1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Cisco Systems. All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * OpenIB.org BSD license below:
12  *
13  *     Redistribution and use in source and binary forms, with or
14  *     without modification, are permitted provided that the following
15  *     conditions are met:
16  *
17  *      - Redistributions of source code must retain the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer.
20  *
21  *      - Redistributions in binary form must reproduce the above
22  *        copyright notice, this list of conditions and the following
23  *        disclaimer in the documentation and/or other materials
24  *        provided with the distribution.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33  * SOFTWARE.
34  *
35  * $Id: mthca_qp.c 1355 2004-12-17 15:23:43Z roland $
36  */
37 
38 #include <linux/string.h>
39 #include <linux/slab.h>
40 #include <linux/sched.h>
41 
42 #include <asm/io.h>
43 
44 #include <rdma/ib_verbs.h>
45 #include <rdma/ib_cache.h>
46 #include <rdma/ib_pack.h>
47 
48 #include "mthca_dev.h"
49 #include "mthca_cmd.h"
50 #include "mthca_memfree.h"
51 #include "mthca_wqe.h"
52 
53 enum {
54 	MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
55 	MTHCA_ACK_REQ_FREQ       = 10,
56 	MTHCA_FLIGHT_LIMIT       = 9,
57 	MTHCA_UD_HEADER_SIZE     = 72, /* largest UD header possible */
58 	MTHCA_INLINE_HEADER_SIZE = 4,  /* data segment overhead for inline */
59 	MTHCA_INLINE_CHUNK_SIZE  = 16  /* inline data segment chunk */
60 };
61 
62 enum {
63 	MTHCA_QP_STATE_RST  = 0,
64 	MTHCA_QP_STATE_INIT = 1,
65 	MTHCA_QP_STATE_RTR  = 2,
66 	MTHCA_QP_STATE_RTS  = 3,
67 	MTHCA_QP_STATE_SQE  = 4,
68 	MTHCA_QP_STATE_SQD  = 5,
69 	MTHCA_QP_STATE_ERR  = 6,
70 	MTHCA_QP_STATE_DRAINING = 7
71 };
72 
73 enum {
74 	MTHCA_QP_ST_RC 	= 0x0,
75 	MTHCA_QP_ST_UC 	= 0x1,
76 	MTHCA_QP_ST_RD 	= 0x2,
77 	MTHCA_QP_ST_UD 	= 0x3,
78 	MTHCA_QP_ST_MLX = 0x7
79 };
80 
81 enum {
82 	MTHCA_QP_PM_MIGRATED = 0x3,
83 	MTHCA_QP_PM_ARMED    = 0x0,
84 	MTHCA_QP_PM_REARM    = 0x1
85 };
86 
87 enum {
88 	/* qp_context flags */
89 	MTHCA_QP_BIT_DE  = 1 <<  8,
90 	/* params1 */
91 	MTHCA_QP_BIT_SRE = 1 << 15,
92 	MTHCA_QP_BIT_SWE = 1 << 14,
93 	MTHCA_QP_BIT_SAE = 1 << 13,
94 	MTHCA_QP_BIT_SIC = 1 <<  4,
95 	MTHCA_QP_BIT_SSC = 1 <<  3,
96 	/* params2 */
97 	MTHCA_QP_BIT_RRE = 1 << 15,
98 	MTHCA_QP_BIT_RWE = 1 << 14,
99 	MTHCA_QP_BIT_RAE = 1 << 13,
100 	MTHCA_QP_BIT_RIC = 1 <<  4,
101 	MTHCA_QP_BIT_RSC = 1 <<  3
102 };
103 
104 enum {
105 	MTHCA_SEND_DOORBELL_FENCE = 1 << 5
106 };
107 
108 struct mthca_qp_path {
109 	__be32 port_pkey;
110 	u8     rnr_retry;
111 	u8     g_mylmc;
112 	__be16 rlid;
113 	u8     ackto;
114 	u8     mgid_index;
115 	u8     static_rate;
116 	u8     hop_limit;
117 	__be32 sl_tclass_flowlabel;
118 	u8     rgid[16];
119 } __attribute__((packed));
120 
121 struct mthca_qp_context {
122 	__be32 flags;
123 	__be32 tavor_sched_queue; /* Reserved on Arbel */
124 	u8     mtu_msgmax;
125 	u8     rq_size_stride;	/* Reserved on Tavor */
126 	u8     sq_size_stride;	/* Reserved on Tavor */
127 	u8     rlkey_arbel_sched_queue;	/* Reserved on Tavor */
128 	__be32 usr_page;
129 	__be32 local_qpn;
130 	__be32 remote_qpn;
131 	u32    reserved1[2];
132 	struct mthca_qp_path pri_path;
133 	struct mthca_qp_path alt_path;
134 	__be32 rdd;
135 	__be32 pd;
136 	__be32 wqe_base;
137 	__be32 wqe_lkey;
138 	__be32 params1;
139 	__be32 reserved2;
140 	__be32 next_send_psn;
141 	__be32 cqn_snd;
142 	__be32 snd_wqe_base_l;	/* Next send WQE on Tavor */
143 	__be32 snd_db_index;	/* (debugging only entries) */
144 	__be32 last_acked_psn;
145 	__be32 ssn;
146 	__be32 params2;
147 	__be32 rnr_nextrecvpsn;
148 	__be32 ra_buff_indx;
149 	__be32 cqn_rcv;
150 	__be32 rcv_wqe_base_l;	/* Next recv WQE on Tavor */
151 	__be32 rcv_db_index;	/* (debugging only entries) */
152 	__be32 qkey;
153 	__be32 srqn;
154 	__be32 rmsn;
155 	__be16 rq_wqe_counter;	/* reserved on Tavor */
156 	__be16 sq_wqe_counter;	/* reserved on Tavor */
157 	u32    reserved3[18];
158 } __attribute__((packed));
159 
160 struct mthca_qp_param {
161 	__be32 opt_param_mask;
162 	u32    reserved1;
163 	struct mthca_qp_context context;
164 	u32    reserved2[62];
165 } __attribute__((packed));
166 
167 enum {
168 	MTHCA_QP_OPTPAR_ALT_ADDR_PATH     = 1 << 0,
169 	MTHCA_QP_OPTPAR_RRE               = 1 << 1,
170 	MTHCA_QP_OPTPAR_RAE               = 1 << 2,
171 	MTHCA_QP_OPTPAR_RWE               = 1 << 3,
172 	MTHCA_QP_OPTPAR_PKEY_INDEX        = 1 << 4,
173 	MTHCA_QP_OPTPAR_Q_KEY             = 1 << 5,
174 	MTHCA_QP_OPTPAR_RNR_TIMEOUT       = 1 << 6,
175 	MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
176 	MTHCA_QP_OPTPAR_SRA_MAX           = 1 << 8,
177 	MTHCA_QP_OPTPAR_RRA_MAX           = 1 << 9,
178 	MTHCA_QP_OPTPAR_PM_STATE          = 1 << 10,
179 	MTHCA_QP_OPTPAR_PORT_NUM          = 1 << 11,
180 	MTHCA_QP_OPTPAR_RETRY_COUNT       = 1 << 12,
181 	MTHCA_QP_OPTPAR_ALT_RNR_RETRY     = 1 << 13,
182 	MTHCA_QP_OPTPAR_ACK_TIMEOUT       = 1 << 14,
183 	MTHCA_QP_OPTPAR_RNR_RETRY         = 1 << 15,
184 	MTHCA_QP_OPTPAR_SCHED_QUEUE       = 1 << 16
185 };
186 
187 static const u8 mthca_opcode[] = {
188 	[IB_WR_SEND]                 = MTHCA_OPCODE_SEND,
189 	[IB_WR_SEND_WITH_IMM]        = MTHCA_OPCODE_SEND_IMM,
190 	[IB_WR_RDMA_WRITE]           = MTHCA_OPCODE_RDMA_WRITE,
191 	[IB_WR_RDMA_WRITE_WITH_IMM]  = MTHCA_OPCODE_RDMA_WRITE_IMM,
192 	[IB_WR_RDMA_READ]            = MTHCA_OPCODE_RDMA_READ,
193 	[IB_WR_ATOMIC_CMP_AND_SWP]   = MTHCA_OPCODE_ATOMIC_CS,
194 	[IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
195 };
196 
197 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
198 {
199 	return qp->qpn >= dev->qp_table.sqp_start &&
200 		qp->qpn <= dev->qp_table.sqp_start + 3;
201 }
202 
203 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
204 {
205 	return qp->qpn >= dev->qp_table.sqp_start &&
206 		qp->qpn <= dev->qp_table.sqp_start + 1;
207 }
208 
209 static void *get_recv_wqe(struct mthca_qp *qp, int n)
210 {
211 	if (qp->is_direct)
212 		return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
213 	else
214 		return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
215 			((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
216 }
217 
218 static void *get_send_wqe(struct mthca_qp *qp, int n)
219 {
220 	if (qp->is_direct)
221 		return qp->queue.direct.buf + qp->send_wqe_offset +
222 			(n << qp->sq.wqe_shift);
223 	else
224 		return qp->queue.page_list[(qp->send_wqe_offset +
225 					    (n << qp->sq.wqe_shift)) >>
226 					   PAGE_SHIFT].buf +
227 			((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
228 			 (PAGE_SIZE - 1));
229 }
230 
231 static void mthca_wq_reset(struct mthca_wq *wq)
232 {
233 	wq->next_ind  = 0;
234 	wq->last_comp = wq->max - 1;
235 	wq->head      = 0;
236 	wq->tail      = 0;
237 }
238 
239 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
240 		    enum ib_event_type event_type)
241 {
242 	struct mthca_qp *qp;
243 	struct ib_event event;
244 
245 	spin_lock(&dev->qp_table.lock);
246 	qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
247 	if (qp)
248 		++qp->refcount;
249 	spin_unlock(&dev->qp_table.lock);
250 
251 	if (!qp) {
252 		mthca_warn(dev, "Async event for bogus QP %08x\n", qpn);
253 		return;
254 	}
255 
256 	if (event_type == IB_EVENT_PATH_MIG)
257 		qp->port = qp->alt_port;
258 
259 	event.device      = &dev->ib_dev;
260 	event.event       = event_type;
261 	event.element.qp  = &qp->ibqp;
262 	if (qp->ibqp.event_handler)
263 		qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
264 
265 	spin_lock(&dev->qp_table.lock);
266 	if (!--qp->refcount)
267 		wake_up(&qp->wait);
268 	spin_unlock(&dev->qp_table.lock);
269 }
270 
271 static int to_mthca_state(enum ib_qp_state ib_state)
272 {
273 	switch (ib_state) {
274 	case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
275 	case IB_QPS_INIT:  return MTHCA_QP_STATE_INIT;
276 	case IB_QPS_RTR:   return MTHCA_QP_STATE_RTR;
277 	case IB_QPS_RTS:   return MTHCA_QP_STATE_RTS;
278 	case IB_QPS_SQD:   return MTHCA_QP_STATE_SQD;
279 	case IB_QPS_SQE:   return MTHCA_QP_STATE_SQE;
280 	case IB_QPS_ERR:   return MTHCA_QP_STATE_ERR;
281 	default:                return -1;
282 	}
283 }
284 
285 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
286 
287 static int to_mthca_st(int transport)
288 {
289 	switch (transport) {
290 	case RC:  return MTHCA_QP_ST_RC;
291 	case UC:  return MTHCA_QP_ST_UC;
292 	case UD:  return MTHCA_QP_ST_UD;
293 	case RD:  return MTHCA_QP_ST_RD;
294 	case MLX: return MTHCA_QP_ST_MLX;
295 	default:  return -1;
296 	}
297 }
298 
299 static void store_attrs(struct mthca_sqp *sqp, const struct ib_qp_attr *attr,
300 			int attr_mask)
301 {
302 	if (attr_mask & IB_QP_PKEY_INDEX)
303 		sqp->pkey_index = attr->pkey_index;
304 	if (attr_mask & IB_QP_QKEY)
305 		sqp->qkey = attr->qkey;
