xref: /linux/drivers/infiniband/hw/mthca/mthca_qp.c (revision de2fe5e07d58424bc286fff3fd3c1b0bf933cd58)
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  * $Id: mthca_qp.c 1355 2004-12-17 15:23:43Z roland $
36  */
37 
38 #include <linux/init.h>
39 #include <linux/string.h>
40 #include <linux/slab.h>
41 
42 #include <rdma/ib_verbs.h>
43 #include <rdma/ib_cache.h>
44 #include <rdma/ib_pack.h>
45 
46 #include "mthca_dev.h"
47 #include "mthca_cmd.h"
48 #include "mthca_memfree.h"
49 #include "mthca_wqe.h"
50 
51 enum {
52 	MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
53 	MTHCA_ACK_REQ_FREQ       = 10,
54 	MTHCA_FLIGHT_LIMIT       = 9,
55 	MTHCA_UD_HEADER_SIZE     = 72, /* largest UD header possible */
56 	MTHCA_INLINE_HEADER_SIZE = 4,  /* data segment overhead for inline */
57 	MTHCA_INLINE_CHUNK_SIZE  = 16  /* inline data segment chunk */
58 };
59 
60 enum {
61 	MTHCA_QP_STATE_RST  = 0,
62 	MTHCA_QP_STATE_INIT = 1,
63 	MTHCA_QP_STATE_RTR  = 2,
64 	MTHCA_QP_STATE_RTS  = 3,
65 	MTHCA_QP_STATE_SQE  = 4,
66 	MTHCA_QP_STATE_SQD  = 5,
67 	MTHCA_QP_STATE_ERR  = 6,
68 	MTHCA_QP_STATE_DRAINING = 7
69 };
70 
71 enum {
72 	MTHCA_QP_ST_RC 	= 0x0,
73 	MTHCA_QP_ST_UC 	= 0x1,
74 	MTHCA_QP_ST_RD 	= 0x2,
75 	MTHCA_QP_ST_UD 	= 0x3,
76 	MTHCA_QP_ST_MLX = 0x7
77 };
78 
79 enum {
80 	MTHCA_QP_PM_MIGRATED = 0x3,
81 	MTHCA_QP_PM_ARMED    = 0x0,
82 	MTHCA_QP_PM_REARM    = 0x1
83 };
84 
85 enum {
86 	/* qp_context flags */
87 	MTHCA_QP_BIT_DE  = 1 <<  8,
88 	/* params1 */
89 	MTHCA_QP_BIT_SRE = 1 << 15,
90 	MTHCA_QP_BIT_SWE = 1 << 14,
91 	MTHCA_QP_BIT_SAE = 1 << 13,
92 	MTHCA_QP_BIT_SIC = 1 <<  4,
93 	MTHCA_QP_BIT_SSC = 1 <<  3,
94 	/* params2 */
95 	MTHCA_QP_BIT_RRE = 1 << 15,
96 	MTHCA_QP_BIT_RWE = 1 << 14,
97 	MTHCA_QP_BIT_RAE = 1 << 13,
98 	MTHCA_QP_BIT_RIC = 1 <<  4,
99 	MTHCA_QP_BIT_RSC = 1 <<  3
100 };
101 
102 struct mthca_qp_path {
103 	__be32 port_pkey;
104 	u8     rnr_retry;
105 	u8     g_mylmc;
106 	__be16 rlid;
107 	u8     ackto;
108 	u8     mgid_index;
109 	u8     static_rate;
110 	u8     hop_limit;
111 	__be32 sl_tclass_flowlabel;
112 	u8     rgid[16];
113 } __attribute__((packed));
114 
115 struct mthca_qp_context {
116 	__be32 flags;
117 	__be32 tavor_sched_queue; /* Reserved on Arbel */
118 	u8     mtu_msgmax;
119 	u8     rq_size_stride;	/* Reserved on Tavor */
120 	u8     sq_size_stride;	/* Reserved on Tavor */
121 	u8     rlkey_arbel_sched_queue;	/* Reserved on Tavor */
122 	__be32 usr_page;
123 	__be32 local_qpn;
124 	__be32 remote_qpn;
125 	u32    reserved1[2];
126 	struct mthca_qp_path pri_path;
127 	struct mthca_qp_path alt_path;
128 	__be32 rdd;
129 	__be32 pd;
130 	__be32 wqe_base;
131 	__be32 wqe_lkey;
132 	__be32 params1;
133 	__be32 reserved2;
134 	__be32 next_send_psn;
135 	__be32 cqn_snd;
136 	__be32 snd_wqe_base_l;	/* Next send WQE on Tavor */
137 	__be32 snd_db_index;	/* (debugging only entries) */
138 	__be32 last_acked_psn;
139 	__be32 ssn;
140 	__be32 params2;
141 	__be32 rnr_nextrecvpsn;
142 	__be32 ra_buff_indx;
143 	__be32 cqn_rcv;
144 	__be32 rcv_wqe_base_l;	/* Next recv WQE on Tavor */
145 	__be32 rcv_db_index;	/* (debugging only entries) */
146 	__be32 qkey;
147 	__be32 srqn;
148 	__be32 rmsn;
149 	__be16 rq_wqe_counter;	/* reserved on Tavor */
150 	__be16 sq_wqe_counter;	/* reserved on Tavor */
151 	u32    reserved3[18];
152 } __attribute__((packed));
153 
154 struct mthca_qp_param {
155 	__be32 opt_param_mask;
156 	u32    reserved1;
157 	struct mthca_qp_context context;
158 	u32    reserved2[62];
159 } __attribute__((packed));
160 
161 enum {
162 	MTHCA_QP_OPTPAR_ALT_ADDR_PATH     = 1 << 0,
163 	MTHCA_QP_OPTPAR_RRE               = 1 << 1,
164 	MTHCA_QP_OPTPAR_RAE               = 1 << 2,
165 	MTHCA_QP_OPTPAR_RWE               = 1 << 3,
166 	MTHCA_QP_OPTPAR_PKEY_INDEX        = 1 << 4,
167 	MTHCA_QP_OPTPAR_Q_KEY             = 1 << 5,
168 	MTHCA_QP_OPTPAR_RNR_TIMEOUT       = 1 << 6,
169 	MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
170 	MTHCA_QP_OPTPAR_SRA_MAX           = 1 << 8,
171 	MTHCA_QP_OPTPAR_RRA_MAX           = 1 << 9,
172 	MTHCA_QP_OPTPAR_PM_STATE          = 1 << 10,
173 	MTHCA_QP_OPTPAR_PORT_NUM          = 1 << 11,
174 	MTHCA_QP_OPTPAR_RETRY_COUNT       = 1 << 12,
175 	MTHCA_QP_OPTPAR_ALT_RNR_RETRY     = 1 << 13,
176 	MTHCA_QP_OPTPAR_ACK_TIMEOUT       = 1 << 14,
177 	MTHCA_QP_OPTPAR_RNR_RETRY         = 1 << 15,
178 	MTHCA_QP_OPTPAR_SCHED_QUEUE       = 1 << 16
179 };
180 
181 static const u8 mthca_opcode[] = {
182 	[IB_WR_SEND]                 = MTHCA_OPCODE_SEND,
183 	[IB_WR_SEND_WITH_IMM]        = MTHCA_OPCODE_SEND_IMM,
184 	[IB_WR_RDMA_WRITE]           = MTHCA_OPCODE_RDMA_WRITE,
185 	[IB_WR_RDMA_WRITE_WITH_IMM]  = MTHCA_OPCODE_RDMA_WRITE_IMM,
186 	[IB_WR_RDMA_READ]            = MTHCA_OPCODE_RDMA_READ,
187 	[IB_WR_ATOMIC_CMP_AND_SWP]   = MTHCA_OPCODE_ATOMIC_CS,
188 	[IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
189 };
190 
191 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
192 {
193 	return qp->qpn >= dev->qp_table.sqp_start &&
194 		qp->qpn <= dev->qp_table.sqp_start + 3;
195 }
196 
197 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
198 {
199 	return qp->qpn >= dev->qp_table.sqp_start &&
200 		qp->qpn <= dev->qp_table.sqp_start + 1;
201 }
202 
203 static void *get_recv_wqe(struct mthca_qp *qp, int n)
204 {
205 	if (qp->is_direct)
206 		return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
207 	else
208 		return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
209 			((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
210 }
211 
212 static void *get_send_wqe(struct mthca_qp *qp, int n)
213 {
214 	if (qp->is_direct)
215 		return qp->queue.direct.buf + qp->send_wqe_offset +
216 			(n << qp->sq.wqe_shift);
217 	else
218 		return qp->queue.page_list[(qp->send_wqe_offset +
219 					    (n << qp->sq.wqe_shift)) >>
220 					   PAGE_SHIFT].buf +
221 			((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
222 			 (PAGE_SIZE - 1));
223 }
224 
225 static void mthca_wq_init(struct mthca_wq *wq)
226 {
227 	spin_lock_init(&wq->lock);
228 	wq->next_ind  = 0;
229 	wq->last_comp = wq->max - 1;
230 	wq->head      = 0;
231 	wq->tail      = 0;
232 }
233 
234 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
235 		    enum ib_event_type event_type)
236 {
237 	struct mthca_qp *qp;
238 	struct ib_event event;
239 
240 	spin_lock(&dev->qp_table.lock);
241 	qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
242 	if (qp)
243 		atomic_inc(&qp->refcount);
244 	spin_unlock(&dev->qp_table.lock);
245 
246 	if (!qp) {
247 		mthca_warn(dev, "Async event for bogus QP %08x\n", qpn);
248 		return;
249 	}
250 
251 	event.device      = &dev->ib_dev;
252 	event.event       = event_type;
253 	event.element.qp  = &qp->ibqp;
254 	if (qp->ibqp.event_handler)
255 		qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
256 
257 	if (atomic_dec_and_test(&qp->refcount))
258 		wake_up(&qp->wait);
259 }
260 
261 static int to_mthca_state(enum ib_qp_state ib_state)
262 {
263 	switch (ib_state) {
264 	case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
265 	case IB_QPS_INIT:  return MTHCA_QP_STATE_INIT;
266 	case IB_QPS_RTR:   return MTHCA_QP_STATE_RTR;
267 	case IB_QPS_RTS:   return MTHCA_QP_STATE_RTS;
268 	case IB_QPS_SQD:   return MTHCA_QP_STATE_SQD;
269 	case IB_QPS_SQE:   return MTHCA_QP_STATE_SQE;
270 	case IB_QPS_ERR:   return MTHCA_QP_STATE_ERR;
271 	default:                return -1;
272 	}
273 }
274 
275 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
276 
277 static int to_mthca_st(int transport)
278 {
279 	switch (transport) {
280 	case RC:  return MTHCA_QP_ST_RC;
281 	case UC:  return MTHCA_QP_ST_UC;
282 	case UD:  return MTHCA_QP_ST_UD;
283 	case RD:  return MTHCA_QP_ST_RD;
284 	case MLX: return MTHCA_QP_ST_MLX;
285 	default:  return -1;
286 	}
287 }
288 
289 static void store_attrs(struct mthca_sqp *sqp, struct ib_qp_attr *attr,
290 			int attr_mask)
291 {
292 	if (attr_mask & IB_QP_PKEY_INDEX)
293 		sqp->pkey_index = attr->pkey_index;
294 	if (attr_mask & IB_QP_QKEY)
295 		sqp->qkey = attr->qkey;
296 	if (attr_mask & IB_QP_SQ_PSN)
297 		sqp->send_psn = attr->sq_psn;
298 }
299 
300 static void init_port(struct mthca_dev *dev, int port)
301 {
302 	int err;
303 	u8 status;
304 	struct mthca_init_ib_param param;
305 
306 	memset(&param, 0, sizeof param);
307 
308 	param.port_width = dev->limits.port_width_cap;
309 	param.vl_cap     = dev->limits.vl_cap;
310 	param.mtu_cap    = dev->limits.mtu_cap;
