1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Cisco Systems. All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * OpenIB.org BSD license below:
12  *
13  *     Redistribution and use in source and binary forms, with or
14  *     without modification, are permitted provided that the following
15  *     conditions are met:
16  *
17  *      - Redistributions of source code must retain the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer.
20  *
21  *      - Redistributions in binary form must reproduce the above
22  *        copyright notice, this list of conditions and the following
23  *        disclaimer in the documentation and/or other materials
24  *        provided with the distribution.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33  * SOFTWARE.
34  */
35 
36 #include <linux/string.h>
37 #include <linux/slab.h>
38 #include <linux/sched.h>
39 
40 #include <asm/io.h>
41 
42 #include <rdma/ib_verbs.h>
43 #include <rdma/ib_cache.h>
44 #include <rdma/ib_pack.h>
45 #include <rdma/uverbs_ioctl.h>
46 
47 #include "mthca_dev.h"
48 #include "mthca_cmd.h"
49 #include "mthca_memfree.h"
50 #include "mthca_wqe.h"
51 
52 enum {
53 	MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
54 	MTHCA_ACK_REQ_FREQ       = 10,
55 	MTHCA_FLIGHT_LIMIT       = 9,
56 	MTHCA_UD_HEADER_SIZE     = 72, /* largest UD header possible */
57 	MTHCA_INLINE_HEADER_SIZE = 4,  /* data segment overhead for inline */
58 	MTHCA_INLINE_CHUNK_SIZE  = 16  /* inline data segment chunk */
59 };
60 
61 enum {
62 	MTHCA_QP_STATE_RST  = 0,
63 	MTHCA_QP_STATE_INIT = 1,
64 	MTHCA_QP_STATE_RTR  = 2,
65 	MTHCA_QP_STATE_RTS  = 3,
66 	MTHCA_QP_STATE_SQE  = 4,
67 	MTHCA_QP_STATE_SQD  = 5,
68 	MTHCA_QP_STATE_ERR  = 6,
69 	MTHCA_QP_STATE_DRAINING = 7
70 };
71 
72 enum {
73 	MTHCA_QP_ST_RC 	= 0x0,
74 	MTHCA_QP_ST_UC 	= 0x1,
75 	MTHCA_QP_ST_RD 	= 0x2,
76 	MTHCA_QP_ST_UD 	= 0x3,
77 	MTHCA_QP_ST_MLX = 0x7
78 };
79 
80 enum {
81 	MTHCA_QP_PM_MIGRATED = 0x3,
82 	MTHCA_QP_PM_ARMED    = 0x0,
83 	MTHCA_QP_PM_REARM    = 0x1
84 };
85 
86 enum {
87 	/* qp_context flags */
88 	MTHCA_QP_BIT_DE  = 1 <<  8,
89 	/* params1 */
90 	MTHCA_QP_BIT_SRE = 1 << 15,
91 	MTHCA_QP_BIT_SWE = 1 << 14,
92 	MTHCA_QP_BIT_SAE = 1 << 13,
93 	MTHCA_QP_BIT_SIC = 1 <<  4,
94 	MTHCA_QP_BIT_SSC = 1 <<  3,
95 	/* params2 */
96 	MTHCA_QP_BIT_RRE = 1 << 15,
97 	MTHCA_QP_BIT_RWE = 1 << 14,
98 	MTHCA_QP_BIT_RAE = 1 << 13,
99 	MTHCA_QP_BIT_RIC = 1 <<  4,
100 	MTHCA_QP_BIT_RSC = 1 <<  3
101 };
102 
103 enum {
104 	MTHCA_SEND_DOORBELL_FENCE = 1 << 5
105 };
106 
107 struct mthca_qp_path {
108 	__be32 port_pkey;
109 	u8     rnr_retry;
110 	u8     g_mylmc;
111 	__be16 rlid;
112 	u8     ackto;
113 	u8     mgid_index;
114 	u8     static_rate;
115 	u8     hop_limit;
116 	__be32 sl_tclass_flowlabel;
117 	u8     rgid[16];
118 } __attribute__((packed));
119 
120 struct mthca_qp_context {
121 	__be32 flags;
122 	__be32 tavor_sched_queue; /* Reserved on Arbel */
123 	u8     mtu_msgmax;
124 	u8     rq_size_stride;	/* Reserved on Tavor */
125 	u8     sq_size_stride;	/* Reserved on Tavor */
126 	u8     rlkey_arbel_sched_queue;	/* Reserved on Tavor */
127 	__be32 usr_page;
128 	__be32 local_qpn;
129 	__be32 remote_qpn;
130 	u32    reserved1[2];
131 	struct mthca_qp_path pri_path;
132 	struct mthca_qp_path alt_path;
133 	__be32 rdd;
134 	__be32 pd;
135 	__be32 wqe_base;
136 	__be32 wqe_lkey;
137 	__be32 params1;
138 	__be32 reserved2;
139 	__be32 next_send_psn;
140 	__be32 cqn_snd;
141 	__be32 snd_wqe_base_l;	/* Next send WQE on Tavor */
142 	__be32 snd_db_index;	/* (debugging only entries) */
143 	__be32 last_acked_psn;
144 	__be32 ssn;
145 	__be32 params2;
146 	__be32 rnr_nextrecvpsn;
147 	__be32 ra_buff_indx;
148 	__be32 cqn_rcv;
149 	__be32 rcv_wqe_base_l;	/* Next recv WQE on Tavor */
150 	__be32 rcv_db_index;	/* (debugging only entries) */
151 	__be32 qkey;
152 	__be32 srqn;
153 	__be32 rmsn;
154 	__be16 rq_wqe_counter;	/* reserved on Tavor */
155 	__be16 sq_wqe_counter;	/* reserved on Tavor */
156 	u32    reserved3[18];
157 } __attribute__((packed));
158 
159 struct mthca_qp_param {
160 	__be32 opt_param_mask;
161 	u32    reserved1;
162 	struct mthca_qp_context context;
163 	u32    reserved2[62];
164 } __attribute__((packed));
165 
166 enum {
167 	MTHCA_QP_OPTPAR_ALT_ADDR_PATH     = 1 << 0,
168 	MTHCA_QP_OPTPAR_RRE               = 1 << 1,
169 	MTHCA_QP_OPTPAR_RAE               = 1 << 2,
170 	MTHCA_QP_OPTPAR_RWE               = 1 << 3,
171 	MTHCA_QP_OPTPAR_PKEY_INDEX        = 1 << 4,
172 	MTHCA_QP_OPTPAR_Q_KEY             = 1 << 5,
173 	MTHCA_QP_OPTPAR_RNR_TIMEOUT       = 1 << 6,
174 	MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
175 	MTHCA_QP_OPTPAR_SRA_MAX           = 1 << 8,
176 	MTHCA_QP_OPTPAR_RRA_MAX           = 1 << 9,
177 	MTHCA_QP_OPTPAR_PM_STATE          = 1 << 10,
178 	MTHCA_QP_OPTPAR_PORT_NUM          = 1 << 11,
179 	MTHCA_QP_OPTPAR_RETRY_COUNT       = 1 << 12,
180 	MTHCA_QP_OPTPAR_ALT_RNR_RETRY     = 1 << 13,
181 	MTHCA_QP_OPTPAR_ACK_TIMEOUT       = 1 << 14,
182 	MTHCA_QP_OPTPAR_RNR_RETRY         = 1 << 15,
183 	MTHCA_QP_OPTPAR_SCHED_QUEUE       = 1 << 16
184 };
185 
186 static const u8 mthca_opcode[] = {
187 	[IB_WR_SEND]                 = MTHCA_OPCODE_SEND,
188 	[IB_WR_SEND_WITH_IMM]        = MTHCA_OPCODE_SEND_IMM,
189 	[IB_WR_RDMA_WRITE]           = MTHCA_OPCODE_RDMA_WRITE,
190 	[IB_WR_RDMA_WRITE_WITH_IMM]  = MTHCA_OPCODE_RDMA_WRITE_IMM,
191 	[IB_WR_RDMA_READ]            = MTHCA_OPCODE_RDMA_READ,
192 	[IB_WR_ATOMIC_CMP_AND_SWP]   = MTHCA_OPCODE_ATOMIC_CS,
193 	[IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
194 };
195 
196 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
197 {
198 	return qp->qpn >= dev->qp_table.sqp_start &&
199 		qp->qpn <= dev->qp_table.sqp_start + 3;
200 }
201 
202 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
203 {
204 	return qp->qpn >= dev->qp_table.sqp_start &&
205 		qp->qpn <= dev->qp_table.sqp_start + 1;
206 }
207 
208 static void *get_recv_wqe(struct mthca_qp *qp, int n)
209 {
210 	if (qp->is_direct)
211 		return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
212 	else
213 		return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
214 			((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
215 }
216 
217 static void *get_send_wqe(struct mthca_qp *qp, int n)
218 {
219 	if (qp->is_direct)
220 		return qp->queue.direct.buf + qp->send_wqe_offset +
221 			(n << qp->sq.wqe_shift);
222 	else
223 		return qp->queue.page_list[(qp->send_wqe_offset +
224 					    (n << qp->sq.wqe_shift)) >>
225 					   PAGE_SHIFT].buf +
226 			((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
227 			 (PAGE_SIZE - 1));
228 }
229 
230 static void mthca_wq_reset(struct mthca_wq *wq)
231 {
232 	wq->next_ind  = 0;
233 	wq->last_comp = wq->max - 1;
234 	wq->head      = 0;
235 	wq->tail      = 0;
236 }
237 
238 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
239 		    enum ib_event_type event_type)
240 {
241 	struct mthca_qp *qp;
242 	struct ib_event event;
243 
244 	spin_lock(&dev->qp_table.lock);
245 	qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
246 	if (qp)
247 		++qp->refcount;
248 	spin_unlock(&dev->qp_table.lock);
249 
250 	if (!qp) {
251 		mthca_warn(dev, "Async event %d for bogus QP %08x\n",
252 			   event_type, qpn);
253 		return;
254 	}
255 
256 	if (event_type == IB_EVENT_PATH_MIG)
257 		qp->port = qp->alt_port;
258 
259 	event.device      = &dev->ib_dev;
260 	event.event       = event_type;
261 	event.element.qp  = &qp->ibqp;
262 	if (qp->ibqp.event_handler)
263 		qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
264 
265 	spin_lock(&dev->qp_table.lock);
266 	if (!--qp->refcount)
267 		wake_up(&qp->wait);
268 	spin_unlock(&dev->qp_table.lock);
269 }
270 
271 static int to_mthca_state(enum ib_qp_state ib_state)
272 {
273 	switch (ib_state) {
274 	case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
275 	case IB_QPS_INIT:  return MTHCA_QP_STATE_INIT;
276 	case IB_QPS_RTR:   return MTHCA_QP_STATE_RTR;
277 	case IB_QPS_RTS:   return MTHCA_QP_STATE_RTS;
278 	case IB_QPS_SQD:   return MTHCA_QP_STATE_SQD;
279 	case IB_QPS_SQE:   return MTHCA_QP_STATE_SQE;
280 	case IB_QPS_ERR:   return MTHCA_QP_STATE_ERR;
281 	default:                return -1;
282 	}
283 }
284 
285 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
286 
287 static int to_mthca_st(int transport)
288 {
289 	switch (transport) {
290 	case RC:  return MTHCA_QP_ST_RC;
291 	case UC:  return MTHCA_QP_ST_UC;
292 	case UD:  return MTHCA_QP_ST_UD;
293 	case RD:  return MTHCA_QP_ST_RD;
294 	case MLX: return MTHCA_QP_ST_MLX;
295 	default:  return -1;
296 	}
297 }
298 
299 static void store_attrs(struct mthca_sqp *sqp, const struct ib_qp_attr *attr,
300 			int attr_mask)
301 {
302 	if (attr_mask & IB_QP_PKEY_INDEX)
303 		sqp->pkey_index = attr->pkey_index;
304 	if (attr_mask & IB_QP_QKEY)
305 		sqp->qkey = attr->qkey;
306 	if (attr_mask & IB_QP_SQ_PSN)
307 		sqp->send_psn = attr->sq_psn;
308 }
309 
