rd_en = attr->qp_access_flags;
1111 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1112 		qp->resp_depth = attr->max_dest_rd_atomic;
1113 	if (attr_mask & IB_QP_PORT)
1114 		qp->port = attr->port_num;
1115 	if (attr_mask & IB_QP_ALT_PATH)
1116 		qp->alt_port = attr->alt_port_num;
1117 
1118 	if (is_sqp(dev, qp))
1119 		store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1120 
1121 	/*
1122 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
1123 	 * QP0 to RESET or ERROR, bring the link back down.
1124 	 */
1125 	if (is_qp0(dev, qp)) {
1126 		if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1127 			if (mlx4_INIT_PORT(dev->dev, qp->port))
1128 				printk(KERN_WARNING "INIT_PORT failed for port %d\n",
1129 				       qp->port);
1130 
1131 		if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1132 		    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1133 			mlx4_CLOSE_PORT(dev->dev, qp->port);
1134 	}
1135 
1136 	/*
1137 	 * If we moved a kernel QP to RESET, clean up all old CQ
1138 	 * entries and reinitialize the QP.
1139 	 */
1140 	if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1141 		mlx4_ib_cq_clean(to_mcq(ibqp->recv_cq), qp->mqp.qpn,
1142 				 ibqp->srq ? to_msrq(ibqp->srq): NULL);
1143 		if (ibqp->send_cq != ibqp->recv_cq)
1144 			mlx4_ib_cq_clean(to_mcq(ibqp->send_cq), qp->mqp.qpn, NULL);
1145 
1146 		qp->rq.head = 0;
1147 		qp->rq.tail = 0;
1148 		qp->sq.head = 0;
1149 		qp->sq.tail = 0;
1150 		qp->sq_next_wqe = 0;
1151 		if (!ibqp->srq)
1152 			*qp->db.db  = 0;
1153 	}
1154 
1155 out:
1156 	kfree(context);
1157 	return err;
1158 }
1159 
1160 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1161 		      int attr_mask, struct ib_udata *udata)
1162 {
1163 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1164 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1165 	enum ib_qp_state cur_state, new_state;
1166 	int err = -EINVAL;
1167 
1168 	mutex_lock(&qp->mutex);
1169 
1170 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1171 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1172 
1173 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask))
1174 		goto out;
1175 
1176 	if ((attr_mask & IB_QP_PORT) &&
1177 	    (attr->port_num == 0 || attr->port_num > dev->dev->caps.num_ports)) {
1178 		goto out;
1179 	}
1180 
1181 	if (attr_mask & IB_QP_PKEY_INDEX) {
1182 		int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1183 		if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p])
1184 			goto out;
1185 	}
1186 
1187 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1188 	    attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
1189 		goto out;
1190 	}
1191 
1192 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1193 	    attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
1194 		goto out;
1195 	}
1196 
1197 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1198 		err = 0;
1199 		goto out;
1200 	}
1201 
1202 	err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1203 
1204 out:
1205 	mutex_unlock(&qp->mutex);
1206 	return err;
1207 }
1208 
1209 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
1210 			    void *wqe, unsigned *mlx_seg_len)
1211 {
1212 	struct ib_device *ib_dev = &to_mdev(sqp->qp.ibqp.device)->ib_dev;
1213 	struct mlx4_wqe_mlx_seg *mlx = wqe;
1214 	struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
1215 	struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
1216 	u16 pkey;
1217 	int send_size;
1218 	int header_size;
1219 	int spc;
1220 	int i;
1221 
1222 	send_size = 0;
1223 	for (i = 0; i < wr->num_sge; ++i)
1224 		send_size += wr->sg_list[i].length;
1225 
1226 	ib_ud_header_init(send_size, mlx4_ib_ah_grh_present(ah), &sqp->ud_header);
1227 
1228 	sqp->ud_header.lrh.service_level   =
1229 		be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 28;
1230 	sqp->ud_header.lrh.destination_lid = ah->av.dlid;