311 	param.gid_cap    = dev->limits.gid_table_len;
312 	param.pkey_cap   = dev->limits.pkey_table_len;
313 
314 	err = mthca_INIT_IB(dev, &param, port, &status);
315 	if (err)
316 		mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
317 	if (status)
318 		mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
319 }
320 
321 static __be32 get_hw_access_flags(struct mthca_qp *qp, struct ib_qp_attr *attr,
322 				  int attr_mask)
323 {
324 	u8 dest_rd_atomic;
325 	u32 access_flags;
326 	u32 hw_access_flags = 0;
327 
328 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
329 		dest_rd_atomic = attr->max_dest_rd_atomic;
330 	else
331 		dest_rd_atomic = qp->resp_depth;
332 
333 	if (attr_mask & IB_QP_ACCESS_FLAGS)
334 		access_flags = attr->qp_access_flags;
335 	else
336 		access_flags = qp->atomic_rd_en;
337 
338 	if (!dest_rd_atomic)
339 		access_flags &= IB_ACCESS_REMOTE_WRITE;
340 
341 	if (access_flags & IB_ACCESS_REMOTE_READ)
342 		hw_access_flags |= MTHCA_QP_BIT_RRE;
343 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
344 		hw_access_flags |= MTHCA_QP_BIT_RAE;
345 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
346 		hw_access_flags |= MTHCA_QP_BIT_RWE;
347 
348 	return cpu_to_be32(hw_access_flags);
349 }
350 
351 static inline enum ib_qp_state to_ib_qp_state(int mthca_state)
352 {
353 	switch (mthca_state) {
354 	case MTHCA_QP_STATE_RST:      return IB_QPS_RESET;
355 	case MTHCA_QP_STATE_INIT:     return IB_QPS_INIT;
356 	case MTHCA_QP_STATE_RTR:      return IB_QPS_RTR;
357 	case MTHCA_QP_STATE_RTS:      return IB_QPS_RTS;
358 	case MTHCA_QP_STATE_DRAINING:
359 	case MTHCA_QP_STATE_SQD:      return IB_QPS_SQD;
360 	case MTHCA_QP_STATE_SQE:      return IB_QPS_SQE;
361 	case MTHCA_QP_STATE_ERR:      return IB_QPS_ERR;
362 	default:                      return -1;
363 	}
364 }
365 
366 static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state)
367 {
368 	switch (mthca_mig_state) {
369 	case 0:  return IB_MIG_ARMED;
370 	case 1:  return IB_MIG_REARM;
371 	case 3:  return IB_MIG_MIGRATED;
372 	default: return -1;
373 	}
374 }
375 
376 static int to_ib_qp_access_flags(int mthca_flags)
377 {
378 	int ib_flags = 0;
379 
380 	if (mthca_flags & MTHCA_QP_BIT_RRE)
381 		ib_flags |= IB_ACCESS_REMOTE_READ;
382 	if (mthca_flags & MTHCA_QP_BIT_RWE)
383 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
384 	if (mthca_flags & MTHCA_QP_BIT_RAE)
385 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
386 
387 	return ib_flags;
388 }
389 
390 static void to_ib_ah_attr(struct mthca_dev *dev, struct ib_ah_attr *ib_ah_attr,
391 				struct mthca_qp_path *path)
392 {
393 	memset(ib_ah_attr, 0, sizeof *path);
394 	ib_ah_attr->port_num 	  = (be32_to_cpu(path->port_pkey) >> 24) & 0x3;
395 	ib_ah_attr->dlid     	  = be16_to_cpu(path->rlid);
396 	ib_ah_attr->sl       	  = be32_to_cpu(path->sl_tclass_flowlabel) >> 28;
397 	ib_ah_attr->src_path_bits = path->g_mylmc & 0x7f;
398 	ib_ah_attr->static_rate   = path->static_rate & 0x7;
399 	ib_ah_attr->ah_flags      = (path->g_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
400 	if (ib_ah_attr->ah_flags) {
401 		ib_ah_attr->grh.sgid_index = path->mgid_index & (dev->limits.gid_table_len - 1);
402 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
403 		ib_ah_attr->grh.traffic_class =
404 			(be32_to_cpu(path->sl_tclass_flowlabel) >> 20) & 0xff;
405 		ib_ah_attr->grh.flow_label =
406 			be32_to_cpu(path->sl_tclass_flowlabel) & 0xfffff;
407 		memcpy(ib_ah_attr->grh.dgid.raw,
408 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
409 	}
410 }
411 
412 int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
413 		   struct ib_qp_init_attr *qp_init_attr)
414 {
415 	struct mthca_dev *dev = to_mdev(ibqp->device);
416 	struct mthca_qp *qp = to_mqp(ibqp);
417 	int err;
418 	struct mthca_mailbox *mailbox;
419 	struct mthca_qp_param *qp_param;
420 	struct mthca_qp_context *context;
421 	int mthca_state;
422 	u8 status;
423 
424 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
425 	if (IS_ERR(mailbox))
426 		return PTR_ERR(mailbox);
427 
428 	err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox, &status);
429 	if (err)
430 		goto out;
431 	if (status) {
432 		mthca_warn(dev, "QUERY_QP returned status %02x\n", status);
433 		err = -EINVAL;
434 		goto out;
435 	}
436 
437 	qp_param    = mailbox->buf;
438 	context     = &qp_param->context;
439 	mthca_state = be32_to_cpu(context->flags) >> 28;
440 
441 	qp_attr->qp_state 	     = to_ib_qp_state(mthca_state);
442 	qp_attr->cur_qp_state 	     = qp_attr->qp_state;
443 	qp_attr->path_mtu 	     = context->mtu_msgmax >> 5;
444 	qp_attr->path_mig_state      =
445 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
446 	qp_attr->qkey 		     = be32_to_cpu(context->qkey);
447 	qp_attr->rq_psn 	     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
448 	qp_attr->sq_psn 	     = be32_to_cpu(context->next_send_psn) & 0xffffff;
449 	qp_attr->dest_qp_num 	     = be32_to_cpu(context->remote_qpn) & 0xffffff;
450 	qp_attr->qp_access_flags     =
451 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
452 	qp_attr->cap.max_send_wr     = qp->sq.max;
453 	qp_attr->cap.max_recv_wr     = qp->rq.max;
454 	qp_attr->cap.max_send_sge    = qp->sq.max_gs;
455 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
456 	qp_attr->cap.max_inline_data = qp->max_inline_data;
457 
458 	to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
459 	to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
460 
461 	qp_attr->pkey_index     = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
462 	qp_attr->alt_pkey_index = be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
463 
464 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
465 	qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
466 
467 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
468 
469 	qp_attr->max_dest_rd_atomic =
470 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
471 	qp_attr->min_rnr_timer 	    =
472 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
473 	qp_attr->port_num 	    = qp_attr->ah_attr.port_num;
474 	qp_attr->timeout 	    = context->pri_path.ackto >> 3;
475 	qp_attr->retry_cnt 	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
476 	qp_attr->rnr_retry 	    = context->pri_path.rnr_retry >> 5;
477 	qp_attr->alt_port_num 	    = qp_attr->alt_ah_attr.port_num;
478 	qp_attr->alt_timeout 	    = context->alt_path.ackto >> 3;
479 	qp_init_attr->cap 	    = qp_attr->cap;
480 
481 out:
482 	mthca_free_mailbox(dev, mailbox);
483 	return err;
484 }
485 
486 static int mthca_path_set(struct mthca_dev *dev, struct ib_ah_attr *ah,
487 			  struct mthca_qp_path *path)
488 {
489 	path->g_mylmc     = ah->src_path_bits & 0x7f;
490 	path->rlid        = cpu_to_be16(ah->dlid);
491 	path->static_rate = !!ah->static_rate;
492 
493 	if (ah->ah_flags & IB_AH_GRH) {
494 		if (ah->grh.sgid_index >= dev->limits.gid_table_len) {
495 			mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
496 				  ah->grh.sgid_index, dev->limits.gid_table_len-1);
497 			return -1;
498 		}
499 
500 		path->g_mylmc   |= 1 << 7;
501 		path->mgid_index = ah->grh.sgid_index;
502 		path->hop_limit  = ah->grh.hop_limit;
503 		path->sl_tclass_flowlabel =
504 			cpu_to_be32((ah->sl << 28)                |
505 				    (ah->grh.traffic_class << 20) |
506 				    (ah->grh.flow_label));
507 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
508 	} else
509 		path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28);
510 
511 	return 0;
512 }
513 
514 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
515 {
516 	struct mthca_dev *dev = to_mdev(ibqp->device);
517 	struct mthca_qp *qp = to_mqp(ibqp);
518 	enum ib_qp_state cur_state, new_state;
519 	struct mthca_mailbox *mailbox;
520 	struct mthca_qp_param *qp_param;
521 	struct mthca_qp_context *qp_context;
522 	u32 sqd_event = 0;
523 	u8 status;
524 	int err;
525 
526 	if (attr_mask & IB_QP_CUR_STATE) {
527 		cur_state = attr->cur_qp_state;
528 	} else {
529 		spin_lock_irq(&qp->sq.lock);
530 		spin_lock(&qp->rq.lock);
531 		cur_state = qp->state;
532 		spin_unlock(&qp->rq.lock);
533 		spin_unlock_irq(&qp->sq.lock);
534 	}
535 
536 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
537 
538 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
539 		mthca_dbg(dev, "Bad QP transition (transport %d) "
540 			  "%d->%d with attr 0x%08x\n",
541 			  qp->transport, cur_state, new_state,
