1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Cisco Systems. All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * OpenIB.org BSD license below:
12  *
13  *     Redistribution and use in source and binary forms, with or
14  *     without modification, are permitted provided that the following
15  *     conditions are met:
16  *
17  *      - Redistributions of source code must retain the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer.
20  *
21  *      - Redistributions in binary form must reproduce the above
22  *        copyright notice, this list of conditions and the following
23  *        disclaimer in the documentation and/or other materials
24  *        provided with the distribution.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33  * SOFTWARE.
34  */
35 
36 #include <linux/string.h>
37 #include <linux/slab.h>
38 #include <linux/sched.h>
39 
40 #include <asm/io.h>
41 
42 #include <rdma/ib_verbs.h>
43 #include <rdma/ib_cache.h>
44 #include <rdma/ib_pack.h>
45 
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 enum {
103 	MTHCA_SEND_DOORBELL_FENCE = 1 << 5
104 };
105 
106 struct mthca_qp_path {
107 	__be32 port_pkey;
108 	u8     rnr_retry;
109 	u8     g_mylmc;
110 	__be16 rlid;
111 	u8     ackto;
112 	u8     mgid_index;
113 	u8     static_rate;
114 	u8     hop_limit;
115 	__be32 sl_tclass_flowlabel;
116 	u8     rgid[16];
117 } __attribute__((packed));
118 
119 struct mthca_qp_context {
120 	__be32 flags;
121 	__be32 tavor_sched_queue; /* Reserved on Arbel */
122 	u8     mtu_msgmax;
123 	u8     rq_size_stride;	/* Reserved on Tavor */
124 	u8     sq_size_stride;	/* Reserved on Tavor */
125 	u8     rlkey_arbel_sched_queue;	/* Reserved on Tavor */
126 	__be32 usr_page;
127 	__be32 local_qpn;
128 	__be32 remote_qpn;
129 	u32    reserved1[2];
130 	struct mthca_qp_path pri_path;
131 	struct mthca_qp_path alt_path;
132 	__be32 rdd;
133 	__be32 pd;
134 	__be32 wqe_base;
135 	__be32 wqe_lkey;
136 	__be32 params1;
137 	__be32 reserved2;
138 	__be32 next_send_psn;
139 	__be32 cqn_snd;
140 	__be32 snd_wqe_base_l;	/* Next send WQE on Tavor */
141 	__be32 snd_db_index;	/* (debugging only entries) */
142 	__be32 last_acked_psn;
143 	__be32 ssn;
144 	__be32 params2;
145 	__be32 rnr_nextrecvpsn;
146 	__be32 ra_buff_indx;
147 	__be32 cqn_rcv;
148 	__be32 rcv_wqe_base_l;	/* Next recv WQE on Tavor */
149 	__be32 rcv_db_index;	/* (debugging only entries) */
150 	__be32 qkey;
151 	__be32 srqn;
152 	__be32 rmsn;
153 	__be16 rq_wqe_counter;	/* reserved on Tavor */
154 	__be16 sq_wqe_counter;	/* reserved on Tavor */
155 	u32    reserved3[18];
156 } __attribute__((packed));
157 
158 struct mthca_qp_param {
159 	__be32 opt_param_mask;
160 	u32    reserved1;
161 	struct mthca_qp_context context;
162 	u32    reserved2[62];
163 } __attribute__((packed));
164 
165 enum {
166 	MTHCA_QP_OPTPAR_ALT_ADDR_PATH     = 1 << 0,
167 	MTHCA_QP_OPTPAR_RRE               = 1 << 1,
168 	MTHCA_QP_OPTPAR_RAE               = 1 << 2,
169 	MTHCA_QP_OPTPAR_RWE               = 1 << 3,
170 	MTHCA_QP_OPTPAR_PKEY_INDEX        = 1 << 4,
171 	MTHCA_QP_OPTPAR_Q_KEY             = 1 << 5,
172 	MTHCA_QP_OPTPAR_RNR_TIMEOUT       = 1 << 6,
173 	MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
174 	MTHCA_QP_OPTPAR_SRA_MAX           = 1 << 8,
175 	MTHCA_QP_OPTPAR_RRA_MAX           = 1 << 9,
176 	MTHCA_QP_OPTPAR_PM_STATE          = 1 << 10,
177 	MTHCA_QP_OPTPAR_PORT_NUM          = 1 << 11,
178 	MTHCA_QP_OPTPAR_RETRY_COUNT       = 1 << 12,
179 	MTHCA_QP_OPTPAR_ALT_RNR_RETRY     = 1 << 13,
180 	MTHCA_QP_OPTPAR_ACK_TIMEOUT       = 1 << 14,
181 	MTHCA_QP_OPTPAR_RNR_RETRY         = 1 << 15,
182 	MTHCA_QP_OPTPAR_SCHED_QUEUE       = 1 << 16
183 };
184 
185 static const u8 mthca_opcode[] = {
186 	[IB_WR_SEND]                 = MTHCA_OPCODE_SEND,
187 	[IB_WR_SEND_WITH_IMM]        = MTHCA_OPCODE_SEND_IMM,
188 	[IB_WR_RDMA_WRITE]           = MTHCA_OPCODE_RDMA_WRITE,
189 	[IB_WR_RDMA_WRITE_WITH_IMM]  = MTHCA_OPCODE_RDMA_WRITE_IMM,
190 	[IB_WR_RDMA_READ]            = MTHCA_OPCODE_RDMA_READ,
191 	[IB_WR_ATOMIC_CMP_AND_SWP]   = MTHCA_OPCODE_ATOMIC_CS,
192 	[IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
193 };
194 
195 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
196 {
197 	return qp->qpn >= dev->qp_table.sqp_start &&
198 		qp->qpn <= dev->qp_table.sqp_start + 3;
199 }
200 
201 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
202 {
203 	return qp->qpn >= dev->qp_table.sqp_start &&
204 		qp->qpn <= dev->qp_table.sqp_start + 1;
205 }
206 
207 static void *get_recv_wqe(struct mthca_qp *qp, int n)
208 {
209 	if (qp->is_direct)
210 		return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
211 	else
212 		return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
213 			((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
214 }
215 
216 static void *get_send_wqe(struct mthca_qp *qp, int n)
217 {
218 	if (qp->is_direct)
219 		return qp->queue.direct.buf + qp->send_wqe_offset +
220 			(n << qp->sq.wqe_shift);
221 	else
222 		return qp->queue.page_list[(qp->send_wqe_offset +
223 					    (n << qp->sq.wqe_shift)) >>
224 					   PAGE_SHIFT].buf +
225 			((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
226 			 (PAGE_SIZE - 1));
227 }
228 
229 static void mthca_wq_reset(struct mthca_wq *wq)
230 {
231 	wq->next_ind  = 0;
232 	wq->last_comp = wq->max - 1;
233 	wq->head      = 0;
234 	wq->tail      = 0;
235 }
236 
237 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
238 		    enum ib_event_type event_type)
239 {
240 	struct mthca_qp *qp;
241 	struct ib_event event;
242 
243 	spin_lock(&dev->qp_table.lock);
244 	qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
245 	if (qp)
246 		++qp->refcount;
247 	spin_unlock(&dev->qp_table.lock);
248 
249 	if (!qp) {
250 		mthca_warn(dev, "Async event for bogus QP %08x\n", qpn);
251 		return;
252 	}
253 
254 	if (event_type == IB_EVENT_PATH_MIG)
255 		qp->port = qp->alt_port;
256 
257 	event.device      = &dev->ib_dev;
258 	event.event       = event_type;
259 	event.element.qp  = &qp->ibqp;
260 	if (qp->ibqp.event_handler)
261 		qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
262 
263 	spin_lock(&dev->qp_table.lock);
264 	if (!--qp->refcount)
265 		wake_up(&qp->wait);
266 	spin_unlock(&dev->qp_table.lock);
267 }
268 
269 static int to_mthca_state(enum ib_qp_state ib_state)
270 {
271 	switch (ib_state) {
272 	case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
273 	case IB_QPS_INIT:  return MTHCA_QP_STATE_INIT;
274 	case IB_QPS_RTR:   return MTHCA_QP_STATE_RTR;
275 	case IB_QPS_RTS:   return MTHCA_QP_STATE_RTS;
276 	case IB_QPS_SQD:   return MTHCA_QP_STATE_SQD;
277 	case IB_QPS_SQE:   return MTHCA_QP_STATE_SQE;
278 	case IB_QPS_ERR:   return MTHCA_QP_STATE_ERR;
279 	default:                return -1;
280 	}
281 }
282 
283 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
284 
285 static int to_mthca_st(int transport)
286 {
287 	switch (transport) {
288 	case RC:  return MTHCA_QP_ST_RC;
289 	case UC:  return MTHCA_QP_ST_UC;
290 	case UD:  return MTHCA_QP_ST_UD;
291 	case RD:  return MTHCA_QP_ST_RD;
292 	case MLX: return MTHCA_QP_ST_MLX;
293 	default:  return -1;
294 	}
295 }
296 
297 static void store_attrs(struct mthca_sqp *sqp, const struct ib_qp_attr *attr,
298 			int attr_mask)
299 {
300 	if (attr_mask & IB_QP_PKEY_INDEX)
301 		sqp->pkey_index = attr->pkey_index;
302 	if (attr_mask & IB_QP_QKEY)
303 		sqp->qkey = attr->qkey;
304 	if (attr_mask & IB_QP_SQ_PSN)
305 		sqp->send_psn = attr->sq_psn;
306 }
307 
308 static void init_port(struct mthca_dev *dev, int port)
309 {
310 	int err;
311 	u8 status;
312 	struct mthca_init_ib_param param;
313 
314 	memset(&param, 0, sizeof param);
315 
316 	param.port_width = dev->limits.port_width_cap;
317 	param.vl_cap     = dev->limits.vl_cap;
318 	param.mtu_cap    = dev->limits.mtu_cap;
319 	param.gid_cap    = dev->limits.gid_table_len;
320 	param.pkey_cap   = dev->limits.pkey_table_len;
321 
322 	err = mthca_INIT_IB(dev, &param, port, &status);
323 	if (err)
324 		mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
325 	if (status)
326 		mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
327 }
328 
329 static __be32 get_hw_access_flags(struct mthca_qp *qp, const struct ib_qp_attr *attr,
330 				  int attr_mask)
331 {
332 	u8 dest_rd_atomic;
333 	u32 access_flags;
334 	u32 hw_access_flags = 0;
335 
336 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
337 		dest_rd_atomic = attr->max_dest_rd_atomic;
338 	else
339 		dest_rd_atomic = qp->resp_depth;
