xref: /linux/drivers/infiniband/hw/mthca/mthca_qp.c (revision 7b12b9137930eb821b68e1bfa11e9de692208620)
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 	if (event_type == IB_EVENT_PATH_MIG)
252 		qp->port = qp->alt_port;
253 
254 	event.device      = &dev->ib_dev;
255 	event.event       = event_type;
256 	event.element.qp  = &qp->ibqp;
257 	if (qp->ibqp.event_handler)
258 		qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
259 
260 	if (atomic_dec_and_test(&qp->refcount))
261 		wake_up(&qp->wait);
262 }
263 
264 static int to_mthca_state(enum ib_qp_state ib_state)
265 {
266 	switch (ib_state) {
267 	case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
268 	case IB_QPS_INIT:  return MTHCA_QP_STATE_INIT;
269 	case IB_QPS_RTR:   return MTHCA_QP_STATE_RTR;
270 	case IB_QPS_RTS:   return MTHCA_QP_STATE_RTS;
271 	case IB_QPS_SQD:   return MTHCA_QP_STATE_SQD;
272 	case IB_QPS_SQE:   return MTHCA_QP_STATE_SQE;
273 	case IB_QPS_ERR:   return MTHCA_QP_STATE_ERR;
274 	default:                return -1;
275 	}
276 }
277 
278 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
279 
280 static int to_mthca_st(int transport)
281 {
282 	switch (transport) {
283 	case RC:  return MTHCA_QP_ST_RC;
284 	case UC:  return MTHCA_QP_ST_UC;
285 	case UD:  return MTHCA_QP_ST_UD;
286 	case RD:  return MTHCA_QP_ST_RD;
287 	case MLX: return MTHCA_QP_ST_MLX;
288 	default:  return -1;
289 	}
290 }
291 
292 static void store_attrs(struct mthca_sqp *sqp, struct ib_qp_attr *attr,
293 			int attr_mask)
294 {
295 	if (attr_mask & IB_QP_PKEY_INDEX)
296 		sqp->pkey_index = attr->pkey_index;
297 	if (attr_mask & IB_QP_QKEY)
298 		sqp->qkey = attr->qkey;
299 	if (attr_mask & IB_QP_SQ_PSN)
300 		sqp->send_psn = attr->sq_psn;
301 }
302 
303 static void init_port(struct mthca_dev *dev, int port)
304 {
305 	int err;
306 	u8 status;
307 	struct mthca_init_ib_param param;
308 
309 	memset(&param, 0, sizeof param);
310 
311 	param.port_width = dev->limits.port_width_cap;
312 	param.vl_cap     = dev->limits.vl_cap;
313 	param.mtu_cap    = dev->limits.mtu_cap;
314 	param.gid_cap    = dev->limits.gid_table_len;
315 	param.pkey_cap   = dev->limits.pkey_table_len;
316 
317 	err = mthca_INIT_IB(dev, &param, port, &status);
318 	if (err)
319 		mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
320 	if (status)
321 		mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
322 }
323 
324 static __be32 get_hw_access_flags(struct mthca_qp *qp, struct ib_qp_attr *attr,
325 				  int attr_mask)
326 {
327 	u8 dest_rd_atomic;
328 	u32 access_flags;
329 	u32 hw_access_flags = 0;
330 
331 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
332 		dest_rd_atomic = attr->max_dest_rd_atomic;
333 	else
334 		dest_rd_atomic = qp->resp_depth;
335 
336 	if (attr_mask & IB_QP_ACCESS_FLAGS)
337 		access_flags = attr->qp_access_flags;
338 	else
339 		access_flags = qp->atomic_rd_en;
340 
341 	if (!dest_rd_atomic)
342 		access_flags &= IB_ACCESS_REMOTE_WRITE;
343 
344 	if (access_flags & IB_ACCESS_REMOTE_READ)
345 		hw_access_flags |= MTHCA_QP_BIT_RRE;
346 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
347 		hw_access_flags |= MTHCA_QP_BIT_RAE;
348 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
349 		hw_access_flags |= MTHCA_QP_BIT_RWE;
350 
351 	return cpu_to_be32(hw_access_flags);
352 }
353 
354 static inline enum ib_qp_state to_ib_qp_state(int mthca_state)
355 {
356 	switch (mthca_state) {
357 	case MTHCA_QP_STATE_RST:      return IB_QPS_RESET;
358 	case MTHCA_QP_STATE_INIT:     return IB_QPS_INIT;
359 	case MTHCA_QP_STATE_RTR:      return IB_QPS_RTR;
360 	case MTHCA_QP_STATE_RTS:      return IB_QPS_RTS;
361 	case MTHCA_QP_STATE_DRAINING:
362 	case MTHCA_QP_STATE_SQD:      return IB_QPS_SQD;
363 	case MTHCA_QP_STATE_SQE:      return IB_QPS_SQE;
364 	case MTHCA_QP_STATE_ERR:      return IB_QPS_ERR;
365 	default:                      return -1;
366 	}
367 }
368 
369 static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state)
370 {
371 	switch (mthca_mig_state) {
372 	case 0:  return IB_MIG_ARMED;
373 	case 1:  return IB_MIG_REARM;
374 	case 3:  return IB_MIG_MIGRATED;
375 	default: return -1;
376 	}
377 }
378 
379 static int to_ib_qp_access_flags(int mthca_flags)
380 {
381 	int ib_flags = 0;
382 
383 	if (mthca_flags & MTHCA_QP_BIT_RRE)
384 		ib_flags |= IB_ACCESS_REMOTE_READ;
385 	if (mthca_flags & MTHCA_QP_BIT_RWE)
386 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
387 	if (mthca_flags & MTHCA_QP_BIT_RAE)
388 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
389 
390 	return ib_flags;
391 }
392 
393 static void to_ib_ah_attr(struct mthca_dev *dev, struct ib_ah_attr *ib_ah_attr,
394 				struct mthca_qp_path *path)
395 {
396 	memset(ib_ah_attr, 0, sizeof *path);
397 	ib_ah_attr->port_num 	  = (be32_to_cpu(path->port_pkey) >> 24) & 0x3;
398 
399 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->limits.num_ports)
400 		return;
401 
402 	ib_ah_attr->dlid     	  = be16_to_cpu(path->rlid);
403 	ib_ah_attr->sl       	  = be32_to_cpu(path->sl_tclass_flowlabel) >> 28;
404 	ib_ah_attr->src_path_bits = path->g_mylmc & 0x7f;
405 	ib_ah_attr->static_rate   = mthca_rate_to_ib(dev,
406 						     path->static_rate & 0x7,
407 						     ib_ah_attr->port_num);
408 	ib_ah_attr->ah_flags      = (path->g_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
409 	if (ib_ah_attr->ah_flags) {
410 		ib_ah_attr->grh.sgid_index = path->mgid_index & (dev->limits.gid_table_len - 1);
411 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
412 		ib_ah_attr->grh.traffic_class =
413 			(be32_to_cpu(path->sl_tclass_flowlabel) >> 20) & 0xff;
414 		ib_ah_attr->grh.flow_label =
415 			be32_to_cpu(path->sl_tclass_flowlabel) & 0xfffff;
416 		memcpy(ib_ah_attr->grh.dgid.raw,
417 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
418 	}
419 }
420 
421 int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
422 		   struct ib_qp_init_attr *qp_init_attr)
423 {
424 	struct mthca_dev *dev = to_mdev(ibqp->device);
425 	struct mthca_qp *qp = to_mqp(ibqp);
426 	int err;
427 	struct mthca_mailbox *mailbox;
428 	struct mthca_qp_param *qp_param;
429 	struct mthca_qp_context *context;
430 	int mthca_state;
431 	u8 status;
432 
433 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
434 	if (IS_ERR(mailbox))
435 		return PTR_ERR(mailbox);
436 
437 	err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox, &status);
438 	if (err)
439 		goto out;
440 	if (status) {
441 		mthca_warn(dev, "QUERY_QP returned status %02x\n", status);
442 		err = -EINVAL;
443 		goto out;
444 	}
445 
446 	qp_param    = mailbox->buf;
447 	context     = &qp_param->context;
448 	mthca_state = be32_to_cpu(context->flags) >> 28;
449 
450 	qp_attr->qp_state 	     = to_ib_qp_state(mthca_state);
451 	qp_attr->cur_qp_state 	     = qp_attr->qp_state;
452 	qp_attr->path_mtu 	     = context->mtu_msgmax >> 5;
453 	qp_attr->path_mig_state      =
454 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
455 	qp_attr->qkey 		     = be32_to_cpu(context->qkey);
456 	qp_attr->rq_psn 	     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
457 	qp_attr->sq_psn 	     = be32_to_cpu(context->next_send_psn) & 0xffffff;
458 	qp_attr->dest_qp_num 	     = be32_to_cpu(context->remote_qpn) & 0xffffff;
459 	qp_attr->qp_access_flags     =
460 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
461 	qp_attr->cap.max_send_wr     = qp->sq.max;
462 	qp_attr->cap.max_recv_wr     = qp->rq.max;
463 	qp_attr->cap.max_send_sge    = qp->sq.max_gs;
464 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
465 	qp_attr->cap.max_inline_data = qp->max_inline_data;
466 
467 	if (qp->transport == RC || qp->transport == UC) {
468 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
469 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
470 	}
471 
472 	qp_attr->pkey_index     = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
473 	qp_attr->alt_pkey_index = be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
474 
475 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
476 	qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
477 
478 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
479 
480 	qp_attr->max_dest_rd_atomic =
481 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
482 	qp_attr->min_rnr_timer 	    =
483 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
