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 	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, &status);
454 	if (err)
455 		goto out_mailbox;
456 	if (status) {
457 		mthca_warn(dev, "QUERY_QP returned status %02x\n", status);
458 		err = -EINVAL;
459 		goto out_mailbox;
460 	}
461 
462 	qp_param    = mailbox->buf;
463 	context     = &qp_param->context;
464 	mthca_state = be32_to_cpu(context->flags) >> 28;
465 
466 	qp->state		     = to_ib_qp_state(mthca_state);
467 	qp_attr->qp_state	     = qp->state;
468 	qp_attr->path_mtu 	     = context->mtu_msgmax >> 5;
469 	qp_attr->path_mig_state      =
470 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
471 	qp_attr->qkey 		     = be32_to_cpu(context->qkey);
472 	qp_attr->rq_psn 	     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
473 	qp_attr->sq_psn 	     = be32_to_cpu(context->next_send_psn) & 0xffffff;
474 	qp_attr->dest_qp_num 	     = be32_to_cpu(context->remote_qpn) & 0xffffff;
475 	qp_attr->qp_access_flags     =
476 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
477 
478 	if (qp->transport == RC || qp->transport == UC) {
479 		to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
480 		to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
481 		qp_attr->alt_pkey_index =
482 			be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
483 		qp_attr->alt_port_num 	= qp_attr->alt_ah_attr.port_num;
484 	}
485 
486 	qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
487 	qp_attr->port_num   =
488 		(be32_to_cpu(context->pri_path.port_pkey) >> 24) & 0x3;
489 
490 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
491 	qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
492 
493 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
494 
495 	qp_attr->max_dest_rd_atomic =
496 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
497 	qp_attr->min_rnr_timer 	    =
498 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
499 	qp_attr->timeout 	    = context->pri_path.ackto >> 3;
500 	qp_attr->retry_cnt 	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
501 	qp_attr->rnr_retry 	    = context->pri_path.rnr_retry >> 5;
502 	qp_attr->alt_timeout 	    = context->alt_path.ackto >> 3;
503 
504 done:
505 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
506 	qp_attr->cap.max_send_wr     = qp->sq.max;
507 	qp_attr->cap.max_recv_wr     = qp->rq.max;
508 	qp_attr->cap.max_send_sge    = qp->sq.max_gs;
509 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
510 	qp_attr->cap.max_inline_data = qp->max_inline_data;
511 
512 	qp_init_attr->cap	     = qp_attr->cap;
513 
514 out_mailbox:
515 	mthca_free_mailbox(dev, mailbox);
516 
517 out:
518 	mutex_unlock(&qp->mutex);
519 	return err;
520 }
521 
522 static int mthca_path_set(struct mthca_dev *dev, const struct ib_ah_attr *ah,
523 			  struct mthca_qp_path *path, u8 port)
524 {
525 	path->g_mylmc     = ah->src_path_bits & 0x7f;
526 	path->rlid        = cpu_to_be16(ah->dlid);
527 	path->static_rate = mthca_get_rate(dev, ah->static_rate, port);
528 
529 	if (ah->ah_flags & IB_AH_GRH) {
530 		if (ah->grh.sgid_index >= dev->limits.gid_table_len) {
531 			mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
532 				  ah->grh.sgid_index, dev->limits.gid_table_len-1);
533 			return -1;
534 		}
535 
536 		path->g_mylmc   |= 1 << 7;
537 		path->mgid_index = ah->grh.sgid_index;
538 		path->hop_limit  = ah->grh.hop_limit;
539 		path->sl_tclass_flowlabel =
540 			cpu_to_be32((ah->sl << 28)                |
541 				    (ah->grh.traffic_class << 20) |
542 				    (ah->grh.flow_label));
543 		memcpy(path->rgid, ah->grh.dgid.raw, 16);
544 	} else
545 		path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28);
546 
547 	return 0;
548 }
549 
550 static int __mthca_modify_qp(struct ib_qp *ibqp,
551 			     const struct ib_qp_attr *attr, int attr_mask,
552 			     enum ib_qp_state cur_state, enum ib_qp_state new_state)
553 {
554 	struct mthca_dev *dev = to_mdev(ibqp->device);
555 	struct mthca_qp *qp = to_mqp(ibqp);
556 	struct mthca_mailbox *mailbox;
557 	struct mthca_qp_param *qp_param;
558 	struct mthca_qp_context *qp_context;
559 	u32 sqd_event = 0;
560 	u8 status;
561 	int err = -EINVAL;
562 
563 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
564 	if (IS_ERR(mailbox)) {
565 		err = PTR_ERR(mailbox);
566 		goto out;
567 	}
568 	qp_param = mailbox->buf;
569 	qp_context = &qp_param->context;
570 	memset(qp_param, 0, sizeof *qp_param);
571 
572 	qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
573 					     (to_mthca_st(qp->transport) << 16));
574 	qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
575 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
576 		qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
577 	else {
578 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
579 		switch (attr->path_mig_state) {
580 		case IB_MIG_MIGRATED:
581 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
582 			break;
583 		case IB_MIG_REARM:
584 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
585 			break;
586 		case IB_MIG_ARMED:
587 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
588 			break;
589 		}
590 	}
591 
592 	/* leave tavor_sched_queue as 0 */
593 
594 	if (qp->transport == MLX || qp->transport == UD)
595 		qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
596 	else if (attr_mask & IB_QP_PATH_MTU) {
597 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
598 			mthca_dbg(dev, "path MTU (%u) is invalid\n",
599 				  attr->path_mtu);
600 			goto out_mailbox;
601 		}
602 		qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
603 	}
604 
605 	if (mthca_is_memfree(dev)) {
606 		if (qp->rq.max)
607 			qp_context->rq_size_stride = ilog2(qp->rq.max) << 3;
608 		qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
609 
610 		if (qp->sq.max)
611 			qp_context->sq_size_stride = ilog2(qp->sq.max) << 3;
612 		qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
613 	}
614 
615 	/* leave arbel_sched_queue as 0 */
616 
617 	if (qp->ibqp.uobject)
618 		qp_context->usr_page =
619 			cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
620 	else
621 		qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
622 	qp_context->local_qpn  = cpu_to_be32(qp->qpn);
623 	if (attr_mask & IB_QP_DEST_QPN) {
624 		qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
625 	}
626 
627 	if (qp->transport == MLX)
628 		qp_context->pri_path.port_pkey |=
629 			cpu_to_be32(qp->port << 24);
630 	else {
631 		if (attr_mask & IB_QP_PORT) {
632 			qp_context->pri_path.port_pkey |=
633 				cpu_to_be32(attr->port_num << 24);
634 			qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
635 		}
636 	}
637 
638 	if (attr_mask & IB_QP_PKEY_INDEX) {
639 		qp_context->pri_path.port_pkey |=
640 			cpu_to_be32(attr->pkey_index);
641 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
642 	}
643 
644 	if (attr_mask & IB_QP_RNR_RETRY) {
645 		qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
646 			attr->rnr_retry << 5;
647 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
648 							MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
649 	}
650 
651 	if (attr_mask & IB_QP_AV) {
652 		if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
653 				   attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
654 			goto out_mailbox;
655 
656 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
657 	}
658 
659 	if (ibqp->qp_type == IB_QPT_RC &&
660 	    cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
661 		u8 sched_queue = ibqp->uobject ? 0x2 : 0x1;
662 
663 		if (mthca_is_memfree(dev))
664 			qp_context->rlkey_arbel_sched_queue |= sched_queue;
665 		else
666 			qp_context->tavor_sched_queue |= cpu_to_be32(sched_queue);
667 
668 		qp_param->opt_param_mask |=
669 			cpu_to_be32(MTHCA_QP_OPTPAR_SCHED_QUEUE);
670 	}
671 
672 	if (attr_mask & IB_QP_TIMEOUT) {
673 		qp_context->pri_path.ackto = attr->timeout << 3;
674 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
675 	}
676 
677 	if (attr_mask & IB_QP_ALT_PATH) {
678 		if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
679 			mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
680 				  attr->alt_pkey_index, dev->limits.pkey_table_len-1);
681 			goto out_mailbox;
682 		}
683 
684 		if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
685 			mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
686 				attr->alt_port_num);
687 			goto out_mailbox;
688 		}
689 
690 		if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
691 				   attr->alt_ah_attr.port_num))
692 			goto out_mailbox;
693 