1231 	sqp->ud_header.lrh.source_lid      = cpu_to_be16(ah->av.g_slid & 0x7f);
1232 	if (mlx4_ib_ah_grh_present(ah)) {
1233 		sqp->ud_header.grh.traffic_class =
1234 			(be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 20) & 0xff;
1235 		sqp->ud_header.grh.flow_label    =
1236 			ah->av.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
1237 		sqp->ud_header.grh.hop_limit     = ah->av.hop_limit;
1238 		ib_get_cached_gid(ib_dev, be32_to_cpu(ah->av.port_pd) >> 24,
1239 				  ah->av.gid_index, &sqp->ud_header.grh.source_gid);
1240 		memcpy(sqp->ud_header.grh.destination_gid.raw,
1241 		       ah->av.dgid, 16);
1242 	}
1243 
1244 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1245 	mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
1246 				  (sqp->ud_header.lrh.destination_lid ==
1247 				   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
1248 				  (sqp->ud_header.lrh.service_level << 8));
1249 	mlx->rlid   = sqp->ud_header.lrh.destination_lid;
1250 
1251 	switch (wr->opcode) {
1252 	case IB_WR_SEND:
1253 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY;
1254 		sqp->ud_header.immediate_present = 0;
1255 		break;
1256 	case IB_WR_SEND_WITH_IMM:
1257 		sqp->ud_header.bth.opcode	 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1258 		sqp->ud_header.immediate_present = 1;
1259 		sqp->ud_header.immediate_data    = wr->ex.imm_data;
1260 		break;
1261 	default:
1262 		return -EINVAL;
1263 	}
1264 
1265 	sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1266 	if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1267 		sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1268 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1269 	if (!sqp->qp.ibqp.qp_num)
1270 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
1271 	else
1272 		ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
1273 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1274 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1275 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1276 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1277 					       sqp->qkey : wr->wr.ud.remote_qkey);
1278 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1279 
1280 	header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
1281 
1282 	if (0) {
1283 		printk(KERN_ERR "built UD header of size %d:\n", header_size);
1284 		for (i = 0; i < header_size / 4; ++i) {
1285 			if (i % 8 == 0)
1286 				printk("  [%02x] ", i * 4);
1287 			printk(" %08x",
1288 			       be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
1289 			if ((i + 1) % 8 == 0)
1290 				printk("\n");
1291 		}
1292 		printk("\n");
1293 	}
1294 
1295 	/*
1296 	 * Inline data segments may not cross a 64 byte boundary.  If
1297 	 * our UD header is bigger than the space available up to the
1298 	 * next 64 byte boundary in the WQE, use two inline data
1299 	 * segments to hold the UD header.
1300 	 */
1301 	spc = MLX4_INLINE_ALIGN -
1302 		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
1303 	if (header_size <= spc) {
1304 		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
1305 		memcpy(inl + 1, sqp->header_buf, header_size);
1306 		i = 1;
1307 	} else {
1308 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
1309 		memcpy(inl + 1, sqp->header_buf, spc);
1310 
1311 		inl = (void *) (inl + 1) + spc;
1312 		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
1313 		/*
1314 		 * Need a barrier here to make sure all the data is
1315 		 * visible before the byte_count field is set.
1316 		 * Otherwise the HCA prefetcher could grab the 64-byte
1317 		 * chunk with this inline segment and get a valid (!=
1318 		 * 0xffffffff) byte count but stale data, and end up
1319 		 * generating a packet with bad headers.
1320 		 *
1321 		 * The first inline segment's byte_count field doesn't
1322 		 * need a barrier, because it comes after a
1323 		 * control/MLX segment and therefore is at an offset
1324 		 * of 16 mod 64.