484 	qp_attr->port_num 	    = qp_attr->ah_attr.port_num;
485 	qp_attr->timeout 	    = context->pri_path.ackto >> 3;
486 	qp_attr->retry_cnt 	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
487 	qp_attr->rnr_retry 	    = context->pri_path.rnr_retry >> 5;
488 	qp_attr->alt_port_num 	    = qp_attr->alt_ah_attr.port_num;
489 	qp_attr->alt_timeout 	    = context->alt_path.ackto >> 3;
490 	qp_init_attr->cap 	    = qp_attr->cap;
491 
492 out:
493 	mthca_free_mailbox(dev, mailbox);
494 	return err;
495 }
496 
497 static int mthca_path_set(struct mthca_dev *dev, struct ib_ah_attr *ah,
498 			  struct mthca_qp_path *path, u8 port)
499 {
500 	path->g_mylmc     = ah->src_path_bits & 0x7f;
501 	path->rlid        = cpu_to_be16(ah->dlid);
502 	path->static_rate = mthca_get_rate(dev, ah->static_rate, port);
503 
504 	if (ah->ah_flags & IB_AH_GRH) {
505 		if (ah->grh.sgid_index >= dev->limits.gid_table_len) {
506 			mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
507 				  ah->grh.sgid_index, dev->limits.gid_table_len-1);
508 			return -1;
509 		}
510 
511 		path->g_mylmc   |= 1 << 7;
512 		path->mgid_index = ah->grh.sgid_index;
513 		path->hop_limit  = ah->grh.hop_limit;
514 		path->sl_tclass_flowlabel =
515 			cpu_to_be32((ah->sl << 28)                |
516 				    (ah->grh.traffic_class << 20) |
517 				    (ah->grh.flow_label));
518 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
519 	} else
520 		path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28);
521 
522 	return 0;
523 }
524 
525 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
526 {
527 	struct mthca_dev *dev = to_mdev(ibqp->device);
528 	struct mthca_qp *qp = to_mqp(ibqp);
529 	enum ib_qp_state cur_state, new_state;
530 	struct mthca_mailbox *mailbox;
531 	struct mthca_qp_param *qp_param;
532 	struct mthca_qp_context *qp_context;
533 	u32 sqd_event = 0;
534 	u8 status;
535 	int err;
536 
537 	if (attr_mask & IB_QP_CUR_STATE) {
538 		cur_state = attr->cur_qp_state;
539 	} else {
540 		spin_lock_irq(&qp->sq.lock);
541 		spin_lock(&qp->rq.lock);
542 		cur_state = qp->state;
543 		spin_unlock(&qp->rq.lock);
544 		spin_unlock_irq(&qp->sq.lock);
545 	}
546 
547 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
548 
549 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
550 		mthca_dbg(dev, "Bad QP transition (transport %d) "
551 			  "%d->%d with attr 0x%08x\n",
552 			  qp->transport, cur_state, new_state,
553 			  attr_mask);
554 		return -EINVAL;
555 	}
556 
557 	if ((attr_mask & IB_QP_PKEY_INDEX) &&
558 	     attr->pkey_index >= dev->limits.pkey_table_len) {
559 		mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
560 			  attr->pkey_index, dev->limits.pkey_table_len-1);
561 		return -EINVAL;
562 	}
563 
564 	if ((attr_mask & IB_QP_PORT) &&
565 	    (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
566 		mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
567 		return -EINVAL;
568 	}
569 
570 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
571 	    attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
572 		mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
573 			  attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
574 		return -EINVAL;
575 	}
576 
577 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
578 	    attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
579 		mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
580 			  attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
581 		return -EINVAL;
582 	}
583 
584 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
585 	if (IS_ERR(mailbox))
586 		return PTR_ERR(mailbox);
587 	qp_param = mailbox->buf;
588 	qp_context = &qp_param->context;
589 	memset(qp_param, 0, sizeof *qp_param);
590 
591 	qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
592 					     (to_mthca_st(qp->transport) << 16));
593 	qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
594 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
595 		qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
596 	else {
597 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
598 		switch (attr->path_mig_state) {
599 		case IB_MIG_MIGRATED:
600 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
601 			break;
602 		case IB_MIG_REARM:
603 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
604 			break;
605 		case IB_MIG_ARMED:
606 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
607 			break;
608 		}
609 	}
610 
611 	/* leave tavor_sched_queue as 0 */
612 
613 	if (qp->transport == MLX || qp->transport == UD)
614 		qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
615 	else if (attr_mask & IB_QP_PATH_MTU) {
616 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
617 			mthca_dbg(dev, "path MTU (%u) is invalid\n",
618 				  attr->path_mtu);
619 			return -EINVAL;
620 		}
621 		qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
622 	}
623 
624 	if (mthca_is_memfree(dev)) {
625 		if (qp->rq.max)
626 			qp_context->rq_size_stride = long_log2(qp->rq.max) << 3;
627 		qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
628 
629 		if (qp->sq.max)
630 			qp_context->sq_size_stride = long_log2(qp->sq.max) << 3;
631 		qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
632 	}
633 
634 	/* leave arbel_sched_queue as 0 */
635 
636 	if (qp->ibqp.uobject)
637 		qp_context->usr_page =
638 			cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
639 	else
640 		qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
641 	qp_context->local_qpn  = cpu_to_be32(qp->qpn);
642 	if (attr_mask & IB_QP_DEST_QPN) {
643 		qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
644 	}
645 
646 	if (qp->transport == MLX)
647 		qp_context->pri_path.port_pkey |=
648 			cpu_to_be32(qp->port << 24);
649 	else {
650 		if (attr_mask & IB_QP_PORT) {
651 			qp_context->pri_path.port_pkey |=
652 				cpu_to_be32(attr->port_num << 24);
653 			qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
654 		}
655 	}
656 
657 	if (attr_mask & IB_QP_PKEY_INDEX) {
658 		qp_context->pri_path.port_pkey |=
659 			cpu_to_be32(attr->pkey_index);
660 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
661 	}
662 
663 	if (attr_mask & IB_QP_RNR_RETRY) {
664 		qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
665 			attr->rnr_retry << 5;
666 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
667 							MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
668 	}
669 
670 	if (attr_mask & IB_QP_AV) {
671 		if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
672 				   attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
673 			return -EINVAL;
674 
675 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
676 	}
677 
678 	if (attr_mask & IB_QP_TIMEOUT) {
679 		qp_context->pri_path.ackto = attr->timeout << 3;
680 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
681 	}
682 
683 	if (attr_mask & IB_QP_ALT_PATH) {
684 		if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
685 			mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
686 				  attr->alt_pkey_index, dev->limits.pkey_table_len-1);
687 			return -EINVAL;
688 		}
689 
690 		if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
691 			mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
692 				attr->alt_port_num);
693 			return -EINVAL;
694 		}
695 
696 		if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
697 				   attr->alt_ah_attr.port_num))
698 			return -EINVAL;
699 
700 		qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
701 							      attr->alt_port_num << 24);
702 		qp_context->alt_path.ackto = attr->alt_timeout << 3;
703 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
704 	}
705 
706 	/* leave rdd as 0 */
707 	qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
708 	/* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
709 	qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
710 	qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
711 					     (MTHCA_FLIGHT_LIMIT << 24) |
712 					     MTHCA_QP_BIT_SWE);
713 	if (qp->sq_policy == IB_SIGNAL_ALL_WR)
714 		qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
715 	if (attr_mask & IB_QP_RETRY_CNT) {
716 		qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
717 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
718 	}
719 
720 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