340 
341 	if (attr_mask & IB_QP_ACCESS_FLAGS)
342 		access_flags = attr->qp_access_flags;
343 	else
344 		access_flags = qp->atomic_rd_en;
345 
346 	if (!dest_rd_atomic)
347 		access_flags &= IB_ACCESS_REMOTE_WRITE;
348 
349 	if (access_flags & IB_ACCESS_REMOTE_READ)
350 		hw_access_flags |= MTHCA_QP_BIT_RRE;
351 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
352 		hw_access_flags |= MTHCA_QP_BIT_RAE;
353 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
354 		hw_access_flags |= MTHCA_QP_BIT_RWE;
355 
356 	return cpu_to_be32(hw_access_flags);
357 }
358 
359 static inline enum ib_qp_state to_ib_qp_state(int mthca_state)
360 {
361 	switch (mthca_state) {
362 	case MTHCA_QP_STATE_RST:      return IB_QPS_RESET;
363 	case MTHCA_QP_STATE_INIT:     return IB_QPS_INIT;
364 	case MTHCA_QP_STATE_RTR:      return IB_QPS_RTR;
365 	case MTHCA_QP_STATE_RTS:      return IB_QPS_RTS;
366 	case MTHCA_QP_STATE_DRAINING:
367 	case MTHCA_QP_STATE_SQD:      return IB_QPS_SQD;
368 	case MTHCA_QP_STATE_SQE:      return IB_QPS_SQE;
369 	case MTHCA_QP_STATE_ERR:      return IB_QPS_ERR;
370 	default:                      return -1;
371 	}
372 }
373 
374 static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state)
375 {
376 	switch (mthca_mig_state) {
377 	case 0:  return IB_MIG_ARMED;
378 	case 1:  return IB_MIG_REARM;
379 	case 3:  return IB_MIG_MIGRATED;
380 	default: return -1;
381 	}
382 }
383 
384 static int to_ib_qp_access_flags(int mthca_flags)
385 {
386 	int ib_flags = 0;
387 
388 	if (mthca_flags & MTHCA_QP_BIT_RRE)
389 		ib_flags |= IB_ACCESS_REMOTE_READ;
390 	if (mthca_flags & MTHCA_QP_BIT_RWE)
391 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
392 	if (mthca_flags & MTHCA_QP_BIT_RAE)
393 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
394 
395 	return ib_flags;
396 }
397 
398 static void to_ib_ah_attr(struct mthca_dev *dev, struct ib_ah_attr *ib_ah_attr,
399 				struct mthca_qp_path *path)
400 {
401 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
402 	ib_ah_attr->port_num 	  = (be32_to_cpu(path->port_pkey) >> 24) & 0x3;
403 
404 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->limits.num_ports)
405 		return;
406 
407 	ib_ah_attr->dlid     	  = be16_to_cpu(path->rlid);
408 	ib_ah_attr->sl       	  = be32_to_cpu(path->sl_tclass_flowlabel) >> 28;
409 	ib_ah_attr->src_path_bits = path->g_mylmc & 0x7f;
410 	ib_ah_attr->static_rate   = mthca_rate_to_ib(dev,
411 						     path->static_rate & 0xf,
412 						     ib_ah_attr->port_num);
413 	ib_ah_attr->ah_flags      = (path->g_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
414 	if (ib_ah_attr->ah_flags) {
415 		ib_ah_attr->grh.sgid_index = path->mgid_index & (dev->limits.gid_table_len - 1);
416 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
417 		ib_ah_attr->grh.traffic_class =
418 			(be32_to_cpu(path->sl_tclass_flowlabel) >> 20) & 0xff;
419 		ib_ah_attr->grh.flow_label =
420 			be32_to_cpu(path->sl_tclass_flowlabel) & 0xfffff;
421 		memcpy(ib_ah_attr->grh.dgid.raw,
422 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
423 	}
424 }
425 
426 int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
427 		   struct ib_qp_init_attr *qp_init_attr)
428 {
429 	struct mthca_dev *dev = to_mdev(ibqp->device);
430 	struct mthca_qp *qp = to_mqp(ibqp);
431 	int err = 0;
432 	struct mthca_mailbox *mailbox = NULL;
433 	struct mthca_qp_param *qp_param;
434 	struct mthca_qp_context *context;
435 	int mthca_state;
436 	u8 status;
437 
438 	mutex_lock(&qp->mutex);
439 
440 	if (qp->state == IB_QPS_RESET) {
441 		qp_attr->qp_state = IB_QPS_RESET;
442 		goto done;
443 	}
444 
445 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
446 	if (IS_ERR(mailbox)) {
447 		err = PTR_ERR(mailbox);
448 		goto out;
449 	}
450 
451 	err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox, &status);
452 	if (err)
453 		goto out_mailbox;
454 	if (status) {
455 		mthca_warn(dev, "QUERY_QP returned status %02x\n", status);
456 		err = -EINVAL;
457 		goto out_mailbox;
458 	}
459 
460 	qp_param    = mailbox->buf;
461 	context     = &qp_param->context;
462 	mthca_state = be32_to_cpu(context->flags) >> 28;
463 
464 	qp->state		     = to_ib_qp_state(mthca_state);
465 	qp_attr->qp_state	     = qp->state;
466 	qp_attr->path_mtu 	     = context->mtu_msgmax >> 5;
467 	qp_attr->path_mig_state      =
468 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
469 	qp_attr->qkey 		     = be32_to_cpu(context->qkey);
470 	qp_attr->rq_psn 	     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
471 	qp_attr->sq_psn 	     = be32_to_cpu(context->next_send_psn) & 0xffffff;
472 	qp_attr->dest_qp_num 	     = be32_to_cpu(context->remote_qpn) & 0xffffff;
473 	qp_attr->qp_access_flags     =
474 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
475 
476 	if (qp->transport == RC || qp->transport == UC) {
477 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
478 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
479 		qp_attr->alt_pkey_index =
480 			be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
481 		qp_attr->alt_port_num 	= qp_attr->alt_ah_attr.port_num;
482 	}
483 
484 	qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
485 	qp_attr->port_num   =
486 		(be32_to_cpu(context->pri_path.port_pkey) >> 24) & 0x3;
487 
488 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
489 	qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
490 
491 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
492 
493 	qp_attr->max_dest_rd_atomic =
494 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
495 	qp_attr->min_rnr_timer 	    =
496 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
497 	qp_attr->timeout 	    = context->pri_path.ackto >> 3;
498 	qp_attr->retry_cnt 	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
499 	qp_attr->rnr_retry 	    = context->pri_path.rnr_retry >> 5;
500 	qp_attr->alt_timeout 	    = context->alt_path.ackto >> 3;
501 
502 done:
503 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
504 	qp_attr->cap.max_send_wr     = qp->sq.max;
505 	qp_attr->cap.max_recv_wr     = qp->rq.max;
506 	qp_attr->cap.max_send_sge    = qp->sq.max_gs;
507 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
508 	qp_attr->cap.max_inline_data = qp->max_inline_data;
509 
510 	qp_init_attr->cap	     = qp_attr->cap;
511 
512 out_mailbox:
513 	mthca_free_mailbox(dev, mailbox);
514 
515 out:
516 	mutex_unlock(&qp->mutex);
517 	return err;
518 }
519 
520 static int mthca_path_set(struct mthca_dev *dev, const struct ib_ah_attr *ah,
521 			  struct mthca_qp_path *path, u8 port)
522 {
523 	path->g_mylmc     = ah->src_path_bits & 0x7f;
524 	path->rlid        = cpu_to_be16(ah->dlid);
525 	path->static_rate = mthca_get_rate(dev, ah->static_rate, port);
526 
527 	if (ah->ah_flags & IB_AH_GRH) {
528 		if (ah->grh.sgid_index >= dev->limits.gid_table_len) {
529 			mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
530 				  ah->grh.sgid_index, dev->limits.gid_table_len-1);
531 			return -1;
532 		}
533 
534 		path->g_mylmc   |= 1 << 7;
535 		path->mgid_index = ah->grh.sgid_index;
536 		path->hop_limit  = ah->grh.hop_limit;
537 		path->sl_tclass_flowlabel =
538 			cpu_to_be32((ah->sl << 28)                |
539 				    (ah->grh.traffic_class << 20) |
540 				    (ah->grh.flow_label));
541 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
542 	} else
543 		path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28);
544 
545 	return 0;
546 }
547 
548 static int __mthca_modify_qp(struct ib_qp *ibqp,
549 			     const struct ib_qp_attr *attr, int attr_mask,
550 			     enum ib_qp_state cur_state, enum ib_qp_state new_state)
551 {
552 	struct mthca_dev *dev = to_mdev(ibqp->device);
553 	struct mthca_qp *qp = to_mqp(ibqp);
554 	struct mthca_mailbox *mailbox;
555 	struct mthca_qp_param *qp_param;
556 	struct mthca_qp_context *qp_context;
557 	u32 sqd_event = 0;
558 	u8 status;
559 	int err = -EINVAL;
560 
561 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
562 	if (IS_ERR(mailbox)) {
563 		err = PTR_ERR(mailbox);
564 		goto out;
565 	}
566 	qp_param = mailbox->buf;
567 	qp_context = &qp_param->context;
568 	memset(qp_param, 0, sizeof *qp_param);
569 
570 	qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
571 					     (to_mthca_st(qp->transport) << 16));
572 	qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
573 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
574 		qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
575 	else {