1325 		 */
1326 		wmb();
1327 		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
1328 		i = 2;
1329 	}
1330 
1331 	*mlx_seg_len =
1332 		ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
1333 	return 0;
1334 }
1335 
1336 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1337 {
1338 	unsigned cur;
1339 	struct mlx4_ib_cq *cq;
1340 
1341 	cur = wq->head - wq->tail;
1342 	if (likely(cur + nreq < wq->max_post))
1343 		return 0;
1344 
1345 	cq = to_mcq(ib_cq);
1346 	spin_lock(&cq->lock);
1347 	cur = wq->head - wq->tail;
1348 	spin_unlock(&cq->lock);
1349 
1350 	return cur + nreq >= wq->max_post;
1351 }
1352 
1353 static __be32 convert_access(int acc)
1354 {
1355 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? cpu_to_be32(MLX4_WQE_FMR_PERM_ATOMIC)       : 0) |
1356 	       (acc & IB_ACCESS_REMOTE_WRITE  ? cpu_to_be32(MLX4_WQE_FMR_PERM_REMOTE_WRITE) : 0) |
1357 	       (acc & IB_ACCESS_REMOTE_READ   ? cpu_to_be32(MLX4_WQE_FMR_PERM_REMOTE_READ)  : 0) |
1358 	       (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
1359 		cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
1360 }
1361 
1362 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr)
1363 {
1364 	struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
1365 	int i;
1366 
1367 	for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i)
1368 		wr->wr.fast_reg.page_list->page_list[i] =
1369 			cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] |
1370 				    MLX4_MTT_FLAG_PRESENT);
1371 
1372 	fseg->flags		= convert_access(wr->wr.fast_reg.access_flags);
1373 	fseg->mem_key		= cpu_to_be32(wr->wr.fast_reg.rkey);
1374 	fseg->buf_list		= cpu_to_be64(mfrpl->map);
1375 	fseg->start_addr	= cpu_to_be64(wr->wr.fast_reg.iova_start);
1376 	fseg->reg_len		= cpu_to_be64(wr->wr.fast_reg.length);
1377 	fseg->offset		= 0; /* XXX -- is this just for ZBVA? */
1378 	fseg->page_size		= cpu_to_be32(wr->wr.fast_reg.page_shift);
1379 	fseg->reserved[0]	= 0;
1380 	fseg->reserved[1]	= 0;
1381 }
1382 
1383 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
1384 {
1385 	iseg->flags	= 0;
1386 	iseg->mem_key	= cpu_to_be32(rkey);
1387 	iseg->guest_id	= 0;
1388 	iseg->pa	= 0;
1389 }
1390 
1391 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
1392 					  u64 remote_addr, u32 rkey)
1393 {
1394 	rseg->raddr    = cpu_to_be64(remote_addr);
1395 	rseg->rkey     = cpu_to_be32(rkey);
1396 	rseg->reserved = 0;
1397 }
1398 
1399 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
1400 {
1401 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1402 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
1403 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
1404 	} else {
1405 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
1406 		aseg->compare  = 0;
1407 	}
1408 
1409 }
1410 
1411 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
1412 			     struct ib_send_wr *wr)
1413 {
1414 	memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
1415 	dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1416 	dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1417 }
1418 
1419 static void set_mlx_icrc_seg(void *dseg)
1420 {
1421 	u32 *t = dseg;
1422 	struct mlx4_wqe_inline_seg *iseg = dseg;
1423 
1424 	t[1] = 0;
1425 
1426 	/*
1427 	 * Need a barrier here before writing the byte_count field to
1428 	 * make sure that all the data is visible before the
1429 	 * byte_count field is set.  Otherwise, if the segment begins
1430 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
1431 	 * chunk and get a valid (!= * 0xffffffff) byte count but
1432 	 * stale data, and end up sending the wrong data.
1433 	 */
1434 	wmb();
1435 
1436 	iseg->byte_count = cpu_to_be32((1 << 31) | 4);
1437 }
1438 
1439 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
1440 {
1441 	dseg->lkey       = cpu_to_be32(sg->lkey);
1442 	dseg->addr       = cpu_to_be64(sg->addr);
1443 
1444 	/*
1445 	 * Need a barrier here before writing the byte_count field to
1446 	 * make sure that all the data is visible before the
1447 	 * byte_count field is set.  Otherwise, if the segment begins
1448 	 * a new cacheline, the HCA prefetcher could grab the 64-byte
1449 	 * chunk and get a valid (!= * 0xffffffff) byte count but
1450 	 * stale data, and end up sending the wrong data.