721 		if (attr->max_rd_atomic) {
722 			qp_context->params1 |=
723 				cpu_to_be32(MTHCA_QP_BIT_SRE |
724 					    MTHCA_QP_BIT_SAE);
725 			qp_context->params1 |=
726 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
727 		}
728 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
729 	}
730 
731 	if (attr_mask & IB_QP_SQ_PSN)
732 		qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
733 	qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
734 
735 	if (mthca_is_memfree(dev)) {
736 		qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
737 		qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
738 	}
739 
740 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
741 		if (attr->max_dest_rd_atomic)
742 			qp_context->params2 |=
743 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
744 
745 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
746 	}
747 
748 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
749 		qp_context->params2      |= get_hw_access_flags(qp, attr, attr_mask);
750 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
751 							MTHCA_QP_OPTPAR_RRE |
752 							MTHCA_QP_OPTPAR_RAE);
753 	}
754 
755 	qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
756 
757 	if (ibqp->srq)
758 		qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
759 
760 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
761 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
762 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
763 	}
764 	if (attr_mask & IB_QP_RQ_PSN)
765 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
766 
767 	qp_context->ra_buff_indx =
768 		cpu_to_be32(dev->qp_table.rdb_base +
769 			    ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
770 			     dev->qp_table.rdb_shift));
771 
772 	qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
773 
774 	if (mthca_is_memfree(dev))
775 		qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
776 
777 	if (attr_mask & IB_QP_QKEY) {
778 		qp_context->qkey = cpu_to_be32(attr->qkey);
779 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
780 	}
781 
782 	if (ibqp->srq)
783 		qp_context->srqn = cpu_to_be32(1 << 24 |
784 					       to_msrq(ibqp->srq)->srqn);
785 
786 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
787 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY		&&
788 	    attr->en_sqd_async_notify)
789 		sqd_event = 1 << 31;
790 
791 	err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
792 			      mailbox, sqd_event, &status);
793 	if (err)
794 		goto out;
795 	if (status) {
796 		mthca_warn(dev, "modify QP %d->%d returned status %02x.\n",
797 			   cur_state, new_state, status);
798 		err = -EINVAL;
799 		goto out;
800 	}
801 
802 	qp->state = new_state;
803 	if (attr_mask & IB_QP_ACCESS_FLAGS)
804 		qp->atomic_rd_en = attr->qp_access_flags;
805 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
806 		qp->resp_depth = attr->max_dest_rd_atomic;
807 	if (attr_mask & IB_QP_PORT)
808 		qp->port = attr->port_num;
809 	if (attr_mask & IB_QP_ALT_PATH)
810 		qp->alt_port = attr->alt_port_num;
811 
812 	if (is_sqp(dev, qp))
813 		store_attrs(to_msqp(qp), attr, attr_mask);
814 
815 	/*
816 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
817 	 * QP0 to RESET or ERROR, bring the link back down.
818 	 */
819 	if (is_qp0(dev, qp)) {
820 		if (cur_state != IB_QPS_RTR &&
821 		    new_state == IB_QPS_RTR)
822 			init_port(dev, qp->port);
823 
824 		if (cur_state != IB_QPS_RESET &&
825 		    cur_state != IB_QPS_ERR &&
826 		    (new_state == IB_QPS_RESET ||
827 		     new_state == IB_QPS_ERR))
828 			mthca_CLOSE_IB(dev, qp->port, &status);
829 	}
830 
831 	/*
832 	 * If we moved a kernel QP to RESET, clean up all old CQ
833 	 * entries and reinitialize the QP.
834 	 */
835 	if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
836 		mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
837 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
838 		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
839 			mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
840 				       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
841 
842 		mthca_wq_init(&qp->sq);
843 		qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
844 
845 		mthca_wq_init(&qp->rq);
846 		qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
847 
848 		if (mthca_is_memfree(dev)) {
849 			*qp->sq.db = 0;
850 			*qp->rq.db = 0;
851 		}
852 	}
853 
854 out:
855 	mthca_free_mailbox(dev, mailbox);
856 	return err;
857 }
858 
859 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
860 {
861 	/*
862 	 * Calculate the maximum size of WQE s/g segments, excluding
863 	 * the next segment and other non-data segments.
864 	 */
865 	int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
866 
867 	switch (qp->transport) {
868 	case MLX:
869 		max_data_size -= 2 * sizeof (struct mthca_data_seg);
870 		break;
871 
872 	case UD:
873 		if (mthca_is_memfree(dev))
874 			max_data_size -= sizeof (struct mthca_arbel_ud_seg);
875 		else
876 			max_data_size -= sizeof (struct mthca_tavor_ud_seg);
877 		break;
878 
879 	default:
880 		max_data_size -= sizeof (struct mthca_raddr_seg);
881 		break;
882 	}
883 
884 	return max_data_size;
885 }
886 
887 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
888 {
889 	/* We don't support inline data for kernel QPs (yet). */
890 	return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
891 }
892 
893 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
894 				 struct mthca_pd *pd,
895 				 struct mthca_qp *qp)
896 {
897 	int max_data_size = mthca_max_data_size(dev, qp,
898 						min(dev->limits.max_desc_sz,
899 						    1 << qp->sq.wqe_shift));
900 
901 	qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
902 
903 	qp->sq.max_gs = min_t(int, dev->limits.max_sg,
904 			      max_data_size / sizeof (struct mthca_data_seg));
905 	qp->rq.max_gs = min_t(int, dev->limits.max_sg,
906 			       (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
907 				sizeof (struct mthca_next_seg)) /
908 			       sizeof (struct mthca_data_seg));
909 }
910 
911 /*
912  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
913  * rq.max_gs and sq.max_gs must all be assigned.
914  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
915  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
916  * queue)
917  */
918 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
919 			       struct mthca_pd *pd,
920 			       struct mthca_qp *qp)
921 {
922 	int size;
923 	int err = -ENOMEM;
924 
925 	size = sizeof (struct mthca_next_seg) +
926 		qp->rq.max_gs * sizeof (struct mthca_data_seg);
927 
928 	if (size > dev->limits.max_desc_sz)
929 		return -EINVAL;
930 
931 	for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
932 	     qp->rq.wqe_shift++)
933 		; /* nothing */
934 
935 	size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
936 	switch (qp->transport) {
937 	case MLX:
938 		size += 2 * sizeof (struct mthca_data_seg);
939 		break;
940 
941 	case UD:
942 		size += mthca_is_memfree(dev) ?
943 			sizeof (struct mthca_arbel_ud_seg) :
944 			sizeof (struct mthca_tavor_ud_seg);
945 		break;
946 
947 	case UC:
948 		size += sizeof (struct mthca_raddr_seg);
949 		break;
950 
951 	case RC:
952 		size += sizeof (struct mthca_raddr_seg);
953 		/*
954 		 * An atomic op will require an atomic segment, a
955 		 * remote address segment and one scatter entry.
956 		 */
957 		size = max_t(int, size,
958 			     sizeof (struct mthca_atomic_seg) +
959 			     sizeof (struct mthca_raddr_seg) +
960 			     sizeof (struct mthca_data_seg));
961 		break;
962 
963 	default:
964 		break;
965 	}
966 
967 	/* Make sure that we have enough space for a bind request */
968 	size = max_t(int, size, sizeof (struct mthca_bind_seg));
969 
970 	size += sizeof (struct mthca_next_seg);
971 
972 	if (size > dev->limits.max_desc_sz)
973 		return -EINVAL;
974 
975 	for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
976 	     qp->sq.wqe_shift++)
977 		; /* nothing */
978 
979 	qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
980 				    1 << qp->sq.wqe_shift);
981 
982 	/*
983 	 * If this is a userspace QP, we don't actually have to
984 	 * allocate anything.  All we need is to calculate the WQE
985 	 * sizes and the send_wqe_offset, so we're done now.