694 		qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
695 							      attr->alt_port_num << 24);
696 		qp_context->alt_path.ackto = attr->alt_timeout << 3;
697 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
698 	}
699 
700 	/* leave rdd as 0 */
701 	qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
702 	/* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
703 	qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
704 	qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
705 					     (MTHCA_FLIGHT_LIMIT << 24) |
706 					     MTHCA_QP_BIT_SWE);
707 	if (qp->sq_policy == IB_SIGNAL_ALL_WR)
708 		qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
709 	if (attr_mask & IB_QP_RETRY_CNT) {
710 		qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
711 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
712 	}
713 
714 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
715 		if (attr->max_rd_atomic) {
716 			qp_context->params1 |=
717 				cpu_to_be32(MTHCA_QP_BIT_SRE |
718 					    MTHCA_QP_BIT_SAE);
719 			qp_context->params1 |=
720 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
721 		}
722 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
723 	}
724 
725 	if (attr_mask & IB_QP_SQ_PSN)
726 		qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
727 	qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
728 
729 	if (mthca_is_memfree(dev)) {
730 		qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
731 		qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
732 	}
733 
734 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
735 		if (attr->max_dest_rd_atomic)
736 			qp_context->params2 |=
737 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
738 
739 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
740 	}
741 
742 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
743 		qp_context->params2      |= get_hw_access_flags(qp, attr, attr_mask);
744 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
745 							MTHCA_QP_OPTPAR_RRE |
746 							MTHCA_QP_OPTPAR_RAE);
747 	}
748 
749 	qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
750 
751 	if (ibqp->srq)
752 		qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
753 
754 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
755 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
756 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
757 	}
758 	if (attr_mask & IB_QP_RQ_PSN)
759 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
760 
761 	qp_context->ra_buff_indx =
762 		cpu_to_be32(dev->qp_table.rdb_base +
763 			    ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
764 			     dev->qp_table.rdb_shift));
765 
766 	qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
767 
768 	if (mthca_is_memfree(dev))
769 		qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
770 
771 	if (attr_mask & IB_QP_QKEY) {
772 		qp_context->qkey = cpu_to_be32(attr->qkey);
773 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
774 	}
775 
776 	if (ibqp->srq)
777 		qp_context->srqn = cpu_to_be32(1 << 24 |
778 					       to_msrq(ibqp->srq)->srqn);
779 
780 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
781 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY		&&
782 	    attr->en_sqd_async_notify)
783 		sqd_event = 1 << 31;
784 
785 	err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
786 			      mailbox, sqd_event, &status);
787 	if (err)
788 		goto out_mailbox;
789 	if (status) {
790 		mthca_warn(dev, "modify QP %d->%d returned status %02x.\n",
791 			   cur_state, new_state, status);
792 		err = -EINVAL;
793 		goto out_mailbox;
794 	}
795 
796 	qp->state = new_state;
797 	if (attr_mask & IB_QP_ACCESS_FLAGS)
798 		qp->atomic_rd_en = attr->qp_access_flags;
799 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
800 		qp->resp_depth = attr->max_dest_rd_atomic;
801 	if (attr_mask & IB_QP_PORT)
802 		qp->port = attr->port_num;
803 	if (attr_mask & IB_QP_ALT_PATH)
804 		qp->alt_port = attr->alt_port_num;
805 
806 	if (is_sqp(dev, qp))
807 		store_attrs(to_msqp(qp), attr, attr_mask);
808 
809 	/*
810 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
811 	 * QP0 to RESET or ERROR, bring the link back down.
812 	 */
813 	if (is_qp0(dev, qp)) {
814 		if (cur_state != IB_QPS_RTR &&
815 		    new_state == IB_QPS_RTR)
816 			init_port(dev, qp->port);
817 
818 		if (cur_state != IB_QPS_RESET &&
819 		    cur_state != IB_QPS_ERR &&
820 		    (new_state == IB_QPS_RESET ||
821 		     new_state == IB_QPS_ERR))
822 			mthca_CLOSE_IB(dev, qp->port, &status);
823 	}
824 
825 	/*
826 	 * If we moved a kernel QP to RESET, clean up all old CQ
827 	 * entries and reinitialize the QP.
828 	 */
829 	if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
830 		mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
831 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
832 		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
833 			mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn, NULL);
834 
835 		mthca_wq_reset(&qp->sq);
836 		qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
837 
838 		mthca_wq_reset(&qp->rq);
839 		qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
840 
841 		if (mthca_is_memfree(dev)) {
842 			*qp->sq.db = 0;
843 			*qp->rq.db = 0;
844 		}
845 	}
846 
847 out_mailbox:
848 	mthca_free_mailbox(dev, mailbox);
849 out:
850 	return err;
851 }
852 
853 static const struct ib_qp_attr dummy_init_attr = { .port_num = 1 };
854 static const int dummy_init_attr_mask[] = {
855 	[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
856 			IB_QP_PORT			|
857 			IB_QP_QKEY),
858 	[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
859 			IB_QP_PORT			|
860 			IB_QP_ACCESS_FLAGS),
861 	[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
862 			IB_QP_PORT			|
863 			IB_QP_ACCESS_FLAGS),
864 	[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
865 			IB_QP_QKEY),
866 	[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
867 			IB_QP_QKEY),
868 };
869 
870 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
871 		    struct ib_udata *udata)
872 {
873 	struct mthca_dev *dev = to_mdev(ibqp->device);
874 	struct mthca_qp *qp = to_mqp(ibqp);
875 	enum ib_qp_state cur_state, new_state;
876 	int err = -EINVAL;
877 
878 	mutex_lock(&qp->mutex);
879 	if (attr_mask & IB_QP_CUR_STATE) {
880 		cur_state = attr->cur_qp_state;
881 	} else {
882 		spin_lock_irq(&qp->sq.lock);
883 		spin_lock(&qp->rq.lock);
884 		cur_state = qp->state;
885 		spin_unlock(&qp->rq.lock);
886 		spin_unlock_irq(&qp->sq.lock);
887 	}
888 
889 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
890 
891 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
892 		mthca_dbg(dev, "Bad QP transition (transport %d) "
893 			  "%d->%d with attr 0x%08x\n",
894 			  qp->transport, cur_state, new_state,
895 			  attr_mask);
896 		goto out;
897 	}
898 
899 	if ((attr_mask & IB_QP_PKEY_INDEX) &&
900 	     attr->pkey_index >= dev->limits.pkey_table_len) {
901 		mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
902 			  attr->pkey_index, dev->limits.pkey_table_len-1);
903 		goto out;
904 	}
905 
906 	if ((attr_mask & IB_QP_PORT) &&
907 	    (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
908 		mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
909 		goto out;
910 	}
911 
912 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
913 	    attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
914 		mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
915 			  attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
916 		goto out;
917 	}
918 
919 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
920 	    attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
921 		mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
922 			  attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
923 		goto out;
924 	}
925 
926 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
927 		err = 0;
928 		goto out;
929 	}
930 
931 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_ERR) {
932 		err = __mthca_modify_qp(ibqp, &dummy_init_attr,
933 					dummy_init_attr_mask[ibqp->qp_type],
934 					IB_QPS_RESET, IB_QPS_INIT);
935 		if (err)
936 			goto out;
937 		cur_state = IB_QPS_INIT;
938 	}
939 
940 	err = __mthca_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
941 
942 out:
943 	mutex_unlock(&qp->mutex);
944 	return err;
945 }
946 
947 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
948 {
949 	/*
950 	 * Calculate the maximum size of WQE s/g segments, excluding
951 	 * the next segment and other non-data segments.