542 			  attr_mask);
543 		return -EINVAL;
544 	}
545 
546 	if ((attr_mask & IB_QP_PKEY_INDEX) &&
547 	     attr->pkey_index >= dev->limits.pkey_table_len) {
548 		mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
549 			  attr->pkey_index, dev->limits.pkey_table_len-1);
550 		return -EINVAL;
551 	}
552 
553 	if ((attr_mask & IB_QP_PORT) &&
554 	    (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
555 		mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
556 		return -EINVAL;
557 	}
558 
559 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
560 	    attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
561 		mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
562 			  attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
563 		return -EINVAL;
564 	}
565 
566 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
567 	    attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
568 		mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
569 			  attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
570 		return -EINVAL;
571 	}
572 
573 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
574 	if (IS_ERR(mailbox))
575 		return PTR_ERR(mailbox);
576 	qp_param = mailbox->buf;
577 	qp_context = &qp_param->context;
578 	memset(qp_param, 0, sizeof *qp_param);
579 
580 	qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
581 					     (to_mthca_st(qp->transport) << 16));
582 	qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
583 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
584 		qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
585 	else {
586 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
587 		switch (attr->path_mig_state) {
588 		case IB_MIG_MIGRATED:
589 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
590 			break;
591 		case IB_MIG_REARM:
592 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
593 			break;
594 		case IB_MIG_ARMED:
595 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
596 			break;
597 		}
598 	}
599 
600 	/* leave tavor_sched_queue as 0 */
601 
602 	if (qp->transport == MLX || qp->transport == UD)
603 		qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
604 	else if (attr_mask & IB_QP_PATH_MTU) {
605 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
606 			mthca_dbg(dev, "path MTU (%u) is invalid\n",
607 				  attr->path_mtu);
608 			return -EINVAL;
609 		}
610 		qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
611 	}
612 
613 	if (mthca_is_memfree(dev)) {
614 		if (qp->rq.max)
615 			qp_context->rq_size_stride = long_log2(qp->rq.max) << 3;
616 		qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
617 
618 		if (qp->sq.max)
619 			qp_context->sq_size_stride = long_log2(qp->sq.max) << 3;
620 		qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
621 	}
622 
623 	/* leave arbel_sched_queue as 0 */
624 
625 	if (qp->ibqp.uobject)
626 		qp_context->usr_page =
627 			cpu_to_be32(to_mucontext(qp->ibqp.uobject->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(to_msqp(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 			return -EINVAL;
662 
663 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
664 	}
665 
666 	if (attr_mask & IB_QP_TIMEOUT) {
667 		qp_context->pri_path.ackto = attr->timeout << 3;
668 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
669 	}
670 
671 	if (attr_mask & IB_QP_ALT_PATH) {
672 		if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
673 			mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
674 				  attr->alt_pkey_index, dev->limits.pkey_table_len-1);
675 			return -EINVAL;
676 		}
677 
678 		if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
679 			mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
680 				attr->alt_port_num);
681 			return -EINVAL;
682 		}
683 
684 		if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path))
685 			return -EINVAL;
686 
687 		qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
688 							      attr->alt_port_num << 24);
689 		qp_context->alt_path.ackto = attr->alt_timeout << 3;
690 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
691 	}
692 
693 	/* leave rdd as 0 */
694 	qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
695 	/* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
696 	qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
697 	qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
698 					     (MTHCA_FLIGHT_LIMIT << 24) |
699 					     MTHCA_QP_BIT_SWE);
700 	if (qp->sq_policy == IB_SIGNAL_ALL_WR)
701 		qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
702 	if (attr_mask & IB_QP_RETRY_CNT) {
703 		qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
704 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
705 	}
706 
707 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
708 		if (attr->max_rd_atomic) {
709 			qp_context->params1 |=
710 				cpu_to_be32(MTHCA_QP_BIT_SRE |
711 					    MTHCA_QP_BIT_SAE);
712 			qp_context->params1 |=
713 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
714 		}
715 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
716 	}
717 
718 	if (attr_mask & IB_QP_SQ_PSN)
719 		qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
720 	qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
721 
722 	if (mthca_is_memfree(dev)) {
723 		qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
724 		qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
725 	}
726 
727 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
728 		if (attr->max_dest_rd_atomic)
729 			qp_context->params2 |=
730 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
731 
732 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
733 	}
734 
735 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
736 		qp_context->params2      |= get_hw_access_flags(qp, attr, attr_mask);
737 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
738 							MTHCA_QP_OPTPAR_RRE |
739 							MTHCA_QP_OPTPAR_RAE);
740 	}
741 
742 	qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
743 
744 	if (ibqp->srq)
745 		qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
746 
747 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
748 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
749 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
750 	}
751 	if (attr_mask & IB_QP_RQ_PSN)
752 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
753 
754 	qp_context->ra_buff_indx =
755 		cpu_to_be32(dev->qp_table.rdb_base +
756 			    ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
757 			     dev->qp_table.rdb_shift));
758 
759 	qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
760 
761 	if (mthca_is_memfree(dev))
762 		qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
763 
764 	if (attr_mask & IB_QP_QKEY) {
765 		qp_context->qkey = cpu_to_be32(attr->qkey);
766 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
767 	}
768 
769 	if (ibqp->srq)
770 		qp_context->srqn = cpu_to_be32(1 << 24 |
771 					       to_msrq(ibqp->srq)->srqn);
772 
773 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
774 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY		&&
775 	    attr->en_sqd_async_notify)
776 		sqd_event = 1 << 31;
777 
778 	err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
779 			      mailbox, sqd_event, &status);
780 	if (err)
781 		goto out;
782 	if (status) {
783 		mthca_warn(dev, "modify QP %d->%d returned status %02x.\n",
784 			   cur_state, new_state, status);
785 		err = -EINVAL;
786 		goto out;
787 	}
788 
789 	qp->state = new_state;
790 	if (attr_mask & IB_QP_ACCESS_FLAGS)
791 		qp->atomic_rd_en = attr->qp_access_flags;
792 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
793 		qp->resp_depth = attr->max_dest_rd_atomic;
794 
795 	if (is_sqp(dev, qp))
796 		store_attrs(to_msqp(qp), attr, attr_mask);
797 
798 	/*
799 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
800 	 * QP0 to RESET or ERROR, bring the link back down.
801 	 */
802 	if (is_qp0(dev, qp)) {
803 		if (cur_state != IB_QPS_RTR &&
804 		    new_state == IB_QPS_RTR)
805 			init_port(dev, to_msqp(qp)->port);
806 
807 		if (cur_state != IB_QPS_RESET &&
808 		    cur_state != IB_QPS_ERR &&
809 		    (new_state == IB_QPS_RESET ||
810 		     new_state == IB_QPS_ERR))
811 			mthca_CLOSE_IB(dev, to_msqp(qp)->port, &status);
812 	}
813 
814 	/*
815 	 * If we moved a kernel QP to RESET, clean up all old CQ
816 	 * entries and reinitialize the QP.