310 static void init_port(struct mthca_dev *dev, int port)
311 {
312 	int err;
313 	struct mthca_init_ib_param param;
314 
315 	memset(&param, 0, sizeof param);
316 
317 	param.port_width = dev->limits.port_width_cap;
318 	param.vl_cap     = dev->limits.vl_cap;
319 	param.mtu_cap    = dev->limits.mtu_cap;
320 	param.gid_cap    = dev->limits.gid_table_len;
321 	param.pkey_cap   = dev->limits.pkey_table_len;
322 
323 	err = mthca_INIT_IB(dev, &param, port);
324 	if (err)
325 		mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
326 }
327 
328 static __be32 get_hw_access_flags(struct mthca_qp *qp, const struct ib_qp_attr *attr,
329 				  int attr_mask)
330 {
331 	u8 dest_rd_atomic;
332 	u32 access_flags;
333 	u32 hw_access_flags = 0;
334 
335 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
336 		dest_rd_atomic = attr->max_dest_rd_atomic;
337 	else
338 		dest_rd_atomic = qp->resp_depth;
339 
340 	if (attr_mask & IB_QP_ACCESS_FLAGS)
341 		access_flags = attr->qp_access_flags;
342 	else
343 		access_flags = qp->atomic_rd_en;
344 
345 	if (!dest_rd_atomic)
346 		access_flags &= IB_ACCESS_REMOTE_WRITE;
347 
348 	if (access_flags & IB_ACCESS_REMOTE_READ)
349 		hw_access_flags |= MTHCA_QP_BIT_RRE;
350 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
351 		hw_access_flags |= MTHCA_QP_BIT_RAE;
352 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
353 		hw_access_flags |= MTHCA_QP_BIT_RWE;
354 
355 	return cpu_to_be32(hw_access_flags);
356 }
357 
358 static inline enum ib_qp_state to_ib_qp_state(int mthca_state)
359 {
360 	switch (mthca_state) {
361 	case MTHCA_QP_STATE_RST:      return IB_QPS_RESET;
362 	case MTHCA_QP_STATE_INIT:     return IB_QPS_INIT;
363 	case MTHCA_QP_STATE_RTR:      return IB_QPS_RTR;
364 	case MTHCA_QP_STATE_RTS:      return IB_QPS_RTS;
365 	case MTHCA_QP_STATE_DRAINING:
366 	case MTHCA_QP_STATE_SQD:      return IB_QPS_SQD;
367 	case MTHCA_QP_STATE_SQE:      return IB_QPS_SQE;
368 	case MTHCA_QP_STATE_ERR:      return IB_QPS_ERR;
369 	default:                      return -1;
370 	}
371 }
372 
373 static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state)
374 {
375 	switch (mthca_mig_state) {
376 	case 0:  return IB_MIG_ARMED;
377 	case 1:  return IB_MIG_REARM;
378 	case 3:  return IB_MIG_MIGRATED;
379 	default: return -1;
380 	}
381 }
382 
383 static int to_ib_qp_access_flags(int mthca_flags)
384 {
385 	int ib_flags = 0;
386 
387 	if (mthca_flags & MTHCA_QP_BIT_RRE)
388 		ib_flags |= IB_ACCESS_REMOTE_READ;
389 	if (mthca_flags & MTHCA_QP_BIT_RWE)
390 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
391 	if (mthca_flags & MTHCA_QP_BIT_RAE)
392 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
393 
394 	return ib_flags;
395 }
396 
397 static void to_rdma_ah_attr(struct mthca_dev *dev,
398 			    struct rdma_ah_attr *ah_attr,
399 			    struct mthca_qp_path *path)
400 {
401 	u8 port_num = (be32_to_cpu(path->port_pkey) >> 24) & 0x3;
402 
403 	memset(ah_attr, 0, sizeof(*ah_attr));
404 
405 	if (port_num == 0 || port_num > dev->limits.num_ports)
406 		return;
407 	ah_attr->type = rdma_ah_find_type(&dev->ib_dev, port_num);
408 	rdma_ah_set_port_num(ah_attr, port_num);
409 
410 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
411 	rdma_ah_set_sl(ah_attr, be32_to_cpu(path->sl_tclass_flowlabel) >> 28);
412 	rdma_ah_set_path_bits(ah_attr, path->g_mylmc & 0x7f);
413 	rdma_ah_set_static_rate(ah_attr,
414 				mthca_rate_to_ib(dev,
415 						 path->static_rate & 0xf,
416 						 port_num));
417 	if (path->g_mylmc & (1 << 7)) {
418 		u32 tc_fl = be32_to_cpu(path->sl_tclass_flowlabel);
419 
420 		rdma_ah_set_grh(ah_attr, NULL,
421 				tc_fl & 0xfffff,
422 				path->mgid_index &
423 				(dev->limits.gid_table_len - 1),
424 				path->hop_limit,
425 				(tc_fl >> 20) & 0xff);
426 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
427 	}
428 }
429 
430 int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
431 		   struct ib_qp_init_attr *qp_init_attr)
432 {
433 	struct mthca_dev *dev = to_mdev(ibqp->device);
434 	struct mthca_qp *qp = to_mqp(ibqp);
435 	int err = 0;
436 	struct mthca_mailbox *mailbox = NULL;
437 	struct mthca_qp_param *qp_param;
438 	struct mthca_qp_context *context;
439 	int mthca_state;
440 
441 	mutex_lock(&qp->mutex);
442 
443 	if (qp->state == IB_QPS_RESET) {
444 		qp_attr->qp_state = IB_QPS_RESET;
445 		goto done;
446 	}
447 
448 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
449 	if (IS_ERR(mailbox)) {
450 		err = PTR_ERR(mailbox);
451 		goto out;
452 	}
453 
454 	err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox);
455 	if (err) {
456 		mthca_warn(dev, "QUERY_QP failed (%d)\n", err);
457 		goto out_mailbox;
458 	}
459 
460 	qp_param    = mailbox->buf;
461 	context     = &qp_param->context;
462 	mthca_state = be32_to_cpu(context->flags) >> 28;
463 
464 	qp->state		     = to_ib_qp_state(mthca_state);
465 	qp_attr->qp_state	     = qp->state;
466 	qp_attr->path_mtu 	     = context->mtu_msgmax >> 5;
467 	qp_attr->path_mig_state      =
468 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
469 	qp_attr->qkey 		     = be32_to_cpu(context->qkey);
470 	qp_attr->rq_psn 	     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
471 	qp_attr->sq_psn 	     = be32_to_cpu(context->next_send_psn) & 0xffffff;
472 	qp_attr->dest_qp_num 	     = be32_to_cpu(context->remote_qpn) & 0xffffff;
473 	qp_attr->qp_access_flags     =
474 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
475 
476 	if (qp->transport == RC || qp->transport == UC) {
477 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
478 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
479 		qp_attr->alt_pkey_index =
480 			be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
481 		qp_attr->alt_port_num	=
482 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
483 	}
484 
485 	qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
486 	qp_attr->port_num   =
487 		(be32_to_cpu(context->pri_path.port_pkey) >> 24) & 0x3;
488 
489 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
490 	qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
491 
492 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
493 
494 	qp_attr->max_dest_rd_atomic =
495 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
496 	qp_attr->min_rnr_timer 	    =
497 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
498 	qp_attr->timeout 	    = context->pri_path.ackto >> 3;
499 	qp_attr->retry_cnt 	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
500 	qp_attr->rnr_retry 	    = context->pri_path.rnr_retry >> 5;
501 	qp_attr->alt_timeout 	    = context->alt_path.ackto >> 3;
502 
503 done:
504 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
505 	qp_attr->cap.max_send_wr     = qp->sq.max;
506 	qp_attr->cap.max_recv_wr     = qp->rq.max;
507 	qp_attr->cap.max_send_sge    = qp->sq.max_gs;
508 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
509 	qp_attr->cap.max_inline_data = qp->max_inline_data;
510 
511 	qp_init_attr->cap	     = qp_attr->cap;
512 	qp_init_attr->sq_sig_type    = qp->sq_policy;
513 
514 out_mailbox:
515 	mthca_free_mailbox(dev, mailbox);
516 
517 out:
518 	mutex_unlock(&qp->mutex);
519 	return err;
520 }
521 
522 static int mthca_path_set(struct mthca_dev *dev, const struct rdma_ah_attr *ah,
523 			  struct mthca_qp_path *path, u8 port)
524 {
525 	path->g_mylmc     = rdma_ah_get_path_bits(ah) & 0x7f;
526 	path->rlid        = cpu_to_be16(rdma_ah_get_dlid(ah));
527 	path->static_rate = mthca_get_rate(dev, rdma_ah_get_static_rate(ah),
528 					   port);
529 
530 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
531 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
532 
533 		if (grh->sgid_index >= dev->limits.gid_table_len) {
534 			mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
535 				  grh->sgid_index,
536 				  dev->limits.gid_table_len - 1);
537 			return -1;
538 		}
539 
540 		path->g_mylmc   |= 1 << 7;
541 		path->mgid_index = grh->sgid_index;
542 		path->hop_limit  = grh->hop_limit;
543 		path->sl_tclass_flowlabel =
544 			cpu_to_be32((rdma_ah_get_sl(ah) << 28) |
545 				    (grh->traffic_class << 20) |
546 				    (grh->flow_label));
547 		memcpy(path->rgid, grh->dgid.raw, 16);
548 	} else {
549 		path->sl_tclass_flowlabel = cpu_to_be32(rdma_ah_get_sl(ah) <<
550 							28);
551 	}
552 
553 	return 0;
554 }
555 
556 static int __mthca_modify_qp(struct ib_qp *ibqp,
557 			     const struct ib_qp_attr *attr, int attr_mask,
558 			     enum ib_qp_state cur_state,
559 			     enum ib_qp_state new_state,
560 			     struct ib_udata *udata)
561 {
562 	struct mthca_dev *dev = to_mdev(ibqp->device);
563 	struct mthca_qp *qp = to_mqp(ibqp);
564 	struct mthca_ucontext *context = rdma_udata_to_drv_context(
565 		udata, struct mthca_ucontext, ibucontext);
566 	struct mthca_mailbox *mailbox;
567 	struct mthca_qp_param *qp_param;
568 	struct mthca_qp_context *qp_context;
569 	u32 sqd_event = 0;
570 	int err = -EINVAL;
571 