1451 	 */
1452 	wmb();
1453 
1454 	dseg->byte_count = cpu_to_be32(sg->length);
1455 }
1456 
1457 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
1458 {
1459 	dseg->byte_count = cpu_to_be32(sg->length);
1460 	dseg->lkey       = cpu_to_be32(sg->lkey);
1461 	dseg->addr       = cpu_to_be64(sg->addr);
1462 }
1463 
1464 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr,
1465 			 struct mlx4_ib_qp *qp, unsigned *lso_seg_len)
1466 {
1467 	unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
1468 
1469 	/*
1470 	 * This is a temporary limitation and will be removed in
1471 	 * a forthcoming FW release:
1472 	 */
1473 	if (unlikely(halign > 64))
1474 		return -EINVAL;
1475 
1476 	if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
1477 		     wr->num_sge > qp->sq.max_gs - (halign >> 4)))
1478 		return -EINVAL;
1479 
1480 	memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
1481 
1482 	/* make sure LSO header is written before overwriting stamping */
1483 	wmb();
1484 
1485 	wqe->mss_hdr_size = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 |
1486 					wr->wr.ud.hlen);
1487 
1488 	*lso_seg_len = halign;
1489 	return 0;
1490 }
1491 
1492 static __be32 send_ieth(struct ib_send_wr *wr)
1493 {
1494 	switch (wr->opcode) {
1495 	case IB_WR_SEND_WITH_IMM:
1496 	case IB_WR_RDMA_WRITE_WITH_IMM:
1497 		return wr->ex.imm_data;
1498 
1499 	case IB_WR_SEND_WITH_INV:
1500 		return cpu_to_be32(wr->ex.invalidate_rkey);
1501 
1502 	default:
1503 		return 0;
1504 	}
1505 }
1506 
1507 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1508 		      struct ib_send_wr **bad_wr)
1509 {
1510 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1511 	void *wqe;
1512 	struct mlx4_wqe_ctrl_seg *ctrl;
1513 	struct mlx4_wqe_data_seg *dseg;
1514 	unsigned long flags;
1515 	int nreq;
1516 	int err = 0;
1517 	unsigned ind;
1518 	int uninitialized_var(stamp);
1519 	int uninitialized_var(size);
1520 	unsigned uninitialized_var(seglen);
1521 	int i;
1522 
1523 	spin_lock_irqsave(&qp->sq.lock, flags);
1524 
1525 	ind = qp->sq_next_wqe;
1526 
1527 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1528 		if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1529 			err = -ENOMEM;
1530 			*bad_wr = wr;
1531 			goto out;
1532 		}
1533 
1534 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
1535 			err = -EINVAL;
1536 			*bad_wr = wr;
1537 			goto out;
1538 		}
1539 
1540 		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
1541 		qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
1542 
1543 		ctrl->srcrb_flags =
1544 			(wr->send_flags & IB_SEND_SIGNALED ?
1545 			 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
1546 			(wr->send_flags & IB_SEND_SOLICITED ?
1547 			 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
1548 			((wr->send_flags & IB_SEND_IP_CSUM) ?