306 	if (attr_mask & IB_QP_SQ_PSN)
307 		sqp->send_psn = attr->sq_psn;
308 }
309 
310 static void init_port(struct mthca_dev *dev, int port)
311 {
312 	int err;
313 	u8 status;
314 	struct mthca_init_ib_param param;
315 
316 	memset(&param, 0, sizeof param);
317 
318 	param.port_width = dev->limits.port_width_cap;
319 	param.vl_cap     = dev->limits.vl_cap;
320 	param.mtu_cap    = dev->limits.mtu_cap;
321 	param.gid_cap    = dev->limits.gid_table_len;
322 	param.pkey_cap   = dev->limits.pkey_table_len;
323 
324 	err = mthca_INIT_IB(dev, &param, port, &status);
325 	if (err)
326 		mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
327 	if (status)
328 		mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
329 }
330 
331 static __be32 get_hw_access_flags(struct mthca_qp *qp, const struct ib_qp_attr *attr,
332 				  int attr_mask)
333 {
334 	u8 dest_rd_atomic;
335 	u32 access_flags;
336 	u32 hw_access_flags = 0;
337 
338 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
339 		dest_rd_atomic = attr->max_dest_rd_atomic;
340 	else
341 		dest_rd_atomic = qp->resp_depth;
342 
343 	if (attr_mask & IB_QP_ACCESS_FLAGS)
344 		access_flags = attr->qp_access_flags;
345 	else
346 		access_flags = qp->atomic_rd_en;
347 
348 	if (!dest_rd_atomic)
349 		access_flags &= IB_ACCESS_REMOTE_WRITE;
350 
351 	if (access_flags & IB_ACCESS_REMOTE_READ)
352 		hw_access_flags |= MTHCA_QP_BIT_RRE;
353 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
354 		hw_access_flags |= MTHCA_QP_BIT_RAE;
355 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
356 		hw_access_flags |= MTHCA_QP_BIT_RWE;
357 
358 	return cpu_to_be32(hw_access_flags);
359 }
360 
361 static inline enum ib_qp_state to_ib_qp_state(int mthca_state)
362 {
363 	switch (mthca_state) {
364 	case MTHCA_QP_STATE_RST:      return IB_QPS_RESET;
365 	case MTHCA_QP_STATE_INIT:     return IB_QPS_INIT;
366 	case MTHCA_QP_STATE_RTR:      return IB_QPS_RTR;
367 	case MTHCA_QP_STATE_RTS:      return IB_QPS_RTS;
368 	case MTHCA_QP_STATE_DRAINING:
369 	case MTHCA_QP_STATE_SQD:      return IB_QPS_SQD;
370 	case MTHCA_QP_STATE_SQE:      return IB_QPS_SQE;
371 	case MTHCA_QP_STATE_ERR:      return IB_QPS_ERR;
372 	default:                      return -1;
373 	}
374 }
375 
376 static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state)
377 {
378 	switch (mthca_mig_state) {
379 	case 0:  return IB_MIG_ARMED;
380 	case 1:  return IB_MIG_REARM;
381 	case 3:  return IB_MIG_MIGRATED;
382 	default: return -1;
383 	}
384 }
385 
386 static int to_ib_qp_access_flags(int mthca_flags)
387 {
388 	int ib_flags = 0;
389 
390 	if (mthca_flags & MTHCA_QP_BIT_RRE)
391 		ib_flags |= IB_ACCESS_REMOTE_READ;
392 	if (mthca_flags & MTHCA_QP_BIT_RWE)
393 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
394 	if (mthca_flags & MTHCA_QP_BIT_RAE)
395 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
396 
397 	return ib_flags;
398 }
399 
400 static void to_ib_ah_attr(struct mthca_dev *dev, struct ib_ah_attr *ib_ah_attr,
401 				struct mthca_qp_path *path)
402 {
403 	memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
404 	ib_ah_attr->port_num 	  = (be32_to_cpu(path->port_pkey) >> 24) & 0x3;
405 
406 	if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->limits.num_ports)
407 		return;
408 
409 	ib_ah_attr->dlid     	  = be16_to_cpu(path->rlid);
410 	ib_ah_attr->sl       	  = be32_to_cpu(path->sl_tclass_flowlabel) >> 28;
411 	ib_ah_attr->src_path_bits = path->g_mylmc & 0x7f;
412 	ib_ah_attr->static_rate   = mthca_rate_to_ib(dev,
413 						     path->static_rate & 0xf,
414 						     ib_ah_attr->port_num);
415 	ib_ah_attr->ah_flags      = (path->g_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
416 	if (ib_ah_attr->ah_flags) {
417 		ib_ah_attr->grh.sgid_index = path->mgid_index & (dev->limits.gid_table_len - 1);
418 		ib_ah_attr->grh.hop_limit  = path->hop_limit;
419 		ib_ah_attr->grh.traffic_class =
420 			(be32_to_cpu(path->sl_tclass_flowlabel) >> 20) & 0xff;
421 		ib_ah_attr->grh.flow_label =
422 			be32_to_cpu(path->sl_tclass_flowlabel) & 0xfffff;
423 		memcpy(ib_ah_attr->grh.dgid.raw,
424 			path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
425 	}
426 }
427 
428 int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
429 		   struct ib_qp_init_attr *qp_init_attr)
430 {
431 	struct mthca_dev *dev = to_mdev(ibqp->device);
432 	struct mthca_qp *qp = to_mqp(ibqp);
433 	int err = 0;
434 	struct mthca_mailbox *mailbox = NULL;
435 	struct mthca_qp_param *qp_param;
436 	struct mthca_qp_context *context;
437 	int mthca_state;
438 	u8 status;
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 		return PTR_ERR(mailbox);
448 
449 	err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox, &status);
450 	if (err)
451 		goto out;
452 	if (status) {
453 		mthca_warn(dev, "QUERY_QP returned status %02x\n", status);
454 		err = -EINVAL;
455 		goto out;
456 	}
457 
458 	qp_param    = mailbox->buf;
459 	context     = &qp_param->context;
460 	mthca_state = be32_to_cpu(context->flags) >> 28;
461 
462 	qp_attr->qp_state 	     = to_ib_qp_state(mthca_state);
463 	qp_attr->path_mtu 	     = context->mtu_msgmax >> 5;
464 	qp_attr->path_mig_state      =
465 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
466 	qp_attr->qkey 		     = be32_to_cpu(context->qkey);
467 	qp_attr->rq_psn 	     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
468 	qp_attr->sq_psn 	     = be32_to_cpu(context->next_send_psn) & 0xffffff;
469 	qp_attr->dest_qp_num 	     = be32_to_cpu(context->remote_qpn) & 0xffffff;
470 	qp_attr->qp_access_flags     =
471 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
472 
473 	if (qp->transport == RC || qp->transport == UC) {
474 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
475 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
476 		qp_attr->alt_pkey_index =
477 			be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
478 		qp_attr->alt_port_num 	= qp_attr->alt_ah_attr.port_num;
479 	}
480 
481 	qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
482 	qp_attr->port_num   =
483 		(be32_to_cpu(context->pri_path.port_pkey) >> 24) & 0x3;
484 
485 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
486 	qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
487 
488 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
489 
490 	qp_attr->max_dest_rd_atomic =
491 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
492 	qp_attr->min_rnr_timer 	    =
493 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
494 	qp_attr->timeout 	    = context->pri_path.ackto >> 3;
495 	qp_attr->retry_cnt 	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
496 	qp_attr->rnr_retry 	    = context->pri_path.rnr_retry >> 5;
497 	qp_attr->alt_timeout 	    = context->alt_path.ackto >> 3;
498 
499 done:
500 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
501 	qp_attr->cap.max_send_wr     = qp->sq.max;
502 	qp_attr->cap.max_recv_wr     = qp->rq.max;
503 	qp_attr->cap.max_send_sge    = qp->sq.max_gs;
504 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
505 	qp_attr->cap.max_inline_data = qp->max_inline_data;
506 
507 	qp_init_attr->cap	     = qp_attr->cap;
508 
509 out:
510 	mthca_free_mailbox(dev, mailbox);
511 	return err;
512 }
513 
514 static int mthca_path_set(struct mthca_dev *dev, const struct ib_ah_attr *ah,
515 			  struct mthca_qp_path *path, u8 port)
516 {
517 	path->g_mylmc     = ah->src_path_bits & 0x7f;
518 	path->rlid        = cpu_to_be16(ah->dlid);
519 	path->static_rate = mthca_get_rate(dev, ah->static_rate, port);
520 
521 	if (ah->ah_flags & IB_AH_GRH) {
522 		if (ah->grh.sgid_index >= dev->limits.gid_table_len) {
523 			mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
524 				  ah->grh.sgid_index, dev->limits.gid_table_len-1);
525 			return -1;
526 		}
527 
528 		path->g_mylmc   |= 1 << 7;
529 		path->mgid_index = ah->grh.sgid_index;
530 		path->hop_limit  = ah->grh.hop_limit;
531 		path->sl_tclass_flowlabel =
532 			cpu_to_be32((ah->sl << 28)                |
533 				    (ah->grh.traffic_class << 20) |
534 				    (ah->grh.flow_label));
535 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
536 	} else
537 		path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28);
538 
539 	return 0;
540 }
541 
542 static int __mthca_modify_qp(struct ib_qp *ibqp,
543 			     const struct ib_qp_attr *attr, int attr_mask,
544 			     enum ib_qp_state cur_state, enum ib_qp_state new_state)
545 {
546 	struct mthca_dev *dev = to_mdev(ibqp->device);
547 	struct mthca_qp *qp = to_mqp(ibqp);
548 	struct mthca_mailbox *mailbox;
549 	struct mthca_qp_param *qp_param;
550 	struct mthca_qp_context *qp_context;
551 	u32 sqd_event = 0;
552 	u8 status;
553 	int err = -EINVAL;
554 
555 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
556 	if (IS_ERR(mailbox)) {
557 		err = PTR_ERR(mailbox);
558 		goto out;
559 	}
560 	qp_param = mailbox->buf;