572 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
573 	if (IS_ERR(mailbox)) {
574 		err = PTR_ERR(mailbox);
575 		goto out;
576 	}
577 	qp_param = mailbox->buf;
578 	qp_context = &qp_param->context;
579 	memset(qp_param, 0, sizeof *qp_param);
580 
581 	qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
582 					     (to_mthca_st(qp->transport) << 16));
583 	qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
584 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
585 		qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
586 	else {
587 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
588 		switch (attr->path_mig_state) {
589 		case IB_MIG_MIGRATED:
590 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
591 			break;
592 		case IB_MIG_REARM:
593 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
594 			break;
595 		case IB_MIG_ARMED:
596 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
597 			break;
598 		}
599 	}
600 
601 	/* leave tavor_sched_queue as 0 */
602 
603 	if (qp->transport == MLX || qp->transport == UD)
604 		qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
605 	else if (attr_mask & IB_QP_PATH_MTU) {
606 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
607 			mthca_dbg(dev, "path MTU (%u) is invalid\n",
608 				  attr->path_mtu);
609 			goto out_mailbox;
610 		}
611 		qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
612 	}
613 
614 	if (mthca_is_memfree(dev)) {
615 		if (qp->rq.max)
616 			qp_context->rq_size_stride = ilog2(qp->rq.max) << 3;
617 		qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
618 
619 		if (qp->sq.max)
620 			qp_context->sq_size_stride = ilog2(qp->sq.max) << 3;
621 		qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
622 	}
623 
624 	/* leave arbel_sched_queue as 0 */
625 
626 	if (qp->ibqp.uobject)
627 		qp_context->usr_page = cpu_to_be32(context->uar.index);
628 	else
629 		qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
630 	qp_context->local_qpn  = cpu_to_be32(qp->qpn);
631 	if (attr_mask & IB_QP_DEST_QPN) {
632 		qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
633 	}
634 
635 	if (qp->transport == MLX)
636 		qp_context->pri_path.port_pkey |=
637 			cpu_to_be32(qp->port << 24);
638 	else {
639 		if (attr_mask & IB_QP_PORT) {
640 			qp_context->pri_path.port_pkey |=
641 				cpu_to_be32(attr->port_num << 24);
642 			qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
643 		}
644 	}
645 
646 	if (attr_mask & IB_QP_PKEY_INDEX) {
647 		qp_context->pri_path.port_pkey |=
648 			cpu_to_be32(attr->pkey_index);
649 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
650 	}
651 
652 	if (attr_mask & IB_QP_RNR_RETRY) {
653 		qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
654 			attr->rnr_retry << 5;
655 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
656 							MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
657 	}
658 
659 	if (attr_mask & IB_QP_AV) {
660 		if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
661 				   attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
662 			goto out_mailbox;
663 
664 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
665 	}
666 
667 	if (ibqp->qp_type == IB_QPT_RC &&
668 	    cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
669 		u8 sched_queue = ibqp->uobject ? 0x2 : 0x1;
670 
671 		if (mthca_is_memfree(dev))
672 			qp_context->rlkey_arbel_sched_queue |= sched_queue;
673 		else
674 			qp_context->tavor_sched_queue |= cpu_to_be32(sched_queue);
675 
676 		qp_param->opt_param_mask |=
677 			cpu_to_be32(MTHCA_QP_OPTPAR_SCHED_QUEUE);
678 	}
679 
680 	if (attr_mask & IB_QP_TIMEOUT) {
681 		qp_context->pri_path.ackto = attr->timeout << 3;
682 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
683 	}
684 
685 	if (attr_mask & IB_QP_ALT_PATH) {
686 		if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
687 			mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
688 				  attr->alt_pkey_index, dev->limits.pkey_table_len-1);
689 			goto out_mailbox;
690 		}
691 
692 		if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
693 			mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
694 				attr->alt_port_num);
695 			goto out_mailbox;
696 		}
697 
698 		if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
699 				   rdma_ah_get_port_num(&attr->alt_ah_attr)))
700 			goto out_mailbox;
701 
702 		qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
703 							      attr->alt_port_num << 24);
704 		qp_context->alt_path.ackto = attr->alt_timeout << 3;
705 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
706 	}
707 
708 	/* leave rdd as 0 */
709 	qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
710 	/* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
711 	qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
712 	qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
713 					     (MTHCA_FLIGHT_LIMIT << 24) |
714 					     MTHCA_QP_BIT_SWE);
715 	if (qp->sq_policy == IB_SIGNAL_ALL_WR)
716 		qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
717 	if (attr_mask & IB_QP_RETRY_CNT) {
718 		qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
719 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
720 	}
721 
722 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
723 		if (attr->max_rd_atomic) {
724 			qp_context->params1 |=
725 				cpu_to_be32(MTHCA_QP_BIT_SRE |
726 					    MTHCA_QP_BIT_SAE);
727 			qp_context->params1 |=
728 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
729 		}
730 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
731 	}
732 
733 	if (attr_mask & IB_QP_SQ_PSN)
734 		qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
735 	qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
736 
737 	if (mthca_is_memfree(dev)) {
738 		qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
739 		qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
740 	}
741 
742 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
743 		if (attr->max_dest_rd_atomic)
744 			qp_context->params2 |=
745 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
746 
747 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
748 	}
749 
750 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
751 		qp_context->params2      |= get_hw_access_flags(qp, attr, attr_mask);
752 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
753 							MTHCA_QP_OPTPAR_RRE |
754 							MTHCA_QP_OPTPAR_RAE);
755 	}
756 
757 	qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
758 
759 	if (ibqp->srq)
760 		qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
761 
762 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
763 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
764 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
765 	}
766 	if (attr_mask & IB_QP_RQ_PSN)
767 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
768 
769 	qp_context->ra_buff_indx =
770 		cpu_to_be32(dev->qp_table.rdb_base +
771 			    ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
772 			     dev->qp_table.rdb_shift));
773 
774 	qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
775 
776 	if (mthca_is_memfree(dev))
777 		qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
778 
779 	if (attr_mask & IB_QP_QKEY) {
780 		qp_context->qkey = cpu_to_be32(attr->qkey);
781 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
782 	}
783 
784 	if (ibqp->srq)
785 		qp_context->srqn = cpu_to_be32(1 << 24 |
786 					       to_msrq(ibqp->srq)->srqn);
787 
788 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
789 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY		&&
790 	    attr->en_sqd_async_notify)
791 		sqd_event = 1 << 31;
792 
793 	err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
794 			      mailbox, sqd_event);
795 	if (err) {
796 		mthca_warn(dev, "modify QP %d->%d returned %d.\n",
797 			   cur_state, new_state, err);
798 		goto out_mailbox;
799 	}
800 
801 	qp->state = new_state;
802 	if (attr_mask & IB_QP_ACCESS_FLAGS)
803 		qp->atomic_rd_en = attr->qp_access_flags;
804 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
805 		qp->resp_depth = attr->max_dest_rd_atomic;
806 	if (attr_mask & IB_QP_PORT)
807 		qp->port = attr->port_num;
808 	if (attr_mask & IB_QP_ALT_PATH)
809 		qp->alt_port = attr->alt_port_num;
810 
811 	if (is_sqp(dev, qp))
812 		store_attrs(to_msqp(qp), attr, attr_mask);
813 
814 	/*
815 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
816 	 * QP0 to RESET or ERROR, bring the link back down.
817 	 */
818 	if (is_qp0(dev, qp)) {
819 		if (cur_state != IB_QPS_RTR &&
820 		    new_state == IB_QPS_RTR)
821 			init_port(dev, qp->port);
822 
823 		if (cur_state != IB_QPS_RESET &&
824 		    cur_state != IB_QPS_ERR &&
825 		    (new_state == IB_QPS_RESET ||
826 		     new_state == IB_QPS_ERR))
827 			mthca_CLOSE_IB(dev, qp->port);
828 	}
829 
830 	/*
831 	 * If we moved a kernel QP to RESET, clean up all old CQ
832 	 * entries and reinitialize the QP.
