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