561 	qp_context = &qp_param->context;
562 	memset(qp_param, 0, sizeof *qp_param);
563 
564 	qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
565 					     (to_mthca_st(qp->transport) << 16));
566 	qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
567 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
568 		qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
569 	else {
570 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
571 		switch (attr->path_mig_state) {
572 		case IB_MIG_MIGRATED:
573 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
574 			break;
575 		case IB_MIG_REARM:
576 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
577 			break;
578 		case IB_MIG_ARMED:
579 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
580 			break;
581 		}
582 	}
583 
584 	/* leave tavor_sched_queue as 0 */
585 
586 	if (qp->transport == MLX || qp->transport == UD)
587 		qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
588 	else if (attr_mask & IB_QP_PATH_MTU) {
589 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
590 			mthca_dbg(dev, "path MTU (%u) is invalid\n",
591 				  attr->path_mtu);
592 			goto out_mailbox;
593 		}
594 		qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
595 	}
596 
597 	if (mthca_is_memfree(dev)) {
598 		if (qp->rq.max)
599 			qp_context->rq_size_stride = ilog2(qp->rq.max) << 3;
600 		qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
601 
602 		if (qp->sq.max)
603 			qp_context->sq_size_stride = ilog2(qp->sq.max) << 3;
604 		qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
605 	}
606 
607 	/* leave arbel_sched_queue as 0 */
608 
609 	if (qp->ibqp.uobject)
610 		qp_context->usr_page =
611 			cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
612 	else
613 		qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
614 	qp_context->local_qpn  = cpu_to_be32(qp->qpn);
615 	if (attr_mask & IB_QP_DEST_QPN) {
616 		qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
617 	}
618 
619 	if (qp->transport == MLX)
620 		qp_context->pri_path.port_pkey |=
621 			cpu_to_be32(qp->port << 24);
622 	else {
623 		if (attr_mask & IB_QP_PORT) {
624 			qp_context->pri_path.port_pkey |=
625 				cpu_to_be32(attr->port_num << 24);
626 			qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
627 		}
628 	}
629 
630 	if (attr_mask & IB_QP_PKEY_INDEX) {
631 		qp_context->pri_path.port_pkey |=
632 			cpu_to_be32(attr->pkey_index);
633 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
634 	}
635 
636 	if (attr_mask & IB_QP_RNR_RETRY) {
637 		qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
638 			attr->rnr_retry << 5;
639 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
640 							MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
641 	}
642 
643 	if (attr_mask & IB_QP_AV) {
644 		if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
645 				   attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
646 			goto out_mailbox;
647 
648 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
649 	}
650 
651 	if (ibqp->qp_type == IB_QPT_RC &&
652 	    cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
653 		u8 sched_queue = ibqp->uobject ? 0x2 : 0x1;
654 
655 		if (mthca_is_memfree(dev))
656 			qp_context->rlkey_arbel_sched_queue |= sched_queue;
657 		else
658 			qp_context->tavor_sched_queue |= cpu_to_be32(sched_queue);
659 
660 		qp_param->opt_param_mask |=
661 			cpu_to_be32(MTHCA_QP_OPTPAR_SCHED_QUEUE);
662 	}
663 
664 	if (attr_mask & IB_QP_TIMEOUT) {
665 		qp_context->pri_path.ackto = attr->timeout << 3;
666 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
667 	}
668 
669 	if (attr_mask & IB_QP_ALT_PATH) {
670 		if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
671 			mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
672 				  attr->alt_pkey_index, dev->limits.pkey_table_len-1);
673 			goto out_mailbox;
674 		}
675 
676 		if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
677 			mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
678 				attr->alt_port_num);
679 			goto out_mailbox;
680 		}
681 
682 		if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
683 				   attr->alt_ah_attr.port_num))
684 			goto out_mailbox;
685 
686 		qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
687 							      attr->alt_port_num << 24);
688 		qp_context->alt_path.ackto = attr->alt_timeout << 3;
689 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
690 	}
691 
692 	/* leave rdd as 0 */
693 	qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
694 	/* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
695 	qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
696 	qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
697 					     (MTHCA_FLIGHT_LIMIT << 24) |
698 					     MTHCA_QP_BIT_SWE);
699 	if (qp->sq_policy == IB_SIGNAL_ALL_WR)
700 		qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
701 	if (attr_mask & IB_QP_RETRY_CNT) {
702 		qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
703 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
704 	}
705 
706 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
707 		if (attr->max_rd_atomic) {
708 			qp_context->params1 |=
709 				cpu_to_be32(MTHCA_QP_BIT_SRE |
710 					    MTHCA_QP_BIT_SAE);
711 			qp_context->params1 |=
712 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
713 		}
714 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
715 	}
716 
717 	if (attr_mask & IB_QP_SQ_PSN)
718 		qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
719 	qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
720 
721 	if (mthca_is_memfree(dev)) {
722 		qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
723 		qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
724 	}
725 
726 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
727 		if (attr->max_dest_rd_atomic)
728 			qp_context->params2 |=
729 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
730 
731 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
732 	}
733 
734 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
735 		qp_context->params2      |= get_hw_access_flags(qp, attr, attr_mask);
736 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
737 							MTHCA_QP_OPTPAR_RRE |
738 							MTHCA_QP_OPTPAR_RAE);
739 	}
740 
741 	qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
742 
743 	if (ibqp->srq)
744 		qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
745 
746 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
747 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
748 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
749 	}
750 	if (attr_mask & IB_QP_RQ_PSN)
751 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
752 
753 	qp_context->ra_buff_indx =
754 		cpu_to_be32(dev->qp_table.rdb_base +
755 			    ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
756 			     dev->qp_table.rdb_shift));
757 
758 	qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
759 
760 	if (mthca_is_memfree(dev))
761 		qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
762 
763 	if (attr_mask & IB_QP_QKEY) {
764 		qp_context->qkey = cpu_to_be32(attr->qkey);
765 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
766 	}
767 
768 	if (ibqp->srq)
769 		qp_context->srqn = cpu_to_be32(1 << 24 |
770 					       to_msrq(ibqp->srq)->srqn);
771 
772 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
773 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY		&&
774 	    attr->en_sqd_async_notify)
775 		sqd_event = 1 << 31;
776 
777 	err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
778 			      mailbox, sqd_event, &status);
779 	if (err)
780 		goto out_mailbox;
781 	if (status) {
782 		mthca_warn(dev, "modify QP %d->%d returned status %02x.\n",
783 			   cur_state, new_state, status);
784 		err = -EINVAL;
785 		goto out_mailbox;
786 	}
787 
788 	qp->state = new_state;
789 	if (attr_mask & IB_QP_ACCESS_FLAGS)
790 		qp->atomic_rd_en = attr->qp_access_flags;
791 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
792 		qp->resp_depth = attr->max_dest_rd_atomic;
793 	if (attr_mask & IB_QP_PORT)
794 		qp->port = attr->port_num;
795 	if (attr_mask & IB_QP_ALT_PATH)
796 		qp->alt_port = attr->alt_port_num;
797 
798 	if (is_sqp(dev, qp))
799 		store_attrs(to_msqp(qp), attr, attr_mask);
800 
801 	/*
802 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
803 	 * QP0 to RESET or ERROR, bring the link back down.
804 	 */
805 	if (is_qp0(dev, qp)) {
806 		if (cur_state != IB_QPS_RTR &&
807 		    new_state == IB_QPS_RTR)
808 			init_port(dev, qp->port);
809 
810 		if (cur_state != IB_QPS_RESET &&
811 		    cur_state != IB_QPS_ERR &&
812 		    (new_state == IB_QPS_RESET ||
813 		     new_state == IB_QPS_ERR))
814 			mthca_CLOSE_IB(dev, qp->port, &status);
815 	}
816 
817 	/*
818 	 * If we moved a kernel QP to RESET, clean up all old CQ
819 	 * entries and reinitialize the QP.