833 	 */
834 	if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
835 		mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
836 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
837 		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
838 			mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn, NULL);
839 
840 		mthca_wq_reset(&qp->sq);
841 		qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
842 
843 		mthca_wq_reset(&qp->rq);
844 		qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
845 
846 		if (mthca_is_memfree(dev)) {
847 			*qp->sq.db = 0;
848 			*qp->rq.db = 0;
849 		}
850 	}
851 
852 out_mailbox:
853 	mthca_free_mailbox(dev, mailbox);
854 out:
855 	return err;
856 }
857 
858 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
859 		    struct ib_udata *udata)
860 {
861 	struct mthca_dev *dev = to_mdev(ibqp->device);
862 	struct mthca_qp *qp = to_mqp(ibqp);
863 	enum ib_qp_state cur_state, new_state;
864 	int err = -EINVAL;
865 
866 	mutex_lock(&qp->mutex);
867 	if (attr_mask & IB_QP_CUR_STATE) {
868 		cur_state = attr->cur_qp_state;
869 	} else {
870 		spin_lock_irq(&qp->sq.lock);
871 		spin_lock(&qp->rq.lock);
872 		cur_state = qp->state;
873 		spin_unlock(&qp->rq.lock);
874 		spin_unlock_irq(&qp->sq.lock);
875 	}
876 
877 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
878 
879 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
880 				attr_mask)) {
881 		mthca_dbg(dev, "Bad QP transition (transport %d) "
882 			  "%d->%d with attr 0x%08x\n",
883 			  qp->transport, cur_state, new_state,
884 			  attr_mask);
885 		goto out;
886 	}
887 
888 	if ((attr_mask & IB_QP_PKEY_INDEX) &&
889 	     attr->pkey_index >= dev->limits.pkey_table_len) {
890 		mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
891 			  attr->pkey_index, dev->limits.pkey_table_len-1);
892 		goto out;
893 	}
894 
895 	if ((attr_mask & IB_QP_PORT) &&
896 	    (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
897 		mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
898 		goto out;
899 	}
900 
901 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
902 	    attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
903 		mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
904 			  attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
905 		goto out;
906 	}
907 
908 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
909 	    attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
910 		mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
911 			  attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
912 		goto out;
913 	}
914 
915 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
916 		err = 0;
917 		goto out;
918 	}
919 
920 	err = __mthca_modify_qp(ibqp, attr, attr_mask, cur_state, new_state,
921 				udata);
922 
923 out:
924 	mutex_unlock(&qp->mutex);
925 	return err;
926 }
927 
928 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
929 {
930 	/*
931 	 * Calculate the maximum size of WQE s/g segments, excluding
932 	 * the next segment and other non-data segments.
933 	 */
934 	int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
935 
936 	switch (qp->transport) {
937 	case MLX:
938 		max_data_size -= 2 * sizeof (struct mthca_data_seg);
939 		break;
940 
941 	case UD:
942 		if (mthca_is_memfree(dev))
943 			max_data_size -= sizeof (struct mthca_arbel_ud_seg);
944 		else
945 			max_data_size -= sizeof (struct mthca_tavor_ud_seg);
946 		break;
947 
948 	default:
949 		max_data_size -= sizeof (struct mthca_raddr_seg);
950 		break;
951 	}
952 
953 	return max_data_size;
954 }
955 
956 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
957 {
958 	/* We don't support inline data for kernel QPs (yet). */
959 	return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
960 }
961 
962 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
963 				 struct mthca_pd *pd,
964 				 struct mthca_qp *qp)
965 {
966 	int max_data_size = mthca_max_data_size(dev, qp,
967 						min(dev->limits.max_desc_sz,
968 						    1 << qp->sq.wqe_shift));
969 
970 	qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
971 
972 	qp->sq.max_gs = min_t(int, dev->limits.max_sg,
973 			      max_data_size / sizeof (struct mthca_data_seg));
974 	qp->rq.max_gs = min_t(int, dev->limits.max_sg,
975 			       (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
976 				sizeof (struct mthca_next_seg)) /
977 			       sizeof (struct mthca_data_seg));
978 }
979 
980 /*
981  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
982  * rq.max_gs and sq.max_gs must all be assigned.
983  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
984  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
985  * queue)
986  */
987 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
988 			       struct mthca_pd *pd,
989 			       struct mthca_qp *qp,
990 			       struct ib_udata *udata)
991 {
992 	int size;
993 	int err = -ENOMEM;
994 
995 	size = sizeof (struct mthca_next_seg) +
996 		qp->rq.max_gs * sizeof (struct mthca_data_seg);
997 
998 	if (size > dev->limits.max_desc_sz)
999 		return -EINVAL;
1000 
1001 	for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
1002 	     qp->rq.wqe_shift++)
1003 		; /* nothing */
1004 
1005 	size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
1006 	switch (qp->transport) {
1007 	case MLX:
1008 		size += 2 * sizeof (struct mthca_data_seg);
1009 		break;
1010 
1011 	case UD:
1012 		size += mthca_is_memfree(dev) ?
1013 			sizeof (struct mthca_arbel_ud_seg) :
1014 			sizeof (struct mthca_tavor_ud_seg);
1015 		break;
1016 
1017 	case UC:
1018 		size += sizeof (struct mthca_raddr_seg);
1019 		break;
1020 
1021 	case RC:
1022 		size += sizeof (struct mthca_raddr_seg);
1023 		/*
1024 		 * An atomic op will require an atomic segment, a
1025 		 * remote address segment and one scatter entry.
1026 		 */
1027 		size = max_t(int, size,
1028 			     sizeof (struct mthca_atomic_seg) +
1029 			     sizeof (struct mthca_raddr_seg) +
1030 			     sizeof (struct mthca_data_seg));
1031 		break;
1032 
1033 	default:
1034 		break;
1035 	}
1036 
1037 	/* Make sure that we have enough space for a bind request */
1038 	size = max_t(int, size, sizeof (struct mthca_bind_seg));
1039 
1040 	size += sizeof (struct mthca_next_seg);
1041 
1042 	if (size > dev->limits.max_desc_sz)
1043 		return -EINVAL;
1044 
1045 	for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
1046 	     qp->sq.wqe_shift++)
1047 		; /* nothing */
1048 
1049 	qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
1050 				    1 << qp->sq.wqe_shift);
1051 
1052 	/*
1053 	 * If this is a userspace QP, we don't actually have to
1054 	 * allocate anything.  All we need is to calculate the WQE
1055 	 * sizes and the send_wqe_offset, so we're done now.
1056 	 */
1057 	if (udata)
1058 		return 0;
1059 
1060 	size = PAGE_ALIGN(qp->send_wqe_offset +
1061 			  (qp->sq.max << qp->sq.wqe_shift));
1062 
1063 	qp->wrid = kmalloc_array(qp->rq.max + qp->sq.max, sizeof(u64),
1064 				 GFP_KERNEL);
1065 	if (!qp->wrid)
1066 		goto err_out;
1067 
1068 	err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1069 			      &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1070 	if (err)
1071 		goto err_out;
1072 
1073 	return 0;
1074 
1075 err_out:
1076 	kfree(qp->wrid);
1077 	return err;
1078 }
1079 
1080 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1081 			       struct mthca_qp *qp)
1082 {
1083 	mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1084 				       (qp->sq.max << qp->sq.wqe_shift)),
1085 		       &qp->queue, qp->is_direct, &qp->mr);
1086 	kfree(qp->wrid);
1087 }
1088 
1089 static int mthca_map_memfree(struct mthca_dev *dev,
1090 			     struct mthca_qp *qp)
1091 {
1092 	int ret;
1093 
1094 	if (mthca_is_memfree(dev)) {
1095 		ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1096 		if (ret)
1097 			return ret;
1098 
1099 		ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1100 		if (ret)
1101 			goto err_qpc;
1102 
1103 		ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1104 				      qp->qpn << dev->qp_table.rdb_shift);
1105 		if (ret)
1106 			goto err_eqpc;
1107 
1108 	}
1109 
1110 	return 0;
1111 
1112 err_eqpc:
1113 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1114 
1115 err_qpc:
1116 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1117 
1118 	return ret;
1119 }
1120 
1121 static void mthca_unmap_memfree(struct mthca_dev *dev,
1122 				struct mthca_qp *qp)
1123 {
1124 	mthca_table_put(dev, dev->qp_table.rdb_table,
1125 			qp->qpn << dev->qp_table.rdb_shift);
1126 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1127 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1128 }
1129 
1130 static int mthca_alloc_memfree(struct mthca_dev *dev,
1131 			       struct mthca_qp *qp)
1132 {
1133 	if (mthca_is_memfree(dev)) {
1134 		qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1135 						 qp->qpn, &qp->rq.db);
1136 		if (qp->rq.db_index < 0)
1137 			return -ENOMEM;
1138 
1139 		qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1140 						 qp->qpn, &qp->sq.db);
1141 		if (qp->sq.db_index < 0) {
1142 			mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1143 			return -ENOMEM;
1144 		}
1145 	}
1146 
1147 	return 0;
1148 }
1149 
1150 static void mthca_free_memfree(struct mthca_dev *dev,
1151 			       struct mthca_qp *qp)
1152 {
1153 	if (mthca_is_memfree(dev)) {
1154 		mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1155 		mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1156 	}
1157 }
1158 
1159 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1160 				 struct mthca_pd *pd,
1161 				 struct mthca_cq *send_cq,
1162 				 struct mthca_cq *recv_cq,
1163 				 enum ib_sig_type send_policy,
1164 				 struct mthca_qp *qp,
1165 				 struct ib_udata *udata)
1166 {
1167 	int ret;
1168 	int i;
1169 	struct mthca_next_seg *next;
1170 
1171 	qp->refcount = 1;
1172 	init_waitqueue_head(&qp->wait);
1173 	mutex_init(&qp->mutex);
1174 	qp->state    	 = IB_QPS_RESET;
1175 	qp->atomic_rd_en = 0;
1176 	qp->resp_depth   = 0;
1177 	qp->sq_policy    = send_policy;
1178 	mthca_wq_reset(&qp->sq);
1179 	mthca_wq_reset(&qp->rq);
1180 
1181 	spin_lock_init(&qp->sq.lock);
1182 	spin_lock_init(&qp->rq.lock);
1183 
1184 	ret = mthca_map_memfree(dev, qp);
1185 	if (ret)
1186 		return ret;
1187 
1188 	ret = mthca_alloc_wqe_buf(dev, pd, qp, udata);
1189 	if (ret) {
1190 		mthca_unmap_memfree(dev, qp);
1191 		return ret;
1192 	}
1193 
1194 	mthca_adjust_qp_caps(dev, pd, qp);
1195 
1196 	/*
1197 	 * If this is a userspace QP, we're done now.  The doorbells
1198 	 * will be allocated and buffers will be initialized in
1199 	 * userspace.