576 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
577 		switch (attr->path_mig_state) {
578 		case IB_MIG_MIGRATED:
579 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
580 			break;
581 		case IB_MIG_REARM:
582 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
583 			break;
584 		case IB_MIG_ARMED:
585 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
586 			break;
587 		}
588 	}
589 
590 	/* leave tavor_sched_queue as 0 */
591 
592 	if (qp->transport == MLX || qp->transport == UD)
593 		qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
594 	else if (attr_mask & IB_QP_PATH_MTU) {
595 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
596 			mthca_dbg(dev, "path MTU (%u) is invalid\n",
597 				  attr->path_mtu);
598 			goto out_mailbox;
599 		}
600 		qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
601 	}
602 
603 	if (mthca_is_memfree(dev)) {
604 		if (qp->rq.max)
605 			qp_context->rq_size_stride = ilog2(qp->rq.max) << 3;
606 		qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
607 
608 		if (qp->sq.max)
609 			qp_context->sq_size_stride = ilog2(qp->sq.max) << 3;
610 		qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
611 	}
612 
613 	/* leave arbel_sched_queue as 0 */
614 
615 	if (qp->ibqp.uobject)
616 		qp_context->usr_page =
617 			cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
618 	else
619 		qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
620 	qp_context->local_qpn  = cpu_to_be32(qp->qpn);
621 	if (attr_mask & IB_QP_DEST_QPN) {
622 		qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
623 	}
624 
625 	if (qp->transport == MLX)
626 		qp_context->pri_path.port_pkey |=
627 			cpu_to_be32(qp->port << 24);
628 	else {
629 		if (attr_mask & IB_QP_PORT) {
630 			qp_context->pri_path.port_pkey |=
631 				cpu_to_be32(attr->port_num << 24);
632 			qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
633 		}
634 	}
635 
636 	if (attr_mask & IB_QP_PKEY_INDEX) {
637 		qp_context->pri_path.port_pkey |=
638 			cpu_to_be32(attr->pkey_index);
639 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
640 	}
641 
642 	if (attr_mask & IB_QP_RNR_RETRY) {
643 		qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
644 			attr->rnr_retry << 5;
645 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
646 							MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
647 	}
648 
649 	if (attr_mask & IB_QP_AV) {
650 		if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
651 				   attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
652 			goto out_mailbox;
653 
654 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
655 	}
656 
657 	if (ibqp->qp_type == IB_QPT_RC &&
658 	    cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
659 		u8 sched_queue = ibqp->uobject ? 0x2 : 0x1;
660 
661 		if (mthca_is_memfree(dev))
662 			qp_context->rlkey_arbel_sched_queue |= sched_queue;
663 		else
664 			qp_context->tavor_sched_queue |= cpu_to_be32(sched_queue);
665 
666 		qp_param->opt_param_mask |=
667 			cpu_to_be32(MTHCA_QP_OPTPAR_SCHED_QUEUE);
668 	}
669 
670 	if (attr_mask & IB_QP_TIMEOUT) {
671 		qp_context->pri_path.ackto = attr->timeout << 3;
672 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
673 	}
674 
675 	if (attr_mask & IB_QP_ALT_PATH) {
676 		if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
677 			mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
678 				  attr->alt_pkey_index, dev->limits.pkey_table_len-1);
679 			goto out_mailbox;
680 		}
681 
682 		if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
683 			mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
684 				attr->alt_port_num);
685 			goto out_mailbox;
686 		}
687 
688 		if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
689 				   attr->alt_ah_attr.port_num))
690 			goto out_mailbox;
691 
692 		qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
693 							      attr->alt_port_num << 24);
694 		qp_context->alt_path.ackto = attr->alt_timeout << 3;
695 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
696 	}
697 
698 	/* leave rdd as 0 */
699 	qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
700 	/* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
701 	qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
702 	qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
703 					     (MTHCA_FLIGHT_LIMIT << 24) |
704 					     MTHCA_QP_BIT_SWE);
705 	if (qp->sq_policy == IB_SIGNAL_ALL_WR)
706 		qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
707 	if (attr_mask & IB_QP_RETRY_CNT) {
708 		qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
709 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
710 	}
711 
712 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
713 		if (attr->max_rd_atomic) {
714 			qp_context->params1 |=
715 				cpu_to_be32(MTHCA_QP_BIT_SRE |
716 					    MTHCA_QP_BIT_SAE);
717 			qp_context->params1 |=
718 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
719 		}
720 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
721 	}
722 
723 	if (attr_mask & IB_QP_SQ_PSN)
724 		qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
725 	qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
726 
727 	if (mthca_is_memfree(dev)) {
728 		qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
729 		qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
730 	}
731 
732 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
733 		if (attr->max_dest_rd_atomic)
734 			qp_context->params2 |=
735 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
736 
737 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
738 	}
739 
740 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
741 		qp_context->params2      |= get_hw_access_flags(qp, attr, attr_mask);
742 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
743 							MTHCA_QP_OPTPAR_RRE |
744 							MTHCA_QP_OPTPAR_RAE);
745 	}
746 
747 	qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
748 
749 	if (ibqp->srq)
750 		qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
751 
752 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
753 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
754 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
755 	}
756 	if (attr_mask & IB_QP_RQ_PSN)
757 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
758 
759 	qp_context->ra_buff_indx =
760 		cpu_to_be32(dev->qp_table.rdb_base +
761 			    ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
762 			     dev->qp_table.rdb_shift));
763 
764 	qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
765 
766 	if (mthca_is_memfree(dev))
767 		qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
768 
769 	if (attr_mask & IB_QP_QKEY) {
770 		qp_context->qkey = cpu_to_be32(attr->qkey);
771 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
772 	}
773 
774 	if (ibqp->srq)
775 		qp_context->srqn = cpu_to_be32(1 << 24 |
776 					       to_msrq(ibqp->srq)->srqn);
777 
778 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
779 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY		&&
780 	    attr->en_sqd_async_notify)
781 		sqd_event = 1 << 31;
782 
783 	err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
784 			      mailbox, sqd_event, &status);
785 	if (err)
786 		goto out_mailbox;
787 	if (status) {
788 		mthca_warn(dev, "modify QP %d->%d returned status %02x.\n",
789 			   cur_state, new_state, status);
790 		err = -EINVAL;
791 		goto out_mailbox;
792 	}
793 
794 	qp->state = new_state;
795 	if (attr_mask & IB_QP_ACCESS_FLAGS)
796 		qp->atomic_rd_en = attr->qp_access_flags;
797 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
798 		qp->resp_depth = attr->max_dest_rd_atomic;
799 	if (attr_mask & IB_QP_PORT)
800 		qp->port = attr->port_num;
801 	if (attr_mask & IB_QP_ALT_PATH)
802 		qp->alt_port = attr->alt_port_num;
803 
804 	if (is_sqp(dev, qp))
805 		store_attrs(to_msqp(qp), attr, attr_mask);
806 
807 	/*
808 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
809 	 * QP0 to RESET or ERROR, bring the link back down.
810 	 */
811 	if (is_qp0(dev, qp)) {
812 		if (cur_state != IB_QPS_RTR &&
813 		    new_state == IB_QPS_RTR)
814 			init_port(dev, qp->port);
815 
816 		if (cur_state != IB_QPS_RESET &&
817 		    cur_state != IB_QPS_ERR &&
818 		    (new_state == IB_QPS_RESET ||
819 		     new_state == IB_QPS_ERR))
820 			mthca_CLOSE_IB(dev, qp->port, &status);
821 	}
822 
823 	/*
824 	 * If we moved a kernel QP to RESET, clean up all old CQ
825 	 * entries and reinitialize the QP.