820 	 */
821 	if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
822 		mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
823 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
824 		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
825 			mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn, NULL);
826 
827 		mthca_wq_reset(&qp->sq);
828 		qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
829 
830 		mthca_wq_reset(&qp->rq);
831 		qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
832 
833 		if (mthca_is_memfree(dev)) {
834 			*qp->sq.db = 0;
835 			*qp->rq.db = 0;
836 		}
837 	}
838 
839 out_mailbox:
840 	mthca_free_mailbox(dev, mailbox);
841 out:
842 	return err;
843 }
844 
845 static const struct ib_qp_attr dummy_init_attr = { .port_num = 1 };
846 static const int dummy_init_attr_mask[] = {
847 	[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
848 			IB_QP_PORT			|
849 			IB_QP_QKEY),
850 	[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
851 			IB_QP_PORT			|
852 			IB_QP_ACCESS_FLAGS),
853 	[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
854 			IB_QP_PORT			|
855 			IB_QP_ACCESS_FLAGS),
856 	[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
857 			IB_QP_QKEY),
858 	[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
859 			IB_QP_QKEY),
860 };
861 
862 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
863 		    struct ib_udata *udata)
864 {
865 	struct mthca_dev *dev = to_mdev(ibqp->device);
866 	struct mthca_qp *qp = to_mqp(ibqp);
867 	enum ib_qp_state cur_state, new_state;
868 	int err = -EINVAL;
869 
870 	mutex_lock(&qp->mutex);
871 	if (attr_mask & IB_QP_CUR_STATE) {
872 		cur_state = attr->cur_qp_state;
873 	} else {
874 		spin_lock_irq(&qp->sq.lock);
875 		spin_lock(&qp->rq.lock);
876 		cur_state = qp->state;
877 		spin_unlock(&qp->rq.lock);
878 		spin_unlock_irq(&qp->sq.lock);
879 	}
880 
881 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
882 
883 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
884 		mthca_dbg(dev, "Bad QP transition (transport %d) "
885 			  "%d->%d with attr 0x%08x\n",
886 			  qp->transport, cur_state, new_state,
887 			  attr_mask);
888 		goto out;
889 	}
890 
891 	if ((attr_mask & IB_QP_PKEY_INDEX) &&
892 	     attr->pkey_index >= dev->limits.pkey_table_len) {
893 		mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
894 			  attr->pkey_index, dev->limits.pkey_table_len-1);
895 		goto out;
896 	}
897 
898 	if ((attr_mask & IB_QP_PORT) &&
899 	    (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
900 		mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
901 		goto out;
902 	}
903 
904 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
905 	    attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
906 		mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
907 			  attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
908 		goto out;
909 	}
910 
911 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
912 	    attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
913 		mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
914 			  attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
915 		goto out;
916 	}
917 
918 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
919 		err = 0;
920 		goto out;
921 	}
922 
923 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_ERR) {
924 		err = __mthca_modify_qp(ibqp, &dummy_init_attr,
925 					dummy_init_attr_mask[ibqp->qp_type],
926 					IB_QPS_RESET, IB_QPS_INIT);
927 		if (err)
928 			goto out;
929 		cur_state = IB_QPS_INIT;
930 	}
931 
932 	err = __mthca_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
933 
934 out:
935 	mutex_unlock(&qp->mutex);
936 	return err;
937 }
938 
939 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
940 {
941 	/*
942 	 * Calculate the maximum size of WQE s/g segments, excluding
943 	 * the next segment and other non-data segments.
944 	 */
945 	int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
946 
947 	switch (qp->transport) {
948 	case MLX:
949 		max_data_size -= 2 * sizeof (struct mthca_data_seg);
950 		break;
951 
952 	case UD:
953 		if (mthca_is_memfree(dev))
954 			max_data_size -= sizeof (struct mthca_arbel_ud_seg);
955 		else
956 			max_data_size -= sizeof (struct mthca_tavor_ud_seg);
957 		break;
958 
959 	default:
960 		max_data_size -= sizeof (struct mthca_raddr_seg);
961 		break;
962 	}
963 
964 	return max_data_size;
965 }
966 
967 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
968 {
969 	/* We don't support inline data for kernel QPs (yet). */
970 	return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
971 }
972 
973 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
974 				 struct mthca_pd *pd,
975 				 struct mthca_qp *qp)
976 {
977 	int max_data_size = mthca_max_data_size(dev, qp,
978 						min(dev->limits.max_desc_sz,
979 						    1 << qp->sq.wqe_shift));
980 
981 	qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
982 
983 	qp->sq.max_gs = min_t(int, dev->limits.max_sg,
984 			      max_data_size / sizeof (struct mthca_data_seg));
985 	qp->rq.max_gs = min_t(int, dev->limits.max_sg,
986 			       (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
987 				sizeof (struct mthca_next_seg)) /
988 			       sizeof (struct mthca_data_seg));
989 }
990 
991 /*
992  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
993  * rq.max_gs and sq.max_gs must all be assigned.
994  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
995  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
996  * queue)
997  */
998 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
999 			       struct mthca_pd *pd,
1000 			       struct mthca_qp *qp)
1001 {
1002 	int size;
1003 	int err = -ENOMEM;
1004 
1005 	size = sizeof (struct mthca_next_seg) +
1006 		qp->rq.max_gs * sizeof (struct mthca_data_seg);
1007 
1008 	if (size > dev->limits.max_desc_sz)
1009 		return -EINVAL;
1010 
1011 	for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
1012 	     qp->rq.wqe_shift++)
1013 		; /* nothing */
1014 
1015 	size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
1016 	switch (qp->transport) {
1017 	case MLX:
1018 		size += 2 * sizeof (struct mthca_data_seg);
1019 		break;
1020 
1021 	case UD:
1022 		size += mthca_is_memfree(dev) ?
1023 			sizeof (struct mthca_arbel_ud_seg) :
1024 			sizeof (struct mthca_tavor_ud_seg);
1025 		break;
1026 
1027 	case UC:
1028 		size += sizeof (struct mthca_raddr_seg);
1029 		break;
1030 
1031 	case RC:
1032 		size += sizeof (struct mthca_raddr_seg);
1033 		/*
1034 		 * An atomic op will require an atomic segment, a
1035 		 * remote address segment and one scatter entry.
1036 		 */
1037 		size = max_t(int, size,
1038 			     sizeof (struct mthca_atomic_seg) +
1039 			     sizeof (struct mthca_raddr_seg) +
1040 			     sizeof (struct mthca_data_seg));
1041 		break;
1042 
1043 	default:
1044 		break;
1045 	}
1046 
1047 	/* Make sure that we have enough space for a bind request */
1048 	size = max_t(int, size, sizeof (struct mthca_bind_seg));
1049 
1050 	size += sizeof (struct mthca_next_seg);
1051 
1052 	if (size > dev->limits.max_desc_sz)
1053 		return -EINVAL;
1054 
1055 	for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
1056 	     qp->sq.wqe_shift++)
1057 		; /* nothing */
1058 
1059 	qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
1060 				    1 << qp->sq.wqe_shift);
1061 
1062 	/*
1063 	 * If this is a userspace QP, we don't actually have to
1064 	 * allocate anything.  All we need is to calculate the WQE
1065 	 * sizes and the send_wqe_offset, so we're done now.
1066 	 */
1067 	if (pd->ibpd.uobject)
1068 		return 0;
1069 
1070 	size = PAGE_ALIGN(qp->send_wqe_offset +
1071 			  (qp->sq.max << qp->sq.wqe_shift));
1072 
1073 	qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1074 			   GFP_KERNEL);
1075 	if (!qp->wrid)
1076 		goto err_out;
1077 
1078 	err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1079 			      &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1080 	if (err)
1081 		goto err_out;
1082 
1083 	return 0;
1084 
1085 err_out:
1086 	kfree(qp->wrid);
1087 	return err;
1088 }
1089 
1090 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1091 			       struct mthca_qp *qp)
1092 {
1093 	mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1094 				       (qp->sq.max << qp->sq.wqe_shift)),
1095 		       &qp->queue, qp->is_direct, &qp->mr);
1096 	kfree(qp->wrid);
1097 }
1098 
1099 static int mthca_map_memfree(struct mthca_dev *dev,
1100 			     struct mthca_qp *qp)
1101 {
1102 	int ret;
1103 
1104 	if (mthca_is_memfree(dev)) {
1105 		ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1106 		if (ret)
1107 			return ret;
1108 
1109 		ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1110 		if (ret)
1111 			goto err_qpc;
1112 
1113 		ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1114 				      qp->qpn << dev->qp_table.rdb_shift);
1115 		if (ret)
1116 			goto err_eqpc;
1117 
1118 	}
1119 
1120 	return 0;
1121 
1122 err_eqpc:
1123 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1124 
1125 err_qpc:
1126 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1127 
1128 	return ret;
1129 }
1130 
1131 static void mthca_unmap_memfree(struct mthca_dev *dev,
1132 				struct mthca_qp *qp)
1133 {
1134 	mthca_table_put(dev, dev->qp_table.rdb_table,
1135 			qp->qpn << dev->qp_table.rdb_shift);
1136 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1137 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1138 }
1139 
1140 static int mthca_alloc_memfree(struct mthca_dev *dev,
1141 			       struct mthca_qp *qp)
1142 {
1143 	if (mthca_is_memfree(dev)) {
1144 		qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1145 						 qp->qpn, &qp->rq.db);
1146 		if (qp->rq.db_index < 0)
1147 			return -ENOMEM;
1148 
1149 		qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1150 						 qp->qpn, &qp->sq.db);
1151 		if (qp->sq.db_index < 0) {
1152 			mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1153 			return -ENOMEM;
1154 		}
1155 	}
1156 
1157 	return 0;
1158 }
1159 
1160 static void mthca_free_memfree(struct mthca_dev *dev,
1161 			       struct mthca_qp *qp)
1162 {
1163 	if (mthca_is_memfree(dev)) {
1164 		mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1165 		mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1166 	}
1167 }
1168 
1169 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1170 				 struct mthca_pd *pd,
1171 				 struct mthca_cq *send_cq,
1172 				 struct mthca_cq *recv_cq,
1173 				 enum ib_sig_type send_policy,
1174 				 struct mthca_qp *qp)
1175 {
1176 	int ret;
1177 	int i;
1178 	struct mthca_next_seg *next;
1179 
1180 	qp->refcount = 1;
1181 	init_waitqueue_head(&qp->wait);
1182 	mutex_init(&qp->mutex);
1183 	qp->state    	 = IB_QPS_RESET;
1184 	qp->atomic_rd_en = 0;
1185 	qp->resp_depth   = 0;
1186 	qp->sq_policy    = send_policy;
1187 	mthca_wq_reset(&qp->sq);
1188 	mthca_wq_reset(&qp->rq);
1189 
1190 	spin_lock_init(&qp->sq.lock);
1191 	spin_lock_init(&qp->rq.lock);
1192 
1193 	ret = mthca_map_memfree(dev, qp);
1194 	if (ret)
1195 		return ret;
1196 
1197 	ret = mthca_alloc_wqe_buf(dev, pd, qp);
1198 	if (ret) {
1199 		mthca_unmap_memfree(dev, qp);
1200 		return ret;
1201 	}
1202 
1203 	mthca_adjust_qp_caps(dev, pd, qp);
1204 
1205 	/*
1206 	 * If this is a userspace QP, we're done now.  The doorbells
1207 	 * will be allocated and buffers will be initialized in
1208 	 * userspace.