952 	 */
953 	int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
954 
955 	switch (qp->transport) {
956 	case MLX:
957 		max_data_size -= 2 * sizeof (struct mthca_data_seg);
958 		break;
959 
960 	case UD:
961 		if (mthca_is_memfree(dev))
962 			max_data_size -= sizeof (struct mthca_arbel_ud_seg);
963 		else
964 			max_data_size -= sizeof (struct mthca_tavor_ud_seg);
965 		break;
966 
967 	default:
968 		max_data_size -= sizeof (struct mthca_raddr_seg);
969 		break;
970 	}
971 
972 	return max_data_size;
973 }
974 
975 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
976 {
977 	/* We don't support inline data for kernel QPs (yet). */
978 	return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
979 }
980 
981 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
982 				 struct mthca_pd *pd,
983 				 struct mthca_qp *qp)
984 {
985 	int max_data_size = mthca_max_data_size(dev, qp,
986 						min(dev->limits.max_desc_sz,
987 						    1 << qp->sq.wqe_shift));
988 
989 	qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
990 
991 	qp->sq.max_gs = min_t(int, dev->limits.max_sg,
992 			      max_data_size / sizeof (struct mthca_data_seg));
993 	qp->rq.max_gs = min_t(int, dev->limits.max_sg,
994 			       (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
995 				sizeof (struct mthca_next_seg)) /
996 			       sizeof (struct mthca_data_seg));
997 }
998 
999 /*
1000  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
1001  * rq.max_gs and sq.max_gs must all be assigned.
1002  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
1003  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
1004  * queue)
1005  */
1006 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
1007 			       struct mthca_pd *pd,
1008 			       struct mthca_qp *qp)
1009 {
1010 	int size;
1011 	int err = -ENOMEM;
1012 
1013 	size = sizeof (struct mthca_next_seg) +
1014 		qp->rq.max_gs * sizeof (struct mthca_data_seg);
1015 
1016 	if (size > dev->limits.max_desc_sz)
1017 		return -EINVAL;
1018 
1019 	for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
1020 	     qp->rq.wqe_shift++)
1021 		; /* nothing */
1022 
1023 	size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
1024 	switch (qp->transport) {
1025 	case MLX:
1026 		size += 2 * sizeof (struct mthca_data_seg);
1027 		break;
1028 
1029 	case UD:
1030 		size += mthca_is_memfree(dev) ?
1031 			sizeof (struct mthca_arbel_ud_seg) :
1032 			sizeof (struct mthca_tavor_ud_seg);
1033 		break;
1034 
1035 	case UC:
1036 		size += sizeof (struct mthca_raddr_seg);
1037 		break;
1038 
1039 	case RC:
1040 		size += sizeof (struct mthca_raddr_seg);
1041 		/*
1042 		 * An atomic op will require an atomic segment, a
1043 		 * remote address segment and one scatter entry.
1044 		 */
1045 		size = max_t(int, size,
1046 			     sizeof (struct mthca_atomic_seg) +
1047 			     sizeof (struct mthca_raddr_seg) +
1048 			     sizeof (struct mthca_data_seg));
1049 		break;
1050 
1051 	default:
1052 		break;
1053 	}
1054 
1055 	/* Make sure that we have enough space for a bind request */
1056 	size = max_t(int, size, sizeof (struct mthca_bind_seg));
1057 
1058 	size += sizeof (struct mthca_next_seg);
1059 
1060 	if (size > dev->limits.max_desc_sz)
1061 		return -EINVAL;
1062 
1063 	for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
1064 	     qp->sq.wqe_shift++)
1065 		; /* nothing */
1066 
1067 	qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
1068 				    1 << qp->sq.wqe_shift);
1069 
1070 	/*
1071 	 * If this is a userspace QP, we don't actually have to
1072 	 * allocate anything.  All we need is to calculate the WQE
1073 	 * sizes and the send_wqe_offset, so we're done now.
1074 	 */
1075 	if (pd->ibpd.uobject)
1076 		return 0;
1077 
1078 	size = PAGE_ALIGN(qp->send_wqe_offset +
1079 			  (qp->sq.max << qp->sq.wqe_shift));
1080 
1081 	qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1082 			   GFP_KERNEL);
1083 	if (!qp->wrid)
1084 		goto err_out;
1085 
1086 	err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1087 			      &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1088 	if (err)
1089 		goto err_out;
1090 
1091 	return 0;
1092 
1093 err_out:
1094 	kfree(qp->wrid);
1095 	return err;
1096 }
1097 
1098 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1099 			       struct mthca_qp *qp)
1100 {
1101 	mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1102 				       (qp->sq.max << qp->sq.wqe_shift)),
1103 		       &qp->queue, qp->is_direct, &qp->mr);
1104 	kfree(qp->wrid);
1105 }
1106 
1107 static int mthca_map_memfree(struct mthca_dev *dev,
1108 			     struct mthca_qp *qp)
1109 {
1110 	int ret;
1111 
1112 	if (mthca_is_memfree(dev)) {
1113 		ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1114 		if (ret)
1115 			return ret;
1116 
1117 		ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1118 		if (ret)
1119 			goto err_qpc;
1120 
1121 		ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1122 				      qp->qpn << dev->qp_table.rdb_shift);
1123 		if (ret)
1124 			goto err_eqpc;
1125 
1126 	}
1127 
1128 	return 0;
1129 
1130 err_eqpc:
1131 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1132 
1133 err_qpc:
1134 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1135 
1136 	return ret;
1137 }
1138 
1139 static void mthca_unmap_memfree(struct mthca_dev *dev,
1140 				struct mthca_qp *qp)
1141 {
1142 	mthca_table_put(dev, dev->qp_table.rdb_table,
1143 			qp->qpn << dev->qp_table.rdb_shift);
1144 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1145 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1146 }
1147 
1148 static int mthca_alloc_memfree(struct mthca_dev *dev,
1149 			       struct mthca_qp *qp)
1150 {
1151 	if (mthca_is_memfree(dev)) {
1152 		qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1153 						 qp->qpn, &qp->rq.db);
1154 		if (qp->rq.db_index < 0)
1155 			return -ENOMEM;
1156 
1157 		qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1158 						 qp->qpn, &qp->sq.db);
1159 		if (qp->sq.db_index < 0) {
1160 			mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1161 			return -ENOMEM;
1162 		}
1163 	}
1164 
1165 	return 0;
1166 }
1167 
1168 static void mthca_free_memfree(struct mthca_dev *dev,
1169 			       struct mthca_qp *qp)
1170 {
1171 	if (mthca_is_memfree(dev)) {
1172 		mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1173 		mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1174 	}
1175 }
1176 
1177 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1178 				 struct mthca_pd *pd,
1179 				 struct mthca_cq *send_cq,
1180 				 struct mthca_cq *recv_cq,
1181 				 enum ib_sig_type send_policy,
1182 				 struct mthca_qp *qp)
1183 {
1184 	int ret;
1185 	int i;
1186 	struct mthca_next_seg *next;
1187 
1188 	qp->refcount = 1;
1189 	init_waitqueue_head(&qp->wait);
1190 	mutex_init(&qp->mutex);
1191 	qp->state    	 = IB_QPS_RESET;
1192 	qp->atomic_rd_en = 0;
1193 	qp->resp_depth   = 0;
1194 	qp->sq_policy    = send_policy;
1195 	mthca_wq_reset(&qp->sq);
1196 	mthca_wq_reset(&qp->rq);
1197 
1198 	spin_lock_init(&qp->sq.lock);
1199 	spin_lock_init(&qp->rq.lock);
1200 
1201 	ret = mthca_map_memfree(dev, qp);
1202 	if (ret)
1203 		return ret;
1204 
1205 	ret = mthca_alloc_wqe_buf(dev, pd, qp);
1206 	if (ret) {
1207 		mthca_unmap_memfree(dev, qp);
1208 		return ret;
1209 	}
1210 
1211 	mthca_adjust_qp_caps(dev, pd, qp);
1212 
1213 	/*
1214 	 * If this is a userspace QP, we're done now.  The doorbells
1215 	 * will be allocated and buffers will be initialized in
1216 	 * userspace.