986 	 */
987 	if (pd->ibpd.uobject)
988 		return 0;
989 
990 	size = PAGE_ALIGN(qp->send_wqe_offset +
991 			  (qp->sq.max << qp->sq.wqe_shift));
992 
993 	qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
994 			   GFP_KERNEL);
995 	if (!qp->wrid)
996 		goto err_out;
997 
998 	err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
999 			      &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1000 	if (err)
1001 		goto err_out;
1002 
1003 	return 0;
1004 
1005 err_out:
1006 	kfree(qp->wrid);
1007 	return err;
1008 }
1009 
1010 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1011 			       struct mthca_qp *qp)
1012 {
1013 	mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1014 				       (qp->sq.max << qp->sq.wqe_shift)),
1015 		       &qp->queue, qp->is_direct, &qp->mr);
1016 	kfree(qp->wrid);
1017 }
1018 
1019 static int mthca_map_memfree(struct mthca_dev *dev,
1020 			     struct mthca_qp *qp)
1021 {
1022 	int ret;
1023 
1024 	if (mthca_is_memfree(dev)) {
1025 		ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1026 		if (ret)
1027 			return ret;
1028 
1029 		ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1030 		if (ret)
1031 			goto err_qpc;
1032 
1033 		ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1034 				      qp->qpn << dev->qp_table.rdb_shift);
1035 		if (ret)
1036 			goto err_eqpc;
1037 
1038 	}
1039 
1040 	return 0;
1041 
1042 err_eqpc:
1043 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1044 
1045 err_qpc:
1046 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1047 
1048 	return ret;
1049 }
1050 
1051 static void mthca_unmap_memfree(struct mthca_dev *dev,
1052 				struct mthca_qp *qp)
1053 {
1054 	mthca_table_put(dev, dev->qp_table.rdb_table,
1055 			qp->qpn << dev->qp_table.rdb_shift);
1056 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1057 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1058 }
1059 
1060 static int mthca_alloc_memfree(struct mthca_dev *dev,
1061 			       struct mthca_qp *qp)
1062 {
1063 	int ret = 0;
1064 
1065 	if (mthca_is_memfree(dev)) {
1066 		qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1067 						 qp->qpn, &qp->rq.db);
1068 		if (qp->rq.db_index < 0)
1069 			return ret;
1070 
1071 		qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1072 						 qp->qpn, &qp->sq.db);
1073 		if (qp->sq.db_index < 0)
1074 			mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1075 	}
1076 
1077 	return ret;
1078 }
1079 
1080 static void mthca_free_memfree(struct mthca_dev *dev,
1081 			       struct mthca_qp *qp)
1082 {
1083 	if (mthca_is_memfree(dev)) {
1084 		mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1085 		mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1086 	}
1087 }
1088 
1089 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1090 				 struct mthca_pd *pd,
1091 				 struct mthca_cq *send_cq,
1092 				 struct mthca_cq *recv_cq,
1093 				 enum ib_sig_type send_policy,
1094 				 struct mthca_qp *qp)
1095 {
1096 	int ret;
1097 	int i;
1098 
1099 	atomic_set(&qp->refcount, 1);
1100 	init_waitqueue_head(&qp->wait);
1101 	qp->state    	 = IB_QPS_RESET;
1102 	qp->atomic_rd_en = 0;
1103 	qp->resp_depth   = 0;
1104 	qp->sq_policy    = send_policy;
1105 	mthca_wq_init(&qp->sq);
1106 	mthca_wq_init(&qp->rq);
1107 
1108 	ret = mthca_map_memfree(dev, qp);
1109 	if (ret)
1110 		return ret;
1111 
1112 	ret = mthca_alloc_wqe_buf(dev, pd, qp);
1113 	if (ret) {
1114 		mthca_unmap_memfree(dev, qp);
1115 		return ret;
1116 	}
1117 
1118 	mthca_adjust_qp_caps(dev, pd, qp);
1119 
1120 	/*
1121 	 * If this is a userspace QP, we're done now.  The doorbells
1122 	 * will be allocated and buffers will be initialized in
1123 	 * userspace.
1124 	 */
1125 	if (pd->ibpd.uobject)
1126 		return 0;
1127 
1128 	ret = mthca_alloc_memfree(dev, qp);
1129 	if (ret) {
1130 		mthca_free_wqe_buf(dev, qp);
1131 		mthca_unmap_memfree(dev, qp);
1132 		return ret;
1133 	}
1134 
1135 	if (mthca_is_memfree(dev)) {
1136 		struct mthca_next_seg *next;
1137 		struct mthca_data_seg *scatter;
1138 		int size = (sizeof (struct mthca_next_seg) +
1139 			    qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1140 
1141 		for (i = 0; i < qp->rq.max; ++i) {
1142 			next = get_recv_wqe(qp, i);
1143 			next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1144 						   qp->rq.wqe_shift);
1145 			next->ee_nds = cpu_to_be32(size);
1146 
1147 			for (scatter = (void *) (next + 1);
1148 			     (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1149 			     ++scatter)
1150 				scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1151 		}
1152 
1153 		for (i = 0; i < qp->sq.max; ++i) {
1154 			next = get_send_wqe(qp, i);
1155 			next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1156 						    qp->sq.wqe_shift) +
1157 						   qp->send_wqe_offset);
1158 		}
1159 	}
1160 
1161 	qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1162 	qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1163 
1164 	return 0;
1165 }
1166 
1167 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1168 			     struct mthca_pd *pd, struct mthca_qp *qp)
1169 {
1170 	int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1171 
1172 	/* Sanity check QP size before proceeding */
1173 	if (cap->max_send_wr  	 > dev->limits.max_wqes ||
1174 	    cap->max_recv_wr  	 > dev->limits.max_wqes ||
1175 	    cap->max_send_sge 	 > dev->limits.max_sg   ||
1176 	    cap->max_recv_sge 	 > dev->limits.max_sg   ||
1177 	    cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1178 		return -EINVAL;
1179 
1180 	/*
1181 	 * For MLX transport we need 2 extra S/G entries:
1182 	 * one for the header and one for the checksum at the end
1183 	 */
1184 	if (qp->transport == MLX && cap->max_recv_sge + 2 > dev->limits.max_sg)
1185 		return -EINVAL;
1186 
1187 	if (mthca_is_memfree(dev)) {
1188 		qp->rq.max = cap->max_recv_wr ?
1189 			roundup_pow_of_two(cap->max_recv_wr) : 0;
1190 		qp->sq.max = cap->max_send_wr ?
1191 			roundup_pow_of_two(cap->max_send_wr) : 0;
1192 	} else {
1193 		qp->rq.max = cap->max_recv_wr;
1194 		qp->sq.max = cap->max_send_wr;
1195 	}
1196 
1197 	qp->rq.max_gs = cap->max_recv_sge;
1198 	qp->sq.max_gs = max_t(int, cap->max_send_sge,
1199 			      ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1200 				    MTHCA_INLINE_CHUNK_SIZE) /
1201 			      sizeof (struct mthca_data_seg));
1202 
1203 	return 0;
1204 }
1205 
1206 int mthca_alloc_qp(struct mthca_dev *dev,
1207 		   struct mthca_pd *pd,
1208 		   struct mthca_cq *send_cq,
1209 		   struct mthca_cq *recv_cq,
1210 		   enum ib_qp_type type,
1211 		   enum ib_sig_type send_policy,
1212 		   struct ib_qp_cap *cap,
1213 		   struct mthca_qp *qp)
1214 {
1215 	int err;
1216 
1217 	switch (type) {
1218 	case IB_QPT_RC: qp->transport = RC; break;
1219 	case IB_QPT_UC: qp->transport = UC; break;
1220 	case IB_QPT_UD: qp->transport = UD; break;
1221 	default: return -EINVAL;
1222 	}
1223 
1224 	err = mthca_set_qp_size(dev, cap, pd, qp);
1225 	if (err)
1226 		return err;
1227 
1228 	qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1229 	if (qp->qpn == -1)
1230 		return -ENOMEM;
1231 
1232 	/* initialize port to zero for error-catching. */
1233 	qp->port = 0;
1234 
1235 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1236 				    send_policy, qp);
1237 	if (err) {
1238 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1239 		return err;
1240 	}
1241 
1242 	spin_lock_irq(&dev->qp_table.lock);
1243 	mthca_array_set(&dev->qp_table.qp,
1244 			qp->qpn & (dev->limits.num_qps - 1), qp);
1245 	spin_unlock_irq(&dev->qp_table.lock);
1246 
1247 	return 0;
1248 }
1249 
1250 int mthca_alloc_sqp(struct mthca_dev *dev,
1251 		    struct mthca_pd *pd,
1252 		    struct mthca_cq *send_cq,
1253 		    struct mthca_cq *recv_cq,
1254 		    enum ib_sig_type send_policy,
1255 		    struct ib_qp_cap *cap,
1256 		    int qpn,
1257 		    int port,
1258 		    struct mthca_sqp *sqp)
1259 {
1260 	u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1261 	int err;
1262 
1263 	sqp->qp.transport = MLX;
1264 	err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1265 	if (err)
1266 		return err;
1267 
1268 	sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1269 	sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1270 					     &sqp->header_dma, GFP_KERNEL);
1271 	if (!sqp->header_buf)
1272 		return -ENOMEM;
1273 
1274 	spin_lock_irq(&dev->qp_table.lock);
1275 	if (mthca_array_get(&dev->qp_table.qp, mqpn))
1276 		err = -EBUSY;
1277 	else
1278 		mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1279 	spin_unlock_irq(&dev->qp_table.lock);
1280 
1281 	if (err)
1282 		goto err_out;
1283 
1284 	sqp->qp.port      = port;
1285 	sqp->qp.qpn       = mqpn;
1286 	sqp->qp.transport = MLX;
1287 
1288 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1289 				    send_policy, &sqp->qp);
1290 	if (err)
1291 		goto err_out_free;
1292 
1293 	atomic_inc(&pd->sqp_count);
1294 
1295 	return 0;
1296 
1297  err_out_free:
1298 	/*
1299 	 * Lock CQs here, so that CQ polling code can do QP lookup
1300 	 * without taking a lock.