1209 	 */
1210 	if (pd->ibpd.uobject)
1211 		return 0;
1212 
1213 	ret = mthca_alloc_memfree(dev, qp);
1214 	if (ret) {
1215 		mthca_free_wqe_buf(dev, qp);
1216 		mthca_unmap_memfree(dev, qp);
1217 		return ret;
1218 	}
1219 
1220 	if (mthca_is_memfree(dev)) {
1221 		struct mthca_data_seg *scatter;
1222 		int size = (sizeof (struct mthca_next_seg) +
1223 			    qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1224 
1225 		for (i = 0; i < qp->rq.max; ++i) {
1226 			next = get_recv_wqe(qp, i);
1227 			next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1228 						   qp->rq.wqe_shift);
1229 			next->ee_nds = cpu_to_be32(size);
1230 
1231 			for (scatter = (void *) (next + 1);
1232 			     (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1233 			     ++scatter)
1234 				scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1235 		}
1236 
1237 		for (i = 0; i < qp->sq.max; ++i) {
1238 			next = get_send_wqe(qp, i);
1239 			next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1240 						    qp->sq.wqe_shift) +
1241 						   qp->send_wqe_offset);
1242 		}
1243 	} else {
1244 		for (i = 0; i < qp->rq.max; ++i) {
1245 			next = get_recv_wqe(qp, i);
1246 			next->nda_op = htonl((((i + 1) % qp->rq.max) <<
1247 					      qp->rq.wqe_shift) | 1);
1248 		}
1249 
1250 	}
1251 
1252 	qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1253 	qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1254 
1255 	return 0;
1256 }
1257 
1258 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1259 			     struct mthca_pd *pd, struct mthca_qp *qp)
1260 {
1261 	int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1262 
1263 	/* Sanity check QP size before proceeding */
1264 	if (cap->max_send_wr  	 > dev->limits.max_wqes ||
1265 	    cap->max_recv_wr  	 > dev->limits.max_wqes ||
1266 	    cap->max_send_sge 	 > dev->limits.max_sg   ||
1267 	    cap->max_recv_sge 	 > dev->limits.max_sg   ||
1268 	    cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1269 		return -EINVAL;
1270 
1271 	/*
1272 	 * For MLX transport we need 2 extra S/G entries:
1273 	 * one for the header and one for the checksum at the end
1274 	 */
1275 	if (qp->transport == MLX && cap->max_recv_sge + 2 > dev->limits.max_sg)
1276 		return -EINVAL;
1277 
1278 	if (mthca_is_memfree(dev)) {
1279 		qp->rq.max = cap->max_recv_wr ?
1280 			roundup_pow_of_two(cap->max_recv_wr) : 0;
1281 		qp->sq.max = cap->max_send_wr ?
1282 			roundup_pow_of_two(cap->max_send_wr) : 0;
1283 	} else {
1284 		qp->rq.max = cap->max_recv_wr;
1285 		qp->sq.max = cap->max_send_wr;
1286 	}
1287 
1288 	qp->rq.max_gs = cap->max_recv_sge;
1289 	qp->sq.max_gs = max_t(int, cap->max_send_sge,
1290 			      ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1291 				    MTHCA_INLINE_CHUNK_SIZE) /
1292 			      sizeof (struct mthca_data_seg));
1293 
1294 	return 0;
1295 }
1296 
1297 int mthca_alloc_qp(struct mthca_dev *dev,
1298 		   struct mthca_pd *pd,
1299 		   struct mthca_cq *send_cq,
1300 		   struct mthca_cq *recv_cq,
1301 		   enum ib_qp_type type,
1302 		   enum ib_sig_type send_policy,
1303 		   struct ib_qp_cap *cap,
1304 		   struct mthca_qp *qp)
1305 {
1306 	int err;
1307 
1308 	switch (type) {
1309 	case IB_QPT_RC: qp->transport = RC; break;
1310 	case IB_QPT_UC: qp->transport = UC; break;
1311 	case IB_QPT_UD: qp->transport = UD; break;
1312 	default: return -EINVAL;
1313 	}
1314 
1315 	err = mthca_set_qp_size(dev, cap, pd, qp);
1316 	if (err)
1317 		return err;
1318 
1319 	qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1320 	if (qp->qpn == -1)
1321 		return -ENOMEM;
1322 
1323 	/* initialize port to zero for error-catching. */
1324 	qp->port = 0;
1325 
1326 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1327 				    send_policy, qp);
1328 	if (err) {
1329 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1330 		return err;
1331 	}
1332 
1333 	spin_lock_irq(&dev->qp_table.lock);
1334 	mthca_array_set(&dev->qp_table.qp,
1335 			qp->qpn & (dev->limits.num_qps - 1), qp);
1336 	spin_unlock_irq(&dev->qp_table.lock);
1337 
1338 	return 0;
1339 }
1340 
1341 static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1342 {
1343 	if (send_cq == recv_cq)
1344 		spin_lock_irq(&send_cq->lock);
1345 	else if (send_cq->cqn < recv_cq->cqn) {
1346 		spin_lock_irq(&send_cq->lock);
1347 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1348 	} else {
1349 		spin_lock_irq(&recv_cq->lock);
1350 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1351 	}
1352 }
1353 
1354 static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1355 {
1356 	if (send_cq == recv_cq)
1357 		spin_unlock_irq(&send_cq->lock);
1358 	else if (send_cq->cqn < recv_cq->cqn) {
1359 		spin_unlock(&recv_cq->lock);
1360 		spin_unlock_irq(&send_cq->lock);
1361 	} else {
1362 		spin_unlock(&send_cq->lock);
1363 		spin_unlock_irq(&recv_cq->lock);
1364 	}
1365 }
1366 
1367 int mthca_alloc_sqp(struct mthca_dev *dev,
1368 		    struct mthca_pd *pd,
1369 		    struct mthca_cq *send_cq,
1370 		    struct mthca_cq *recv_cq,
1371 		    enum ib_sig_type send_policy,
1372 		    struct ib_qp_cap *cap,
1373 		    int qpn,
1374 		    int port,
1375 		    struct mthca_sqp *sqp)
1376 {
1377 	u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1378 	int err;
1379 
1380 	sqp->qp.transport = MLX;
1381 	err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1382 	if (err)
1383 		return err;
1384 
1385 	sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1386 	sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1387 					     &sqp->header_dma, GFP_KERNEL);
1388 	if (!sqp->header_buf)
1389 		return -ENOMEM;
1390 
1391 	spin_lock_irq(&dev->qp_table.lock);
1392 	if (mthca_array_get(&dev->qp_table.qp, mqpn))
1393 		err = -EBUSY;
1394 	else
1395 		mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1396 	spin_unlock_irq(&dev->qp_table.lock);
1397 
1398 	if (err)
1399 		goto err_out;
1400 
1401 	sqp->qp.port      = port;
1402 	sqp->qp.qpn       = mqpn;
1403 	sqp->qp.transport = MLX;
1404 
1405 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1406 				    send_policy, &sqp->qp);
1407 	if (err)
1408 		goto err_out_free;
1409 
1410 	atomic_inc(&pd->sqp_count);
1411 
1412 	return 0;
1413 
1414  err_out_free:
1415 	/*
1416 	 * Lock CQs here, so that CQ polling code can do QP lookup
1417 	 * without taking a lock.
1418 	 */
1419 	mthca_lock_cqs(send_cq, recv_cq);
1420 
1421 	spin_lock(&dev->qp_table.lock);
1422 	mthca_array_clear(&dev->qp_table.qp, mqpn);
1423 	spin_unlock(&dev->qp_table.lock);
1424 
1425 	mthca_unlock_cqs(send_cq, recv_cq);
1426 
1427  err_out:
1428 	dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1429 			  sqp->header_buf, sqp->header_dma);
1430 
1431 	return err;
1432 }
1433 
1434 static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp)
1435 {
1436 	int c;
1437 
1438 	spin_lock_irq(&dev->qp_table.lock);
1439 	c = qp->refcount;
1440 	spin_unlock_irq(&dev->qp_table.lock);
1441 
1442 	return c;
1443 }
1444 
1445 void mthca_free_qp(struct mthca_dev *dev,
1446 		   struct mthca_qp *qp)
1447 {
1448 	u8 status;
1449 	struct mthca_cq *send_cq;
1450 	struct mthca_cq *recv_cq;
1451 
1452 	send_cq = to_mcq(qp->ibqp.send_cq);
1453 	recv_cq = to_mcq(qp->ibqp.recv_cq);
1454 
1455 	/*
1456 	 * Lock CQs here, so that CQ polling code can do QP lookup
1457 	 * without taking a lock.