817 	 */
818 	if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
819 		mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
820 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
821 		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
822 			mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
823 				       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
824 
825 		mthca_wq_init(&qp->sq);
826 		qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
827 
828 		mthca_wq_init(&qp->rq);
829 		qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
830 
831 		if (mthca_is_memfree(dev)) {
832 			*qp->sq.db = 0;
833 			*qp->rq.db = 0;
834 		}
835 	}
836 
837 out:
838 	mthca_free_mailbox(dev, mailbox);
839 	return err;
840 }
841 
842 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
843 {
844 	/*
845 	 * Calculate the maximum size of WQE s/g segments, excluding
846 	 * the next segment and other non-data segments.
847 	 */
848 	int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
849 
850 	switch (qp->transport) {
851 	case MLX:
852 		max_data_size -= 2 * sizeof (struct mthca_data_seg);
853 		break;
854 
855 	case UD:
856 		if (mthca_is_memfree(dev))
857 			max_data_size -= sizeof (struct mthca_arbel_ud_seg);
858 		else
859 			max_data_size -= sizeof (struct mthca_tavor_ud_seg);
860 		break;
861 
862 	default:
863 		max_data_size -= sizeof (struct mthca_raddr_seg);
864 		break;
865 	}
866 
867 	return max_data_size;
868 }
869 
870 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
871 {
872 	/* We don't support inline data for kernel QPs (yet). */
873 	return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
874 }
875 
876 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
877 				 struct mthca_pd *pd,
878 				 struct mthca_qp *qp)
879 {
880 	int max_data_size = mthca_max_data_size(dev, qp,
881 						min(dev->limits.max_desc_sz,
882 						    1 << qp->sq.wqe_shift));
883 
884 	qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
885 
886 	qp->sq.max_gs = min_t(int, dev->limits.max_sg,
887 			      max_data_size / sizeof (struct mthca_data_seg));
888 	qp->rq.max_gs = min_t(int, dev->limits.max_sg,
889 			       (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
890 				sizeof (struct mthca_next_seg)) /
891 			       sizeof (struct mthca_data_seg));
892 }
893 
894 /*
895  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
896  * rq.max_gs and sq.max_gs must all be assigned.
897  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
898  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
899  * queue)
900  */
901 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
902 			       struct mthca_pd *pd,
903 			       struct mthca_qp *qp)
904 {
905 	int size;
906 	int err = -ENOMEM;
907 
908 	size = sizeof (struct mthca_next_seg) +
909 		qp->rq.max_gs * sizeof (struct mthca_data_seg);
910 
911 	if (size > dev->limits.max_desc_sz)
912 		return -EINVAL;
913 
914 	for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
915 	     qp->rq.wqe_shift++)
916 		; /* nothing */
917 
918 	size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
919 	switch (qp->transport) {
920 	case MLX:
921 		size += 2 * sizeof (struct mthca_data_seg);
922 		break;
923 
924 	case UD:
925 		size += mthca_is_memfree(dev) ?
926 			sizeof (struct mthca_arbel_ud_seg) :
927 			sizeof (struct mthca_tavor_ud_seg);
928 		break;
929 
930 	case UC:
931 		size += sizeof (struct mthca_raddr_seg);
932 		break;
933 
934 	case RC:
935 		size += sizeof (struct mthca_raddr_seg);
936 		/*
937 		 * An atomic op will require an atomic segment, a
938 		 * remote address segment and one scatter entry.
939 		 */
940 		size = max_t(int, size,
941 			     sizeof (struct mthca_atomic_seg) +
942 			     sizeof (struct mthca_raddr_seg) +
943 			     sizeof (struct mthca_data_seg));
944 		break;
945 
946 	default:
947 		break;
948 	}
949 
950 	/* Make sure that we have enough space for a bind request */
951 	size = max_t(int, size, sizeof (struct mthca_bind_seg));
952 
953 	size += sizeof (struct mthca_next_seg);
954 
955 	if (size > dev->limits.max_desc_sz)
956 		return -EINVAL;
957 
958 	for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
959 	     qp->sq.wqe_shift++)
960 		; /* nothing */
961 
962 	qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
963 				    1 << qp->sq.wqe_shift);
964 
965 	/*
966 	 * If this is a userspace QP, we don't actually have to
967 	 * allocate anything.  All we need is to calculate the WQE
968 	 * sizes and the send_wqe_offset, so we're done now.
969 	 */
970 	if (pd->ibpd.uobject)
971 		return 0;
972 
973 	size = PAGE_ALIGN(qp->send_wqe_offset +
974 			  (qp->sq.max << qp->sq.wqe_shift));
975 
976 	qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
977 			   GFP_KERNEL);
978 	if (!qp->wrid)
979 		goto err_out;
980 
981 	err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
982 			      &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
983 	if (err)
984 		goto err_out;
985 
986 	return 0;
987 
988 err_out:
989 	kfree(qp->wrid);
990 	return err;
991 }
992 
993 static void mthca_free_wqe_buf(struct mthca_dev *dev,
994 			       struct mthca_qp *qp)
995 {
996 	mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
997 				       (qp->sq.max << qp->sq.wqe_shift)),
998 		       &qp->queue, qp->is_direct, &qp->mr);
999 	kfree(qp->wrid);
1000 }
1001 
1002 static int mthca_map_memfree(struct mthca_dev *dev,
1003 			     struct mthca_qp *qp)
1004 {
1005 	int ret;
1006 
1007 	if (mthca_is_memfree(dev)) {
1008 		ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1009 		if (ret)
1010 			return ret;
1011 
1012 		ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1013 		if (ret)
1014 			goto err_qpc;
1015 
1016 		ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1017 				      qp->qpn << dev->qp_table.rdb_shift);
1018 		if (ret)
1019 			goto err_eqpc;
1020 
1021 	}
1022 
1023 	return 0;
1024 
1025 err_eqpc:
1026 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1027 
1028 err_qpc:
1029 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1030 
1031 	return ret;
1032 }
1033 
1034 static void mthca_unmap_memfree(struct mthca_dev *dev,
1035 				struct mthca_qp *qp)
1036 {
1037 	mthca_table_put(dev, dev->qp_table.rdb_table,
1038 			qp->qpn << dev->qp_table.rdb_shift);
1039 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1040 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1041 }
1042 
1043 static int mthca_alloc_memfree(struct mthca_dev *dev,
1044 			       struct mthca_qp *qp)
1045 {
1046 	int ret = 0;
1047 
1048 	if (mthca_is_memfree(dev)) {
1049 		qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1050 						 qp->qpn, &qp->rq.db);
1051 		if (qp->rq.db_index < 0)
1052 			return ret;
1053 
1054 		qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1055 						 qp->qpn, &qp->sq.db);
1056 		if (qp->sq.db_index < 0)
1057 			mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1058 	}
1059 
1060 	return ret;
1061 }
1062 
1063 static void mthca_free_memfree(struct mthca_dev *dev,
1064 			       struct mthca_qp *qp)
1065 {
1066 	if (mthca_is_memfree(dev)) {
1067 		mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1068 		mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1069 	}
1070 }
1071 
1072 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1073 				 struct mthca_pd *pd,
1074 				 struct mthca_cq *send_cq,
1075 				 struct mthca_cq *recv_cq,
1076 				 enum ib_sig_type send_policy,
1077 				 struct mthca_qp *qp)
1078 {
1079 	int ret;
1080 	int i;
1081 
1082 	atomic_set(&qp->refcount, 1);
1083 	init_waitqueue_head(&qp->wait);
1084 	qp->state    	 = IB_QPS_RESET;
1085 	qp->atomic_rd_en = 0;
1086 	qp->resp_depth   = 0;
1087 	qp->sq_policy    = send_policy;
1088 	mthca_wq_init(&qp->sq);
1089 	mthca_wq_init(&qp->rq);
1090 
1091 	ret = mthca_map_memfree(dev, qp);
1092 	if (ret)
1093 		return ret;
1094 
1095 	ret = mthca_alloc_wqe_buf(dev, pd, qp);
1096 	if (ret) {
1097 		mthca_unmap_memfree(dev, qp);
1098 		return ret;
1099 	}
1100 
1101 	mthca_adjust_qp_caps(dev, pd, qp);
1102 
1103 	/*
1104 	 * If this is a userspace QP, we're done now.  The doorbells
1105 	 * will be allocated and buffers will be initialized in
1106 	 * userspace.