1549 			 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
1550 				     MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
1551 			qp->sq_signal_bits;
1552 
1553 		ctrl->imm = send_ieth(wr);
1554 
1555 		wqe += sizeof *ctrl;
1556 		size = sizeof *ctrl / 16;
1557 
1558 		switch (ibqp->qp_type) {
1559 		case IB_QPT_RC:
1560 		case IB_QPT_UC:
1561 			switch (wr->opcode) {
1562 			case IB_WR_ATOMIC_CMP_AND_SWP:
1563 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1564 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
1565 					      wr->wr.atomic.rkey);
1566 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
1567 
1568 				set_atomic_seg(wqe, wr);
1569 				wqe  += sizeof (struct mlx4_wqe_atomic_seg);
1570 
1571 				size += (sizeof (struct mlx4_wqe_raddr_seg) +
1572 					 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
1573 
1574 				break;
1575 
1576 			case IB_WR_RDMA_READ:
1577 			case IB_WR_RDMA_WRITE:
1578 			case IB_WR_RDMA_WRITE_WITH_IMM:
1579 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1580 					      wr->wr.rdma.rkey);
1581 				wqe  += sizeof (struct mlx4_wqe_raddr_seg);
1582 				size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
1583 				break;
1584 
1585 			case IB_WR_LOCAL_INV:
1586 				set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
1587 				wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
1588 				size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
1589 				break;
1590 
1591 			case IB_WR_FAST_REG_MR:
1592 				set_fmr_seg(wqe, wr);
1593 				wqe  += sizeof (struct mlx4_wqe_fmr_seg);
1594 				size += sizeof (struct mlx4_wqe_fmr_seg) / 16;
1595 				break;
1596 
1597 			default:
1598 				/* No extra segments required for sends */
1599 				break;
1600 			}
1601 			break;
1602 
1603 		case IB_QPT_UD:
1604 			set_datagram_seg(wqe, wr);
1605 			wqe  += sizeof (struct mlx4_wqe_datagram_seg);
1606 			size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
1607 
1608 			if (wr->opcode == IB_WR_LSO) {
1609 				err = build_lso_seg(wqe, wr, qp, &seglen);
1610 				if (unlikely(err)) {
1611 					*bad_wr = wr;
1612 					goto out;
1613 				}
1614 				wqe  += seglen;
1615 				size += seglen / 16;
1616 			}
1617 			break;
1618 
1619 		case IB_QPT_SMI:
1620 		case IB_QPT_GSI:
1621 			err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
1622 			if (unlikely(err)) {
1623 				*bad_wr = wr;
1624 				goto out;
1625 			}
1626 			wqe  += seglen;
1627 			size += seglen / 16;
1628 			break;
1629 
1630 		default:
1631 			break;
1632 		}
1633 
1634 		/*
1635 		 * Write data segments in reverse order, so as to
1636 		 * overwrite cacheline stamp last within each
1637 		 * cacheline.  This avoids issues with WQE
1638 		 * prefetching.
1639 		 */
1640 
1641 		dseg = wqe;
1642 		dseg += wr->num_sge - 1;
1643 		size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
1644 
1645 		/* Add one more inline data segment for ICRC for MLX sends */
1646 		if (unlikely(qp->ibqp.qp_type == IB_QPT_SMI ||
1647 			     qp->ibqp.qp_type == IB_QPT_GSI)) {
1648 			set_mlx_icrc_seg(dseg + 1);
1649 			size += sizeof (struct mlx4_wqe_data_seg) / 16;
1650 		}
1651 
1652 		for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
1653 			set_data_seg(dseg, wr->sg_list + i);
1654 
1655 		ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
1656 				    MLX4_WQE_CTRL_FENCE : 0) | size;
1657 
1658 		/*
1659 		 * Make sure descriptor is fully written before
1660 		 * setting ownership bit (because HW can start
1661 		 * executing as soon as we do).
1662 		 */
1663 		wmb();
1664 
1665 		if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
1666 			err = -EINVAL;
1667 			goto out;
1668 		}
1669 
1670 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
1671 			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
1672 
1673 		stamp = ind + qp->sq_spare_wqes;
1674 		ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
1675 
1676 		/*
1677 		 * We can improve latency by not stamping the last
1678 		 * send queue WQE until after ringing the doorbell, so
1679 		 * only stamp here if there are still more WQEs to post.
1680 		 *
1681 		 * Same optimization applies to padding with NOP wqe
1682 		 * in case of WQE shrinking (used to prevent wrap-around
1683 		 * in the middle of WR).
1684 		 */
1685 		if (wr->next) {
1686 			stamp_send_wqe(qp, stamp, size * 16);
1687 			ind = pad_wraparound(qp, ind);
1688 		}
1689 
1690 	}
1691 
1692 out:
1693 	if (likely(nreq)) {
1694 		qp->sq.head += nreq;
1695 
1696 		/*
1697 		 * Make sure that descriptors are written before
1698 		 * doorbell record.