1458 	 */
1459 	mthca_lock_cqs(send_cq, recv_cq);
1460 
1461 	spin_lock(&dev->qp_table.lock);
1462 	mthca_array_clear(&dev->qp_table.qp,
1463 			  qp->qpn & (dev->limits.num_qps - 1));
1464 	--qp->refcount;
1465 	spin_unlock(&dev->qp_table.lock);
1466 
1467 	mthca_unlock_cqs(send_cq, recv_cq);
1468 
1469 	wait_event(qp->wait, !get_qp_refcount(dev, qp));
1470 
1471 	if (qp->state != IB_QPS_RESET)
1472 		mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1473 				NULL, 0, &status);
1474 
1475 	/*
1476 	 * If this is a userspace QP, the buffers, MR, CQs and so on
1477 	 * will be cleaned up in userspace, so all we have to do is
1478 	 * unref the mem-free tables and free the QPN in our table.
1479 	 */
1480 	if (!qp->ibqp.uobject) {
1481 		mthca_cq_clean(dev, recv_cq, qp->qpn,
1482 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1483 		if (send_cq != recv_cq)
1484 			mthca_cq_clean(dev, send_cq, qp->qpn, NULL);
1485 
1486 		mthca_free_memfree(dev, qp);
1487 		mthca_free_wqe_buf(dev, qp);
1488 	}
1489 
1490 	mthca_unmap_memfree(dev, qp);
1491 
1492 	if (is_sqp(dev, qp)) {
1493 		atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1494 		dma_free_coherent(&dev->pdev->dev,
1495 				  to_msqp(qp)->header_buf_size,
1496 				  to_msqp(qp)->header_buf,
1497 				  to_msqp(qp)->header_dma);
1498 	} else
1499 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1500 }
1501 
1502 /* Create UD header for an MLX send and build a data segment for it */
1503 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1504 			    int ind, struct ib_send_wr *wr,
1505 			    struct mthca_mlx_seg *mlx,
1506 			    struct mthca_data_seg *data)
1507 {
1508 	int header_size;
1509 	int err;
1510 	u16 pkey;
1511 
1512 	ib_ud_header_init(256, /* assume a MAD */
1513 			  mthca_ah_grh_present(to_mah(wr->wr.ud.ah)),
1514 			  &sqp->ud_header);
1515 
1516 	err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1517 	if (err)
1518 		return err;
1519 	mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1520 	mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1521 				  (sqp->ud_header.lrh.destination_lid ==
1522 				   IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1523 				  (sqp->ud_header.lrh.service_level << 8));
1524 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
1525 	mlx->vcrc = 0;
1526 
1527 	switch (wr->opcode) {
1528 	case IB_WR_SEND:
1529 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1530 		sqp->ud_header.immediate_present = 0;
1531 		break;
1532 	case IB_WR_SEND_WITH_IMM:
1533 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1534 		sqp->ud_header.immediate_present = 1;
1535 		sqp->ud_header.immediate_data = wr->imm_data;
1536 		break;
1537 	default:
1538 		return -EINVAL;
1539 	}
1540 
1541 	sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1542 	if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1543 		sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1544 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1545 	if (!sqp->qp.ibqp.qp_num)
1546 		ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1547 				   sqp->pkey_index, &pkey);
1548 	else
1549 		ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1550 				   wr->wr.ud.pkey_index, &pkey);
1551 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1552 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1553 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1554 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1555 					       sqp->qkey : wr->wr.ud.remote_qkey);
1556 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1557 
1558 	header_size = ib_ud_header_pack(&sqp->ud_header,
1559 					sqp->header_buf +
1560 					ind * MTHCA_UD_HEADER_SIZE);
1561 
1562 	data->byte_count = cpu_to_be32(header_size);
1563 	data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1564 	data->addr       = cpu_to_be64(sqp->header_dma +
1565 				       ind * MTHCA_UD_HEADER_SIZE);
1566 
1567 	return 0;
1568 }
1569 
1570 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1571 				    struct ib_cq *ib_cq)
1572 {
1573 	unsigned cur;
1574 	struct mthca_cq *cq;
1575 
1576 	cur = wq->head - wq->tail;
1577 	if (likely(cur + nreq < wq->max))
1578 		return 0;
1579 
1580 	cq = to_mcq(ib_cq);
1581 	spin_lock(&cq->lock);
1582 	cur = wq->head - wq->tail;
1583 	spin_unlock(&cq->lock);
1584 
1585 	return cur + nreq >= wq->max;
1586 }
1587 
1588 static __always_inline void set_raddr_seg(struct mthca_raddr_seg *rseg,
1589 					  u64 remote_addr, u32 rkey)
1590 {
1591 	rseg->raddr    = cpu_to_be64(remote_addr);
1592 	rseg->rkey     = cpu_to_be32(rkey);
1593 	rseg->reserved = 0;
1594 }
1595 
1596 static __always_inline void set_atomic_seg(struct mthca_atomic_seg *aseg,
1597 					   struct ib_send_wr *wr)
1598 {
1599 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1600 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
1601 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
1602 	} else {
1603 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
1604 		aseg->compare  = 0;
1605 	}
1606 
1607 }
1608 
1609 static void set_tavor_ud_seg(struct mthca_tavor_ud_seg *useg,
1610 			     struct ib_send_wr *wr)
1611 {
1612 	useg->lkey    = cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1613 	useg->av_addr =	cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1614 	useg->dqpn    =	cpu_to_be32(wr->wr.ud.remote_qpn);
1615 	useg->qkey    =	cpu_to_be32(wr->wr.ud.remote_qkey);
1616 
1617 }
1618 
1619 static void set_arbel_ud_seg(struct mthca_arbel_ud_seg *useg,
1620 			     struct ib_send_wr *wr)
1621 {
1622 	memcpy(useg->av, to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1623 	useg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1624 	useg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1625 }
1626 
1627 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1628 			  struct ib_send_wr **bad_wr)
1629 {
1630 	struct mthca_dev *dev = to_mdev(ibqp->device);
1631 	struct mthca_qp *qp = to_mqp(ibqp);
1632 	void *wqe;
1633 	void *prev_wqe;
1634 	unsigned long flags;
1635 	int err = 0;
1636 	int nreq;
1637 	int i;
1638 	int size;
1639 	/*
1640 	 * f0 and size0 are only used if nreq != 0, and they will
1641 	 * always be initialized the first time through the main loop
1642 	 * before nreq is incremented.  So nreq cannot become non-zero
1643 	 * without initializing f0 and size0, and they are in fact
1644 	 * never used uninitialized.
1645 	 */
1646 	int uninitialized_var(size0);
1647 	u32 uninitialized_var(f0);
1648 	int ind;
1649 	u8 op0 = 0;
1650 
1651 	spin_lock_irqsave(&qp->sq.lock, flags);
1652 
1653 	/* XXX check that state is OK to post send */
1654 
1655 	ind = qp->sq.next_ind;
1656 
1657 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1658 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1659 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1660 					" %d max, %d nreq)\n", qp->qpn,
1661 					qp->sq.head, qp->sq.tail,
1662 					qp->sq.max, nreq);
1663 			err = -ENOMEM;
1664 			*bad_wr = wr;
1665 			goto out;
1666 		}
1667 
1668 		wqe = get_send_wqe(qp, ind);
1669 		prev_wqe = qp->sq.last;
1670 		qp->sq.last = wqe;
1671 
1672 		((struct mthca_next_seg *) wqe)->nda_op = 0;
1673 		((struct mthca_next_seg *) wqe)->ee_nds = 0;
1674 		((struct mthca_next_seg *) wqe)->flags =
1675 			((wr->send_flags & IB_SEND_SIGNALED) ?
1676 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1677 			((wr->send_flags & IB_SEND_SOLICITED) ?
1678 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1679 			cpu_to_be32(1);
1680 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1681 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1682 			((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1683 
1684 		wqe += sizeof (struct mthca_next_seg);
1685 		size = sizeof (struct mthca_next_seg) / 16;
1686 
1687 		switch (qp->transport) {
1688 		case RC:
1689 			switch (wr->opcode) {
1690 			case IB_WR_ATOMIC_CMP_AND_SWP:
1691 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1692 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
1693 					      wr->wr.atomic.rkey);
1694 				wqe += sizeof (struct mthca_raddr_seg);
1695 
1696 				set_atomic_seg(wqe, wr);
1697 				wqe += sizeof (struct mthca_atomic_seg);
1698 				size += (sizeof (struct mthca_raddr_seg) +
1699 					 sizeof (struct mthca_atomic_seg)) / 16;
1700 				break;
1701 
1702 			case IB_WR_RDMA_WRITE:
1703 			case IB_WR_RDMA_WRITE_WITH_IMM:
1704 			case IB_WR_RDMA_READ:
1705 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1706 					      wr->wr.rdma.rkey);
1707 				wqe  += sizeof (struct mthca_raddr_seg);
1708 				size += sizeof (struct mthca_raddr_seg) / 16;
1709 				break;
1710 
1711 			default:
1712 				/* No extra segments required for sends */
1713 				break;
1714 			}
1715 
1716 			break;
1717 
1718 		case UC:
1719 			switch (wr->opcode) {
1720 			case IB_WR_RDMA_WRITE:
1721 			case IB_WR_RDMA_WRITE_WITH_IMM:
1722 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1723 					      wr->wr.rdma.rkey);
1724 				wqe  += sizeof (struct mthca_raddr_seg);
1725 				size += sizeof (struct mthca_raddr_seg) / 16;
1726 				break;
1727 
1728 			default:
1729 				/* No extra segments required for sends */
1730 				break;
1731 			}
1732 
1733 			break;
1734 
1735 		case UD:
1736 			set_tavor_ud_seg(wqe, wr);
1737 			wqe  += sizeof (struct mthca_tavor_ud_seg);
1738 			size += sizeof (struct mthca_tavor_ud_seg) / 16;
1739 			break;
1740 
1741 		case MLX:
1742 			err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1743 					       wqe - sizeof (struct mthca_next_seg),
1744 					       wqe);
1745 			if (err) {
1746 				*bad_wr = wr;
1747 				goto out;
1748 			}
1749 			wqe += sizeof (struct mthca_data_seg);
1750 			size += sizeof (struct mthca_data_seg) / 16;
1751 			break;
1752 		}
1753 
1754 		if (wr->num_sge > qp->sq.max_gs) {
1755 			mthca_err(dev, "too many gathers\n");
1756 			err = -EINVAL;
1757 			*bad_wr = wr;
1758 			goto out;
1759 		}
1760 
1761 		for (i = 0; i < wr->num_sge; ++i) {
1762 			mthca_set_data_seg(wqe, wr->sg_list + i);
1763 			wqe  += sizeof (struct mthca_data_seg);
1764 			size += sizeof (struct mthca_data_seg) / 16;
1765 		}
1766 
1767 		/* Add one more inline data segment for ICRC */
1768 		if (qp->transport == MLX) {
1769 			((struct mthca_data_seg *) wqe)->byte_count =
1770 				cpu_to_be32((1 << 31) | 4);
1771 			((u32 *) wqe)[1] = 0;
1772 			wqe += sizeof (struct mthca_data_seg);
1773 			size += sizeof (struct mthca_data_seg) / 16;
1774 		}
1775 
1776 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
1777 
1778 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1779 			mthca_err(dev, "opcode invalid\n");
1780 			err = -EINVAL;
1781 			*bad_wr = wr;
1782 			goto out;
1783 		}
1784 
1785 		((struct mthca_next_seg *) prev_wqe)->nda_op =
1786 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
1787 				     qp->send_wqe_offset) |
1788 				    mthca_opcode[wr->opcode]);
1789 		wmb();
1790 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1791 			cpu_to_be32((nreq ? 0 : MTHCA_NEXT_DBD) | size |
1792 				    ((wr->send_flags & IB_SEND_FENCE) ?