1107 	 */
1108 	if (pd->ibpd.uobject)
1109 		return 0;
1110 
1111 	ret = mthca_alloc_memfree(dev, qp);
1112 	if (ret) {
1113 		mthca_free_wqe_buf(dev, qp);
1114 		mthca_unmap_memfree(dev, qp);
1115 		return ret;
1116 	}
1117 
1118 	if (mthca_is_memfree(dev)) {
1119 		struct mthca_next_seg *next;
1120 		struct mthca_data_seg *scatter;
1121 		int size = (sizeof (struct mthca_next_seg) +
1122 			    qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1123 
1124 		for (i = 0; i < qp->rq.max; ++i) {
1125 			next = get_recv_wqe(qp, i);
1126 			next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1127 						   qp->rq.wqe_shift);
1128 			next->ee_nds = cpu_to_be32(size);
1129 
1130 			for (scatter = (void *) (next + 1);
1131 			     (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1132 			     ++scatter)
1133 				scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1134 		}
1135 
1136 		for (i = 0; i < qp->sq.max; ++i) {
1137 			next = get_send_wqe(qp, i);
1138 			next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1139 						    qp->sq.wqe_shift) +
1140 						   qp->send_wqe_offset);
1141 		}
1142 	}
1143 
1144 	qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1145 	qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1146 
1147 	return 0;
1148 }
1149 
1150 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1151 			     struct mthca_pd *pd, struct mthca_qp *qp)
1152 {
1153 	int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1154 
1155 	/* Sanity check QP size before proceeding */
1156 	if (cap->max_send_wr  	 > dev->limits.max_wqes ||
1157 	    cap->max_recv_wr  	 > dev->limits.max_wqes ||
1158 	    cap->max_send_sge 	 > dev->limits.max_sg   ||
1159 	    cap->max_recv_sge 	 > dev->limits.max_sg   ||
1160 	    cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1161 		return -EINVAL;
1162 
1163 	/*
1164 	 * For MLX transport we need 2 extra S/G entries:
1165 	 * one for the header and one for the checksum at the end
1166 	 */
1167 	if (qp->transport == MLX && cap->max_recv_sge + 2 > dev->limits.max_sg)
1168 		return -EINVAL;
1169 
1170 	if (mthca_is_memfree(dev)) {
1171 		qp->rq.max = cap->max_recv_wr ?
1172 			roundup_pow_of_two(cap->max_recv_wr) : 0;
1173 		qp->sq.max = cap->max_send_wr ?
1174 			roundup_pow_of_two(cap->max_send_wr) : 0;
1175 	} else {
1176 		qp->rq.max = cap->max_recv_wr;
1177 		qp->sq.max = cap->max_send_wr;
1178 	}
1179 
1180 	qp->rq.max_gs = cap->max_recv_sge;
1181 	qp->sq.max_gs = max_t(int, cap->max_send_sge,
1182 			      ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1183 				    MTHCA_INLINE_CHUNK_SIZE) /
1184 			      sizeof (struct mthca_data_seg));
1185 
1186 	return 0;
1187 }
1188 
1189 int mthca_alloc_qp(struct mthca_dev *dev,
1190 		   struct mthca_pd *pd,
1191 		   struct mthca_cq *send_cq,
1192 		   struct mthca_cq *recv_cq,
1193 		   enum ib_qp_type type,
1194 		   enum ib_sig_type send_policy,
1195 		   struct ib_qp_cap *cap,
1196 		   struct mthca_qp *qp)
1197 {
1198 	int err;
1199 
1200 	switch (type) {
1201 	case IB_QPT_RC: qp->transport = RC; break;
1202 	case IB_QPT_UC: qp->transport = UC; break;
1203 	case IB_QPT_UD: qp->transport = UD; break;
1204 	default: return -EINVAL;
1205 	}
1206 
1207 	err = mthca_set_qp_size(dev, cap, pd, qp);
1208 	if (err)
1209 		return err;
1210 
1211 	qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1212 	if (qp->qpn == -1)
1213 		return -ENOMEM;
1214 
1215 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1216 				    send_policy, qp);
1217 	if (err) {
1218 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1219 		return err;
1220 	}
1221 
1222 	spin_lock_irq(&dev->qp_table.lock);
1223 	mthca_array_set(&dev->qp_table.qp,
1224 			qp->qpn & (dev->limits.num_qps - 1), qp);
1225 	spin_unlock_irq(&dev->qp_table.lock);
1226 
1227 	return 0;
1228 }
1229 
1230 int mthca_alloc_sqp(struct mthca_dev *dev,
1231 		    struct mthca_pd *pd,
1232 		    struct mthca_cq *send_cq,
1233 		    struct mthca_cq *recv_cq,
1234 		    enum ib_sig_type send_policy,
1235 		    struct ib_qp_cap *cap,
1236 		    int qpn,
1237 		    int port,
1238 		    struct mthca_sqp *sqp)
1239 {
1240 	u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1241 	int err;
1242 
1243 	sqp->qp.transport = MLX;
1244 	err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1245 	if (err)
1246 		return err;
1247 
1248 	sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1249 	sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1250 					     &sqp->header_dma, GFP_KERNEL);
1251 	if (!sqp->header_buf)
1252 		return -ENOMEM;
1253 
1254 	spin_lock_irq(&dev->qp_table.lock);
1255 	if (mthca_array_get(&dev->qp_table.qp, mqpn))
1256 		err = -EBUSY;
1257 	else
1258 		mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1259 	spin_unlock_irq(&dev->qp_table.lock);
1260 
1261 	if (err)
1262 		goto err_out;
1263 
1264 	sqp->port = port;
1265 	sqp->qp.qpn       = mqpn;
1266 	sqp->qp.transport = MLX;
1267 
1268 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1269 				    send_policy, &sqp->qp);
1270 	if (err)
1271 		goto err_out_free;
1272 
1273 	atomic_inc(&pd->sqp_count);
1274 
1275 	return 0;
1276 
1277  err_out_free:
1278 	/*
1279 	 * Lock CQs here, so that CQ polling code can do QP lookup
1280 	 * without taking a lock.
1281 	 */
1282 	spin_lock_irq(&send_cq->lock);
1283 	if (send_cq != recv_cq)
1284 		spin_lock(&recv_cq->lock);
1285 
1286 	spin_lock(&dev->qp_table.lock);
1287 	mthca_array_clear(&dev->qp_table.qp, mqpn);
1288 	spin_unlock(&dev->qp_table.lock);
1289 
1290 	if (send_cq != recv_cq)
1291 		spin_unlock(&recv_cq->lock);
1292 	spin_unlock_irq(&send_cq->lock);
1293 
1294  err_out:
1295 	dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1296 			  sqp->header_buf, sqp->header_dma);
1297 
1298 	return err;
1299 }
1300 
1301 void mthca_free_qp(struct mthca_dev *dev,
1302 		   struct mthca_qp *qp)
1303 {
1304 	u8 status;
1305 	struct mthca_cq *send_cq;
1306 	struct mthca_cq *recv_cq;
1307 
1308 	send_cq = to_mcq(qp->ibqp.send_cq);
1309 	recv_cq = to_mcq(qp->ibqp.recv_cq);
1310 
1311 	/*
1312 	 * Lock CQs here, so that CQ polling code can do QP lookup
1313 	 * without taking a lock.
1314 	 */
1315 	spin_lock_irq(&send_cq->lock);
1316 	if (send_cq != recv_cq)
1317 		spin_lock(&recv_cq->lock);
1318 
1319 	spin_lock(&dev->qp_table.lock);
1320 	mthca_array_clear(&dev->qp_table.qp,
1321 			  qp->qpn & (dev->limits.num_qps - 1));
1322 	spin_unlock(&dev->qp_table.lock);
1323 
1324 	if (send_cq != recv_cq)
1325 		spin_unlock(&recv_cq->lock);
1326 	spin_unlock_irq(&send_cq->lock);
1327 
1328 	atomic_dec(&qp->refcount);
1329 	wait_event(qp->wait, !atomic_read(&qp->refcount));
1330 
1331 	if (qp->state != IB_QPS_RESET)
1332 		mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1333 				NULL, 0, &status);
1334 
1335 	/*
1336 	 * If this is a userspace QP, the buffers, MR, CQs and so on
1337 	 * will be cleaned up in userspace, so all we have to do is
1338 	 * unref the mem-free tables and free the QPN in our table.