1217 	 */
1218 	if (pd->ibpd.uobject)
1219 		return 0;
1220 
1221 	ret = mthca_alloc_memfree(dev, qp);
1222 	if (ret) {
1223 		mthca_free_wqe_buf(dev, qp);
1224 		mthca_unmap_memfree(dev, qp);
1225 		return ret;
1226 	}
1227 
1228 	if (mthca_is_memfree(dev)) {
1229 		struct mthca_data_seg *scatter;
1230 		int size = (sizeof (struct mthca_next_seg) +
1231 			    qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1232 
1233 		for (i = 0; i < qp->rq.max; ++i) {
1234 			next = get_recv_wqe(qp, i);
1235 			next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1236 						   qp->rq.wqe_shift);
1237 			next->ee_nds = cpu_to_be32(size);
1238 
1239 			for (scatter = (void *) (next + 1);
1240 			     (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1241 			     ++scatter)
1242 				scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1243 		}
1244 
1245 		for (i = 0; i < qp->sq.max; ++i) {
1246 			next = get_send_wqe(qp, i);
1247 			next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1248 						    qp->sq.wqe_shift) +
1249 						   qp->send_wqe_offset);
1250 		}
1251 	} else {
1252 		for (i = 0; i < qp->rq.max; ++i) {
1253 			next = get_recv_wqe(qp, i);
1254 			next->nda_op = htonl((((i + 1) % qp->rq.max) <<
1255 					      qp->rq.wqe_shift) | 1);
1256 		}
1257 
1258 	}
1259 
1260 	qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1261 	qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1262 
1263 	return 0;
1264 }
1265 
1266 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1267 			     struct mthca_pd *pd, struct mthca_qp *qp)
1268 {
1269 	int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1270 
1271 	/* Sanity check QP size before proceeding */
1272 	if (cap->max_send_wr  	 > dev->limits.max_wqes ||
1273 	    cap->max_recv_wr  	 > dev->limits.max_wqes ||
1274 	    cap->max_send_sge 	 > dev->limits.max_sg   ||
1275 	    cap->max_recv_sge 	 > dev->limits.max_sg   ||
1276 	    cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1277 		return -EINVAL;
1278 
1279 	/*
1280 	 * For MLX transport we need 2 extra S/G entries:
1281 	 * one for the header and one for the checksum at the end
1282 	 */
1283 	if (qp->transport == MLX && cap->max_recv_sge + 2 > dev->limits.max_sg)
1284 		return -EINVAL;
1285 
1286 	if (mthca_is_memfree(dev)) {
1287 		qp->rq.max = cap->max_recv_wr ?
1288 			roundup_pow_of_two(cap->max_recv_wr) : 0;
1289 		qp->sq.max = cap->max_send_wr ?
1290 			roundup_pow_of_two(cap->max_send_wr) : 0;
1291 	} else {
1292 		qp->rq.max = cap->max_recv_wr;
1293 		qp->sq.max = cap->max_send_wr;
1294 	}
1295 
1296 	qp->rq.max_gs = cap->max_recv_sge;
1297 	qp->sq.max_gs = max_t(int, cap->max_send_sge,
1298 			      ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1299 				    MTHCA_INLINE_CHUNK_SIZE) /
1300 			      sizeof (struct mthca_data_seg));
1301 
1302 	return 0;
1303 }
1304 
1305 int mthca_alloc_qp(struct mthca_dev *dev,
1306 		   struct mthca_pd *pd,
1307 		   struct mthca_cq *send_cq,
1308 		   struct mthca_cq *recv_cq,
1309 		   enum ib_qp_type type,
1310 		   enum ib_sig_type send_policy,
1311 		   struct ib_qp_cap *cap,
1312 		   struct mthca_qp *qp)
1313 {
1314 	int err;
1315 
1316 	switch (type) {
1317 	case IB_QPT_RC: qp->transport = RC; break;
1318 	case IB_QPT_UC: qp->transport = UC; break;
1319 	case IB_QPT_UD: qp->transport = UD; break;
1320 	default: return -EINVAL;
1321 	}
1322 
1323 	err = mthca_set_qp_size(dev, cap, pd, qp);
1324 	if (err)
1325 		return err;
1326 
1327 	qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1328 	if (qp->qpn == -1)
1329 		return -ENOMEM;
1330 
1331 	/* initialize port to zero for error-catching. */
1332 	qp->port = 0;
1333 
1334 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1335 				    send_policy, qp);
1336 	if (err) {
1337 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1338 		return err;
1339 	}
1340 
1341 	spin_lock_irq(&dev->qp_table.lock);
1342 	mthca_array_set(&dev->qp_table.qp,
1343 			qp->qpn & (dev->limits.num_qps - 1), qp);
1344 	spin_unlock_irq(&dev->qp_table.lock);
1345 
1346 	return 0;
1347 }
1348 
1349 static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1350 {
1351 	if (send_cq == recv_cq)
1352 		spin_lock_irq(&send_cq->lock);
1353 	else if (send_cq->cqn < recv_cq->cqn) {
1354 		spin_lock_irq(&send_cq->lock);
1355 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1356 	} else {
1357 		spin_lock_irq(&recv_cq->lock);
1358 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1359 	}
1360 }
1361 
1362 static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1363 {
1364 	if (send_cq == recv_cq)
1365 		spin_unlock_irq(&send_cq->lock);
1366 	else if (send_cq->cqn < recv_cq->cqn) {
1367 		spin_unlock(&recv_cq->lock);
1368 		spin_unlock_irq(&send_cq->lock);
1369 	} else {
1370 		spin_unlock(&send_cq->lock);
1371 		spin_unlock_irq(&recv_cq->lock);
1372 	}
1373 }
1374 
1375 int mthca_alloc_sqp(struct mthca_dev *dev,
1376 		    struct mthca_pd *pd,
1377 		    struct mthca_cq *send_cq,
1378 		    struct mthca_cq *recv_cq,
1379 		    enum ib_sig_type send_policy,
1380 		    struct ib_qp_cap *cap,
1381 		    int qpn,
1382 		    int port,
1383 		    struct mthca_sqp *sqp)
1384 {
1385 	u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1386 	int err;
1387 
1388 	sqp->qp.transport = MLX;
1389 	err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1390 	if (err)
1391 		return err;
1392 
1393 	sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1394 	sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1395 					     &sqp->header_dma, GFP_KERNEL);
1396 	if (!sqp->header_buf)
1397 		return -ENOMEM;
1398 
1399 	spin_lock_irq(&dev->qp_table.lock);
1400 	if (mthca_array_get(&dev->qp_table.qp, mqpn))
1401 		err = -EBUSY;
1402 	else
1403 		mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1404 	spin_unlock_irq(&dev->qp_table.lock);
1405 
1406 	if (err)
1407 		goto err_out;
1408 
1409 	sqp->qp.port      = port;
1410 	sqp->qp.qpn       = mqpn;
1411 	sqp->qp.transport = MLX;
1412 
1413 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1414 				    send_policy, &sqp->qp);
1415 	if (err)
1416 		goto err_out_free;
1417 
1418 	atomic_inc(&pd->sqp_count);
1419 
1420 	return 0;
1421 
1422  err_out_free:
1423 	/*
1424 	 * Lock CQs here, so that CQ polling code can do QP lookup
1425 	 * without taking a lock.