1793 				    MTHCA_NEXT_FENCE : 0));
1794 
1795 		if (!nreq) {
1796 			size0 = size;
1797 			op0   = mthca_opcode[wr->opcode];
1798 			f0    = wr->send_flags & IB_SEND_FENCE ?
1799 				MTHCA_SEND_DOORBELL_FENCE : 0;
1800 		}
1801 
1802 		++ind;
1803 		if (unlikely(ind >= qp->sq.max))
1804 			ind -= qp->sq.max;
1805 	}
1806 
1807 out:
1808 	if (likely(nreq)) {
1809 		wmb();
1810 
1811 		mthca_write64(((qp->sq.next_ind << qp->sq.wqe_shift) +
1812 			       qp->send_wqe_offset) | f0 | op0,
1813 			      (qp->qpn << 8) | size0,
1814 			      dev->kar + MTHCA_SEND_DOORBELL,
1815 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1816 		/*
1817 		 * Make sure doorbells don't leak out of SQ spinlock
1818 		 * and reach the HCA out of order:
1819 		 */
1820 		mmiowb();
1821 	}
1822 
1823 	qp->sq.next_ind = ind;
1824 	qp->sq.head    += nreq;
1825 
1826 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1827 	return err;
1828 }
1829 
1830 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1831 			     struct ib_recv_wr **bad_wr)
1832 {
1833 	struct mthca_dev *dev = to_mdev(ibqp->device);
1834 	struct mthca_qp *qp = to_mqp(ibqp);
1835 	unsigned long flags;
1836 	int err = 0;
1837 	int nreq;
1838 	int i;
1839 	int size;
1840 	/*
1841 	 * size0 is only used if nreq != 0, and it will always be
1842 	 * initialized the first time through the main loop before
1843 	 * nreq is incremented.  So nreq cannot become non-zero
1844 	 * without initializing size0, and it is in fact never used
1845 	 * uninitialized.
1846 	 */
1847 	int uninitialized_var(size0);
1848 	int ind;
1849 	void *wqe;
1850 	void *prev_wqe;
1851 
1852 	spin_lock_irqsave(&qp->rq.lock, flags);
1853 
1854 	/* XXX check that state is OK to post receive */
1855 
1856 	ind = qp->rq.next_ind;
1857 
1858 	for (nreq = 0; wr; wr = wr->next) {
1859 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1860 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1861 					" %d max, %d nreq)\n", qp->qpn,
1862 					qp->rq.head, qp->rq.tail,
1863 					qp->rq.max, nreq);
1864 			err = -ENOMEM;
1865 			*bad_wr = wr;
1866 			goto out;
1867 		}
1868 
1869 		wqe = get_recv_wqe(qp, ind);
1870 		prev_wqe = qp->rq.last;
1871 		qp->rq.last = wqe;
1872 
1873 		((struct mthca_next_seg *) wqe)->ee_nds =
1874 			cpu_to_be32(MTHCA_NEXT_DBD);
1875 		((struct mthca_next_seg *) wqe)->flags = 0;
1876 
1877 		wqe += sizeof (struct mthca_next_seg);
1878 		size = sizeof (struct mthca_next_seg) / 16;
1879 
1880 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1881 			err = -EINVAL;
1882 			*bad_wr = wr;
1883 			goto out;
1884 		}
1885 
1886 		for (i = 0; i < wr->num_sge; ++i) {
1887 			mthca_set_data_seg(wqe, wr->sg_list + i);
1888 			wqe  += sizeof (struct mthca_data_seg);
1889 			size += sizeof (struct mthca_data_seg) / 16;
1890 		}
1891 
1892 		qp->wrid[ind] = wr->wr_id;
1893 
1894 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1895 			cpu_to_be32(MTHCA_NEXT_DBD | size);
1896 
1897 		if (!nreq)
1898 			size0 = size;
1899 
1900 		++ind;
1901 		if (unlikely(ind >= qp->rq.max))
1902 			ind -= qp->rq.max;
1903 
1904 		++nreq;
1905 		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1906 			nreq = 0;
1907 
1908 			wmb();
1909 
1910 			mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1911 				      qp->qpn << 8, dev->kar + MTHCA_RECEIVE_DOORBELL,
1912 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1913 
1914 			qp->rq.next_ind = ind;
1915 			qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1916 		}
1917 	}
1918 
1919 out:
1920 	if (likely(nreq)) {
1921 		wmb();
1922 
1923 		mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1924 			      qp->qpn << 8 | nreq, dev->kar + MTHCA_RECEIVE_DOORBELL,
1925 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1926 	}
1927 
1928 	qp->rq.next_ind = ind;
1929 	qp->rq.head    += nreq;
1930 
1931 	/*
1932 	 * Make sure doorbells don't leak out of RQ spinlock and reach
1933 	 * the HCA out of order:
1934 	 */
1935 	mmiowb();
1936 
1937 	spin_unlock_irqrestore(&qp->rq.lock, flags);
1938 	return err;
1939 }
1940 
1941 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1942 			  struct ib_send_wr **bad_wr)
1943 {
1944 	struct mthca_dev *dev = to_mdev(ibqp->device);
1945 	struct mthca_qp *qp = to_mqp(ibqp);
1946 	u32 dbhi;
1947 	void *wqe;
1948 	void *prev_wqe;
1949 	unsigned long flags;
1950 	int err = 0;
1951 	int nreq;
1952 	int i;
1953 	int size;
1954 	/*
1955 	 * f0 and size0 are only used if nreq != 0, and they will
1956 	 * always be initialized the first time through the main loop
1957 	 * before nreq is incremented.  So nreq cannot become non-zero
1958 	 * without initializing f0 and size0, and they are in fact
1959 	 * never used uninitialized.
1960 	 */
1961 	int uninitialized_var(size0);
1962 	u32 uninitialized_var(f0);
1963 	int ind;
1964 	u8 op0 = 0;
1965 
1966 	spin_lock_irqsave(&qp->sq.lock, flags);
1967 
1968 	/* XXX check that state is OK to post send */
1969 
1970 	ind = qp->sq.head & (qp->sq.max - 1);
1971 
1972 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1973 		if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1974 			nreq = 0;
1975 
1976 			dbhi = (MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1977 				((qp->sq.head & 0xffff) << 8) | f0 | op0;
1978 
1979 			qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1980 
1981 			/*
1982 			 * Make sure that descriptors are written before
1983 			 * doorbell record.
1984 			 */
1985 			wmb();
1986 			*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1987 
1988 			/*
1989 			 * Make sure doorbell record is written before we
1990 			 * write MMIO send doorbell.
1991 			 */
1992 			wmb();
1993 
1994 			mthca_write64(dbhi, (qp->qpn << 8) | size0,
1995 				      dev->kar + MTHCA_SEND_DOORBELL,
1996 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1997 		}
1998 
1999 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2000 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
2001 					" %d max, %d nreq)\n", qp->qpn,
2002 					qp->sq.head, qp->sq.tail,
2003 					qp->sq.max, nreq);
2004 			err = -ENOMEM;
2005 			*bad_wr = wr;
2006 			goto out;
2007 		}
2008 
2009 		wqe = get_send_wqe(qp, ind);
2010 		prev_wqe = qp->sq.last;
2011 		qp->sq.last = wqe;
2012 
2013 		((struct mthca_next_seg *) wqe)->flags =
2014 			((wr->send_flags & IB_SEND_SIGNALED) ?
2015 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
2016 			((wr->send_flags & IB_SEND_SOLICITED) ?