826 	 */
827 	if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
828 		mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
829 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
830 		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
831 			mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn, NULL);
832 
833 		mthca_wq_reset(&qp->sq);
834 		qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
835 
836 		mthca_wq_reset(&qp->rq);
837 		qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
838 
839 		if (mthca_is_memfree(dev)) {
840 			*qp->sq.db = 0;
841 			*qp->rq.db = 0;
842 		}
843 	}
844 
845 out_mailbox:
846 	mthca_free_mailbox(dev, mailbox);
847 out:
848 	return err;
849 }
850 
851 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
852 		    struct ib_udata *udata)
853 {
854 	struct mthca_dev *dev = to_mdev(ibqp->device);
855 	struct mthca_qp *qp = to_mqp(ibqp);
856 	enum ib_qp_state cur_state, new_state;
857 	int err = -EINVAL;
858 
859 	mutex_lock(&qp->mutex);
860 	if (attr_mask & IB_QP_CUR_STATE) {
861 		cur_state = attr->cur_qp_state;
862 	} else {
863 		spin_lock_irq(&qp->sq.lock);
864 		spin_lock(&qp->rq.lock);
865 		cur_state = qp->state;
866 		spin_unlock(&qp->rq.lock);
867 		spin_unlock_irq(&qp->sq.lock);
868 	}
869 
870 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
871 
872 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
873 		mthca_dbg(dev, "Bad QP transition (transport %d) "
874 			  "%d->%d with attr 0x%08x\n",
875 			  qp->transport, cur_state, new_state,
876 			  attr_mask);
877 		goto out;
878 	}
879 
880 	if ((attr_mask & IB_QP_PKEY_INDEX) &&
881 	     attr->pkey_index >= dev->limits.pkey_table_len) {
882 		mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
883 			  attr->pkey_index, dev->limits.pkey_table_len-1);
884 		goto out;
885 	}
886 
887 	if ((attr_mask & IB_QP_PORT) &&
888 	    (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
889 		mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
890 		goto out;
891 	}
892 
893 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
894 	    attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
895 		mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
896 			  attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
897 		goto out;
898 	}
899 
900 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
901 	    attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
902 		mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
903 			  attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
904 		goto out;
905 	}
906 
907 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
908 		err = 0;
909 		goto out;
910 	}
911 
912 	err = __mthca_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
913 
914 out:
915 	mutex_unlock(&qp->mutex);
916 	return err;
917 }
918 
919 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
920 {
921 	/*
922 	 * Calculate the maximum size of WQE s/g segments, excluding
923 	 * the next segment and other non-data segments.
924 	 */
925 	int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
926 
927 	switch (qp->transport) {
928 	case MLX:
929 		max_data_size -= 2 * sizeof (struct mthca_data_seg);
930 		break;
931 
932 	case UD:
933 		if (mthca_is_memfree(dev))
934 			max_data_size -= sizeof (struct mthca_arbel_ud_seg);
935 		else
936 			max_data_size -= sizeof (struct mthca_tavor_ud_seg);
937 		break;
938 
939 	default:
940 		max_data_size -= sizeof (struct mthca_raddr_seg);
941 		break;
942 	}
943 
944 	return max_data_size;
945 }
946 
947 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
948 {
949 	/* We don't support inline data for kernel QPs (yet). */
950 	return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
951 }
952 
953 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
954 				 struct mthca_pd *pd,
955 				 struct mthca_qp *qp)
956 {
957 	int max_data_size = mthca_max_data_size(dev, qp,
958 						min(dev->limits.max_desc_sz,
959 						    1 << qp->sq.wqe_shift));
960 
961 	qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
962 
963 	qp->sq.max_gs = min_t(int, dev->limits.max_sg,
964 			      max_data_size / sizeof (struct mthca_data_seg));
965 	qp->rq.max_gs = min_t(int, dev->limits.max_sg,
966 			       (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
967 				sizeof (struct mthca_next_seg)) /
968 			       sizeof (struct mthca_data_seg));
969 }
970 
971 /*
972  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
973  * rq.max_gs and sq.max_gs must all be assigned.
974  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
975  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
976  * queue)
977  */
978 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
979 			       struct mthca_pd *pd,
980 			       struct mthca_qp *qp)
981 {
982 	int size;
983 	int err = -ENOMEM;
984 
985 	size = sizeof (struct mthca_next_seg) +
986 		qp->rq.max_gs * sizeof (struct mthca_data_seg);
987 
988 	if (size > dev->limits.max_desc_sz)
989 		return -EINVAL;
990 
991 	for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
992 	     qp->rq.wqe_shift++)
993 		; /* nothing */
994 
995 	size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
996 	switch (qp->transport) {
997 	case MLX:
998 		size += 2 * sizeof (struct mthca_data_seg);
999 		break;
1000 
1001 	case UD:
1002 		size += mthca_is_memfree(dev) ?
1003 			sizeof (struct mthca_arbel_ud_seg) :
1004 			sizeof (struct mthca_tavor_ud_seg);
1005 		break;
1006 
1007 	case UC:
1008 		size += sizeof (struct mthca_raddr_seg);
1009 		break;
1010 
1011 	case RC:
1012 		size += sizeof (struct mthca_raddr_seg);
1013 		/*
1014 		 * An atomic op will require an atomic segment, a
1015 		 * remote address segment and one scatter entry.
1016 		 */
1017 		size = max_t(int, size,
1018 			     sizeof (struct mthca_atomic_seg) +
1019 			     sizeof (struct mthca_raddr_seg) +
1020 			     sizeof (struct mthca_data_seg));
1021 		break;
1022 
1023 	default:
1024 		break;
1025 	}
1026 
1027 	/* Make sure that we have enough space for a bind request */
1028 	size = max_t(int, size, sizeof (struct mthca_bind_seg));
1029 
1030 	size += sizeof (struct mthca_next_seg);
1031 
1032 	if (size > dev->limits.max_desc_sz)
1033 		return -EINVAL;
1034 
1035 	for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
1036 	     qp->sq.wqe_shift++)
1037 		; /* nothing */
1038 
1039 	qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
1040 				    1 << qp->sq.wqe_shift);
1041 
1042 	/*
1043 	 * If this is a userspace QP, we don't actually have to
1044 	 * allocate anything.  All we need is to calculate the WQE
1045 	 * sizes and the send_wqe_offset, so we're done now.
1046 	 */
1047 	if (pd->ibpd.uobject)
1048 		return 0;
1049 
1050 	size = PAGE_ALIGN(qp->send_wqe_offset +
1051 			  (qp->sq.max << qp->sq.wqe_shift));
1052 
1053 	qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1054 			   GFP_KERNEL);
1055 	if (!qp->wrid)
1056 		goto err_out;
1057 
1058 	err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1059 			      &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1060 	if (err)
1061 		goto err_out;
1062 
1063 	return 0;
1064 
1065 err_out:
1066 	kfree(qp->wrid);
1067 	return err;
1068 }
1069 
1070 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1071 			       struct mthca_qp *qp)
1072 {
1073 	mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1074 				       (qp->sq.max << qp->sq.wqe_shift)),
1075 		       &qp->queue, qp->is_direct, &qp->mr);
1076 	kfree(qp->wrid);
1077 }
1078 
1079 static int mthca_map_memfree(struct mthca_dev *dev,
1080 			     struct mthca_qp *qp)
1081 {
1082 	int ret;
1083 
1084 	if (mthca_is_memfree(dev)) {
1085 		ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1086 		if (ret)
1087 			return ret;
1088 
1089 		ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1090 		if (ret)
1091 			goto err_qpc;
1092 
1093 		ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1094 				      qp->qpn << dev->qp_table.rdb_shift);
1095 		if (ret)
1096 			goto err_eqpc;
1097 
1098 	}
1099 
1100 	return 0;
1101 
1102 err_eqpc:
1103 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1104 
1105 err_qpc:
1106 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1107 
1108 	return ret;
1109 }
1110 
1111 static void mthca_unmap_memfree(struct mthca_dev *dev,
1112 				struct mthca_qp *qp)
1113 {
1114 	mthca_table_put(dev, dev->qp_table.rdb_table,
1115 			qp->qpn << dev->qp_table.rdb_shift);
1116 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1117 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1118 }
1119 
1120 static int mthca_alloc_memfree(struct mthca_dev *dev,
1121 			       struct mthca_qp *qp)
1122 {
1123 	if (mthca_is_memfree(dev)) {
1124 		qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1125 						 qp->qpn, &qp->rq.db);
1126 		if (qp->rq.db_index < 0)
1127 			return -ENOMEM;
1128 
1129 		qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1130 						 qp->qpn, &qp->sq.db);
1131 		if (qp->sq.db_index < 0) {
1132 			mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1133 			return -ENOMEM;
1134 		}
1135 	}
1136 
1137 	return 0;
1138 }
1139 
1140 static void mthca_free_memfree(struct mthca_dev *dev,
1141 			       struct mthca_qp *qp)
1142 {
1143 	if (mthca_is_memfree(dev)) {
1144 		mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1145 		mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1146 	}
1147 }
1148 
1149 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1150 				 struct mthca_pd *pd,
1151 				 struct mthca_cq *send_cq,
1152 				 struct mthca_cq *recv_cq,
1153 				 enum ib_sig_type send_policy,
1154 				 struct mthca_qp *qp)
1155 {
1156 	int ret;
1157 	int i;
1158 	struct mthca_next_seg *next;
1159 
1160 	qp->refcount = 1;
1161 	init_waitqueue_head(&qp->wait);
1162 	mutex_init(&qp->mutex);
1163 	qp->state    	 = IB_QPS_RESET;
1164 	qp->atomic_rd_en = 0;
1165 	qp->resp_depth   = 0;
1166 	qp->sq_policy    = send_policy;
1167 	mthca_wq_reset(&qp->sq);
1168 	mthca_wq_reset(&qp->rq);
1169 
1170 	spin_lock_init(&qp->sq.lock);
1171 	spin_lock_init(&qp->rq.lock);
1172 
1173 	ret = mthca_map_memfree(dev, qp);
1174 	if (ret)
1175 		return ret;
1176 
1177 	ret = mthca_alloc_wqe_buf(dev, pd, qp);
1178 	if (ret) {
1179 		mthca_unmap_memfree(dev, qp);
1180 		return ret;
1181 	}
1182 
1183 	mthca_adjust_qp_caps(dev, pd, qp);
1184 
1185 	/*
1186 	 * If this is a userspace QP, we're done now.  The doorbells
1187 	 * will be allocated and buffers will be initialized in
1188 	 * userspace.