1339 	 */
1340 	if (!qp->ibqp.uobject) {
1341 		mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
1342 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1343 		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
1344 			mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
1345 				       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1346 
1347 		mthca_free_memfree(dev, qp);
1348 		mthca_free_wqe_buf(dev, qp);
1349 	}
1350 
1351 	mthca_unmap_memfree(dev, qp);
1352 
1353 	if (is_sqp(dev, qp)) {
1354 		atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1355 		dma_free_coherent(&dev->pdev->dev,
1356 				  to_msqp(qp)->header_buf_size,
1357 				  to_msqp(qp)->header_buf,
1358 				  to_msqp(qp)->header_dma);
1359 	} else
1360 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1361 }
1362 
1363 /* Create UD header for an MLX send and build a data segment for it */
1364 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1365 			    int ind, struct ib_send_wr *wr,
1366 			    struct mthca_mlx_seg *mlx,
1367 			    struct mthca_data_seg *data)
1368 {
1369 	int header_size;
1370 	int err;
1371 	u16 pkey;
1372 
1373 	ib_ud_header_init(256, /* assume a MAD */
1374 			  mthca_ah_grh_present(to_mah(wr->wr.ud.ah)),
1375 			  &sqp->ud_header);
1376 
1377 	err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1378 	if (err)
1379 		return err;
1380 	mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1381 	mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1382 				  (sqp->ud_header.lrh.destination_lid ==
1383 				   IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1384 				  (sqp->ud_header.lrh.service_level << 8));
1385 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
1386 	mlx->vcrc = 0;
1387 
1388 	switch (wr->opcode) {
1389 	case IB_WR_SEND:
1390 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1391 		sqp->ud_header.immediate_present = 0;
1392 		break;
1393 	case IB_WR_SEND_WITH_IMM:
1394 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1395 		sqp->ud_header.immediate_present = 1;
1396 		sqp->ud_header.immediate_data = wr->imm_data;
1397 		break;
1398 	default:
1399 		return -EINVAL;
1400 	}
1401 
1402 	sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1403 	if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1404 		sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1405 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1406 	if (!sqp->qp.ibqp.qp_num)
1407 		ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1408 				   sqp->pkey_index, &pkey);
1409 	else
1410 		ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1411 				   wr->wr.ud.pkey_index, &pkey);
1412 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1413 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1414 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1415 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1416 					       sqp->qkey : wr->wr.ud.remote_qkey);
1417 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1418 
1419 	header_size = ib_ud_header_pack(&sqp->ud_header,
1420 					sqp->header_buf +
1421 					ind * MTHCA_UD_HEADER_SIZE);
1422 
1423 	data->byte_count = cpu_to_be32(header_size);
1424 	data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1425 	data->addr       = cpu_to_be64(sqp->header_dma +
1426 				       ind * MTHCA_UD_HEADER_SIZE);
1427 
1428 	return 0;
1429 }
1430 
1431 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1432 				    struct ib_cq *ib_cq)
1433 {
1434 	unsigned cur;
1435 	struct mthca_cq *cq;
1436 
1437 	cur = wq->head - wq->tail;
1438 	if (likely(cur + nreq < wq->max))
1439 		return 0;
1440 
1441 	cq = to_mcq(ib_cq);
1442 	spin_lock(&cq->lock);
1443 	cur = wq->head - wq->tail;
1444 	spin_unlock(&cq->lock);
1445 
1446 	return cur + nreq >= wq->max;
1447 }
1448 
1449 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1450 			  struct ib_send_wr **bad_wr)
1451 {
1452 	struct mthca_dev *dev = to_mdev(ibqp->device);
1453 	struct mthca_qp *qp = to_mqp(ibqp);
1454 	void *wqe;
1455 	void *prev_wqe;
1456 	unsigned long flags;
1457 	int err = 0;
1458 	int nreq;
1459 	int i;
1460 	int size;
1461 	int size0 = 0;
1462 	u32 f0 = 0;
1463 	int ind;
1464 	u8 op0 = 0;
1465 
1466 	spin_lock_irqsave(&qp->sq.lock, flags);
1467 
1468 	/* XXX check that state is OK to post send */
1469 
1470 	ind = qp->sq.next_ind;
1471 
1472 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1473 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1474 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1475 					" %d max, %d nreq)\n", qp->qpn,
1476 					qp->sq.head, qp->sq.tail,
1477 					qp->sq.max, nreq);
1478 			err = -ENOMEM;
1479 			*bad_wr = wr;
1480 			goto out;
1481 		}
1482 
1483 		wqe = get_send_wqe(qp, ind);
1484 		prev_wqe = qp->sq.last;
1485 		qp->sq.last = wqe;
1486 
1487 		((struct mthca_next_seg *) wqe)->nda_op = 0;
1488 		((struct mthca_next_seg *) wqe)->ee_nds = 0;
1489 		((struct mthca_next_seg *) wqe)->flags =
1490 			((wr->send_flags & IB_SEND_SIGNALED) ?
1491 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1492 			((wr->send_flags & IB_SEND_SOLICITED) ?
1493 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1494 			cpu_to_be32(1);
1495 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1496 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1497 			((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1498 
1499 		wqe += sizeof (struct mthca_next_seg);
1500 		size = sizeof (struct mthca_next_seg) / 16;
1501 
1502 		switch (qp->transport) {
1503 		case RC:
1504 			switch (wr->opcode) {
1505 			case IB_WR_ATOMIC_CMP_AND_SWP:
1506 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1507 				((struct mthca_raddr_seg *) wqe)->raddr =
1508 					cpu_to_be64(wr->wr.atomic.remote_addr);
1509 				((struct mthca_raddr_seg *) wqe)->rkey =
1510 					cpu_to_be32(wr->wr.atomic.rkey);
1511 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1512 
1513 				wqe += sizeof (struct mthca_raddr_seg);
1514 
1515 				if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1516 					((struct mthca_atomic_seg *) wqe)->swap_add =
1517 						cpu_to_be64(wr->wr.atomic.swap);
1518 					((struct mthca_atomic_seg *) wqe)->compare =
1519 						cpu_to_be64(wr->wr.atomic.compare_add);
1520 				} else {
1521 					((struct mthca_atomic_seg *) wqe)->swap_add =
1522 						cpu_to_be64(wr->wr.atomic.compare_add);
1523 					((struct mthca_atomic_seg *) wqe)->compare = 0;
1524 				}
1525 
1526 				wqe += sizeof (struct mthca_atomic_seg);
1527 				size += (sizeof (struct mthca_raddr_seg) +
1528 					 sizeof (struct mthca_atomic_seg)) / 16;
1529 				break;
1530 
1531 			case IB_WR_RDMA_WRITE:
1532 			case IB_WR_RDMA_WRITE_WITH_IMM:
1533 			case IB_WR_RDMA_READ:
1534 				((struct mthca_raddr_seg *) wqe)->raddr =
1535 					cpu_to_be64(wr->wr.rdma.remote_addr);
1536 				((struct mthca_raddr_seg *) wqe)->rkey =
1537 					cpu_to_be32(wr->wr.rdma.rkey);
1538 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1539 				wqe += sizeof (struct mthca_raddr_seg);
1540 				size += sizeof (struct mthca_raddr_seg) / 16;
1541 				break;
1542 
1543 			default:
1544 				/* No extra segments required for sends */
1545 				break;
1546 			}
1547 
1548 			break;
1549 
1550 		case UC:
1551 			switch (wr->opcode) {
1552 			case IB_WR_RDMA_WRITE:
1553 			case IB_WR_RDMA_WRITE_WITH_IMM:
1554 				((struct mthca_raddr_seg *) wqe)->raddr =
1555 					cpu_to_be64(wr->wr.rdma.remote_addr);
1556 				((struct mthca_raddr_seg *) wqe)->rkey =
1557 					cpu_to_be32(wr->wr.rdma.rkey);
1558 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1559 				wqe += sizeof (struct mthca_raddr_seg);
1560 				size += sizeof (struct mthca_raddr_seg) / 16;
1561 				break;
1562 
1563 			default:
1564 				/* No extra segments required for sends */
1565 				break;
1566 			}
1567 
1568 			break;
1569 
1570 		case UD:
1571 			((struct mthca_tavor_ud_seg *) wqe)->lkey =
1572 				cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1573 			((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1574 				cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1575 			((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1576 				cpu_to_be32(wr->wr.ud.remote_qpn);
1577 			((struct mthca_tavor_ud_seg *) wqe)->qkey =
1578 				cpu_to_be32(wr->wr.ud.remote_qkey);
1579 
1580 			wqe += sizeof (struct mthca_tavor_ud_seg);
1581 			size += sizeof (struct mthca_tavor_ud_seg) / 16;
1582 			break;
1583 
1584 		case MLX:
1585 			err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1586 					       wqe - sizeof (struct mthca_next_seg),
1587 					       wqe);
1588 			if (err) {
1589 				*bad_wr = wr;
1590 				goto out;
1591 			}
1592 			wqe += sizeof (struct mthca_data_seg);
1593 			size += sizeof (struct mthca_data_seg) / 16;
1594 			break;
1595 		}
1596 
1597 		if (wr->num_sge > qp->sq.max_gs) {
1598 			mthca_err(dev, "too many gathers\n");
1599 			err = -EINVAL;
1600 			*bad_wr = wr;
1601 			goto out;
1602 		}
1603 
1604 		for (i = 0; i < wr->num_sge; ++i) {
1605 			((struct mthca_data_seg *) wqe)->byte_count =
1606 				cpu_to_be32(wr->sg_list[i].length);
1607 			((struct mthca_data_seg *) wqe)->lkey =
1608 				cpu_to_be32(wr->sg_list[i].lkey);
1609 			((struct mthca_data_seg *) wqe)->addr =
1610 				cpu_to_be64(wr->sg_list[i].addr);
1611 			wqe += sizeof (struct mthca_data_seg);
1612 			size += sizeof (struct mthca_data_seg) / 16;
1613 		}
1614 
1615 		/* Add one more inline data segment for ICRC */
1616 		if (qp->transport == MLX) {
1617 			((struct mthca_data_seg *) wqe)->byte_count =
1618 				cpu_to_be32((1 << 31) | 4);
1619 			((u32 *) wqe)[1] = 0;
1620 			wqe += sizeof (struct mthca_data_seg);
1621 			size += sizeof (struct mthca_data_seg) / 16;
1622 		}
1623 
1624 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
1625 
1626 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1627 			mthca_err(dev, "opcode invalid\n");
1628 			err = -EINVAL;
1629 			*bad_wr = wr;
1630 			goto out;
1631 		}
1632 
1633 		((struct mthca_next_seg *) prev_wqe)->nda_op =
1634 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
1635 				     qp->send_wqe_offset) |
1636 				    mthca_opcode[wr->opcode]);
1637 		wmb();
1638 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1639 			cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size |
1640 				    ((wr->send_flags & IB_SEND_FENCE) ?