1301 	 */
1302 	spin_lock_irq(&send_cq->lock);
1303 	if (send_cq != recv_cq)
1304 		spin_lock(&recv_cq->lock);
1305 
1306 	spin_lock(&dev->qp_table.lock);
1307 	mthca_array_clear(&dev->qp_table.qp, mqpn);
1308 	spin_unlock(&dev->qp_table.lock);
1309 
1310 	if (send_cq != recv_cq)
1311 		spin_unlock(&recv_cq->lock);
1312 	spin_unlock_irq(&send_cq->lock);
1313 
1314  err_out:
1315 	dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1316 			  sqp->header_buf, sqp->header_dma);
1317 
1318 	return err;
1319 }
1320 
1321 void mthca_free_qp(struct mthca_dev *dev,
1322 		   struct mthca_qp *qp)
1323 {
1324 	u8 status;
1325 	struct mthca_cq *send_cq;
1326 	struct mthca_cq *recv_cq;
1327 
1328 	send_cq = to_mcq(qp->ibqp.send_cq);
1329 	recv_cq = to_mcq(qp->ibqp.recv_cq);
1330 
1331 	/*
1332 	 * Lock CQs here, so that CQ polling code can do QP lookup
1333 	 * without taking a lock.
1334 	 */
1335 	spin_lock_irq(&send_cq->lock);
1336 	if (send_cq != recv_cq)
1337 		spin_lock(&recv_cq->lock);
1338 
1339 	spin_lock(&dev->qp_table.lock);
1340 	mthca_array_clear(&dev->qp_table.qp,
1341 			  qp->qpn & (dev->limits.num_qps - 1));
1342 	spin_unlock(&dev->qp_table.lock);
1343 
1344 	if (send_cq != recv_cq)
1345 		spin_unlock(&recv_cq->lock);
1346 	spin_unlock_irq(&send_cq->lock);
1347 
1348 	atomic_dec(&qp->refcount);
1349 	wait_event(qp->wait, !atomic_read(&qp->refcount));
1350 
1351 	if (qp->state != IB_QPS_RESET)
1352 		mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1353 				NULL, 0, &status);
1354 
1355 	/*
1356 	 * If this is a userspace QP, the buffers, MR, CQs and so on
1357 	 * will be cleaned up in userspace, so all we have to do is
1358 	 * unref the mem-free tables and free the QPN in our table.
1359 	 */
1360 	if (!qp->ibqp.uobject) {
1361 		mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
1362 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1363 		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
1364 			mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
1365 				       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1366 
1367 		mthca_free_memfree(dev, qp);
1368 		mthca_free_wqe_buf(dev, qp);
1369 	}
1370 
1371 	mthca_unmap_memfree(dev, qp);
1372 
1373 	if (is_sqp(dev, qp)) {
1374 		atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1375 		dma_free_coherent(&dev->pdev->dev,
1376 				  to_msqp(qp)->header_buf_size,
1377 				  to_msqp(qp)->header_buf,
1378 				  to_msqp(qp)->header_dma);
1379 	} else
1380 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1381 }
1382 
1383 /* Create UD header for an MLX send and build a data segment for it */
1384 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1385 			    int ind, struct ib_send_wr *wr,
1386 			    struct mthca_mlx_seg *mlx,
1387 			    struct mthca_data_seg *data)
1388 {
1389 	int header_size;
1390 	int err;
1391 	u16 pkey;
1392 
1393 	ib_ud_header_init(256, /* assume a MAD */
1394 			  mthca_ah_grh_present(to_mah(wr->wr.ud.ah)),
1395 			  &sqp->ud_header);
1396 
1397 	err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1398 	if (err)
1399 		return err;
1400 	mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1401 	mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1402 				  (sqp->ud_header.lrh.destination_lid ==
1403 				   IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1404 				  (sqp->ud_header.lrh.service_level << 8));
1405 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
1406 	mlx->vcrc = 0;
1407 
1408 	switch (wr->opcode) {
1409 	case IB_WR_SEND:
1410 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1411 		sqp->ud_header.immediate_present = 0;
1412 		break;
1413 	case IB_WR_SEND_WITH_IMM:
1414 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1415 		sqp->ud_header.immediate_present = 1;
1416 		sqp->ud_header.immediate_data = wr->imm_data;
1417 		break;
1418 	default:
1419 		return -EINVAL;
1420 	}
1421 
1422 	sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1423 	if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1424 		sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1425 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1426 	if (!sqp->qp.ibqp.qp_num)
1427 		ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1428 				   sqp->pkey_index, &pkey);
1429 	else
1430 		ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1431 				   wr->wr.ud.pkey_index, &pkey);
1432 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1433 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1434 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1435 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1436 					       sqp->qkey : wr->wr.ud.remote_qkey);
1437 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1438 
1439 	header_size = ib_ud_header_pack(&sqp->ud_header,
1440 					sqp->header_buf +
1441 					ind * MTHCA_UD_HEADER_SIZE);
1442 
1443 	data->byte_count = cpu_to_be32(header_size);
1444 	data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1445 	data->addr       = cpu_to_be64(sqp->header_dma +
1446 				       ind * MTHCA_UD_HEADER_SIZE);
1447 
1448 	return 0;
1449 }
1450 
1451 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1452 				    struct ib_cq *ib_cq)
1453 {
1454 	unsigned cur;
1455 	struct mthca_cq *cq;
1456 
1457 	cur = wq->head - wq->tail;
1458 	if (likely(cur + nreq < wq->max))
1459 		return 0;
1460 
1461 	cq = to_mcq(ib_cq);
1462 	spin_lock(&cq->lock);
1463 	cur = wq->head - wq->tail;
1464 	spin_unlock(&cq->lock);
1465 
1466 	return cur + nreq >= wq->max;
1467 }
1468 
1469 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1470 			  struct ib_send_wr **bad_wr)
1471 {
1472 	struct mthca_dev *dev = to_mdev(ibqp->device);
1473 	struct mthca_qp *qp = to_mqp(ibqp);
1474 	void *wqe;
1475 	void *prev_wqe;
1476 	unsigned long flags;
1477 	int err = 0;
1478 	int nreq;
1479 	int i;
1480 	int size;
1481 	int size0 = 0;
1482 	u32 f0 = 0;
1483 	int ind;
1484 	u8 op0 = 0;
1485 
1486 	spin_lock_irqsave(&qp->sq.lock, flags);
1487 
1488 	/* XXX check that state is OK to post send */
1489 
1490 	ind = qp->sq.next_ind;
1491 
1492 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1493 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1494 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1495 					" %d max, %d nreq)\n", qp->qpn,
1496 					qp->sq.head, qp->sq.tail,
1497 					qp->sq.max, nreq);
1498 			err = -ENOMEM;
1499 			*bad_wr = wr;
1500 			goto out;
1501 		}
1502 
1503 		wqe = get_send_wqe(qp, ind);
1504 		prev_wqe = qp->sq.last;
1505 		qp->sq.last = wqe;
1506 
1507 		((struct mthca_next_seg *) wqe)->nda_op = 0;
1508 		((struct mthca_next_seg *) wqe)->ee_nds = 0;
1509 		((struct mthca_next_seg *) wqe)->flags =
1510 			((wr->send_flags & IB_SEND_SIGNALED) ?