1200 	 */
1201 	if (udata)
1202 		return 0;
1203 
1204 	ret = mthca_alloc_memfree(dev, qp);
1205 	if (ret) {
1206 		mthca_free_wqe_buf(dev, qp);
1207 		mthca_unmap_memfree(dev, qp);
1208 		return ret;
1209 	}
1210 
1211 	if (mthca_is_memfree(dev)) {
1212 		struct mthca_data_seg *scatter;
1213 		int size = (sizeof (struct mthca_next_seg) +
1214 			    qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1215 
1216 		for (i = 0; i < qp->rq.max; ++i) {
1217 			next = get_recv_wqe(qp, i);
1218 			next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1219 						   qp->rq.wqe_shift);
1220 			next->ee_nds = cpu_to_be32(size);
1221 
1222 			for (scatter = (void *) (next + 1);
1223 			     (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1224 			     ++scatter)
1225 				scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1226 		}
1227 
1228 		for (i = 0; i < qp->sq.max; ++i) {
1229 			next = get_send_wqe(qp, i);
1230 			next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1231 						    qp->sq.wqe_shift) +
1232 						   qp->send_wqe_offset);
1233 		}
1234 	} else {
1235 		for (i = 0; i < qp->rq.max; ++i) {
1236 			next = get_recv_wqe(qp, i);
1237 			next->nda_op = htonl((((i + 1) % qp->rq.max) <<
1238 					      qp->rq.wqe_shift) | 1);
1239 		}
1240 
1241 	}
1242 
1243 	qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1244 	qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1245 
1246 	return 0;
1247 }
1248 
1249 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1250 			     struct mthca_pd *pd, struct mthca_qp *qp)
1251 {
1252 	int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1253 
1254 	/* Sanity check QP size before proceeding */
1255 	if (cap->max_send_wr  	 > dev->limits.max_wqes ||
1256 	    cap->max_recv_wr  	 > dev->limits.max_wqes ||
1257 	    cap->max_send_sge 	 > dev->limits.max_sg   ||
1258 	    cap->max_recv_sge 	 > dev->limits.max_sg   ||
1259 	    cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1260 		return -EINVAL;
1261 
1262 	/*
1263 	 * For MLX transport we need 2 extra send gather entries:
1264 	 * one for the header and one for the checksum at the end
1265 	 */
1266 	if (qp->transport == MLX && cap->max_send_sge + 2 > dev->limits.max_sg)
1267 		return -EINVAL;
1268 
1269 	if (mthca_is_memfree(dev)) {
1270 		qp->rq.max = cap->max_recv_wr ?
1271 			roundup_pow_of_two(cap->max_recv_wr) : 0;
1272 		qp->sq.max = cap->max_send_wr ?
1273 			roundup_pow_of_two(cap->max_send_wr) : 0;
1274 	} else {
1275 		qp->rq.max = cap->max_recv_wr;
1276 		qp->sq.max = cap->max_send_wr;
1277 	}
1278 
1279 	qp->rq.max_gs = cap->max_recv_sge;
1280 	qp->sq.max_gs = max_t(int, cap->max_send_sge,
1281 			      ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1282 				    MTHCA_INLINE_CHUNK_SIZE) /
1283 			      sizeof (struct mthca_data_seg));
1284 
1285 	return 0;
1286 }
1287 
1288 int mthca_alloc_qp(struct mthca_dev *dev,
1289 		   struct mthca_pd *pd,
1290 		   struct mthca_cq *send_cq,
1291 		   struct mthca_cq *recv_cq,
1292 		   enum ib_qp_type type,
1293 		   enum ib_sig_type send_policy,
1294 		   struct ib_qp_cap *cap,
1295 		   struct mthca_qp *qp,
1296 		   struct ib_udata *udata)
1297 {
1298 	int err;
1299 
1300 	switch (type) {
1301 	case IB_QPT_RC: qp->transport = RC; break;
1302 	case IB_QPT_UC: qp->transport = UC; break;
1303 	case IB_QPT_UD: qp->transport = UD; break;
1304 	default: return -EINVAL;
1305 	}
1306 
1307 	err = mthca_set_qp_size(dev, cap, pd, qp);
1308 	if (err)
1309 		return err;
1310 
1311 	qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1312 	if (qp->qpn == -1)
1313 		return -ENOMEM;
1314 
1315 	/* initialize port to zero for error-catching. */
1316 	qp->port = 0;
1317 
1318 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1319 				    send_policy, qp, udata);
1320 	if (err) {
1321 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1322 		return err;
1323 	}
1324 
1325 	spin_lock_irq(&dev->qp_table.lock);
1326 	mthca_array_set(&dev->qp_table.qp,
1327 			qp->qpn & (dev->limits.num_qps - 1), qp);
1328 	spin_unlock_irq(&dev->qp_table.lock);
1329 
1330 	return 0;
1331 }
1332 
1333 static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1334 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1335 {
1336 	if (send_cq == recv_cq) {
1337 		spin_lock_irq(&send_cq->lock);
1338 		__acquire(&recv_cq->lock);
1339 	} else if (send_cq->cqn < recv_cq->cqn) {
1340 		spin_lock_irq(&send_cq->lock);
1341 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1342 	} else {
1343 		spin_lock_irq(&recv_cq->lock);
1344 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1345 	}
1346 }
1347 
1348 static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1349 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1350 {
1351 	if (send_cq == recv_cq) {
1352 		__release(&recv_cq->lock);
1353 		spin_unlock_irq(&send_cq->lock);
1354 	} else if (send_cq->cqn < recv_cq->cqn) {
1355 		spin_unlock(&recv_cq->lock);
1356 		spin_unlock_irq(&send_cq->lock);
1357 	} else {
1358 		spin_unlock(&send_cq->lock);
1359 		spin_unlock_irq(&recv_cq->lock);
1360 	}
1361 }
1362 
1363 int mthca_alloc_sqp(struct mthca_dev *dev,
1364 		    struct mthca_pd *pd,
1365 		    struct mthca_cq *send_cq,
1366 		    struct mthca_cq *recv_cq,
1367 		    enum ib_sig_type send_policy,
1368 		    struct ib_qp_cap *cap,
1369 		    int qpn,
1370 		    int port,
1371 		    struct mthca_sqp *sqp,
1372 		    struct ib_udata *udata)
1373 {
1374 	u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1375 	int err;
1376 
1377 	sqp->qp.transport = MLX;
1378 	err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1379 	if (err)
1380 		return err;
1381 
1382 	sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1383 	sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1384 					     &sqp->header_dma, GFP_KERNEL);
1385 	if (!sqp->header_buf)
1386 		return -ENOMEM;
1387 
1388 	spin_lock_irq(&dev->qp_table.lock);
1389 	if (mthca_array_get(&dev->qp_table.qp, mqpn))
1390 		err = -EBUSY;
1391 	else
1392 		mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1393 	spin_unlock_irq(&dev->qp_table.lock);
1394 
1395 	if (err)
1396 		goto err_out;
1397 
1398 	sqp->qp.port      = port;
1399 	sqp->qp.qpn       = mqpn;
1400 	sqp->qp.transport = MLX;
1401 
1402 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1403 				    send_policy, &sqp->qp, udata);
1404 	if (err)
1405 		goto err_out_free;
1406 
1407 	atomic_inc(&pd->sqp_count);
1408 
1409 	return 0;
1410 
1411  err_out_free:
1412 	/*
1413 	 * Lock CQs here, so that CQ polling code can do QP lookup
1414 	 * without taking a lock.
1415 	 */
1416 	mthca_lock_cqs(send_cq, recv_cq);
1417 
1418 	spin_lock(&dev->qp_table.lock);
1419 	mthca_array_clear(&dev->qp_table.qp, mqpn);
1420 	spin_unlock(&dev->qp_table.lock);
1421 
1422 	mthca_unlock_cqs(send_cq, recv_cq);
1423 
1424  err_out:
1425 	dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1426 			  sqp->header_buf, sqp->header_dma);
1427 
1428 	return err;
1429 }
1430 
1431 static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp)
1432 {
1433 	int c;
1434 
1435 	spin_lock_irq(&dev->qp_table.lock);
1436 	c = qp->refcount;
1437 	spin_unlock_irq(&dev->qp_table.lock);
1438 
1439 	return c;
1440 }
1441 
1442 void mthca_free_qp(struct mthca_dev *dev,
1443 		   struct mthca_qp *qp)
1444 {
1445 	struct mthca_cq *send_cq;
1446 	struct mthca_cq *recv_cq;
1447 
1448 	send_cq = to_mcq(qp->ibqp.send_cq);
1449 	recv_cq = to_mcq(qp->ibqp.recv_cq);
1450 
1451 	/*
1452 	 * Lock CQs here, so that CQ polling code can do QP lookup
1453 	 * without taking a lock.
1454 	 */
1455 	mthca_lock_cqs(send_cq, recv_cq);
1456 
1457 	spin_lock(&dev->qp_table.lock);
1458 	mthca_array_clear(&dev->qp_table.qp,
1459 			  qp->qpn & (dev->limits.num_qps - 1));
1460 	--qp->refcount;
1461 	spin_unlock(&dev->qp_table.lock);
1462 
1463 	mthca_unlock_cqs(send_cq, recv_cq);
1464 
1465 	wait_event(qp->wait, !get_qp_refcount(dev, qp));
1466 
1467 	if (qp->state != IB_QPS_RESET)
1468 		mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1469 				NULL, 0);
1470 
1471 	/*
1472 	 * If this is a userspace QP, the buffers, MR, CQs and so on
1473 	 * will be cleaned up in userspace, so all we have to do is
1474 	 * unref the mem-free tables and free the QPN in our table.