1699 		 */
1700 		wmb();
1701 
1702 		writel(qp->doorbell_qpn,
1703 		       to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
1704 
1705 		/*
1706 		 * Make sure doorbells don't leak out of SQ spinlock
1707 		 * and reach the HCA out of order.
1708 		 */
1709 		mmiowb();
1710 
1711 		stamp_send_wqe(qp, stamp, size * 16);
1712 
1713 		ind = pad_wraparound(qp, ind);
1714 		qp->sq_next_wqe = ind;
1715 	}
1716 
1717 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1718 
1719 	return err;
1720 }
1721 
1722 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1723 		      struct ib_recv_wr **bad_wr)
1724 {
1725 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1726 	struct mlx4_wqe_data_seg *scat;
1727 	unsigned long flags;
1728 	int err = 0;
1729 	int nreq;
1730 	int ind;
1731 	int i;
1732 
1733 	spin_lock_irqsave(&qp->rq.lock, flags);
1734 
1735 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
1736 
1737 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1738 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.send_cq)) {
1739 			err = -ENOMEM;
1740 			*bad_wr = wr;
1741 			goto out;
1742 		}
1743 
1744 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1745 			err = -EINVAL;
1746 			*bad_wr = wr;
1747 			goto out;
1748 		}
1749 
1750 		scat = get_recv_wqe(qp, ind);
1751 
1752 		for (i = 0; i < wr->num_sge; ++i)
1753 			__set_data_seg(scat + i, wr->sg_list + i);
1754 
1755 		if (i < qp->rq.max_gs) {
1756 			scat[i].byte_count = 0;
1757 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
1758 			scat[i].addr       = 0;
1759 		}
1760 
1761 		qp->rq.wrid[ind] = wr->wr_id;
1762 
1763 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
1764 	}
1765 
1766 out:
1767 	if (likely(nreq)) {
1768 		qp->rq.head += nreq;
1769 
1770 		/*
1771 		 * Make sure that descriptors are written before
1772 		 * doorbell record.
1773 		 */
1774 		wmb();
1775 
1776 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
1777 	}
1778 
1779 	spin_unlock_irqrestore(&qp->rq.lock, flags);
1780 
1781 	return err;
1782 }
1783 
1784 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
1785 {
1786 	switch (mlx4_state) {
1787 	case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
1788 	case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
1789 	case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
1790 	case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
1791 	case MLX4_QP_STATE_SQ_DRAINING:
1792 	case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
1793 	case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
1794 	case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
1795 	default:		     return -1;
1796 	}
1797 }
1798 
1799 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
1800 {
1801 	switch (mlx4_mig_state) {
1802 	case MLX4_QP_PM_ARMED:		return IB_MIG_ARMED;
1803 	case MLX4_QP_PM_REARM:		return IB_MIG_REARM;
1804 	case MLX4_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
1805 	default: return -1;
1806 	}
1807 }
1808 
1809 static int to_ib_qp_access_flags(int mlx4_flags)
1810 {
1811 	int ib_flags = 0;
1812 
1813 	if (mlx4_flags & MLX4_QP_BIT_RRE)
1814 		ib_flags |= IB_ACCESS_REMOTE_READ;
1815 	if (mlx4_flags & MLX4_QP_BIT_RWE)
1816 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
1817 	if (mlx4_flags & MLX4_QP_BIT_RAE)
1818 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
1819 
1820 	return ib_flags;
1821 }
1822 
1823 static void to_ib_ah_attr(struct mlx4_dev *dev, struct ib_ah_attr *ib_ah_attr,
1824 				struct mlx4_qp_path *path)
1825 {
1826 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
1827 	ib_ah_attr->port_num	  = path->sched_queue & 0x40 ? 2 : 1;
1828 
1829 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
1830 		return;
1831 
1832 	ib_ah_attr->dlid	  = be16_to_cpu(path->rlid);
1833 	ib_ah_attr->sl		  = (path->sched_queue >> 2) & 0xf;
1834 	ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
1835 	ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
1836 	ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