1426 	 */
1427 	mthca_lock_cqs(send_cq, recv_cq);
1428 
1429 	spin_lock(&dev->qp_table.lock);
1430 	mthca_array_clear(&dev->qp_table.qp, mqpn);
1431 	spin_unlock(&dev->qp_table.lock);
1432 
1433 	mthca_unlock_cqs(send_cq, recv_cq);
1434 
1435  err_out:
1436 	dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1437 			  sqp->header_buf, sqp->header_dma);
1438 
1439 	return err;
1440 }
1441 
1442 static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp)
1443 {
1444 	int c;
1445 
1446 	spin_lock_irq(&dev->qp_table.lock);
1447 	c = qp->refcount;
1448 	spin_unlock_irq(&dev->qp_table.lock);
1449 
1450 	return c;
1451 }
1452 
1453 void mthca_free_qp(struct mthca_dev *dev,
1454 		   struct mthca_qp *qp)
1455 {
1456 	u8 status;
1457 	struct mthca_cq *send_cq;
1458 	struct mthca_cq *recv_cq;
1459 
1460 	send_cq = to_mcq(qp->ibqp.send_cq);
1461 	recv_cq = to_mcq(qp->ibqp.recv_cq);
1462 
1463 	/*
1464 	 * Lock CQs here, so that CQ polling code can do QP lookup
1465 	 * without taking a lock.
1466 	 */
1467 	mthca_lock_cqs(send_cq, recv_cq);
1468 
1469 	spin_lock(&dev->qp_table.lock);
1470 	mthca_array_clear(&dev->qp_table.qp,
1471 			  qp->qpn & (dev->limits.num_qps - 1));
1472 	--qp->refcount;
1473 	spin_unlock(&dev->qp_table.lock);
1474 
1475 	mthca_unlock_cqs(send_cq, recv_cq);
1476 
1477 	wait_event(qp->wait, !get_qp_refcount(dev, qp));
1478 
1479 	if (qp->state != IB_QPS_RESET)
1480 		mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1481 				NULL, 0, &status);
1482 
1483 	/*
1484 	 * If this is a userspace QP, the buffers, MR, CQs and so on
1485 	 * will be cleaned up in userspace, so all we have to do is
1486 	 * unref the mem-free tables and free the QPN in our table.
1487 	 */
1488 	if (!qp->ibqp.uobject) {
1489 		mthca_cq_clean(dev, recv_cq, qp->qpn,
1490 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1491 		if (send_cq != recv_cq)
1492 			mthca_cq_clean(dev, send_cq, qp->qpn, NULL);
1493 
1494 		mthca_free_memfree(dev, qp);
1495 		mthca_free_wqe_buf(dev, qp);
1496 	}
1497 
1498 	mthca_unmap_memfree(dev, qp);
1499 
1500 	if (is_sqp(dev, qp)) {
1501 		atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1502 		dma_free_coherent(&dev->pdev->dev,
1503 				  to_msqp(qp)->header_buf_size,
1504 				  to_msqp(qp)->header_buf,
1505 				  to_msqp(qp)->header_dma);
1506 	} else
1507 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1508 }
1509 
1510 /* Create UD header for an MLX send and build a data segment for it */
1511 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1512 			    int ind, struct ib_send_wr *wr,
1513 			    struct mthca_mlx_seg *mlx,
1514 			    struct mthca_data_seg *data)
1515 {
1516 	int header_size;
1517 	int err;
1518 	u16 pkey;
1519 
1520 	ib_ud_header_init(256, /* assume a MAD */
1521 			  mthca_ah_grh_present(to_mah(wr->wr.ud.ah)),
1522 			  &sqp->ud_header);
1523 
1524 	err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1525 	if (err)
1526 		return err;
1527 	mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1528 	mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1529 				  (sqp->ud_header.lrh.destination_lid ==
1530 				   IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1531 				  (sqp->ud_header.lrh.service_level << 8));
1532 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
1533 	mlx->vcrc = 0;
1534 
1535 	switch (wr->opcode) {
1536 	case IB_WR_SEND:
1537 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1538 		sqp->ud_header.immediate_present = 0;
1539 		break;
1540 	case IB_WR_SEND_WITH_IMM:
1541 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1542 		sqp->ud_header.immediate_present = 1;
1543 		sqp->ud_header.immediate_data = wr->ex.imm_data;
1544 		break;
1545 	default:
1546 		return -EINVAL;
1547 	}
1548 
1549 	sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1550 	if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1551 		sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1552 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1553 	if (!sqp->qp.ibqp.qp_num)
1554 		ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1555 				   sqp->pkey_index, &pkey);
1556 	else
1557 		ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1558 				   wr->wr.ud.pkey_index, &pkey);
1559 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1560 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1561 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1562 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1563 					       sqp->qkey : wr->wr.ud.remote_qkey);
1564 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1565 
1566 	header_size = ib_ud_header_pack(&sqp->ud_header,
1567 					sqp->header_buf +
1568 					ind * MTHCA_UD_HEADER_SIZE);
1569 
1570 	data->byte_count = cpu_to_be32(header_size);
1571 	data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1572 	data->addr       = cpu_to_be64(sqp->header_dma +
1573 				       ind * MTHCA_UD_HEADER_SIZE);
1574 
1575 	return 0;
1576 }
1577 
1578 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1579 				    struct ib_cq *ib_cq)
1580 {
1581 	unsigned cur;
1582 	struct mthca_cq *cq;
1583 
1584 	cur = wq->head - wq->tail;
1585 	if (likely(cur + nreq < wq->max))
1586 		return 0;
1587 
1588 	cq = to_mcq(ib_cq);
1589 	spin_lock(&cq->lock);
1590 	cur = wq->head - wq->tail;
1591 	spin_unlock(&cq->lock);
1592 
1593 	return cur + nreq >= wq->max;
1594 }
1595 
1596 static __always_inline void set_raddr_seg(struct mthca_raddr_seg *rseg,
1597 					  u64 remote_addr, u32 rkey)
1598 {
1599 	rseg->raddr    = cpu_to_be64(remote_addr);
1600 	rseg->rkey     = cpu_to_be32(rkey);
1601 	rseg->reserved = 0;
1602 }
1603 
1604 static __always_inline void set_atomic_seg(struct mthca_atomic_seg *aseg,
1605 					   struct ib_send_wr *wr)
1606 {
1607 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1608 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
1609 		aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
1610 	} else {
1611 		aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
1612 		aseg->compare  = 0;
1613 	}
1614 
1615 }
1616 
1617 static void set_tavor_ud_seg(struct mthca_tavor_ud_seg *useg,
1618 			     struct ib_send_wr *wr)
1619 {
1620 	useg->lkey    = cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1621 	useg->av_addr =	cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1622 	useg->dqpn    =	cpu_to_be32(wr->wr.ud.remote_qpn);
1623 	useg->qkey    =	cpu_to_be32(wr->wr.ud.remote_qkey);
1624 
1625 }
1626 
1627 static void set_arbel_ud_seg(struct mthca_arbel_ud_seg *useg,
1628 			     struct ib_send_wr *wr)
1629 {
1630 	memcpy(useg->av, to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1631 	useg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1632 	useg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1633 }
1634 
1635 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1636 			  struct ib_send_wr **bad_wr)
1637 {
1638 	struct mthca_dev *dev = to_mdev(ibqp->device);
1639 	struct mthca_qp *qp = to_mqp(ibqp);
1640 	void *wqe;
1641 	void *prev_wqe;
1642 	unsigned long flags;
1643 	int err = 0;
1644 	int nreq;
1645 	int i;
1646 	int size;
1647 	/*
1648 	 * f0 and size0 are only used if nreq != 0, and they will
1649 	 * always be initialized the first time through the main loop
1650 	 * before nreq is incremented.  So nreq cannot become non-zero
1651 	 * without initializing f0 and size0, and they are in fact
1652 	 * never used uninitialized.