1189 	 */
1190 	if (pd->ibpd.uobject)
1191 		return 0;
1192 
1193 	ret = mthca_alloc_memfree(dev, qp);
1194 	if (ret) {
1195 		mthca_free_wqe_buf(dev, qp);
1196 		mthca_unmap_memfree(dev, qp);
1197 		return ret;
1198 	}
1199 
1200 	if (mthca_is_memfree(dev)) {
1201 		struct mthca_data_seg *scatter;
1202 		int size = (sizeof (struct mthca_next_seg) +
1203 			    qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1204 
1205 		for (i = 0; i < qp->rq.max; ++i) {
1206 			next = get_recv_wqe(qp, i);
1207 			next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1208 						   qp->rq.wqe_shift);
1209 			next->ee_nds = cpu_to_be32(size);
1210 
1211 			for (scatter = (void *) (next + 1);
1212 			     (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1213 			     ++scatter)
1214 				scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1215 		}
1216 
1217 		for (i = 0; i < qp->sq.max; ++i) {
1218 			next = get_send_wqe(qp, i);
1219 			next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1220 						    qp->sq.wqe_shift) +
1221 						   qp->send_wqe_offset);
1222 		}
1223 	} else {
1224 		for (i = 0; i < qp->rq.max; ++i) {
1225 			next = get_recv_wqe(qp, i);
1226 			next->nda_op = htonl((((i + 1) % qp->rq.max) <<
1227 					      qp->rq.wqe_shift) | 1);
1228 		}
1229 
1230 	}
1231 
1232 	qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1233 	qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1234 
1235 	return 0;
1236 }
1237 
1238 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1239 			     struct mthca_pd *pd, struct mthca_qp *qp)
1240 {
1241 	int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1242 
1243 	/* Sanity check QP size before proceeding */
1244 	if (cap->max_send_wr  	 > dev->limits.max_wqes ||
1245 	    cap->max_recv_wr  	 > dev->limits.max_wqes ||
1246 	    cap->max_send_sge 	 > dev->limits.max_sg   ||
1247 	    cap->max_recv_sge 	 > dev->limits.max_sg   ||
1248 	    cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1249 		return -EINVAL;
1250 
1251 	/*
1252 	 * For MLX transport we need 2 extra send gather entries:
1253 	 * one for the header and one for the checksum at the end
1254 	 */
1255 	if (qp->transport == MLX && cap->max_send_sge + 2 > dev->limits.max_sg)
1256 		return -EINVAL;
1257 
1258 	if (mthca_is_memfree(dev)) {
1259 		qp->rq.max = cap->max_recv_wr ?
1260 			roundup_pow_of_two(cap->max_recv_wr) : 0;
1261 		qp->sq.max = cap->max_send_wr ?
1262 			roundup_pow_of_two(cap->max_send_wr) : 0;
1263 	} else {
1264 		qp->rq.max = cap->max_recv_wr;
1265 		qp->sq.max = cap->max_send_wr;
1266 	}
1267 
1268 	qp->rq.max_gs = cap->max_recv_sge;
1269 	qp->sq.max_gs = max_t(int, cap->max_send_sge,
1270 			      ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1271 				    MTHCA_INLINE_CHUNK_SIZE) /
1272 			      sizeof (struct mthca_data_seg));
1273 
1274 	return 0;
1275 }
1276 
1277 int mthca_alloc_qp(struct mthca_dev *dev,
1278 		   struct mthca_pd *pd,
1279 		   struct mthca_cq *send_cq,
1280 		   struct mthca_cq *recv_cq,
1281 		   enum ib_qp_type type,
1282 		   enum ib_sig_type send_policy,
1283 		   struct ib_qp_cap *cap,
1284 		   struct mthca_qp *qp)
1285 {
1286 	int err;
1287 
1288 	switch (type) {
1289 	case IB_QPT_RC: qp->transport = RC; break;
1290 	case IB_QPT_UC: qp->transport = UC; break;
1291 	case IB_QPT_UD: qp->transport = UD; break;
1292 	default: return -EINVAL;
1293 	}
1294 
1295 	err = mthca_set_qp_size(dev, cap, pd, qp);
1296 	if (err)
1297 		return err;
1298 
1299 	qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1300 	if (qp->qpn == -1)
1301 		return -ENOMEM;
1302 
1303 	/* initialize port to zero for error-catching. */
1304 	qp->port = 0;
1305 
1306 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1307 				    send_policy, qp);
1308 	if (err) {
1309 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1310 		return err;
1311 	}
1312 
1313 	spin_lock_irq(&dev->qp_table.lock);
1314 	mthca_array_set(&dev->qp_table.qp,
1315 			qp->qpn & (dev->limits.num_qps - 1), qp);
1316 	spin_unlock_irq(&dev->qp_table.lock);
1317 
1318 	return 0;
1319 }
1320 
1321 static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1322 {
1323 	if (send_cq == recv_cq)
1324 		spin_lock_irq(&send_cq->lock);
1325 	else if (send_cq->cqn < recv_cq->cqn) {
1326 		spin_lock_irq(&send_cq->lock);
1327 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1328 	} else {
1329 		spin_lock_irq(&recv_cq->lock);
1330 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1331 	}
1332 }
1333 
1334 static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1335 {
1336 	if (send_cq == recv_cq)
1337 		spin_unlock_irq(&send_cq->lock);
1338 	else if (send_cq->cqn < recv_cq->cqn) {
1339 		spin_unlock(&recv_cq->lock);
1340 		spin_unlock_irq(&send_cq->lock);
1341 	} else {
1342 		spin_unlock(&send_cq->lock);
1343 		spin_unlock_irq(&recv_cq->lock);
1344 	}
1345 }
1346 
1347 int mthca_alloc_sqp(struct mthca_dev *dev,
1348 		    struct mthca_pd *pd,
1349 		    struct mthca_cq *send_cq,
1350 		    struct mthca_cq *recv_cq,
1351 		    enum ib_sig_type send_policy,
1352 		    struct ib_qp_cap *cap,
1353 		    int qpn,
1354 		    int port,
1355 		    struct mthca_sqp *sqp)
1356 {
1357 	u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1358 	int err;
1359 
1360 	sqp->qp.transport = MLX;
1361 	err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1362 	if (err)
1363 		return err;
1364 
1365 	sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1366 	sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1367 					     &sqp->header_dma, GFP_KERNEL);
1368 	if (!sqp->header_buf)
1369 		return -ENOMEM;
1370 
1371 	spin_lock_irq(&dev->qp_table.lock);
1372 	if (mthca_array_get(&dev->qp_table.qp, mqpn))
1373 		err = -EBUSY;
1374 	else
1375 		mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1376 	spin_unlock_irq(&dev->qp_table.lock);
1377 
1378 	if (err)
1379 		goto err_out;
1380 
1381 	sqp->qp.port      = port;
1382 	sqp->qp.qpn       = mqpn;
1383 	sqp->qp.transport = MLX;
1384 
1385 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1386 				    send_policy, &sqp->qp);
1387 	if (err)
1388 		goto err_out_free;
1389 
1390 	atomic_inc(&pd->sqp_count);
1391 
1392 	return 0;
1393 
1394  err_out_free:
1395 	/*
1396 	 * Lock CQs here, so that CQ polling code can do QP lookup
1397 	 * without taking a lock.
1398 	 */
1399 	mthca_lock_cqs(send_cq, recv_cq);
1400 
1401 	spin_lock(&dev->qp_table.lock);
1402 	mthca_array_clear(&dev->qp_table.qp, mqpn);
1403 	spin_unlock(&dev->qp_table.lock);
1404 
1405 	mthca_unlock_cqs(send_cq, recv_cq);
1406 
1407  err_out:
1408 	dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1409 			  sqp->header_buf, sqp->header_dma);
1410 
1411 	return err;
1412 }
1413 
1414 static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp)
1415 {
1416 	int c;
1417 
1418 	spin_lock_irq(&dev->qp_table.lock);
1419 	c = qp->refcount;
1420 	spin_unlock_irq(&dev->qp_table.lock);
1421 
1422 	return c;
1423 }
1424 
1425 void mthca_free_qp(struct mthca_dev *dev,
1426 		   struct mthca_qp *qp)
1427 {
1428 	u8 status;
1429 	struct mthca_cq *send_cq;
1430 	struct mthca_cq *recv_cq;
1431 
1432 	send_cq = to_mcq(qp->ibqp.send_cq);
1433 	recv_cq = to_mcq(qp->ibqp.recv_cq);
1434 
1435 	/*
1436 	 * Lock CQs here, so that CQ polling code can do QP lookup
1437 	 * without taking a lock.
1438 	 */
1439 	mthca_lock_cqs(send_cq, recv_cq);
1440 
1441 	spin_lock(&dev->qp_table.lock);
1442 	mthca_array_clear(&dev->qp_table.qp,
1443 			  qp->qpn & (dev->limits.num_qps - 1));
1444 	--qp->refcount;
1445 	spin_unlock(&dev->qp_table.lock);
1446 
1447 	mthca_unlock_cqs(send_cq, recv_cq);
1448 
1449 	wait_event(qp->wait, !get_qp_refcount(dev, qp));
1450 
1451 	if (qp->state != IB_QPS_RESET)
1452 		mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1453 				NULL, 0, &status);
1454 
1455 	/*
1456 	 * If this is a userspace QP, the buffers, MR, CQs and so on
1457 	 * will be cleaned up in userspace, so all we have to do is
1458 	 * unref the mem-free tables and free the QPN in our table.