2017 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
2018 			cpu_to_be32(1);
2019 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
2020 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
2021 			((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
2022 
2023 		wqe += sizeof (struct mthca_next_seg);
2024 		size = sizeof (struct mthca_next_seg) / 16;
2025 
2026 		switch (qp->transport) {
2027 		case RC:
2028 			switch (wr->opcode) {
2029 			case IB_WR_ATOMIC_CMP_AND_SWP:
2030 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2031 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2032 					      wr->wr.atomic.rkey);
2033 				wqe += sizeof (struct mthca_raddr_seg);
2034 
2035 				set_atomic_seg(wqe, wr);
2036 				wqe  += sizeof (struct mthca_atomic_seg);
2037 				size += (sizeof (struct mthca_raddr_seg) +
2038 					 sizeof (struct mthca_atomic_seg)) / 16;
2039 				break;
2040 
2041 			case IB_WR_RDMA_READ:
2042 			case IB_WR_RDMA_WRITE:
2043 			case IB_WR_RDMA_WRITE_WITH_IMM:
2044 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2045 					      wr->wr.rdma.rkey);
2046 				wqe  += sizeof (struct mthca_raddr_seg);
2047 				size += sizeof (struct mthca_raddr_seg) / 16;
2048 				break;
2049 
2050 			default:
2051 				/* No extra segments required for sends */
2052 				break;
2053 			}
2054 
2055 			break;
2056 
2057 		case UC:
2058 			switch (wr->opcode) {
2059 			case IB_WR_RDMA_WRITE:
2060 			case IB_WR_RDMA_WRITE_WITH_IMM:
2061 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2062 					      wr->wr.rdma.rkey);
2063 				wqe  += sizeof (struct mthca_raddr_seg);
2064 				size += sizeof (struct mthca_raddr_seg) / 16;
2065 				break;
2066 
2067 			default:
2068 				/* No extra segments required for sends */
2069 				break;
2070 			}
2071 
2072 			break;
2073 
2074 		case UD:
2075 			set_arbel_ud_seg(wqe, wr);
2076 			wqe  += sizeof (struct mthca_arbel_ud_seg);
2077 			size += sizeof (struct mthca_arbel_ud_seg) / 16;
2078 			break;
2079 
2080 		case MLX:
2081 			err = build_mlx_header(dev, to_msqp(qp), ind, wr,
2082 					       wqe - sizeof (struct mthca_next_seg),
2083 					       wqe);
2084 			if (err) {
2085 				*bad_wr = wr;
2086 				goto out;
2087 			}
2088 			wqe += sizeof (struct mthca_data_seg);
2089 			size += sizeof (struct mthca_data_seg) / 16;
2090 			break;
2091 		}
2092 
2093 		if (wr->num_sge > qp->sq.max_gs) {
2094 			mthca_err(dev, "too many gathers\n");
2095 			err = -EINVAL;
2096 			*bad_wr = wr;
2097 			goto out;
2098 		}
2099 
2100 		for (i = 0; i < wr->num_sge; ++i) {
2101 			mthca_set_data_seg(wqe, wr->sg_list + i);
2102 			wqe  += sizeof (struct mthca_data_seg);
2103 			size += sizeof (struct mthca_data_seg) / 16;
2104 		}
2105 
2106 		/* Add one more inline data segment for ICRC */
2107 		if (qp->transport == MLX) {
2108 			((struct mthca_data_seg *) wqe)->byte_count =
2109 				cpu_to_be32((1 << 31) | 4);
2110 			((u32 *) wqe)[1] = 0;
2111 			wqe += sizeof (struct mthca_data_seg);
2112 			size += sizeof (struct mthca_data_seg) / 16;
2113 		}
2114 
2115 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
2116 
2117 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2118 			mthca_err(dev, "opcode invalid\n");
2119 			err = -EINVAL;
2120 			*bad_wr = wr;
2121 			goto out;
2122 		}
2123 
2124 		((struct mthca_next_seg *) prev_wqe)->nda_op =
2125 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
2126 				     qp->send_wqe_offset) |
2127 				    mthca_opcode[wr->opcode]);
2128 		wmb();
2129 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
2130 			cpu_to_be32(MTHCA_NEXT_DBD | size |
2131 				    ((wr->send_flags & IB_SEND_FENCE) ?
2132 				     MTHCA_NEXT_FENCE : 0));
2133 
2134 		if (!nreq) {
2135 			size0 = size;
2136 			op0   = mthca_opcode[wr->opcode];
2137 			f0    = wr->send_flags & IB_SEND_FENCE ?
2138 				MTHCA_SEND_DOORBELL_FENCE : 0;
2139 		}
2140 
2141 		++ind;
2142 		if (unlikely(ind >= qp->sq.max))
2143 			ind -= qp->sq.max;
2144 	}
2145 
2146 out:
2147 	if (likely(nreq)) {
2148 		dbhi = (nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0;
2149 
2150 		qp->sq.head += nreq;
2151 
2152 		/*
2153 		 * Make sure that descriptors are written before
2154 		 * doorbell record.
2155 		 */
2156 		wmb();
2157 		*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2158 
2159 		/*
2160 		 * Make sure doorbell record is written before we
2161 		 * write MMIO send doorbell.
2162 		 */
2163 		wmb();
2164 
2165 		mthca_write64(dbhi, (qp->qpn << 8) | size0, dev->kar + MTHCA_SEND_DOORBELL,
2166 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2167 	}
2168 
2169 	/*
2170 	 * Make sure doorbells don't leak out of SQ spinlock and reach
2171 	 * the HCA out of order:
2172 	 */
2173 	mmiowb();
2174 
2175 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2176 	return err;
2177 }
2178 
2179 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2180 			     struct ib_recv_wr **bad_wr)
2181 {
2182 	struct mthca_dev *dev = to_mdev(ibqp->device);
2183 	struct mthca_qp *qp = to_mqp(ibqp);
2184 	unsigned long flags;
2185 	int err = 0;
2186 	int nreq;
2187 	int ind;
2188 	int i;
2189 	void *wqe;
2190 
2191 	spin_lock_irqsave(&qp->rq.lock, flags);
2192 
2193 	/* XXX check that state is OK to post receive */
2194 
2195 	ind = qp->rq.head & (qp->rq.max - 1);
2196 
2197 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2198 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2199 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2200 					" %d max, %d nreq)\n", qp->qpn,
2201 					qp->rq.head, qp->rq.tail,
2202 					qp->rq.max, nreq);
2203 			err = -ENOMEM;
2204 			*bad_wr = wr;
2205 			goto out;
2206 		}
2207 
2208 		wqe = get_recv_wqe(qp, ind);
2209 
2210 		((struct mthca_next_seg *) wqe)->flags = 0;
2211 
2212 		wqe += sizeof (struct mthca_next_seg);
2213 
2214 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2215 			err = -EINVAL;
2216 			*bad_wr = wr;
2217 			goto out;
2218 		}
2219 
2220 		for (i = 0; i < wr->num_sge; ++i) {
2221 			mthca_set_data_seg(wqe, wr->sg_list + i);
2222 			wqe += sizeof (struct mthca_data_seg);
2223 		}
2224 
2225 		if (i < qp->rq.max_gs)
2226 			mthca_set_data_seg_inval(wqe);
2227 
2228 		qp->wrid[ind] = wr->wr_id;
2229 
2230 		++ind;
2231 		if (unlikely(ind >= qp->rq.max))
2232 			ind -= qp->rq.max;
2233 	}
2234 out:
2235 	if (likely(nreq)) {
2236 		qp->rq.head += nreq;
2237 
2238 		/*
2239 		 * Make sure that descriptors are written before
2240 		 * doorbell record.
2241 		 */
2242 		wmb();
2243 		*qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2244 	}
2245 
2246 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2247 	return err;
2248 }
2249 
2250 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2251 			int index, int *dbd, __be32 *new_wqe)
2252 {
2253 	struct mthca_next_seg *next;
2254 
2255 	/*
2256 	 * For SRQs, all receive WQEs generate a CQE, so we're always
2257 	 * at the end of the doorbell chain.
2258 	 */
2259 	if (qp->ibqp.srq && !is_send) {
2260 		*new_wqe = 0;
2261 		return;
2262 	}
2263 
2264 	if (is_send)
2265 		next = get_send_wqe(qp, index);
2266 	else
2267 		next = get_recv_wqe(qp, index);
2268 
2269 	*dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2270 	if (next->ee_nds & cpu_to_be32(0x3f))
2271 		*new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2272 			(next->ee_nds & cpu_to_be32(0x3f));
2273 	else
2274 		*new_wqe = 0;
2275 }
2276 
2277 int mthca_init_qp_table(struct mthca_dev *dev)
2278 {
2279 	int err;
2280 	u8 status;
2281 	int i;
2282 
2283 	spin_lock_init(&dev->qp_table.lock);
2284 
2285 	/*
2286 	 * We reserve 2 extra QPs per port for the special QPs.  The
2287 	 * special QP for port 1 has to be even, so round up.
2288 	 */
2289 	dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2290 	err = mthca_alloc_init(&dev->qp_table.alloc,
2291 			       dev->limits.num_qps,
2292 			       (1 << 24) - 1,
2293 			       dev->qp_table.sqp_start +
2294 			       MTHCA_MAX_PORTS * 2);
2295 	if (err)
2296 		return err;
2297 
2298 	err = mthca_array_init(&dev->qp_table.qp,
2299 			       dev->limits.num_qps);
2300 	if (err) {
2301 		mthca_alloc_cleanup(&dev->qp_table.alloc);
2302 		return err;
2303 	}
2304 
2305 	for (i = 0; i < 2; ++i) {
2306 		err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2307 					    dev->qp_table.sqp_start + i * 2,
2308 					    &status);
2309 		if (err)
2310 			goto err_out;
2311 		if (status) {
2312 			mthca_warn(dev, "CONF_SPECIAL_QP returned "
2313 				   "status %02x, aborting.\n",
2314 				   status);
2315 			err = -EINVAL;
2316 			goto err_out;
2317 		}
2318 	}
2319 	return 0;
2320 
2321  err_out:
2322 	for (i = 0; i < 2; ++i)
2323 		mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2324 
2325 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2326 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2327 
2328 	return err;
2329 }
2330 
2331 void mthca_cleanup_qp_table(struct mthca_dev *dev)
2332 {
2333 	int i;
2334 	u8 status;
2335 
2336 	for (i = 0; i < 2; ++i)
2337 		mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2338 
2339 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2340 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2341 }
2342