1511 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1512 			((wr->send_flags & IB_SEND_SOLICITED) ?
1513 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1514 			cpu_to_be32(1);
1515 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1516 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1517 			((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1518 
1519 		wqe += sizeof (struct mthca_next_seg);
1520 		size = sizeof (struct mthca_next_seg) / 16;
1521 
1522 		switch (qp->transport) {
1523 		case RC:
1524 			switch (wr->opcode) {
1525 			case IB_WR_ATOMIC_CMP_AND_SWP:
1526 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1527 				((struct mthca_raddr_seg *) wqe)->raddr =
1528 					cpu_to_be64(wr->wr.atomic.remote_addr);
1529 				((struct mthca_raddr_seg *) wqe)->rkey =
1530 					cpu_to_be32(wr->wr.atomic.rkey);
1531 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1532 
1533 				wqe += sizeof (struct mthca_raddr_seg);
1534 
1535 				if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1536 					((struct mthca_atomic_seg *) wqe)->swap_add =
1537 						cpu_to_be64(wr->wr.atomic.swap);
1538 					((struct mthca_atomic_seg *) wqe)->compare =
1539 						cpu_to_be64(wr->wr.atomic.compare_add);
1540 				} else {
1541 					((struct mthca_atomic_seg *) wqe)->swap_add =
1542 						cpu_to_be64(wr->wr.atomic.compare_add);
1543 					((struct mthca_atomic_seg *) wqe)->compare = 0;
1544 				}
1545 
1546 				wqe += sizeof (struct mthca_atomic_seg);
1547 				size += (sizeof (struct mthca_raddr_seg) +
1548 					 sizeof (struct mthca_atomic_seg)) / 16;
1549 				break;
1550 
1551 			case IB_WR_RDMA_WRITE:
1552 			case IB_WR_RDMA_WRITE_WITH_IMM:
1553 			case IB_WR_RDMA_READ:
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 UC:
1571 			switch (wr->opcode) {
1572 			case IB_WR_RDMA_WRITE:
1573 			case IB_WR_RDMA_WRITE_WITH_IMM:
1574 				((struct mthca_raddr_seg *) wqe)->raddr =
1575 					cpu_to_be64(wr->wr.rdma.remote_addr);
1576 				((struct mthca_raddr_seg *) wqe)->rkey =
1577 					cpu_to_be32(wr->wr.rdma.rkey);
1578 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1579 				wqe += sizeof (struct mthca_raddr_seg);
1580 				size += sizeof (struct mthca_raddr_seg) / 16;
1581 				break;
1582 
1583 			default:
1584 				/* No extra segments required for sends */
1585 				break;
1586 			}
1587 
1588 			break;
1589 
1590 		case UD:
1591 			((struct mthca_tavor_ud_seg *) wqe)->lkey =
1592 				cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1593 			((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1594 				cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1595 			((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1596 				cpu_to_be32(wr->wr.ud.remote_qpn);
1597 			((struct mthca_tavor_ud_seg *) wqe)->qkey =
1598 				cpu_to_be32(wr->wr.ud.remote_qkey);
1599 
1600 			wqe += sizeof (struct mthca_tavor_ud_seg);
1601 			size += sizeof (struct mthca_tavor_ud_seg) / 16;
1602 			break;
1603 
1604 		case MLX:
1605 			err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1606 					       wqe - sizeof (struct mthca_next_seg),
1607 					       wqe);
1608 			if (err) {
1609 				*bad_wr = wr;
1610 				goto out;
1611 			}
1612 			wqe += sizeof (struct mthca_data_seg);
1613 			size += sizeof (struct mthca_data_seg) / 16;
1614 			break;
1615 		}
1616 
1617 		if (wr->num_sge > qp->sq.max_gs) {
1618 			mthca_err(dev, "too many gathers\n");
1619 			err = -EINVAL;
1620 			*bad_wr = wr;
1621 			goto out;
1622 		}
1623 
1624 		for (i = 0; i < wr->num_sge; ++i) {
1625 			((struct mthca_data_seg *) wqe)->byte_count =
1626 				cpu_to_be32(wr->sg_list[i].length);
1627 			((struct mthca_data_seg *) wqe)->lkey =
1628 				cpu_to_be32(wr->sg_list[i].lkey);
1629 			((struct mthca_data_seg *) wqe)->addr =
1630 				cpu_to_be64(wr->sg_list[i].addr);
1631 			wqe += sizeof (struct mthca_data_seg);
1632 			size += sizeof (struct mthca_data_seg) / 16;
1633 		}
1634 
1635 		/* Add one more inline data segment for ICRC */
1636 		if (qp->transport == MLX) {
1637 			((struct mthca_data_seg *) wqe)->byte_count =
1638 				cpu_to_be32((1 << 31) | 4);
1639 			((u32 *) wqe)[1] = 0;
1640 			wqe += sizeof (struct mthca_data_seg);
1641 			size += sizeof (struct mthca_data_seg) / 16;
1642 		}
1643 
1644 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
1645 
1646 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1647 			mthca_err(dev, "opcode invalid\n");
1648 			err = -EINVAL;
1649 			*bad_wr = wr;
1650 			goto out;
1651 		}
1652 
1653 		((struct mthca_next_seg *) prev_wqe)->nda_op =
1654 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
1655 				     qp->send_wqe_offset) |
1656 				    mthca_opcode[wr->opcode]);
1657 		wmb();
1658 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1659 			cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size |
1660 				    ((wr->send_flags & IB_SEND_FENCE) ?