1475 	 */
1476 	if (!qp->ibqp.uobject) {
1477 		mthca_cq_clean(dev, recv_cq, qp->qpn,
1478 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1479 		if (send_cq != recv_cq)
1480 			mthca_cq_clean(dev, send_cq, qp->qpn, NULL);
1481 
1482 		mthca_free_memfree(dev, qp);
1483 		mthca_free_wqe_buf(dev, qp);
1484 	}
1485 
1486 	mthca_unmap_memfree(dev, qp);
1487 
1488 	if (is_sqp(dev, qp)) {
1489 		atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1490 		dma_free_coherent(&dev->pdev->dev,
1491 				  to_msqp(qp)->header_buf_size,
1492 				  to_msqp(qp)->header_buf,
1493 				  to_msqp(qp)->header_dma);
1494 	} else
1495 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1496 }
1497 
1498 /* Create UD header for an MLX send and build a data segment for it */
1499 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1500 			    int ind, const struct ib_ud_wr *wr,
1501 			    struct mthca_mlx_seg *mlx,
1502 			    struct mthca_data_seg *data)
1503 {
1504 	int header_size;
1505 	int err;
1506 	u16 pkey;
1507 
1508 	ib_ud_header_init(256, /* assume a MAD */ 1, 0, 0,
1509 			  mthca_ah_grh_present(to_mah(wr->ah)), 0, 0, 0,
1510 			  &sqp->ud_header);
1511 
1512 	err = mthca_read_ah(dev, to_mah(wr->ah), &sqp->ud_header);
1513 	if (err)
1514 		return err;
1515 	mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1516 	mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1517 				  (sqp->ud_header.lrh.destination_lid ==
1518 				   IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1519 				  (sqp->ud_header.lrh.service_level << 8));
1520 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
1521 	mlx->vcrc = 0;
1522 
1523 	switch (wr->wr.opcode) {
1524 	case IB_WR_SEND:
1525 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1526 		sqp->ud_header.immediate_present = 0;
1527 		break;
1528 	case IB_WR_SEND_WITH_IMM:
1529 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1530 		sqp->ud_header.immediate_present = 1;
1531 		sqp->ud_header.immediate_data = wr->wr.ex.imm_data;
1532 		break;
1533 	default:
1534 		return -EINVAL;
1535 	}
1536 
1537 	sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1538 	if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1539 		sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1540 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
1541 	if (!sqp->qp.ibqp.qp_num)
1542 		ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1543 				   sqp->pkey_index, &pkey);
1544 	else
1545 		ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1546 				   wr->pkey_index, &pkey);
1547 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1548 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
1549 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1550 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
1551 					       sqp->qkey : wr->remote_qkey);
1552 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1553 
1554 	header_size = ib_ud_header_pack(&sqp->ud_header,
1555 					sqp->header_buf +
1556 					ind * MTHCA_UD_HEADER_SIZE);
1557 
1558 	data->byte_count = cpu_to_be32(header_size);
1559 	data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1560 	data->addr       = cpu_to_be64(sqp->header_dma +
1561 				       ind * MTHCA_UD_HEADER_SIZE);
1562 
1563 	return 0;
1564 }
1565 
1566 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1567 				    struct ib_cq *ib_cq)
1568 {
1569 	unsigned cur;
1570 	struct mthca_cq *cq;
1571 
1572 	cur = wq->head - wq->tail;
1573 	if (likely(cur + nreq < wq->max))
1574 		return 0;
1575 
1576 	cq = to_mcq(ib_cq);
1577 	spin_lock(&cq->lock);
1578 	cur = wq->head - wq->tail;
1579 	spin_unlock(&cq->lock);
1580 
1581 	return cur + nreq >= wq->max;
1582 }
1583 
1584 static __always_inline void set_raddr_seg(struct mthca_raddr_seg *rseg,
1585 					  u64 remote_addr, u32 rkey)
1586 {
1587 	rseg->raddr    = cpu_to_be64(remote_addr);
1588 	rseg->rkey     = cpu_to_be32(rkey);
1589 	rseg->reserved = 0;
1590 }
1591 
1592 static __always_inline void set_atomic_seg(struct mthca_atomic_seg *aseg,
1593 					   const struct ib_atomic_wr *wr)
1594 {
1595 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1596 		aseg->swap_add = cpu_to_be64(wr->swap);
1597 		aseg->compare  = cpu_to_be64(wr->compare_add);
1598 	} else {
1599 		aseg->swap_add = cpu_to_be64(wr->compare_add);
1600 		aseg->compare  = 0;
1601 	}
1602 
1603 }
1604 
1605 static void set_tavor_ud_seg(struct mthca_tavor_ud_seg *useg,
1606 			     const struct ib_ud_wr *wr)
1607 {
1608 	useg->lkey    = cpu_to_be32(to_mah(wr->ah)->key);
1609 	useg->av_addr =	cpu_to_be64(to_mah(wr->ah)->avdma);
1610 	useg->dqpn    =	cpu_to_be32(wr->remote_qpn);
1611 	useg->qkey    =	cpu_to_be32(wr->remote_qkey);
1612 
1613 }
1614 
1615 static void set_arbel_ud_seg(struct mthca_arbel_ud_seg *useg,
1616 			     const struct ib_ud_wr *wr)
1617 {
1618 	memcpy(useg->av, to_mah(wr->ah)->av, MTHCA_AV_SIZE);
1619 	useg->dqpn = cpu_to_be32(wr->remote_qpn);
1620 	useg->qkey = cpu_to_be32(wr->remote_qkey);
1621 }
1622 
1623 int mthca_tavor_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
1624 			  const struct ib_send_wr **bad_wr)
1625 {
1626 	struct mthca_dev *dev = to_mdev(ibqp->device);
1627 	struct mthca_qp *qp = to_mqp(ibqp);
1628 	void *wqe;
1629 	void *prev_wqe;
1630 	unsigned long flags;
1631 	int err = 0;
1632 	int nreq;
1633 	int i;
1634 	int size;
1635 	/*
1636 	 * f0 and size0 are only used if nreq != 0, and they will
1637 	 * always be initialized the first time through the main loop
1638 	 * before nreq is incremented.  So nreq cannot become non-zero
1639 	 * without initializing f0 and size0, and they are in fact
1640 	 * never used uninitialized.
1641 	 */
1642 	int uninitialized_var(size0);
1643 	u32 uninitialized_var(f0);
1644 	int ind;
1645 	u8 op0 = 0;
1646 
1647 	spin_lock_irqsave(&qp->sq.lock, flags);
1648 
1649 	/* XXX check that state is OK to post send */
1650 
1651 	ind = qp->sq.next_ind;
1652 
1653 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1654 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1655 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1656 					" %d max, %d nreq)\n", qp->qpn,
1657 					qp->sq.head, qp->sq.tail,
1658 					qp->sq.max, nreq);
1659 			err = -ENOMEM;
1660 			*bad_wr = wr;
1661 			goto out;
1662 		}
1663 
1664 		wqe = get_send_wqe(qp, ind);
1665 		prev_wqe = qp->sq.last;
1666 		qp->sq.last = wqe;
1667 
1668 		((struct mthca_next_seg *) wqe)->nda_op = 0;
1669 		((struct mthca_next_seg *) wqe)->ee_nds = 0;
1670 		((struct mthca_next_seg *) wqe)->flags =
1671 			((wr->send_flags & IB_SEND_SIGNALED) ?
1672 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1673 			((wr->send_flags & IB_SEND_SOLICITED) ?
1674 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1675 			cpu_to_be32(1);
1676 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1677 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1678 			((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
1679 
1680 		wqe += sizeof (struct mthca_next_seg);
1681 		size = sizeof (struct mthca_next_seg) / 16;
1682 
1683 		switch (qp->transport) {
1684 		case RC:
1685 			switch (wr->opcode) {
1686 			case IB_WR_ATOMIC_CMP_AND_SWP:
1687 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1688 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
1689 					      atomic_wr(wr)->rkey);
1690 				wqe += sizeof (struct mthca_raddr_seg);
1691 
1692 				set_atomic_seg(wqe, atomic_wr(wr));
1693 				wqe += sizeof (struct mthca_atomic_seg);
1694 				size += (sizeof (struct mthca_raddr_seg) +
1695 					 sizeof (struct mthca_atomic_seg)) / 16;
1696 				break;
1697 
1698 			case IB_WR_RDMA_WRITE:
1699 			case IB_WR_RDMA_WRITE_WITH_IMM:
1700 			case IB_WR_RDMA_READ:
1701 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
1702 					      rdma_wr(wr)->rkey);
1703 				wqe  += sizeof (struct mthca_raddr_seg);
1704 				size += sizeof (struct mthca_raddr_seg) / 16;
1705 				break;
1706 
1707 			default:
1708 				/* No extra segments required for sends */
1709 				break;
1710 			}
1711 
1712 			break;
1713 
1714 		case UC:
1715 			switch (wr->opcode) {
1716 			case IB_WR_RDMA_WRITE:
1717 			case IB_WR_RDMA_WRITE_WITH_IMM:
1718 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
1719 					      rdma_wr(wr)->rkey);
1720 				wqe  += sizeof (struct mthca_raddr_seg);
1721 				size += sizeof (struct mthca_raddr_seg) / 16;
1722 				break;
1723 
1724 			default:
1725 				/* No extra segments required for sends */
1726 				break;
1727 			}
1728 
1729 			break;
1730 
1731 		case UD:
1732 			set_tavor_ud_seg(wqe, ud_wr(wr));
1733 			wqe  += sizeof (struct mthca_tavor_ud_seg);
1734 			size += sizeof (struct mthca_tavor_ud_seg) / 16;
1735 			break;
1736 
1737 		case MLX:
1738 			err = build_mlx_header(dev, to_msqp(qp), ind, ud_wr(wr),
1739 					       wqe - sizeof (struct mthca_next_seg),
1740 					       wqe);
1741 			if (err) {
1742 				*bad_wr = wr;
1743 				goto out;
1744 			}
1745 			wqe += sizeof (struct mthca_data_seg);
1746 			size += sizeof (struct mthca_data_seg) / 16;
1747 			break;
1748 		}
1749 
1750 		if (wr->num_sge > qp->sq.max_gs) {
1751 			mthca_err(dev, "too many gathers\n");
1752 			err = -EINVAL;
1753 			*bad_wr = wr;
1754 			goto out;
1755 		}
1756 
1757 		for (i = 0; i < wr->num_sge; ++i) {
1758 			mthca_set_data_seg(wqe, wr->sg_list + i);
1759 			wqe  += sizeof (struct mthca_data_seg);
1760 			size += sizeof (struct mthca_data_seg) / 16;
1761 		}
1762 
1763 		/* Add one more inline data segment for ICRC */
1764 		if (qp->transport == MLX) {
1765 			((struct mthca_data_seg *) wqe)->byte_count =
1766 				cpu_to_be32((1 << 31) | 4);
1767 			((u32 *) wqe)[1] = 0;
1768 			wqe += sizeof (struct mthca_data_seg);
1769 			size += sizeof (struct mthca_data_seg) / 16;
1770 		}
1771 
1772 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
1773 
1774 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1775 			mthca_err(dev, "opcode invalid\n");
1776 			err = -EINVAL;
1777 			*bad_wr = wr;
1778 			goto out;
1779 		}
1780 
1781 		((struct mthca_next_seg *) prev_wqe)->nda_op =
1782 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
1783 				     qp->send_wqe_offset) |
1784 				    mthca_opcode[wr->opcode]);
1785 		wmb();
1786 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1787 			cpu_to_be32((nreq ? 0 : MTHCA_NEXT_DBD) | size |
1788 				    ((wr->send_flags & IB_SEND_FENCE) ?