1459 	 */
1460 	if (!qp->ibqp.uobject) {
1461 		mthca_cq_clean(dev, recv_cq, qp->qpn,
1462 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1463 		if (send_cq != recv_cq)
1464 			mthca_cq_clean(dev, send_cq, qp->qpn, NULL);
1465 
1466 		mthca_free_memfree(dev, qp);
1467 		mthca_free_wqe_buf(dev, qp);
1468 	}
1469 
1470 	mthca_unmap_memfree(dev, qp);
1471 
1472 	if (is_sqp(dev, qp)) {
1473 		atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1474 		dma_free_coherent(&dev->pdev->dev,
1475 				  to_msqp(qp)->header_buf_size,
1476 				  to_msqp(qp)->header_buf,
1477 				  to_msqp(qp)->header_dma);
1478 	} else
1479 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1480 }
1481 
1482 /* Create UD header for an MLX send and build a data segment for it */
1483 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1484 			    int ind, struct ib_send_wr *wr,
1485 			    struct mthca_mlx_seg *mlx,
1486 			    struct mthca_data_seg *data)
1487 {
1488 	int header_size;
1489 	int err;
1490 	u16 pkey;
1491 
1492 	ib_ud_header_init(256, /* assume a MAD */
1493 			  mthca_ah_grh_present(to_mah(wr->wr.ud.ah)),
1494 			  &sqp->ud_header);
1495 
1496 	err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1497 	if (err)
1498 		return err;
1499 	mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1500 	mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1501 				  (sqp->ud_header.lrh.destination_lid ==
1502 				   IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1503 				  (sqp->ud_header.lrh.service_level << 8));
1504 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
1505 	mlx->vcrc = 0;
1506 
1507 	switch (wr->opcode) {
1508 	case IB_WR_SEND:
1509 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1510 		sqp->ud_header.immediate_present = 0;
1511 		break;
1512 	case IB_WR_SEND_WITH_IMM:
1513 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1514 		sqp->ud_header.immediate_present = 1;
1515 		sqp->ud_header.immediate_data = wr->ex.imm_data;
1516 		break;
1517 	default:
1518 		return -EINVAL;
1519 	}
1520 
1521 	sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1522 	if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1523 		sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1524 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1525 	if (!sqp->qp.ibqp.qp_num)
1526 		ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1527 				   sqp->pkey_index, &pkey);
1528 	else
1529 		ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1530 				   wr->wr.ud.pkey_index, &pkey);
1531 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1532 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1533 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1534 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1535 					       sqp->qkey : wr->wr.ud.remote_qkey);
1536 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1537 
1538 	header_size = ib_ud_header_pack(&sqp->ud_header,
1539 					sqp->header_buf +
1540 					ind * MTHCA_UD_HEADER_SIZE);
1541 
1542 	data->byte_count = cpu_to_be32(header_size);
1543 	data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1544 	data->addr       = cpu_to_be64(sqp->header_dma +
1545 				       ind * MTHCA_UD_HEADER_SIZE);
1546 
1547 	return 0;
1548 }
1549 
1550 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1551 				    struct ib_cq *ib_cq)
1552 {
1553 	unsigned cur;
1554 	struct mthca_cq *cq;
1555 
1556 	cur = wq->head - wq->tail;
1557 	if (likely(cur + nreq < wq->max))
1558 		return 0;
1559 
1560 	cq = to_mcq(ib_cq);
1561 	spin_lock(&cq->lock);
1562 	cur = wq->head - wq->tail;
1563 	spin_unlock(&cq->lock);
1564 
1565 	return cur + nreq >= wq->max;
1566 }
1567 
1568 static __always_inline void set_raddr_seg(struct mthca_raddr_seg *rseg,
1569 					  u64 remote_addr, u32 rkey)
1570 {
1571 	rseg->raddr    = cpu_to_be64(remote_addr);
1572 	rseg->rkey     = cpu_to_be32(rkey);
1573 	rseg->reserved = 0;
1574 }
1575 
1576 static __always_inline void set_atomic_seg(struct mthca_atomic_seg *aseg,
1577 					   struct ib_send_wr *wr)
1578 {
1579 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1580 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
1581 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
1582 	} else {
1583 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
1584 		aseg->compare  = 0;
1585 	}
1586 
1587 }
1588 
1589 static void set_tavor_ud_seg(struct mthca_tavor_ud_seg *useg,
1590 			     struct ib_send_wr *wr)
1591 {
1592 	useg->lkey    = cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1593 	useg->av_addr =	cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1594 	useg->dqpn    =	cpu_to_be32(wr->wr.ud.remote_qpn);
1595 	useg->qkey    =	cpu_to_be32(wr->wr.ud.remote_qkey);
1596 
1597 }
1598 
1599 static void set_arbel_ud_seg(struct mthca_arbel_ud_seg *useg,
1600 			     struct ib_send_wr *wr)
1601 {
1602 	memcpy(useg->av, to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1603 	useg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1604 	useg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1605 }
1606 
1607 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1608 			  struct ib_send_wr **bad_wr)
1609 {
1610 	struct mthca_dev *dev = to_mdev(ibqp->device);
1611 	struct mthca_qp *qp = to_mqp(ibqp);
1612 	void *wqe;
1613 	void *prev_wqe;
1614 	unsigned long flags;
1615 	int err = 0;
1616 	int nreq;
1617 	int i;
1618 	int size;
1619 	/*
1620 	 * f0 and size0 are only used if nreq != 0, and they will
1621 	 * always be initialized the first time through the main loop
1622 	 * before nreq is incremented.  So nreq cannot become non-zero
1623 	 * without initializing f0 and size0, and they are in fact
1624 	 * never used uninitialized.
1625 	 */
1626 	int uninitialized_var(size0);
1627 	u32 uninitialized_var(f0);
1628 	int ind;
1629 	u8 op0 = 0;
1630 
1631 	spin_lock_irqsave(&qp->sq.lock, flags);
1632 
1633 	/* XXX check that state is OK to post send */
1634 
1635 	ind = qp->sq.next_ind;
1636 
1637 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1638 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1639 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1640 					" %d max, %d nreq)\n", qp->qpn,
1641 					qp->sq.head, qp->sq.tail,
1642 					qp->sq.max, nreq);
1643 			err = -ENOMEM;
1644 			*bad_wr = wr;
1645 			goto out;
1646 		}
1647 
1648 		wqe = get_send_wqe(qp, ind);
1649 		prev_wqe = qp->sq.last;
1650 		qp->sq.last = wqe;
1651 
1652 		((struct mthca_next_seg *) wqe)->nda_op = 0;
1653 		((struct mthca_next_seg *) wqe)->ee_nds = 0;
1654 		((struct mthca_next_seg *) wqe)->flags =
1655 			((wr->send_flags & IB_SEND_SIGNALED) ?
1656 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1657 			((wr->send_flags & IB_SEND_SOLICITED) ?
1658 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1659 			cpu_to_be32(1);
1660 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1661 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1662 			((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
1663 
1664 		wqe += sizeof (struct mthca_next_seg);
1665 		size = sizeof (struct mthca_next_seg) / 16;
1666 
1667 		switch (qp->transport) {
1668 		case RC:
1669 			switch (wr->opcode) {
1670 			case IB_WR_ATOMIC_CMP_AND_SWP:
1671 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1672 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
1673 					      wr->wr.atomic.rkey);
1674 				wqe += sizeof (struct mthca_raddr_seg);
1675 
1676 				set_atomic_seg(wqe, wr);
1677 				wqe += sizeof (struct mthca_atomic_seg);
1678 				size += (sizeof (struct mthca_raddr_seg) +
1679 					 sizeof (struct mthca_atomic_seg)) / 16;
1680 				break;
1681 
1682 			case IB_WR_RDMA_WRITE:
1683 			case IB_WR_RDMA_WRITE_WITH_IMM:
1684 			case IB_WR_RDMA_READ:
1685 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1686 					      wr->wr.rdma.rkey);
1687 				wqe  += sizeof (struct mthca_raddr_seg);
1688 				size += sizeof (struct mthca_raddr_seg) / 16;
1689 				break;
1690 
1691 			default:
1692 				/* No extra segments required for sends */
1693 				break;
1694 			}
1695 
1696 			break;
1697 
1698 		case UC:
1699 			switch (wr->opcode) {
1700 			case IB_WR_RDMA_WRITE:
1701 			case IB_WR_RDMA_WRITE_WITH_IMM:
1702 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1703 					      wr->wr.rdma.rkey);
1704 				wqe  += sizeof (struct mthca_raddr_seg);
1705 				size += sizeof (struct mthca_raddr_seg) / 16;
1706 				break;
1707 
1708 			default:
1709 				/* No extra segments required for sends */
1710 				break;
1711 			}
1712 
1713 			break;
1714 
1715 		case UD:
1716 			set_tavor_ud_seg(wqe, wr);
1717 			wqe  += sizeof (struct mthca_tavor_ud_seg);
1718 			size += sizeof (struct mthca_tavor_ud_seg) / 16;
1719 			break;
1720 
1721 		case MLX:
1722 			err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1723 					       wqe - sizeof (struct mthca_next_seg),
1724 					       wqe);
1725 			if (err) {
1726 				*bad_wr = wr;
1727 				goto out;
1728 			}
1729 			wqe += sizeof (struct mthca_data_seg);
1730 			size += sizeof (struct mthca_data_seg) / 16;
1731 			break;
1732 		}
1733 
1734 		if (wr->num_sge > qp->sq.max_gs) {
1735 			mthca_err(dev, "too many gathers\n");
1736 			err = -EINVAL;
1737 			*bad_wr = wr;
1738 			goto out;
1739 		}
1740 
1741 		for (i = 0; i < wr->num_sge; ++i) {
1742 			mthca_set_data_seg(wqe, wr->sg_list + i);
1743 			wqe  += sizeof (struct mthca_data_seg);
1744 			size += sizeof (struct mthca_data_seg) / 16;
1745 		}
1746 
1747 		/* Add one more inline data segment for ICRC */
1748 		if (qp->transport == MLX) {
1749 			((struct mthca_data_seg *) wqe)->byte_count =
1750 				cpu_to_be32((1 << 31) | 4);
1751 			((u32 *) wqe)[1] = 0;
1752 			wqe += sizeof (struct mthca_data_seg);
1753 			size += sizeof (struct mthca_data_seg) / 16;
1754 		}
1755 
1756 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
1757 
1758 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1759 			mthca_err(dev, "opcode invalid\n");
1760 			err = -EINVAL;
1761 			*bad_wr = wr;
1762 			goto out;
1763 		}
1764 
1765 		((struct mthca_next_seg *) prev_wqe)->nda_op =
1766 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
1767 				     qp->send_wqe_offset) |
1768 				    mthca_opcode[wr->opcode]);
1769 		wmb();
1770 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1771 			cpu_to_be32((nreq ? 0 : MTHCA_NEXT_DBD) | size |
1772 				    ((wr->send_flags & IB_SEND_FENCE) ?