1661 				    MTHCA_NEXT_FENCE : 0));
1662 
1663 		if (!size0) {
1664 			size0 = size;
1665 			op0   = mthca_opcode[wr->opcode];
1666 		}
1667 
1668 		++ind;
1669 		if (unlikely(ind >= qp->sq.max))
1670 			ind -= qp->sq.max;
1671 	}
1672 
1673 out:
1674 	if (likely(nreq)) {
1675 		__be32 doorbell[2];
1676 
1677 		doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1678 					   qp->send_wqe_offset) | f0 | op0);
1679 		doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1680 
1681 		wmb();
1682 
1683 		mthca_write64(doorbell,
1684 			      dev->kar + MTHCA_SEND_DOORBELL,
1685 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1686 	}
1687 
1688 	qp->sq.next_ind = ind;
1689 	qp->sq.head    += nreq;
1690 
1691 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1692 	return err;
1693 }
1694 
1695 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1696 			     struct ib_recv_wr **bad_wr)
1697 {
1698 	struct mthca_dev *dev = to_mdev(ibqp->device);
1699 	struct mthca_qp *qp = to_mqp(ibqp);
1700 	__be32 doorbell[2];
1701 	unsigned long flags;
1702 	int err = 0;
1703 	int nreq;
1704 	int i;
1705 	int size;
1706 	int size0 = 0;
1707 	int ind;
1708 	void *wqe;
1709 	void *prev_wqe;
1710 
1711 	spin_lock_irqsave(&qp->rq.lock, flags);
1712 
1713 	/* XXX check that state is OK to post receive */
1714 
1715 	ind = qp->rq.next_ind;
1716 
1717 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1718 		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1719 			nreq = 0;
1720 
1721 			doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1722 			doorbell[1] = cpu_to_be32(qp->qpn << 8);
1723 
1724 			wmb();
1725 
1726 			mthca_write64(doorbell,
1727 				      dev->kar + MTHCA_RECEIVE_DOORBELL,
1728 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1729 
1730 			qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1731 			size0 = 0;
1732 		}
1733 
1734 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1735 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1736 					" %d max, %d nreq)\n", qp->qpn,
1737 					qp->rq.head, qp->rq.tail,
1738 					qp->rq.max, nreq);
1739 			err = -ENOMEM;
1740 			*bad_wr = wr;
1741 			goto out;
1742 		}
1743 
1744 		wqe = get_recv_wqe(qp, ind);
1745 		prev_wqe = qp->rq.last;
1746 		qp->rq.last = wqe;
1747 
1748 		((struct mthca_next_seg *) wqe)->nda_op = 0;
1749 		((struct mthca_next_seg *) wqe)->ee_nds =
1750 			cpu_to_be32(MTHCA_NEXT_DBD);
1751 		((struct mthca_next_seg *) wqe)->flags = 0;
1752 
1753 		wqe += sizeof (struct mthca_next_seg);
1754 		size = sizeof (struct mthca_next_seg) / 16;
1755 
1756 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1757 			err = -EINVAL;
1758 			*bad_wr = wr;
1759 			goto out;
1760 		}
1761 
1762 		for (i = 0; i < wr->num_sge; ++i) {
1763 			((struct mthca_data_seg *) wqe)->byte_count =
1764 				cpu_to_be32(wr->sg_list[i].length);
1765 			((struct mthca_data_seg *) wqe)->lkey =
1766 				cpu_to_be32(wr->sg_list[i].lkey);
1767 			((struct mthca_data_seg *) wqe)->addr =
1768 				cpu_to_be64(wr->sg_list[i].addr);
1769 			wqe += sizeof (struct mthca_data_seg);
1770 			size += sizeof (struct mthca_data_seg) / 16;
1771 		}
1772 
1773 		qp->wrid[ind] = wr->wr_id;
1774 
1775 		((struct mthca_next_seg *) prev_wqe)->nda_op =
1776 			cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1777 		wmb();
1778 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1779 			cpu_to_be32(MTHCA_NEXT_DBD | size);
1780 
1781 		if (!size0)
1782 			size0 = size;
1783 
1784 		++ind;
1785 		if (unlikely(ind >= qp->rq.max))
1786 			ind -= qp->rq.max;
1787 	}
1788 
1789 out:
1790 	if (likely(nreq)) {
1791 		doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1792 		doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
1793 
1794 		wmb();
1795 
1796 		mthca_write64(doorbell,
1797 			      dev->kar + MTHCA_RECEIVE_DOORBELL,
1798 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1799 	}
1800 
1801 	qp->rq.next_ind = ind;
1802 	qp->rq.head    += nreq;
1803 
1804 	spin_unlock_irqrestore(&qp->rq.lock, flags);
1805 	return err;
1806 }
1807 
1808 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1809 			  struct ib_send_wr **bad_wr)
1810 {
1811 	struct mthca_dev *dev = to_mdev(ibqp->device);
1812 	struct mthca_qp *qp = to_mqp(ibqp);
1813 	__be32 doorbell[2];
1814 	void *wqe;
1815 	void *prev_wqe;
1816 	unsigned long flags;
1817 	int err = 0;
1818 	int nreq;
1819 	int i;
1820 	int size;
1821 	int size0 = 0;
1822 	u32 f0 = 0;
1823 	int ind;
1824 	u8 op0 = 0;
1825 
1826 	spin_lock_irqsave(&qp->sq.lock, flags);
1827 
1828 	/* XXX check that state is OK to post send */
1829 
1830 	ind = qp->sq.head & (qp->sq.max - 1);
1831 
1832 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1833 		if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1834 			nreq = 0;
1835 
1836 			doorbell[0] = cpu_to_be32((MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1837 						  ((qp->sq.head & 0xffff) << 8) |
1838 						  f0 | op0);
1839 			doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1840 
1841 			qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1842 			size0 = 0;
1843 
1844 			/*
1845 			 * Make sure that descriptors are written before
1846 			 * doorbell record.
1847 			 */
1848 			wmb();
1849 			*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1850 
1851 			/*
1852 			 * Make sure doorbell record is written before we
1853 			 * write MMIO send doorbell.
1854 			 */
1855 			wmb();
1856 			mthca_write64(doorbell,
1857 				      dev->kar + MTHCA_SEND_DOORBELL,
1858 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1859 		}
1860 
1861 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1862 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1863 					" %d max, %d nreq)\n", qp->qpn,
1864 					qp->sq.head, qp->sq.tail,
1865 					qp->sq.max, nreq);
1866 			err = -ENOMEM;
1867 			*bad_wr = wr;
1868 			goto out;
1869 		}
1870 
1871 		wqe = get_send_wqe(qp, ind);
1872 		prev_wqe = qp->sq.last;
1873 		qp->sq.last = wqe;
1874 
1875 		((struct mthca_next_seg *) wqe)->flags =
1876 			((wr->send_flags & IB_SEND_SIGNALED) ?
1877 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1878 			((wr->send_flags & IB_SEND_SOLICITED) ?
1879 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1880 			cpu_to_be32(1);
1881 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1882 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1883 			((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1884 
1885 		wqe += sizeof (struct mthca_next_seg);
1886 		size = sizeof (struct mthca_next_seg) / 16;
1887 
1888 		switch (qp->transport) {
1889 		case RC:
1890 			switch (wr->opcode) {
1891 			case IB_WR_ATOMIC_CMP_AND_SWP:
1892 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1893 				((struct mthca_raddr_seg *) wqe)->raddr =
1894 					cpu_to_be64(wr->wr.atomic.remote_addr);
1895 				((struct mthca_raddr_seg *) wqe)->rkey =
1896 					cpu_to_be32(wr->wr.atomic.rkey);
1897 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1898 
1899 				wqe += sizeof (struct mthca_raddr_seg);
1900 
1901 				if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1902 					((struct mthca_atomic_seg *) wqe)->swap_add =
1903 						cpu_to_be64(wr->wr.atomic.swap);
1904 					((struct mthca_atomic_seg *) wqe)->compare =
1905 						cpu_to_be64(wr->wr.atomic.compare_add);
1906 				} else {
1907 					((struct mthca_atomic_seg *) wqe)->swap_add =
1908 						cpu_to_be64(wr->wr.atomic.compare_add);
1909 					((struct mthca_atomic_seg *) wqe)->compare = 0;
1910 				}
1911 
1912 				wqe += sizeof (struct mthca_atomic_seg);
1913 				size += (sizeof (struct mthca_raddr_seg) +
1914 					 sizeof (struct mthca_atomic_seg)) / 16;
1915 				break;
1916 
1917 			case IB_WR_RDMA_READ:
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 UC:
1937 			switch (wr->opcode) {
1938 			case IB_WR_RDMA_WRITE:
1939 			case IB_WR_RDMA_WRITE_WITH_IMM:
1940 				((struct mthca_raddr_seg *) wqe)->raddr =
1941 					cpu_to_be64(wr->wr.rdma.remote_addr);
1942 				((struct mthca_raddr_seg *) wqe)->rkey =
1943 					cpu_to_be32(wr->wr.rdma.rkey);
1944 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1945 				wqe += sizeof (struct mthca_raddr_seg);
1946 				size += sizeof (struct mthca_raddr_seg) / 16;
1947 				break;
1948 
1949 			default:
1950 				/* No extra segments required for sends */
1951 				break;
1952 			}
1953 
1954 			break;
1955 
1956 		case UD:
1957 			memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
1958 			       to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1959 			((struct mthca_arbel_ud_seg *) wqe)->dqpn =
1960 				cpu_to_be32(wr->wr.ud.remote_qpn);
1961 			((struct mthca_arbel_ud_seg *) wqe)->qkey =
1962 				cpu_to_be32(wr->wr.ud.remote_qkey);
1963 
1964 			wqe += sizeof (struct mthca_arbel_ud_seg);
1965 			size += sizeof (struct mthca_arbel_ud_seg) / 16;
1966 			break;
1967 
1968 		case MLX:
1969 			err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1970 					       wqe - sizeof (struct mthca_next_seg),
1971 					       wqe);
1972 			if (err) {
1973 				*bad_wr = wr;
1974 				goto out;
1975 			}
1976 			wqe += sizeof (struct mthca_data_seg);
1977 			size += sizeof (struct mthca_data_seg) / 16;
1978 			break;
1979 		}
1980 
1981 		if (wr->num_sge > qp->sq.max_gs) {
1982 			mthca_err(dev, "too many gathers\n");
1983 			err = -EINVAL;
1984 			*bad_wr = wr;
1985 			goto out;
1986 		}
1987 
1988 		for (i = 0; i < wr->num_sge; ++i) {
1989 			((struct mthca_data_seg *) wqe)->byte_count =
1990 				cpu_to_be32(wr->sg_list[i].length);
1991 			((struct mthca_data_seg *) wqe)->lkey =
1992 				cpu_to_be32(wr->sg_list[i].lkey);
1993 			((struct mthca_data_seg *) wqe)->addr =
1994 				cpu_to_be64(wr->sg_list[i].addr);
1995 			wqe += sizeof (struct mthca_data_seg);
1996 			size += sizeof (struct mthca_data_seg) / 16;
1997 		}
1998 
1999 		/* Add one more inline data segment for ICRC */
2000 		if (qp->transport == MLX) {
2001 			((struct mthca_data_seg *) wqe)->byte_count =
2002 				cpu_to_be32((1 << 31) | 4);
2003 			((u32 *) wqe)[1] = 0;
2004 			wqe += sizeof (struct mthca_data_seg);
2005 			size += sizeof (struct mthca_data_seg) / 16;
2006 		}
2007 
2008 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
2009 
2010 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2011 			mthca_err(dev, "opcode invalid\n");
2012 			err = -EINVAL;
2013 			*bad_wr = wr;
2014 			goto out;
2015 		}
2016 
2017 		((struct mthca_next_seg *) prev_wqe)->nda_op =
2018 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
2019 				     qp->send_wqe_offset) |
2020 				    mthca_opcode[wr->opcode]);
2021 		wmb();
2022 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
2023 			cpu_to_be32(MTHCA_NEXT_DBD | size |
2024 				    ((wr->send_flags & IB_SEND_FENCE) ?