1641 				    MTHCA_NEXT_FENCE : 0));
1642 
1643 		if (!size0) {
1644 			size0 = size;
1645 			op0   = mthca_opcode[wr->opcode];
1646 		}
1647 
1648 		++ind;
1649 		if (unlikely(ind >= qp->sq.max))
1650 			ind -= qp->sq.max;
1651 	}
1652 
1653 out:
1654 	if (likely(nreq)) {
1655 		__be32 doorbell[2];
1656 
1657 		doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1658 					   qp->send_wqe_offset) | f0 | op0);
1659 		doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1660 
1661 		wmb();
1662 
1663 		mthca_write64(doorbell,
1664 			      dev->kar + MTHCA_SEND_DOORBELL,
1665 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1666 	}
1667 
1668 	qp->sq.next_ind = ind;
1669 	qp->sq.head    += nreq;
1670 
1671 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1672 	return err;
1673 }
1674 
1675 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1676 			     struct ib_recv_wr **bad_wr)
1677 {
1678 	struct mthca_dev *dev = to_mdev(ibqp->device);
1679 	struct mthca_qp *qp = to_mqp(ibqp);
1680 	__be32 doorbell[2];
1681 	unsigned long flags;
1682 	int err = 0;
1683 	int nreq;
1684 	int i;
1685 	int size;
1686 	int size0 = 0;
1687 	int ind;
1688 	void *wqe;
1689 	void *prev_wqe;
1690 
1691 	spin_lock_irqsave(&qp->rq.lock, flags);
1692 
1693 	/* XXX check that state is OK to post receive */
1694 
1695 	ind = qp->rq.next_ind;
1696 
1697 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1698 		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1699 			nreq = 0;
1700 
1701 			doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1702 			doorbell[1] = cpu_to_be32(qp->qpn << 8);
1703 
1704 			wmb();
1705 
1706 			mthca_write64(doorbell,
1707 				      dev->kar + MTHCA_RECEIVE_DOORBELL,
1708 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1709 
1710 			qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1711 			size0 = 0;
1712 		}
1713 
1714 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1715 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1716 					" %d max, %d nreq)\n", qp->qpn,
1717 					qp->rq.head, qp->rq.tail,
1718 					qp->rq.max, nreq);
1719 			err = -ENOMEM;
1720 			*bad_wr = wr;
1721 			goto out;
1722 		}
1723 
1724 		wqe = get_recv_wqe(qp, ind);
1725 		prev_wqe = qp->rq.last;
1726 		qp->rq.last = wqe;
1727 
1728 		((struct mthca_next_seg *) wqe)->nda_op = 0;
1729 		((struct mthca_next_seg *) wqe)->ee_nds =
1730 			cpu_to_be32(MTHCA_NEXT_DBD);
1731 		((struct mthca_next_seg *) wqe)->flags = 0;
1732 
1733 		wqe += sizeof (struct mthca_next_seg);
1734 		size = sizeof (struct mthca_next_seg) / 16;
1735 
1736 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1737 			err = -EINVAL;
1738 			*bad_wr = wr;
1739 			goto out;
1740 		}
1741 
1742 		for (i = 0; i < wr->num_sge; ++i) {
1743 			((struct mthca_data_seg *) wqe)->byte_count =
1744 				cpu_to_be32(wr->sg_list[i].length);
1745 			((struct mthca_data_seg *) wqe)->lkey =
1746 				cpu_to_be32(wr->sg_list[i].lkey);
1747 			((struct mthca_data_seg *) wqe)->addr =
1748 				cpu_to_be64(wr->sg_list[i].addr);
1749 			wqe += sizeof (struct mthca_data_seg);
1750 			size += sizeof (struct mthca_data_seg) / 16;
1751 		}
1752 
1753 		qp->wrid[ind] = wr->wr_id;
1754 
1755 		((struct mthca_next_seg *) prev_wqe)->nda_op =
1756 			cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1757 		wmb();
1758 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1759 			cpu_to_be32(MTHCA_NEXT_DBD | size);
1760 
1761 		if (!size0)
1762 			size0 = size;
1763 
1764 		++ind;
1765 		if (unlikely(ind >= qp->rq.max))
1766 			ind -= qp->rq.max;
1767 	}
1768 
1769 out:
1770 	if (likely(nreq)) {
1771 		doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1772 		doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
1773 
1774 		wmb();
1775 
1776 		mthca_write64(doorbell,
1777 			      dev->kar + MTHCA_RECEIVE_DOORBELL,
1778 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1779 	}
1780 
1781 	qp->rq.next_ind = ind;
1782 	qp->rq.head    += nreq;
1783 
1784 	spin_unlock_irqrestore(&qp->rq.lock, flags);
1785 	return err;
1786 }
1787 
1788 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1789 			  struct ib_send_wr **bad_wr)
1790 {
1791 	struct mthca_dev *dev = to_mdev(ibqp->device);
1792 	struct mthca_qp *qp = to_mqp(ibqp);
1793 	__be32 doorbell[2];
1794 	void *wqe;
1795 	void *prev_wqe;
1796 	unsigned long flags;
1797 	int err = 0;
1798 	int nreq;
1799 	int i;
1800 	int size;
1801 	int size0 = 0;
1802 	u32 f0 = 0;
1803 	int ind;
1804 	u8 op0 = 0;
1805 
1806 	spin_lock_irqsave(&qp->sq.lock, flags);
1807 
1808 	/* XXX check that state is OK to post send */
1809 
1810 	ind = qp->sq.head & (qp->sq.max - 1);
1811 
1812 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1813 		if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1814 			nreq = 0;
1815 
1816 			doorbell[0] = cpu_to_be32((MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1817 						  ((qp->sq.head & 0xffff) << 8) |
1818 						  f0 | op0);
1819 			doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1820 
1821 			qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1822 			size0 = 0;
1823 
1824 			/*
1825 			 * Make sure that descriptors are written before
1826 			 * doorbell record.
1827 			 */
1828 			wmb();
1829 			*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1830 
1831 			/*
1832 			 * Make sure doorbell record is written before we
1833 			 * write MMIO send doorbell.
1834 			 */
1835 			wmb();
1836 			mthca_write64(doorbell,
1837 				      dev->kar + MTHCA_SEND_DOORBELL,
1838 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1839 		}
1840 
1841 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1842 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1843 					" %d max, %d nreq)\n", qp->qpn,
1844 					qp->sq.head, qp->sq.tail,
1845 					qp->sq.max, nreq);
1846 			err = -ENOMEM;
1847 			*bad_wr = wr;
1848 			goto out;
1849 		}
1850 
1851 		wqe = get_send_wqe(qp, ind);
1852 		prev_wqe = qp->sq.last;
1853 		qp->sq.last = wqe;
1854 
1855 		((struct mthca_next_seg *) wqe)->flags =
1856 			((wr->send_flags & IB_SEND_SIGNALED) ?
1857 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1858 			((wr->send_flags & IB_SEND_SOLICITED) ?
1859 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1860 			cpu_to_be32(1);
1861 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1862 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1863 			((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1864 
1865 		wqe += sizeof (struct mthca_next_seg);
1866 		size = sizeof (struct mthca_next_seg) / 16;
1867 
1868 		switch (qp->transport) {
1869 		case RC:
1870 			switch (wr->opcode) {
1871 			case IB_WR_ATOMIC_CMP_AND_SWP:
1872 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1873 				((struct mthca_raddr_seg *) wqe)->raddr =
1874 					cpu_to_be64(wr->wr.atomic.remote_addr);
1875 				((struct mthca_raddr_seg *) wqe)->rkey =
1876 					cpu_to_be32(wr->wr.atomic.rkey);
1877 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1878 
1879 				wqe += sizeof (struct mthca_raddr_seg);
1880 
1881 				if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1882 					((struct mthca_atomic_seg *) wqe)->swap_add =
1883 						cpu_to_be64(wr->wr.atomic.swap);
1884 					((struct mthca_atomic_seg *) wqe)->compare =
1885 						cpu_to_be64(wr->wr.atomic.compare_add);
1886 				} else {
1887 					((struct mthca_atomic_seg *) wqe)->swap_add =
1888 						cpu_to_be64(wr->wr.atomic.compare_add);
1889 					((struct mthca_atomic_seg *) wqe)->compare = 0;
1890 				}
1891 
1892 				wqe += sizeof (struct mthca_atomic_seg);
1893 				size += (sizeof (struct mthca_raddr_seg) +
1894 					 sizeof (struct mthca_atomic_seg)) / 16;
1895 				break;
1896 
1897 			case IB_WR_RDMA_READ:
1898 			case IB_WR_RDMA_WRITE:
1899 			case IB_WR_RDMA_WRITE_WITH_IMM:
1900 				((struct mthca_raddr_seg *) wqe)->raddr =
1901 					cpu_to_be64(wr->wr.rdma.remote_addr);
1902 				((struct mthca_raddr_seg *) wqe)->rkey =
1903 					cpu_to_be32(wr->wr.rdma.rkey);
1904 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1905 				wqe += sizeof (struct mthca_raddr_seg);
1906 				size += sizeof (struct mthca_raddr_seg) / 16;
1907 				break;
1908 
1909 			default:
1910 				/* No extra segments required for sends */
1911 				break;
1912 			}
1913 
1914 			break;
1915 
1916 		case UC:
1917 			switch (wr->opcode) {
1918 			case IB_WR_RDMA_WRITE:
1919 			case IB_WR_RDMA_WRITE_WITH_IMM:
1920 				((struct mthca_raddr_seg *) wqe)->raddr =
1921 					cpu_to_be64(wr->wr.rdma.remote_addr);
1922 				((struct mthca_raddr_seg *) wqe)->rkey =