1789 				    MTHCA_NEXT_FENCE : 0));
1790 
1791 		if (!nreq) {
1792 			size0 = size;
1793 			op0   = mthca_opcode[wr->opcode];
1794 			f0    = wr->send_flags & IB_SEND_FENCE ?
1795 				MTHCA_SEND_DOORBELL_FENCE : 0;
1796 		}
1797 
1798 		++ind;
1799 		if (unlikely(ind >= qp->sq.max))
1800 			ind -= qp->sq.max;
1801 	}
1802 
1803 out:
1804 	if (likely(nreq)) {
1805 		wmb();
1806 
1807 		mthca_write64(((qp->sq.next_ind << qp->sq.wqe_shift) +
1808 			       qp->send_wqe_offset) | f0 | op0,
1809 			      (qp->qpn << 8) | size0,
1810 			      dev->kar + MTHCA_SEND_DOORBELL,
1811 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1812 		/*
1813 		 * Make sure doorbells don't leak out of SQ spinlock
1814 		 * and reach the HCA out of order:
1815 		 */
1816 		mmiowb();
1817 	}
1818 
1819 	qp->sq.next_ind = ind;
1820 	qp->sq.head    += nreq;
1821 
1822 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1823 	return err;
1824 }
1825 
1826 int mthca_tavor_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
1827 			     const struct ib_recv_wr **bad_wr)
1828 {
1829 	struct mthca_dev *dev = to_mdev(ibqp->device);
1830 	struct mthca_qp *qp = to_mqp(ibqp);
1831 	unsigned long flags;
1832 	int err = 0;
1833 	int nreq;
1834 	int i;
1835 	int size;
1836 	/*
1837 	 * size0 is only used if nreq != 0, and it will always be
1838 	 * initialized the first time through the main loop before
1839 	 * nreq is incremented.  So nreq cannot become non-zero
1840 	 * without initializing size0, and it is in fact never used
1841 	 * uninitialized.
1842 	 */
1843 	int uninitialized_var(size0);
1844 	int ind;
1845 	void *wqe;
1846 	void *prev_wqe;
1847 
1848 	spin_lock_irqsave(&qp->rq.lock, flags);
1849 
1850 	/* XXX check that state is OK to post receive */
1851 
1852 	ind = qp->rq.next_ind;
1853 
1854 	for (nreq = 0; wr; wr = wr->next) {
1855 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1856 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1857 					" %d max, %d nreq)\n", qp->qpn,
1858 					qp->rq.head, qp->rq.tail,
1859 					qp->rq.max, nreq);
1860 			err = -ENOMEM;
1861 			*bad_wr = wr;
1862 			goto out;
1863 		}
1864 
1865 		wqe = get_recv_wqe(qp, ind);
1866 		prev_wqe = qp->rq.last;
1867 		qp->rq.last = wqe;
1868 
1869 		((struct mthca_next_seg *) wqe)->ee_nds =
1870 			cpu_to_be32(MTHCA_NEXT_DBD);
1871 		((struct mthca_next_seg *) wqe)->flags = 0;
1872 
1873 		wqe += sizeof (struct mthca_next_seg);
1874 		size = sizeof (struct mthca_next_seg) / 16;
1875 
1876 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1877 			err = -EINVAL;
1878 			*bad_wr = wr;
1879 			goto out;
1880 		}
1881 
1882 		for (i = 0; i < wr->num_sge; ++i) {
1883 			mthca_set_data_seg(wqe, wr->sg_list + i);
1884 			wqe  += sizeof (struct mthca_data_seg);
1885 			size += sizeof (struct mthca_data_seg) / 16;
1886 		}
1887 
1888 		qp->wrid[ind] = wr->wr_id;
1889 
1890 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1891 			cpu_to_be32(MTHCA_NEXT_DBD | size);
1892 
1893 		if (!nreq)
1894 			size0 = size;
1895 
1896 		++ind;
1897 		if (unlikely(ind >= qp->rq.max))
1898 			ind -= qp->rq.max;
1899 
1900 		++nreq;
1901 		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1902 			nreq = 0;
1903 
1904 			wmb();
1905 
1906 			mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1907 				      qp->qpn << 8, dev->kar + MTHCA_RECEIVE_DOORBELL,
1908 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1909 
1910 			qp->rq.next_ind = ind;
1911 			qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1912 		}
1913 	}
1914 
1915 out:
1916 	if (likely(nreq)) {
1917 		wmb();
1918 
1919 		mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1920 			      qp->qpn << 8 | nreq, dev->kar + MTHCA_RECEIVE_DOORBELL,
1921 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1922 	}
1923 
1924 	qp->rq.next_ind = ind;
1925 	qp->rq.head    += nreq;
1926 
1927 	/*
1928 	 * Make sure doorbells don't leak out of RQ spinlock and reach
1929 	 * the HCA out of order:
1930 	 */
1931 	mmiowb();
1932 
1933 	spin_unlock_irqrestore(&qp->rq.lock, flags);
1934 	return err;
1935 }
1936 
1937 int mthca_arbel_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
1938 			  const struct ib_send_wr **bad_wr)
1939 {
1940 	struct mthca_dev *dev = to_mdev(ibqp->device);
1941 	struct mthca_qp *qp = to_mqp(ibqp);
1942 	u32 dbhi;
1943 	void *wqe;
1944 	void *prev_wqe;
1945 	unsigned long flags;
1946 	int err = 0;
1947 	int nreq;
1948 	int i;
1949 	int size;
1950 	/*
1951 	 * f0 and size0 are only used if nreq != 0, and they will
1952 	 * always be initialized the first time through the main loop
1953 	 * before nreq is incremented.  So nreq cannot become non-zero
1954 	 * without initializing f0 and size0, and they are in fact
1955 	 * never used uninitialized.
1956 	 */
1957 	int uninitialized_var(size0);
1958 	u32 uninitialized_var(f0);
1959 	int ind;
1960 	u8 op0 = 0;
1961 
1962 	spin_lock_irqsave(&qp->sq.lock, flags);
1963 
1964 	/* XXX check that state is OK to post send */
1965 
1966 	ind = qp->sq.head & (qp->sq.max - 1);
1967 
1968 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1969 		if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1970 			nreq = 0;
1971 
1972 			dbhi = (MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1973 				((qp->sq.head & 0xffff) << 8) | f0 | op0;
1974 
1975 			qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1976 
1977 			/*
1978 			 * Make sure that descriptors are written before
1979 			 * doorbell record.
1980 			 */
1981 			wmb();
1982 			*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1983 
1984 			/*
1985 			 * Make sure doorbell record is written before we
1986 			 * write MMIO send doorbell.
1987 			 */
1988 			wmb();
1989 
1990 			mthca_write64(dbhi, (qp->qpn << 8) | size0,
1991 				      dev->kar + MTHCA_SEND_DOORBELL,
1992 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1993 		}
1994 
1995 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1996 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1997 					" %d max, %d nreq)\n", qp->qpn,
1998 					qp->sq.head, qp->sq.tail,
1999 					qp->sq.max, nreq);
2000 			err = -ENOMEM;
2001 			*bad_wr = wr;
2002 			goto out;
2003 		}
2004 
2005 		wqe = get_send_wqe(qp, ind);
2006 		prev_wqe = qp->sq.last;
2007 		qp->sq.last = wqe;
2008 
2009 		((struct mthca_next_seg *) wqe)->flags =
2010 			((wr->send_flags & IB_SEND_SIGNALED) ?
2011 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
2012 			((wr->send_flags & IB_SEND_SOLICITED) ?
2013 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
2014 			((wr->send_flags & IB_SEND_IP_CSUM) ?