2025 				     MTHCA_NEXT_FENCE : 0));
2026 
2027 		if (!size0) {
2028 			size0 = size;
2029 			op0   = mthca_opcode[wr->opcode];
2030 		}
2031 
2032 		++ind;
2033 		if (unlikely(ind >= qp->sq.max))
2034 			ind -= qp->sq.max;
2035 	}
2036 
2037 out:
2038 	if (likely(nreq)) {
2039 		doorbell[0] = cpu_to_be32((nreq << 24)                  |
2040 					  ((qp->sq.head & 0xffff) << 8) |
2041 					  f0 | op0);
2042 		doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
2043 
2044 		qp->sq.head += nreq;
2045 
2046 		/*
2047 		 * Make sure that descriptors are written before
2048 		 * doorbell record.
2049 		 */
2050 		wmb();
2051 		*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2052 
2053 		/*
2054 		 * Make sure doorbell record is written before we
2055 		 * write MMIO send doorbell.
2056 		 */
2057 		wmb();
2058 		mthca_write64(doorbell,
2059 			      dev->kar + MTHCA_SEND_DOORBELL,
2060 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2061 	}
2062 
2063 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2064 	return err;
2065 }
2066 
2067 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2068 			     struct ib_recv_wr **bad_wr)
2069 {
2070 	struct mthca_dev *dev = to_mdev(ibqp->device);
2071 	struct mthca_qp *qp = to_mqp(ibqp);
2072 	unsigned long flags;
2073 	int err = 0;
2074 	int nreq;
2075 	int ind;
2076 	int i;
2077 	void *wqe;
2078 
2079 	spin_lock_irqsave(&qp->rq.lock, flags);
2080 
2081 	/* XXX check that state is OK to post receive */
2082 
2083 	ind = qp->rq.head & (qp->rq.max - 1);
2084 
2085 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2086 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2087 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2088 					" %d max, %d nreq)\n", qp->qpn,
2089 					qp->rq.head, qp->rq.tail,
2090 					qp->rq.max, nreq);
2091 			err = -ENOMEM;
2092 			*bad_wr = wr;
2093 			goto out;
2094 		}
2095 
2096 		wqe = get_recv_wqe(qp, ind);
2097 
2098 		((struct mthca_next_seg *) wqe)->flags = 0;
2099 
2100 		wqe += sizeof (struct mthca_next_seg);
2101 
2102 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2103 			err = -EINVAL;
2104 			*bad_wr = wr;
2105 			goto out;
2106 		}
2107 
2108 		for (i = 0; i < wr->num_sge; ++i) {
2109 			((struct mthca_data_seg *) wqe)->byte_count =
2110 				cpu_to_be32(wr->sg_list[i].length);
2111 			((struct mthca_data_seg *) wqe)->lkey =
2112 				cpu_to_be32(wr->sg_list[i].lkey);
2113 			((struct mthca_data_seg *) wqe)->addr =
2114 				cpu_to_be64(wr->sg_list[i].addr);
2115 			wqe += sizeof (struct mthca_data_seg);
2116 		}
2117 
2118 		if (i < qp->rq.max_gs) {
2119 			((struct mthca_data_seg *) wqe)->byte_count = 0;
2120 			((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
2121 			((struct mthca_data_seg *) wqe)->addr = 0;
2122 		}
2123 
2124 		qp->wrid[ind] = wr->wr_id;
2125 
2126 		++ind;
2127 		if (unlikely(ind >= qp->rq.max))
2128 			ind -= qp->rq.max;
2129 	}
2130 out:
2131 	if (likely(nreq)) {
2132 		qp->rq.head += nreq;
2133 
2134 		/*
2135 		 * Make sure that descriptors are written before
2136 		 * doorbell record.
2137 		 */
2138 		wmb();
2139 		*qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2140 	}
2141 
2142 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2143 	return err;
2144 }
2145 
2146 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2147 			int index, int *dbd, __be32 *new_wqe)
2148 {
2149 	struct mthca_next_seg *next;
2150 
2151 	/*
2152 	 * For SRQs, all WQEs generate a CQE, so we're always at the
2153 	 * end of the doorbell chain.
2154 	 */
2155 	if (qp->ibqp.srq) {
2156 		*new_wqe = 0;
2157 		return;
2158 	}
2159 
2160 	if (is_send)
2161 		next = get_send_wqe(qp, index);
2162 	else
2163 		next = get_recv_wqe(qp, index);
2164 
2165 	*dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2166 	if (next->ee_nds & cpu_to_be32(0x3f))
2167 		*new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2168 			(next->ee_nds & cpu_to_be32(0x3f));
2169 	else
2170 		*new_wqe = 0;
2171 }
2172 
2173 int __devinit mthca_init_qp_table(struct mthca_dev *dev)
2174 {
2175 	int err;
2176 	u8 status;
2177 	int i;
2178 
2179 	spin_lock_init(&dev->qp_table.lock);
2180 
2181 	/*
2182 	 * We reserve 2 extra QPs per port for the special QPs.  The
2183 	 * special QP for port 1 has to be even, so round up.
2184 	 */
2185 	dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2186 	err = mthca_alloc_init(&dev->qp_table.alloc,
2187 			       dev->limits.num_qps,
2188 			       (1 << 24) - 1,
2189 			       dev->qp_table.sqp_start +
2190 			       MTHCA_MAX_PORTS * 2);
2191 	if (err)
2192 		return err;
2193 
2194 	err = mthca_array_init(&dev->qp_table.qp,
2195 			       dev->limits.num_qps);
2196 	if (err) {
2197 		mthca_alloc_cleanup(&dev->qp_table.alloc);
2198 		return err;
2199 	}
2200 
2201 	for (i = 0; i < 2; ++i) {
2202 		err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2203 					    dev->qp_table.sqp_start + i * 2,
2204 					    &status);
2205 		if (err)
2206 			goto err_out;
2207 		if (status) {
2208 			mthca_warn(dev, "CONF_SPECIAL_QP returned "
2209 				   "status %02x, aborting.\n",
2210 				   status);
2211 			err = -EINVAL;
2212 			goto err_out;
2213 		}
2214 	}
2215 	return 0;
2216 
2217  err_out:
2218 	for (i = 0; i < 2; ++i)
2219 		mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2220 
2221 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2222 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2223 
2224 	return err;
2225 }
2226 
2227 void mthca_cleanup_qp_table(struct mthca_dev *dev)
2228 {
2229 	int i;
2230 	u8 status;
2231 
2232 	for (i = 0; i < 2; ++i)
2233 		mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2234 
2235 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2236 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2237 }
2238