1653 	 */
1654 	int uninitialized_var(size0);
1655 	u32 uninitialized_var(f0);
1656 	int ind;
1657 	u8 op0 = 0;
1658 
1659 	spin_lock_irqsave(&qp->sq.lock, flags);
1660 
1661 	/* XXX check that state is OK to post send */
1662 
1663 	ind = qp->sq.next_ind;
1664 
1665 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1666 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1667 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1668 					" %d max, %d nreq)\n", qp->qpn,
1669 					qp->sq.head, qp->sq.tail,
1670 					qp->sq.max, nreq);
1671 			err = -ENOMEM;
1672 			*bad_wr = wr;
1673 			goto out;
1674 		}
1675 
1676 		wqe = get_send_wqe(qp, ind);
1677 		prev_wqe = qp->sq.last;
1678 		qp->sq.last = wqe;
1679 
1680 		((struct mthca_next_seg *) wqe)->nda_op = 0;
1681 		((struct mthca_next_seg *) wqe)->ee_nds = 0;
1682 		((struct mthca_next_seg *) wqe)->flags =
1683 			((wr->send_flags & IB_SEND_SIGNALED) ?
1684 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1685 			((wr->send_flags & IB_SEND_SOLICITED) ?
1686 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1687 			cpu_to_be32(1);
1688 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1689 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1690 			((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
1691 
1692 		wqe += sizeof (struct mthca_next_seg);
1693 		size = sizeof (struct mthca_next_seg) / 16;
1694 
1695 		switch (qp->transport) {
1696 		case RC:
1697 			switch (wr->opcode) {
1698 			case IB_WR_ATOMIC_CMP_AND_SWP:
1699 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1700 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
1701 					      wr->wr.atomic.rkey);
1702 				wqe += sizeof (struct mthca_raddr_seg);
1703 
1704 				set_atomic_seg(wqe, wr);
1705 				wqe += sizeof (struct mthca_atomic_seg);
1706 				size += (sizeof (struct mthca_raddr_seg) +
1707 					 sizeof (struct mthca_atomic_seg)) / 16;
1708 				break;
1709 
1710 			case IB_WR_RDMA_WRITE:
1711 			case IB_WR_RDMA_WRITE_WITH_IMM:
1712 			case IB_WR_RDMA_READ:
1713 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1714 					      wr->wr.rdma.rkey);
1715 				wqe  += sizeof (struct mthca_raddr_seg);
1716 				size += sizeof (struct mthca_raddr_seg) / 16;
1717 				break;
1718 
1719 			default:
1720 				/* No extra segments required for sends */
1721 				break;
1722 			}
1723 
1724 			break;
1725 
1726 		case UC:
1727 			switch (wr->opcode) {
1728 			case IB_WR_RDMA_WRITE:
1729 			case IB_WR_RDMA_WRITE_WITH_IMM:
1730 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
1731 					      wr->wr.rdma.rkey);
1732 				wqe  += sizeof (struct mthca_raddr_seg);
1733 				size += sizeof (struct mthca_raddr_seg) / 16;
1734 				break;
1735 
1736 			default:
1737 				/* No extra segments required for sends */
1738 				break;
1739 			}
1740 
1741 			break;
1742 
1743 		case UD:
1744 			set_tavor_ud_seg(wqe, wr);
1745 			wqe  += sizeof (struct mthca_tavor_ud_seg);
1746 			size += sizeof (struct mthca_tavor_ud_seg) / 16;
1747 			break;
1748 
1749 		case MLX:
1750 			err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1751 					       wqe - sizeof (struct mthca_next_seg),
1752 					       wqe);
1753 			if (err) {
1754 				*bad_wr = wr;
1755 				goto out;
1756 			}
1757 			wqe += sizeof (struct mthca_data_seg);
1758 			size += sizeof (struct mthca_data_seg) / 16;
1759 			break;
1760 		}
1761 
1762 		if (wr->num_sge > qp->sq.max_gs) {
1763 			mthca_err(dev, "too many gathers\n");
1764 			err = -EINVAL;
1765 			*bad_wr = wr;
1766 			goto out;
1767 		}
1768 
1769 		for (i = 0; i < wr->num_sge; ++i) {
1770 			mthca_set_data_seg(wqe, wr->sg_list + i);
1771 			wqe  += sizeof (struct mthca_data_seg);
1772 			size += sizeof (struct mthca_data_seg) / 16;
1773 		}
1774 
1775 		/* Add one more inline data segment for ICRC */
1776 		if (qp->transport == MLX) {
1777 			((struct mthca_data_seg *) wqe)->byte_count =
1778 				cpu_to_be32((1 << 31) | 4);
1779 			((u32 *) wqe)[1] = 0;
1780 			wqe += sizeof (struct mthca_data_seg);
1781 			size += sizeof (struct mthca_data_seg) / 16;
1782 		}
1783 
1784 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
1785 
1786 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1787 			mthca_err(dev, "opcode invalid\n");
1788 			err = -EINVAL;
1789 			*bad_wr = wr;
1790 			goto out;
1791 		}
1792 
1793 		((struct mthca_next_seg *) prev_wqe)->nda_op =
1794 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
1795 				     qp->send_wqe_offset) |
1796 				    mthca_opcode[wr->opcode]);
1797 		wmb();
1798 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1799 			cpu_to_be32((nreq ? 0 : MTHCA_NEXT_DBD) | size |
1800 				    ((wr->send_flags & IB_SEND_FENCE) ?
1801 				    MTHCA_NEXT_FENCE : 0));
1802 
1803 		if (!nreq) {
1804 			size0 = size;
1805 			op0   = mthca_opcode[wr->opcode];
1806 			f0    = wr->send_flags & IB_SEND_FENCE ?
1807 				MTHCA_SEND_DOORBELL_FENCE : 0;
1808 		}
1809 
1810 		++ind;
1811 		if (unlikely(ind >= qp->sq.max))
1812 			ind -= qp->sq.max;
1813 	}
1814 
1815 out:
1816 	if (likely(nreq)) {
1817 		wmb();
1818 
1819 		mthca_write64(((qp->sq.next_ind << qp->sq.wqe_shift) +
1820 			       qp->send_wqe_offset) | f0 | op0,
1821 			      (qp->qpn << 8) | size0,
1822 			      dev->kar + MTHCA_SEND_DOORBELL,
1823 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1824 		/*
1825 		 * Make sure doorbells don't leak out of SQ spinlock
1826 		 * and reach the HCA out of order:
1827 		 */
1828 		mmiowb();
1829 	}
1830 
1831 	qp->sq.next_ind = ind;
1832 	qp->sq.head    += nreq;
1833 
1834 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1835 	return err;
1836 }
1837 
1838 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1839 			     struct ib_recv_wr **bad_wr)
1840 {
1841 	struct mthca_dev *dev = to_mdev(ibqp->device);
1842 	struct mthca_qp *qp = to_mqp(ibqp);
1843 	unsigned long flags;
1844 	int err = 0;
1845 	int nreq;
1846 	int i;
1847 	int size;
1848 	/*
1849 	 * size0 is only used if nreq != 0, and it will always be
1850 	 * initialized the first time through the main loop before
1851 	 * nreq is incremented.  So nreq cannot become non-zero
1852 	 * without initializing size0, and it is in fact never used
1853 	 * uninitialized.