1837 	if (ib_ah_attr->ah_flags) {
1838 		ib_ah_attr->grh.sgid_index = path->mgid_index;
1839 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
1840 		ib_ah_attr->grh.traffic_class =
1841 			(be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
1842 		ib_ah_attr->grh.flow_label =
1843 			be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
1844 		memcpy(ib_ah_attr->grh.dgid.raw,
1845 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
1846 	}
1847 }
1848 
1849 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
1850 		     struct ib_qp_init_attr *qp_init_attr)
1851 {
1852 	struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1853 	struct mlx4_ib_qp *qp = to_mqp(ibqp);
1854 	struct mlx4_qp_context context;
1855 	int mlx4_state;
1856 	int err = 0;
1857 
1858 	mutex_lock(&qp->mutex);
1859 
1860 	if (qp->state == IB_QPS_RESET) {
1861 		qp_attr->qp_state = IB_QPS_RESET;
1862 		goto done;
1863 	}
1864 
1865 	err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
1866 	if (err) {
1867 		err = -EINVAL;
1868 		goto out;
1869 	}
1870 
1871 	mlx4_state = be32_to_cpu(context.flags) >> 28;
1872 
1873 	qp->state		     = to_ib_qp_state(mlx4_state);
1874 	qp_attr->qp_state	     = qp->state;
1875 	qp_attr->path_mtu	     = context.mtu_msgmax >> 5;
1876 	qp_attr->path_mig_state	     =
1877 		to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
1878 	qp_attr->qkey		     = be32_to_cpu(context.qkey);
1879 	qp_attr->rq_psn		     = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
1880 	qp_attr->sq_psn		     = be32_to_cpu(context.next_send_psn) & 0xffffff;
1881 	qp_attr->dest_qp_num	     = be32_to_cpu(context.remote_qpn) & 0xffffff;
1882 	qp_attr->qp_access_flags     =
1883 		to_ib_qp_access_flags(be32_to_cpu(context.params2));
1884 
1885 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
1886 		to_ib_ah_attr(dev->dev, &qp_attr->ah_attr, &context.pri_path);
1887 		to_ib_ah_attr(dev->dev, &qp_attr->alt_ah_attr, &context.alt_path);
1888 		qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
1889 		qp_attr->alt_port_num	= qp_attr->alt_ah_attr.port_num;
1890 	}
1891 
1892 	qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
1893 	if (qp_attr->qp_state == IB_QPS_INIT)
1894 		qp_attr->port_num = qp->port;
1895 	else
1896 		qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
1897 
1898 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
1899 	qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
1900 
1901 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
1902 
1903 	qp_attr->max_dest_rd_atomic =
1904 		1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
1905 	qp_attr->min_rnr_timer	    =
1906 		(be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
1907 	qp_attr->timeout	    = context.pri_path.ackto >> 3;
1908 	qp_attr->retry_cnt	    = (be32_to_cpu(context.params1) >> 16) & 0x7;
1909 	qp_attr->rnr_retry	    = (be32_to_cpu(context.params1) >> 13) & 0x7;
1910 	qp_attr->alt_timeout	    = context.alt_path.ackto >> 3;
1911 
1912 done:
1913 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
1914 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
1915 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
1916 
1917 	if (!ibqp->uobject) {
1918 		qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
1919 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
1920 	} else {
1921 		qp_attr->cap.max_send_wr  = 0;
1922 		qp_attr->cap.max_send_sge = 0;
1923 	}
1924 
1925 	/*
1926 	 * We don't support inline sends for kernel QPs (yet), and we
1927 	 * don't know what userspace's value should be.
1928 	 */
1929 	qp_attr->cap.max_inline_data = 0;
1930 
1931 	qp_init_attr->cap	     = qp_attr->cap;
1932 
1933 	qp_init_attr->create_flags = 0;
1934 	if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
1935 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
1936 
1937 	if (qp->flags & MLX4_IB_QP_LSO)
1938 		qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
1939 
1940 out:
1941 	mutex_unlock(&qp->mutex);
1942 	return err;
1943 }
1944 
1945