1773 				    MTHCA_NEXT_FENCE : 0));
1774 
1775 		if (!nreq) {
1776 			size0 = size;
1777 			op0   = mthca_opcode[wr->opcode];
1778 			f0    = wr->send_flags & IB_SEND_FENCE ?
1779 				MTHCA_SEND_DOORBELL_FENCE : 0;
1780 		}
1781 
1782 		++ind;
1783 		if (unlikely(ind >= qp->sq.max))
1784 			ind -= qp->sq.max;
1785 	}
1786 
1787 out:
1788 	if (likely(nreq)) {
1789 		wmb();
1790 
1791 		mthca_write64(((qp->sq.next_ind << qp->sq.wqe_shift) +
1792 			       qp->send_wqe_offset) | f0 | op0,
1793 			      (qp->qpn << 8) | size0,
1794 			      dev->kar + MTHCA_SEND_DOORBELL,
1795 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1796 		/*
1797 		 * Make sure doorbells don't leak out of SQ spinlock
1798 		 * and reach the HCA out of order:
1799 		 */
1800 		mmiowb();
1801 	}
1802 
1803 	qp->sq.next_ind = ind;
1804 	qp->sq.head    += nreq;
1805 
1806 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1807 	return err;
1808 }
1809 
1810 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1811 			     struct ib_recv_wr **bad_wr)
1812 {
1813 	struct mthca_dev *dev = to_mdev(ibqp->device);
1814 	struct mthca_qp *qp = to_mqp(ibqp);
1815 	unsigned long flags;
1816 	int err = 0;
1817 	int nreq;
1818 	int i;
1819 	int size;
1820 	/*
1821 	 * size0 is only used if nreq != 0, and it will always be
1822 	 * initialized the first time through the main loop before
1823 	 * nreq is incremented.  So nreq cannot become non-zero
1824 	 * without initializing size0, and it is in fact never used
1825 	 * uninitialized.
1826 	 */
1827 	int uninitialized_var(size0);
1828 	int ind;
1829 	void *wqe;
1830 	void *prev_wqe;
1831 
1832 	spin_lock_irqsave(&qp->rq.lock, flags);
1833 
1834 	/* XXX check that state is OK to post receive */
1835 
1836 	ind = qp->rq.next_ind;
1837 
1838 	for (nreq = 0; wr; wr = wr->next) {
1839 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1840 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1841 					" %d max, %d nreq)\n", qp->qpn,
1842 					qp->rq.head, qp->rq.tail,
1843 					qp->rq.max, nreq);
1844 			err = -ENOMEM;
1845 			*bad_wr = wr;
1846 			goto out;
1847 		}
1848 
1849 		wqe = get_recv_wqe(qp, ind);
1850 		prev_wqe = qp->rq.last;
1851 		qp->rq.last = wqe;
1852 
1853 		((struct mthca_next_seg *) wqe)->ee_nds =
1854 			cpu_to_be32(MTHCA_NEXT_DBD);
1855 		((struct mthca_next_seg *) wqe)->flags = 0;
1856 
1857 		wqe += sizeof (struct mthca_next_seg);
1858 		size = sizeof (struct mthca_next_seg) / 16;
1859 
1860 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1861 			err = -EINVAL;
1862 			*bad_wr = wr;
1863 			goto out;
1864 		}
1865 
1866 		for (i = 0; i < wr->num_sge; ++i) {
1867 			mthca_set_data_seg(wqe, wr->sg_list + i);
1868 			wqe  += sizeof (struct mthca_data_seg);
1869 			size += sizeof (struct mthca_data_seg) / 16;
1870 		}
1871 
1872 		qp->wrid[ind] = wr->wr_id;
1873 
1874 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1875 			cpu_to_be32(MTHCA_NEXT_DBD | size);
1876 
1877 		if (!nreq)
1878 			size0 = size;
1879 
1880 		++ind;
1881 		if (unlikely(ind >= qp->rq.max))
1882 			ind -= qp->rq.max;
1883 
1884 		++nreq;
1885 		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1886 			nreq = 0;
1887 
1888 			wmb();
1889 
1890 			mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1891 				      qp->qpn << 8, dev->kar + MTHCA_RECEIVE_DOORBELL,
1892 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1893 
1894 			qp->rq.next_ind = ind;
1895 			qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1896 		}
1897 	}
1898 
1899 out:
1900 	if (likely(nreq)) {
1901 		wmb();
1902 
1903 		mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1904 			      qp->qpn << 8 | nreq, dev->kar + MTHCA_RECEIVE_DOORBELL,
1905 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1906 	}
1907 
1908 	qp->rq.next_ind = ind;
1909 	qp->rq.head    += nreq;
1910 
1911 	/*
1912 	 * Make sure doorbells don't leak out of RQ spinlock and reach
1913 	 * the HCA out of order:
1914 	 */
1915 	mmiowb();
1916 
1917 	spin_unlock_irqrestore(&qp->rq.lock, flags);
1918 	return err;
1919 }
1920 
1921 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1922 			  struct ib_send_wr **bad_wr)
1923 {
1924 	struct mthca_dev *dev = to_mdev(ibqp->device);
1925 	struct mthca_qp *qp = to_mqp(ibqp);
1926 	u32 dbhi;
1927 	void *wqe;
1928 	void *prev_wqe;
1929 	unsigned long flags;
1930 	int err = 0;
1931 	int nreq;
1932 	int i;
1933 	int size;
1934 	/*
1935 	 * f0 and size0 are only used if nreq != 0, and they will
1936 	 * always be initialized the first time through the main loop
1937 	 * before nreq is incremented.  So nreq cannot become non-zero
1938 	 * without initializing f0 and size0, and they are in fact
1939 	 * never used uninitialized.
1940 	 */
1941 	int uninitialized_var(size0);
1942 	u32 uninitialized_var(f0);
1943 	int ind;
1944 	u8 op0 = 0;
1945 
1946 	spin_lock_irqsave(&qp->sq.lock, flags);
1947 
1948 	/* XXX check that state is OK to post send */
1949 
1950 	ind = qp->sq.head & (qp->sq.max - 1);
1951 
1952 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1953 		if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1954 			nreq = 0;
1955 
1956 			dbhi = (MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1957 				((qp->sq.head & 0xffff) << 8) | f0 | op0;
1958 
1959 			qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1960 
1961 			/*
1962 			 * Make sure that descriptors are written before
1963 			 * doorbell record.
1964 			 */
1965 			wmb();
1966 			*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1967 
1968 			/*
1969 			 * Make sure doorbell record is written before we
1970 			 * write MMIO send doorbell.
1971 			 */
1972 			wmb();
1973 
1974 			mthca_write64(dbhi, (qp->qpn << 8) | size0,
1975 				      dev->kar + MTHCA_SEND_DOORBELL,
1976 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1977 		}
1978 
1979 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1980 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1981 					" %d max, %d nreq)\n", qp->qpn,
1982 					qp->sq.head, qp->sq.tail,
1983 					qp->sq.max, nreq);
1984 			err = -ENOMEM;
1985 			*bad_wr = wr;
1986 			goto out;
1987 		}
1988 
1989 		wqe = get_send_wqe(qp, ind);
1990 		prev_wqe = qp->sq.last;
1991 		qp->sq.last = wqe;
1992 
1993 		((struct mthca_next_seg *) wqe)->flags =
1994 			((wr->send_flags & IB_SEND_SIGNALED) ?
1995 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1996 			((wr->send_flags & IB_SEND_SOLICITED) ?
1997 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1998 			((wr->send_flags & IB_SEND_IP_CSUM) ?