1854 	 */
1855 	int uninitialized_var(size0);
1856 	int ind;
1857 	void *wqe;
1858 	void *prev_wqe;
1859 
1860 	spin_lock_irqsave(&qp->rq.lock, flags);
1861 
1862 	/* XXX check that state is OK to post receive */
1863 
1864 	ind = qp->rq.next_ind;
1865 
1866 	for (nreq = 0; wr; wr = wr->next) {
1867 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1868 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1869 					" %d max, %d nreq)\n", qp->qpn,
1870 					qp->rq.head, qp->rq.tail,
1871 					qp->rq.max, nreq);
1872 			err = -ENOMEM;
1873 			*bad_wr = wr;
1874 			goto out;
1875 		}
1876 
1877 		wqe = get_recv_wqe(qp, ind);
1878 		prev_wqe = qp->rq.last;
1879 		qp->rq.last = wqe;
1880 
1881 		((struct mthca_next_seg *) wqe)->ee_nds =
1882 			cpu_to_be32(MTHCA_NEXT_DBD);
1883 		((struct mthca_next_seg *) wqe)->flags = 0;
1884 
1885 		wqe += sizeof (struct mthca_next_seg);
1886 		size = sizeof (struct mthca_next_seg) / 16;
1887 
1888 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1889 			err = -EINVAL;
1890 			*bad_wr = wr;
1891 			goto out;
1892 		}
1893 
1894 		for (i = 0; i < wr->num_sge; ++i) {
1895 			mthca_set_data_seg(wqe, wr->sg_list + i);
1896 			wqe  += sizeof (struct mthca_data_seg);
1897 			size += sizeof (struct mthca_data_seg) / 16;
1898 		}
1899 
1900 		qp->wrid[ind] = wr->wr_id;
1901 
1902 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1903 			cpu_to_be32(MTHCA_NEXT_DBD | size);
1904 
1905 		if (!nreq)
1906 			size0 = size;
1907 
1908 		++ind;
1909 		if (unlikely(ind >= qp->rq.max))
1910 			ind -= qp->rq.max;
1911 
1912 		++nreq;
1913 		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1914 			nreq = 0;
1915 
1916 			wmb();
1917 
1918 			mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1919 				      qp->qpn << 8, dev->kar + MTHCA_RECEIVE_DOORBELL,
1920 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1921 
1922 			qp->rq.next_ind = ind;
1923 			qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1924 		}
1925 	}
1926 
1927 out:
1928 	if (likely(nreq)) {
1929 		wmb();
1930 
1931 		mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1932 			      qp->qpn << 8 | nreq, dev->kar + MTHCA_RECEIVE_DOORBELL,
1933 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1934 	}
1935 
1936 	qp->rq.next_ind = ind;
1937 	qp->rq.head    += nreq;
1938 
1939 	/*
1940 	 * Make sure doorbells don't leak out of RQ spinlock and reach
1941 	 * the HCA out of order:
1942 	 */
1943 	mmiowb();
1944 
1945 	spin_unlock_irqrestore(&qp->rq.lock, flags);
1946 	return err;
1947 }
1948 
1949 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1950 			  struct ib_send_wr **bad_wr)
1951 {
1952 	struct mthca_dev *dev = to_mdev(ibqp->device);
1953 	struct mthca_qp *qp = to_mqp(ibqp);
1954 	u32 dbhi;
1955 	void *wqe;
1956 	void *prev_wqe;
1957 	unsigned long flags;
1958 	int err = 0;
1959 	int nreq;
1960 	int i;
1961 	int size;
1962 	/*
1963 	 * f0 and size0 are only used if nreq != 0, and they will
1964 	 * always be initialized the first time through the main loop
1965 	 * before nreq is incremented.  So nreq cannot become non-zero
1966 	 * without initializing f0 and size0, and they are in fact
1967 	 * never used uninitialized.
1968 	 */
1969 	int uninitialized_var(size0);
1970 	u32 uninitialized_var(f0);
1971 	int ind;
1972 	u8 op0 = 0;
1973 
1974 	spin_lock_irqsave(&qp->sq.lock, flags);
1975 
1976 	/* XXX check that state is OK to post send */
1977 
1978 	ind = qp->sq.head & (qp->sq.max - 1);
1979 
1980 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1981 		if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1982 			nreq = 0;
1983 
1984 			dbhi = (MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1985 				((qp->sq.head & 0xffff) << 8) | f0 | op0;
1986 
1987 			qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1988 
1989 			/*
1990 			 * Make sure that descriptors are written before
1991 			 * doorbell record.
1992 			 */
1993 			wmb();
1994 			*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1995 
1996 			/*
1997 			 * Make sure doorbell record is written before we
1998 			 * write MMIO send doorbell.
1999 			 */
2000 			wmb();
2001 
2002 			mthca_write64(dbhi, (qp->qpn << 8) | size0,
2003 				      dev->kar + MTHCA_SEND_DOORBELL,
2004 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2005 		}
2006 
2007 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2008 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
2009 					" %d max, %d nreq)\n", qp->qpn,
2010 					qp->sq.head, qp->sq.tail,
2011 					qp->sq.max, nreq);
2012 			err = -ENOMEM;
2013 			*bad_wr = wr;
2014 			goto out;
2015 		}
2016 
2017 		wqe = get_send_wqe(qp, ind);
2018 		prev_wqe = qp->sq.last;
2019 		qp->sq.last = wqe;
2020 
2021 		((struct mthca_next_seg *) wqe)->flags =
2022 			((wr->send_flags & IB_SEND_SIGNALED) ?
2023 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
2024 			((wr->send_flags & IB_SEND_SOLICITED) ?
2025 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
2026 			((wr->send_flags & IB_SEND_IP_CSUM) ?
2027 			 cpu_to_be32(MTHCA_NEXT_IP_CSUM | MTHCA_NEXT_TCP_UDP_CSUM) : 0) |
2028 			cpu_to_be32(1);
2029 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
2030 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
2031 			((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
2032 
2033 		wqe += sizeof (struct mthca_next_seg);
2034 		size = sizeof (struct mthca_next_seg) / 16;
2035 
2036 		switch (qp->transport) {
2037 		case RC:
2038 			switch (wr->opcode) {
2039 			case IB_WR_ATOMIC_CMP_AND_SWP:
2040 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2041 				set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2042 					      wr->wr.atomic.rkey);
2043 				wqe += sizeof (struct mthca_raddr_seg);
2044 
2045 				set_atomic_seg(wqe, wr);
2046 				wqe  += sizeof (struct mthca_atomic_seg);
2047 				size += (sizeof (struct mthca_raddr_seg) +
2048 					 sizeof (struct mthca_atomic_seg)) / 16;
2049 				break;
2050 
2051 			case IB_WR_RDMA_READ:
2052 			case IB_WR_RDMA_WRITE:
2053 			case IB_WR_RDMA_WRITE_WITH_IMM:
2054 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2055 					      wr->wr.rdma.rkey);
2056 				wqe  += sizeof (struct mthca_raddr_seg);
2057 				size += sizeof (struct mthca_raddr_seg) / 16;
2058 				break;
2059 
2060 			default:
2061 				/* No extra segments required for sends */
2062 				break;
2063 			}
2064 
2065 			break;
2066 
2067 		case UC:
2068 			switch (wr->opcode) {
2069 			case IB_WR_RDMA_WRITE:
2070 			case IB_WR_RDMA_WRITE_WITH_IMM:
2071 				set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2072 					      wr->wr.rdma.rkey);
2073 				wqe  += sizeof (struct mthca_raddr_seg);
2074 				size += sizeof (struct mthca_raddr_seg) / 16;
2075 				break;
2076 
2077 			default:
2078 				/* No extra segments required for sends */
2079 				break;
2080 			}
2081 
2082 			break;
2083 
2084 		case UD:
2085 			set_arbel_ud_seg(wqe, wr);
2086 			wqe  += sizeof (struct mthca_arbel_ud_seg);
2087 			size += sizeof (struct mthca_arbel_ud_seg) / 16;
2088 			break;
2089 
2090 		case MLX:
2091 			err = build_mlx_header(dev, to_msqp(qp), ind, wr,
2092 					       wqe - sizeof (struct mthca_next_seg),
2093 					       wqe);
2094 			if (err) {
2095 				*bad_wr = wr;
2096 				goto out;
2097 			}
2098 			wqe += sizeof (struct mthca_data_seg);
2099 			size += sizeof (struct mthca_data_seg) / 16;
2100 			break;
2101 		}
2102 
2103 		if (wr->num_sge > qp->sq.max_gs) {
2104 			mthca_err(dev, "too many gathers\n");
2105 			err = -EINVAL;
2106 			*bad_wr = wr;
2107 			goto out;
2108 		}
2109 
2110 		for (i = 0; i < wr->num_sge; ++i) {
2111 			mthca_set_data_seg(wqe, wr->sg_list + i);
2112 			wqe  += sizeof (struct mthca_data_seg);
2113 			size += sizeof (struct mthca_data_seg) / 16;
2114 		}
2115 
2116 		/* Add one more inline data segment for ICRC */
2117 		if (qp->transport == MLX) {
2118 			((struct mthca_data_seg *) wqe)->byte_count =
2119 				cpu_to_be32((1 << 31) | 4);
2120 			((u32 *) wqe)[1] = 0;
2121 			wqe += sizeof (struct mthca_data_seg);
2122 			size += sizeof (struct mthca_data_seg) / 16;
2123 		}
2124 
2125 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
2126 
2127 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2128 			mthca_err(dev, "opcode invalid\n");
2129 			err = -EINVAL;
2130 			*bad_wr = wr;
2131 			goto out;
2132 		}
2133 
2134 		((struct mthca_next_seg *) prev_wqe)->nda_op =
2135 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
2136 				     qp->send_wqe_offset) |
2137 				    mthca_opcode[wr->opcode]);
2138 		wmb();
2139 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
2140 			cpu_to_be32(MTHCA_NEXT_DBD | size |
2141 				    ((wr->send_flags & IB_SEND_FENCE) ?