1923 					cpu_to_be32(wr->wr.rdma.rkey);
1924 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1925 				wqe += sizeof (struct mthca_raddr_seg);
1926 				size += sizeof (struct mthca_raddr_seg) / 16;
1927 				break;
1928 
1929 			default:
1930 				/* No extra segments required for sends */
1931 				break;
1932 			}
1933 
1934 			break;
1935 
1936 		case UD:
1937 			memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
1938 			       to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1939 			((struct mthca_arbel_ud_seg *) wqe)->dqpn =
1940 				cpu_to_be32(wr->wr.ud.remote_qpn);
1941 			((struct mthca_arbel_ud_seg *) wqe)->qkey =
1942 				cpu_to_be32(wr->wr.ud.remote_qkey);
1943 
1944 			wqe += sizeof (struct mthca_arbel_ud_seg);
1945 			size += sizeof (struct mthca_arbel_ud_seg) / 16;
1946 			break;
1947 
1948 		case MLX:
1949 			err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1950 					       wqe - sizeof (struct mthca_next_seg),
1951 					       wqe);
1952 			if (err) {
1953 				*bad_wr = wr;
1954 				goto out;
1955 			}
1956 			wqe += sizeof (struct mthca_data_seg);
1957 			size += sizeof (struct mthca_data_seg) / 16;
1958 			break;
1959 		}
1960 
1961 		if (wr->num_sge > qp->sq.max_gs) {
1962 			mthca_err(dev, "too many gathers\n");
1963 			err = -EINVAL;
1964 			*bad_wr = wr;
1965 			goto out;
1966 		}
1967 
1968 		for (i = 0; i < wr->num_sge; ++i) {
1969 			((struct mthca_data_seg *) wqe)->byte_count =
1970 				cpu_to_be32(wr->sg_list[i].length);
1971 			((struct mthca_data_seg *) wqe)->lkey =
1972 				cpu_to_be32(wr->sg_list[i].lkey);
1973 			((struct mthca_data_seg *) wqe)->addr =
1974 				cpu_to_be64(wr->sg_list[i].addr);
1975 			wqe += sizeof (struct mthca_data_seg);
1976 			size += sizeof (struct mthca_data_seg) / 16;
1977 		}
1978 
1979 		/* Add one more inline data segment for ICRC */
1980 		if (qp->transport == MLX) {
1981 			((struct mthca_data_seg *) wqe)->byte_count =
1982 				cpu_to_be32((1 << 31) | 4);
1983 			((u32 *) wqe)[1] = 0;
1984 			wqe += sizeof (struct mthca_data_seg);
1985 			size += sizeof (struct mthca_data_seg) / 16;
1986 		}
1987 
1988 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
1989 
1990 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1991 			mthca_err(dev, "opcode invalid\n");
1992 			err = -EINVAL;
1993 			*bad_wr = wr;
1994 			goto out;
1995 		}
1996 
1997 		((struct mthca_next_seg *) prev_wqe)->nda_op =
1998 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
1999 				     qp->send_wqe_offset) |
2000 				    mthca_opcode[wr->opcode]);
2001 		wmb();
2002 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
2003 			cpu_to_be32(MTHCA_NEXT_DBD | size |
2004 				    ((wr->send_flags & IB_SEND_FENCE) ?
2005 				     MTHCA_NEXT_FENCE : 0));
2006 
2007 		if (!size0) {
2008 			size0 = size;
2009 			op0   = mthca_opcode[wr->opcode];
2010 		}
2011 
2012 		++ind;
2013 		if (unlikely(ind >= qp->sq.max))
2014 			ind -= qp->sq.max;
2015 	}
2016 
2017 out:
2018 	if (likely(nreq)) {
2019 		doorbell[0] = cpu_to_be32((nreq << 24)                  |
2020 					  ((qp->sq.head & 0xffff) << 8) |
2021 					  f0 | op0);
2022 		doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
2023 
2024 		qp->sq.head += nreq;
2025 
2026 		/*
2027 		 * Make sure that descriptors are written before
2028 		 * doorbell record.
2029 		 */
2030 		wmb();
2031 		*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2032 
2033 		/*
2034 		 * Make sure doorbell record is written before we
2035 		 * write MMIO send doorbell.
2036 		 */
2037 		wmb();
2038 		mthca_write64(doorbell,
2039 			      dev->kar + MTHCA_SEND_DOORBELL,
2040 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2041 	}
2042 
2043 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2044 	return err;
2045 }
2046 
2047 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2048 			     struct ib_recv_wr **bad_wr)
2049 {
2050 	struct mthca_dev *dev = to_mdev(ibqp->device);
2051 	struct mthca_qp *qp = to_mqp(ibqp);
2052 	unsigned long flags;
2053 	int err = 0;
2054 	int nreq;
2055 	int ind;
2056 	int i;
2057 	void *wqe;
2058 
2059 	spin_lock_irqsave(&qp->rq.lock, flags);
2060 
2061 	/* XXX check that state is OK to post receive */
2062 
2063 	ind = qp->rq.head & (qp->rq.max - 1);
2064 
2065 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2066 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2067 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2068 					" %d max, %d nreq)\n", qp->qpn,
2069 					qp->rq.head, qp->rq.tail,
2070 					qp->rq.max, nreq);
2071 			err = -ENOMEM;
2072 			*bad_wr = wr;
2073 			goto out;
2074 		}
2075 
2076 		wqe = get_recv_wqe(qp, ind);
2077 
2078 		((struct mthca_next_seg *) wqe)->flags = 0;
2079 
2080 		wqe += sizeof (struct mthca_next_seg);
2081 
2082 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2083 			err = -EINVAL;
2084 			*bad_wr = wr;
2085 			goto out;
2086 		}
2087 
2088 		for (i = 0; i < wr->num_sge; ++i) {
2089 			((struct mthca_data_seg *) wqe)->byte_count =
2090 				cpu_to_be32(wr->sg_list[i].length);
2091 			((struct mthca_data_seg *) wqe)->lkey =
2092 				cpu_to_be32(wr->sg_list[i].lkey);
2093 			((struct mthca_data_seg *) wqe)->addr =
2094 				cpu_to_be64(wr->sg_list[i].addr);
2095 			wqe += sizeof (struct mthca_data_seg);
2096 		}
2097 
2098 		if (i < qp->rq.max_gs) {
2099 			((struct mthca_data_seg *) wqe)->byte_count = 0;
2100 			((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
2101 			((struct mthca_data_seg *) wqe)->addr = 0;
2102 		}
2103 
2104 		qp->wrid[ind] = wr->wr_id;
2105 
2106 		++ind;
2107 		if (unlikely(ind >= qp->rq.max))
2108 			ind -= qp->rq.max;
2109 	}
2110 out:
2111 	if (likely(nreq)) {
2112 		qp->rq.head += nreq;
2113 
2114 		/*
2115 		 * Make sure that descriptors are written before
2116 		 * doorbell record.
2117 		 */
2118 		wmb();
2119 		*qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2120 	}
2121 
2122 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2123 	return err;
2124 }
2125 
2126 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2127 			int index, int *dbd, __be32 *new_wqe)
2128 {
2129 	struct mthca_next_seg *next;
2130 
2131 	/*
2132 	 * For SRQs, all WQEs generate a CQE, so we're always at the
2133 	 * end of the doorbell chain.
2134 	 */
2135 	if (qp->ibqp.srq) {
2136 		*new_wqe = 0;
2137 		return;
2138 	}
2139 
2140 	if (is_send)
2141 		next = get_send_wqe(qp, index);
2142 	else
2143 		next = get_recv_wqe(qp, index);
2144 
2145 	*dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2146 	if (next->ee_nds & cpu_to_be32(0x3f))
2147 		*new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2148 			(next->ee_nds & cpu_to_be32(0x3f));
2149 	else
2150 		*new_wqe = 0;
2151 }
2152 
2153 int __devinit mthca_init_qp_table(struct mthca_dev *dev)
2154 {
2155 	int err;
2156 	u8 status;
2157 	int i;
2158 
2159 	spin_lock_init(&dev->qp_table.lock);
2160 
2161 	/*
2162 	 * We reserve 2 extra QPs per port for the special QPs.  The
2163 	 * special QP for port 1 has to be even, so round up.
2164 	 */
2165 	dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2166 	err = mthca_alloc_init(&dev->qp_table.alloc,
2167 			       dev->limits.num_qps,
2168 			       (1 << 24) - 1,
2169 			       dev->qp_table.sqp_start +
2170 			       MTHCA_MAX_PORTS * 2);
2171 	if (err)
2172 		return err;
2173 
2174 	err = mthca_array_init(&dev->qp_table.qp,
2175 			       dev->limits.num_qps);
2176 	if (err) {
2177 		mthca_alloc_cleanup(&dev->qp_table.alloc);
2178 		return err;
2179 	}
2180 
2181 	for (i = 0; i < 2; ++i) {
2182 		err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2183 					    dev->qp_table.sqp_start + i * 2,
2184 					    &status);
2185 		if (err)
2186 			goto err_out;
2187 		if (status) {
2188 			mthca_warn(dev, "CONF_SPECIAL_QP returned "
2189 				   "status %02x, aborting.\n",
2190 				   status);
2191 			err = -EINVAL;
2192 			goto err_out;
2193 		}
2194 	}
2195 	return 0;
2196 
2197  err_out:
2198 	for (i = 0; i < 2; ++i)
2199 		mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2200 
2201 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2202 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2203 
2204 	return err;
2205 }
2206 
2207 void mthca_cleanup_qp_table(struct mthca_dev *dev)
2208 {
2209 	int i;
2210 	u8 status;
2211 
2212 	for (i = 0; i < 2; ++i)
2213 		mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2214 
2215 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2216 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2217 }
2218