1999 			 cpu_to_be32(MTHCA_NEXT_IP_CSUM | MTHCA_NEXT_TCP_UDP_CSUM) : 0) |
2000 			cpu_to_be32(1);
2001 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
2002 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
2003 			((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
2004 
2005 		wqe += sizeof (struct mthca_next_seg);
2006 		size = sizeof (struct mthca_next_seg) / 16;
2007 
2008 		switch (qp->transport) {
2009 		case RC:
2010 			switch (wr->opcode) {
2011 			case IB_WR_ATOMIC_CMP_AND_SWP:
2012 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2013 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2014 					      wr->wr.atomic.rkey);
2015 				wqe += sizeof (struct mthca_raddr_seg);
2016 
2017 				set_atomic_seg(wqe, wr);
2018 				wqe  += sizeof (struct mthca_atomic_seg);
2019 				size += (sizeof (struct mthca_raddr_seg) +
2020 					 sizeof (struct mthca_atomic_seg)) / 16;
2021 				break;
2022 
2023 			case IB_WR_RDMA_READ:
2024 			case IB_WR_RDMA_WRITE:
2025 			case IB_WR_RDMA_WRITE_WITH_IMM:
2026 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2027 					      wr->wr.rdma.rkey);
2028 				wqe  += sizeof (struct mthca_raddr_seg);
2029 				size += sizeof (struct mthca_raddr_seg) / 16;
2030 				break;
2031 
2032 			default:
2033 				/* No extra segments required for sends */
2034 				break;
2035 			}
2036 
2037 			break;
2038 
2039 		case UC:
2040 			switch (wr->opcode) {
2041 			case IB_WR_RDMA_WRITE:
2042 			case IB_WR_RDMA_WRITE_WITH_IMM:
2043 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2044 					      wr->wr.rdma.rkey);
2045 				wqe  += sizeof (struct mthca_raddr_seg);
2046 				size += sizeof (struct mthca_raddr_seg) / 16;
2047 				break;
2048 
2049 			default:
2050 				/* No extra segments required for sends */
2051 				break;
2052 			}
2053 
2054 			break;
2055 
2056 		case UD:
2057 			set_arbel_ud_seg(wqe, wr);
2058 			wqe  += sizeof (struct mthca_arbel_ud_seg);
2059 			size += sizeof (struct mthca_arbel_ud_seg) / 16;
2060 			break;
2061 
2062 		case MLX:
2063 			err = build_mlx_header(dev, to_msqp(qp), ind, wr,
2064 					       wqe - sizeof (struct mthca_next_seg),
2065 					       wqe);
2066 			if (err) {
2067 				*bad_wr = wr;
2068 				goto out;
2069 			}
2070 			wqe += sizeof (struct mthca_data_seg);
2071 			size += sizeof (struct mthca_data_seg) / 16;
2072 			break;
2073 		}
2074 
2075 		if (wr->num_sge > qp->sq.max_gs) {
2076 			mthca_err(dev, "too many gathers\n");
2077 			err = -EINVAL;
2078 			*bad_wr = wr;
2079 			goto out;
2080 		}
2081 
2082 		for (i = 0; i < wr->num_sge; ++i) {
2083 			mthca_set_data_seg(wqe, wr->sg_list + i);
2084 			wqe  += sizeof (struct mthca_data_seg);
2085 			size += sizeof (struct mthca_data_seg) / 16;
2086 		}
2087 
2088 		/* Add one more inline data segment for ICRC */
2089 		if (qp->transport == MLX) {
2090 			((struct mthca_data_seg *) wqe)->byte_count =
2091 				cpu_to_be32((1 << 31) | 4);
2092 			((u32 *) wqe)[1] = 0;
2093 			wqe += sizeof (struct mthca_data_seg);
2094 			size += sizeof (struct mthca_data_seg) / 16;
2095 		}
2096 
2097 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
2098 
2099 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2100 			mthca_err(dev, "opcode invalid\n");
2101 			err = -EINVAL;
2102 			*bad_wr = wr;
2103 			goto out;
2104 		}
2105 
2106 		((struct mthca_next_seg *) prev_wqe)->nda_op =
2107 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
2108 				     qp->send_wqe_offset) |
2109 				    mthca_opcode[wr->opcode]);
2110 		wmb();
2111 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
2112 			cpu_to_be32(MTHCA_NEXT_DBD | size |
2113 				    ((wr->send_flags & IB_SEND_FENCE) ?
2114 				     MTHCA_NEXT_FENCE : 0));
2115 
2116 		if (!nreq) {
2117 			size0 = size;
2118 			op0   = mthca_opcode[wr->opcode];
2119 			f0    = wr->send_flags & IB_SEND_FENCE ?
2120 				MTHCA_SEND_DOORBELL_FENCE : 0;
2121 		}
2122 
2123 		++ind;
2124 		if (unlikely(ind >= qp->sq.max))
2125 			ind -= qp->sq.max;
2126 	}
2127 
2128 out:
2129 	if (likely(nreq)) {
2130 		dbhi = (nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0;
2131 
2132 		qp->sq.head += nreq;
2133 
2134 		/*
2135 		 * Make sure that descriptors are written before
2136 		 * doorbell record.
2137 		 */
2138 		wmb();
2139 		*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2140 
2141 		/*
2142 		 * Make sure doorbell record is written before we
2143 		 * write MMIO send doorbell.
2144 		 */
2145 		wmb();
2146 
2147 		mthca_write64(dbhi, (qp->qpn << 8) | size0, dev->kar + MTHCA_SEND_DOORBELL,
2148 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2149 	}
2150 
2151 	/*
2152 	 * Make sure doorbells don't leak out of SQ spinlock and reach
2153 	 * the HCA out of order:
2154 	 */
2155 	mmiowb();
2156 
2157 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2158 	return err;
2159 }
2160 
2161 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2162 			     struct ib_recv_wr **bad_wr)
2163 {
2164 	struct mthca_dev *dev = to_mdev(ibqp->device);
2165 	struct mthca_qp *qp = to_mqp(ibqp);
2166 	unsigned long flags;
2167 	int err = 0;
2168 	int nreq;
2169 	int ind;
2170 	int i;
2171 	void *wqe;
2172 
2173 	spin_lock_irqsave(&qp->rq.lock, flags);
2174 
2175 	/* XXX check that state is OK to post receive */
2176 
2177 	ind = qp->rq.head & (qp->rq.max - 1);
2178 
2179 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2180 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2181 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2182 					" %d max, %d nreq)\n", qp->qpn,
2183 					qp->rq.head, qp->rq.tail,
2184 					qp->rq.max, nreq);
2185 			err = -ENOMEM;
2186 			*bad_wr = wr;
2187 			goto out;
2188 		}
2189 
2190 		wqe = get_recv_wqe(qp, ind);
2191 
2192 		((struct mthca_next_seg *) wqe)->flags = 0;
2193 
2194 		wqe += sizeof (struct mthca_next_seg);
2195 
2196 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2197 			err = -EINVAL;
2198 			*bad_wr = wr;
2199 			goto out;
2200 		}
2201 
2202 		for (i = 0; i < wr->num_sge; ++i) {
2203 			mthca_set_data_seg(wqe, wr->sg_list + i);
2204 			wqe += sizeof (struct mthca_data_seg);
2205 		}
2206 
2207 		if (i < qp->rq.max_gs)
2208 			mthca_set_data_seg_inval(wqe);
2209 
2210 		qp->wrid[ind] = wr->wr_id;
2211 
2212 		++ind;
2213 		if (unlikely(ind >= qp->rq.max))
2214 			ind -= qp->rq.max;
2215 	}
2216 out:
2217 	if (likely(nreq)) {
2218 		qp->rq.head += nreq;
2219 
2220 		/*
2221 		 * Make sure that descriptors are written before
2222 		 * doorbell record.
2223 		 */
2224 		wmb();
2225 		*qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2226 	}
2227 
2228 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2229 	return err;
2230 }
2231 
2232 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2233 			int index, int *dbd, __be32 *new_wqe)
2234 {
2235 	struct mthca_next_seg *next;
2236 
2237 	/*
2238 	 * For SRQs, all receive WQEs generate a CQE, so we're always
2239 	 * at the end of the doorbell chain.
2240 	 */
2241 	if (qp->ibqp.srq && !is_send) {
2242 		*new_wqe = 0;
2243 		return;
2244 	}
2245 
2246 	if (is_send)
2247 		next = get_send_wqe(qp, index);
2248 	else
2249 		next = get_recv_wqe(qp, index);
2250 
2251 	*dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2252 	if (next->ee_nds & cpu_to_be32(0x3f))
2253 		*new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2254 			(next->ee_nds & cpu_to_be32(0x3f));
2255 	else
2256 		*new_wqe = 0;
2257 }
2258 
2259 int mthca_init_qp_table(struct mthca_dev *dev)
2260 {
2261 	int err;
2262 	u8 status;
2263 	int i;
2264 
2265 	spin_lock_init(&dev->qp_table.lock);
2266 
2267 	/*
2268 	 * We reserve 2 extra QPs per port for the special QPs.  The
2269 	 * special QP for port 1 has to be even, so round up.
2270 	 */
2271 	dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2272 	err = mthca_alloc_init(&dev->qp_table.alloc,
2273 			       dev->limits.num_qps,
2274 			       (1 << 24) - 1,
2275 			       dev->qp_table.sqp_start +
2276 			       MTHCA_MAX_PORTS * 2);
2277 	if (err)
2278 		return err;
2279 
2280 	err = mthca_array_init(&dev->qp_table.qp,
2281 			       dev->limits.num_qps);
2282 	if (err) {
2283 		mthca_alloc_cleanup(&dev->qp_table.alloc);
2284 		return err;
2285 	}
2286 
2287 	for (i = 0; i < 2; ++i) {
2288 		err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2289 					    dev->qp_table.sqp_start + i * 2,
2290 					    &status);
2291 		if (err)
2292 			goto err_out;
2293 		if (status) {
2294 			mthca_warn(dev, "CONF_SPECIAL_QP returned "
2295 				   "status %02x, aborting.\n",
2296 				   status);
2297 			err = -EINVAL;
2298 			goto err_out;
2299 		}
2300 	}
2301 	return 0;
2302 
2303  err_out:
2304 	for (i = 0; i < 2; ++i)
2305 		mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2306 
2307 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2308 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2309 
2310 	return err;
2311 }
2312 
2313 void mthca_cleanup_qp_table(struct mthca_dev *dev)
2314 {
2315 	int i;
2316 	u8 status;
2317 
2318 	for (i = 0; i < 2; ++i)
2319 		mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2320 
2321 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2322 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2323 }
2324