2015 			 cpu_to_be32(MTHCA_NEXT_IP_CSUM | MTHCA_NEXT_TCP_UDP_CSUM) : 0) |
2016 			cpu_to_be32(1);
2017 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
2018 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
2019 			((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
2020 
2021 		wqe += sizeof (struct mthca_next_seg);
2022 		size = sizeof (struct mthca_next_seg) / 16;
2023 
2024 		switch (qp->transport) {
2025 		case RC:
2026 			switch (wr->opcode) {
2027 			case IB_WR_ATOMIC_CMP_AND_SWP:
2028 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2029 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
2030 					      atomic_wr(wr)->rkey);
2031 				wqe += sizeof (struct mthca_raddr_seg);
2032 
2033 				set_atomic_seg(wqe, atomic_wr(wr));
2034 				wqe  += sizeof (struct mthca_atomic_seg);
2035 				size += (sizeof (struct mthca_raddr_seg) +
2036 					 sizeof (struct mthca_atomic_seg)) / 16;
2037 				break;
2038 
2039 			case IB_WR_RDMA_READ:
2040 			case IB_WR_RDMA_WRITE:
2041 			case IB_WR_RDMA_WRITE_WITH_IMM:
2042 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
2043 					      rdma_wr(wr)->rkey);
2044 				wqe  += sizeof (struct mthca_raddr_seg);
2045 				size += sizeof (struct mthca_raddr_seg) / 16;
2046 				break;
2047 
2048 			default:
2049 				/* No extra segments required for sends */
2050 				break;
2051 			}
2052 
2053 			break;
2054 
2055 		case UC:
2056 			switch (wr->opcode) {
2057 			case IB_WR_RDMA_WRITE:
2058 			case IB_WR_RDMA_WRITE_WITH_IMM:
2059 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
2060 					      rdma_wr(wr)->rkey);
2061 				wqe  += sizeof (struct mthca_raddr_seg);
2062 				size += sizeof (struct mthca_raddr_seg) / 16;
2063 				break;
2064 
2065 			default:
2066 				/* No extra segments required for sends */
2067 				break;
2068 			}
2069 
2070 			break;
2071 
2072 		case UD:
2073 			set_arbel_ud_seg(wqe, ud_wr(wr));
2074 			wqe  += sizeof (struct mthca_arbel_ud_seg);
2075 			size += sizeof (struct mthca_arbel_ud_seg) / 16;
2076 			break;
2077 
2078 		case MLX:
2079 			err = build_mlx_header(dev, to_msqp(qp), ind, ud_wr(wr),
2080 					       wqe - sizeof (struct mthca_next_seg),
2081 					       wqe);
2082 			if (err) {
2083 				*bad_wr = wr;
2084 				goto out;
2085 			}
2086 			wqe += sizeof (struct mthca_data_seg);
2087 			size += sizeof (struct mthca_data_seg) / 16;
2088 			break;
2089 		}
2090 
2091 		if (wr->num_sge > qp->sq.max_gs) {
2092 			mthca_err(dev, "too many gathers\n");
2093 			err = -EINVAL;
2094 			*bad_wr = wr;
2095 			goto out;
2096 		}
2097 
2098 		for (i = 0; i < wr->num_sge; ++i) {
2099 			mthca_set_data_seg(wqe, wr->sg_list + i);
2100 			wqe  += sizeof (struct mthca_data_seg);
2101 			size += sizeof (struct mthca_data_seg) / 16;
2102 		}
2103 
2104 		/* Add one more inline data segment for ICRC */
2105 		if (qp->transport == MLX) {
2106 			((struct mthca_data_seg *) wqe)->byte_count =
2107 				cpu_to_be32((1 << 31) | 4);
2108 			((u32 *) wqe)[1] = 0;
2109 			wqe += sizeof (struct mthca_data_seg);
2110 			size += sizeof (struct mthca_data_seg) / 16;
2111 		}
2112 
2113 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
2114 
2115 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2116 			mthca_err(dev, "opcode invalid\n");
2117 			err = -EINVAL;
2118 			*bad_wr = wr;
2119 			goto out;
2120 		}
2121 
2122 		((struct mthca_next_seg *) prev_wqe)->nda_op =
2123 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
2124 				     qp->send_wqe_offset) |
2125 				    mthca_opcode[wr->opcode]);
2126 		wmb();
2127 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
2128 			cpu_to_be32(MTHCA_NEXT_DBD | size |
2129 				    ((wr->send_flags & IB_SEND_FENCE) ?
2130 				     MTHCA_NEXT_FENCE : 0));
2131 
2132 		if (!nreq) {
2133 			size0 = size;
2134 			op0   = mthca_opcode[wr->opcode];
2135 			f0    = wr->send_flags & IB_SEND_FENCE ?
2136 				MTHCA_SEND_DOORBELL_FENCE : 0;
2137 		}
2138 
2139 		++ind;
2140 		if (unlikely(ind >= qp->sq.max))
2141 			ind -= qp->sq.max;
2142 	}
2143 
2144 out:
2145 	if (likely(nreq)) {
2146 		dbhi = (nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0;
2147 
2148 		qp->sq.head += nreq;
2149 
2150 		/*
2151 		 * Make sure that descriptors are written before
2152 		 * doorbell record.
2153 		 */
2154 		wmb();
2155 		*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2156 
2157 		/*
2158 		 * Make sure doorbell record is written before we
2159 		 * write MMIO send doorbell.
2160 		 */
2161 		wmb();
2162 
2163 		mthca_write64(dbhi, (qp->qpn << 8) | size0, dev->kar + MTHCA_SEND_DOORBELL,
2164 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2165 	}
2166 
2167 	/*
2168 	 * Make sure doorbells don't leak out of SQ spinlock and reach
2169 	 * the HCA out of order:
2170 	 */
2171 	mmiowb();
2172 
2173 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2174 	return err;
2175 }
2176 
2177 int mthca_arbel_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
2178 			     const struct ib_recv_wr **bad_wr)
2179 {
2180 	struct mthca_dev *dev = to_mdev(ibqp->device);
2181 	struct mthca_qp *qp = to_mqp(ibqp);
2182 	unsigned long flags;
2183 	int err = 0;
2184 	int nreq;
2185 	int ind;
2186 	int i;
2187 	void *wqe;
2188 
2189 	spin_lock_irqsave(&qp->rq.lock, flags);
2190 
2191 	/* XXX check that state is OK to post receive */
2192 
2193 	ind = qp->rq.head & (qp->rq.max - 1);
2194 
2195 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2196 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2197 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2198 					" %d max, %d nreq)\n", qp->qpn,
2199 					qp->rq.head, qp->rq.tail,
2200 					qp->rq.max, nreq);
2201 			err = -ENOMEM;
2202 			*bad_wr = wr;
2203 			goto out;
2204 		}
2205 
2206 		wqe = get_recv_wqe(qp, ind);
2207 
2208 		((struct mthca_next_seg *) wqe)->flags = 0;
2209 
2210 		wqe += sizeof (struct mthca_next_seg);
2211 
2212 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2213 			err = -EINVAL;
2214 			*bad_wr = wr;
2215 			goto out;
2216 		}
2217 
2218 		for (i = 0; i < wr->num_sge; ++i) {
2219 			mthca_set_data_seg(wqe, wr->sg_list + i);
2220 			wqe += sizeof (struct mthca_data_seg);
2221 		}
2222 
2223 		if (i < qp->rq.max_gs)
2224 			mthca_set_data_seg_inval(wqe);
2225 
2226 		qp->wrid[ind] = wr->wr_id;
2227 
2228 		++ind;
2229 		if (unlikely(ind >= qp->rq.max))
2230 			ind -= qp->rq.max;
2231 	}
2232 out:
2233 	if (likely(nreq)) {
2234 		qp->rq.head += nreq;
2235 
2236 		/*
2237 		 * Make sure that descriptors are written before
2238 		 * doorbell record.
2239 		 */
2240 		wmb();
2241 		*qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2242 	}
2243 
2244 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2245 	return err;
2246 }
2247 
2248 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2249 			int index, int *dbd, __be32 *new_wqe)
2250 {
2251 	struct mthca_next_seg *next;
2252 
2253 	/*
2254 	 * For SRQs, all receive WQEs generate a CQE, so we're always
2255 	 * at the end of the doorbell chain.
2256 	 */
2257 	if (qp->ibqp.srq && !is_send) {
2258 		*new_wqe = 0;
2259 		return;
2260 	}
2261 
2262 	if (is_send)
2263 		next = get_send_wqe(qp, index);
2264 	else
2265 		next = get_recv_wqe(qp, index);
2266 
2267 	*dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2268 	if (next->ee_nds & cpu_to_be32(0x3f))
2269 		*new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2270 			(next->ee_nds & cpu_to_be32(0x3f));
2271 	else
2272 		*new_wqe = 0;
2273 }
2274 
2275 int mthca_init_qp_table(struct mthca_dev *dev)
2276 {
2277 	int err;
2278 	int i;
2279 
2280 	spin_lock_init(&dev->qp_table.lock);
2281 
2282 	/*
2283 	 * We reserve 2 extra QPs per port for the special QPs.  The
2284 	 * special QP for port 1 has to be even, so round up.
2285 	 */
2286 	dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2287 	err = mthca_alloc_init(&dev->qp_table.alloc,
2288 			       dev->limits.num_qps,
2289 			       (1 << 24) - 1,
2290 			       dev->qp_table.sqp_start +
2291 			       MTHCA_MAX_PORTS * 2);
2292 	if (err)
2293 		return err;
2294 
2295 	err = mthca_array_init(&dev->qp_table.qp,
2296 			       dev->limits.num_qps);
2297 	if (err) {
2298 		mthca_alloc_cleanup(&dev->qp_table.alloc);
2299 		return err;
2300 	}
2301 
2302 	for (i = 0; i < 2; ++i) {
2303 		err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2304 				    dev->qp_table.sqp_start + i * 2);
2305 		if (err) {
2306 			mthca_warn(dev, "CONF_SPECIAL_QP returned "
2307 				   "%d, aborting.\n", err);
2308 			goto err_out;
2309 		}
2310 	}
2311 	return 0;
2312 
2313  err_out:
2314 	for (i = 0; i < 2; ++i)
2315 		mthca_CONF_SPECIAL_QP(dev, i, 0);
2316 
2317 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2318 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2319 
2320 	return err;
2321 }
2322 
2323 void mthca_cleanup_qp_table(struct mthca_dev *dev)
2324 {
2325 	int i;
2326 
2327 	for (i = 0; i < 2; ++i)
2328 		mthca_CONF_SPECIAL_QP(dev, i, 0);
2329 
2330 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2331 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2332 }
2333