2142 				     MTHCA_NEXT_FENCE : 0));
2143 
2144 		if (!nreq) {
2145 			size0 = size;
2146 			op0   = mthca_opcode[wr->opcode];
2147 			f0    = wr->send_flags & IB_SEND_FENCE ?
2148 				MTHCA_SEND_DOORBELL_FENCE : 0;
2149 		}
2150 
2151 		++ind;
2152 		if (unlikely(ind >= qp->sq.max))
2153 			ind -= qp->sq.max;
2154 	}
2155 
2156 out:
2157 	if (likely(nreq)) {
2158 		dbhi = (nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0;
2159 
2160 		qp->sq.head += nreq;
2161 
2162 		/*
2163 		 * Make sure that descriptors are written before
2164 		 * doorbell record.
2165 		 */
2166 		wmb();
2167 		*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2168 
2169 		/*
2170 		 * Make sure doorbell record is written before we
2171 		 * write MMIO send doorbell.
2172 		 */
2173 		wmb();
2174 
2175 		mthca_write64(dbhi, (qp->qpn << 8) | size0, dev->kar + MTHCA_SEND_DOORBELL,
2176 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2177 	}
2178 
2179 	/*
2180 	 * Make sure doorbells don't leak out of SQ spinlock and reach
2181 	 * the HCA out of order:
2182 	 */
2183 	mmiowb();
2184 
2185 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2186 	return err;
2187 }
2188 
2189 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2190 			     struct ib_recv_wr **bad_wr)
2191 {
2192 	struct mthca_dev *dev = to_mdev(ibqp->device);
2193 	struct mthca_qp *qp = to_mqp(ibqp);
2194 	unsigned long flags;
2195 	int err = 0;
2196 	int nreq;
2197 	int ind;
2198 	int i;
2199 	void *wqe;
2200 
2201 	spin_lock_irqsave(&qp->rq.lock, flags);
2202 
2203 	/* XXX check that state is OK to post receive */
2204 
2205 	ind = qp->rq.head & (qp->rq.max - 1);
2206 
2207 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2208 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2209 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2210 					" %d max, %d nreq)\n", qp->qpn,
2211 					qp->rq.head, qp->rq.tail,
2212 					qp->rq.max, nreq);
2213 			err = -ENOMEM;
2214 			*bad_wr = wr;
2215 			goto out;
2216 		}
2217 
2218 		wqe = get_recv_wqe(qp, ind);
2219 
2220 		((struct mthca_next_seg *) wqe)->flags = 0;
2221 
2222 		wqe += sizeof (struct mthca_next_seg);
2223 
2224 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2225 			err = -EINVAL;
2226 			*bad_wr = wr;
2227 			goto out;
2228 		}
2229 
2230 		for (i = 0; i < wr->num_sge; ++i) {
2231 			mthca_set_data_seg(wqe, wr->sg_list + i);
2232 			wqe += sizeof (struct mthca_data_seg);
2233 		}
2234 
2235 		if (i < qp->rq.max_gs)
2236 			mthca_set_data_seg_inval(wqe);
2237 
2238 		qp->wrid[ind] = wr->wr_id;
2239 
2240 		++ind;
2241 		if (unlikely(ind >= qp->rq.max))
2242 			ind -= qp->rq.max;
2243 	}
2244 out:
2245 	if (likely(nreq)) {
2246 		qp->rq.head += nreq;
2247 
2248 		/*
2249 		 * Make sure that descriptors are written before
2250 		 * doorbell record.
2251 		 */
2252 		wmb();
2253 		*qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2254 	}
2255 
2256 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2257 	return err;
2258 }
2259 
2260 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2261 			int index, int *dbd, __be32 *new_wqe)
2262 {
2263 	struct mthca_next_seg *next;
2264 
2265 	/*
2266 	 * For SRQs, all receive WQEs generate a CQE, so we're always
2267 	 * at the end of the doorbell chain.
2268 	 */
2269 	if (qp->ibqp.srq && !is_send) {
2270 		*new_wqe = 0;
2271 		return;
2272 	}
2273 
2274 	if (is_send)
2275 		next = get_send_wqe(qp, index);
2276 	else
2277 		next = get_recv_wqe(qp, index);
2278 
2279 	*dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2280 	if (next->ee_nds & cpu_to_be32(0x3f))
2281 		*new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2282 			(next->ee_nds & cpu_to_be32(0x3f));
2283 	else
2284 		*new_wqe = 0;
2285 }
2286 
2287 int mthca_init_qp_table(struct mthca_dev *dev)
2288 {
2289 	int err;
2290 	u8 status;
2291 	int i;
2292 
2293 	spin_lock_init(&dev->qp_table.lock);
2294 
2295 	/*
2296 	 * We reserve 2 extra QPs per port for the special QPs.  The
2297 	 * special QP for port 1 has to be even, so round up.
2298 	 */
2299 	dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2300 	err = mthca_alloc_init(&dev->qp_table.alloc,
2301 			       dev->limits.num_qps,
2302 			       (1 << 24) - 1,
2303 			       dev->qp_table.sqp_start +
2304 			       MTHCA_MAX_PORTS * 2);
2305 	if (err)
2306 		return err;
2307 
2308 	err = mthca_array_init(&dev->qp_table.qp,
2309 			       dev->limits.num_qps);
2310 	if (err) {
2311 		mthca_alloc_cleanup(&dev->qp_table.alloc);
2312 		return err;
2313 	}
2314 
2315 	for (i = 0; i < 2; ++i) {
2316 		err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2317 					    dev->qp_table.sqp_start + i * 2,
2318 					    &status);
2319 		if (err)
2320 			goto err_out;
2321 		if (status) {
2322 			mthca_warn(dev, "CONF_SPECIAL_QP returned "
2323 				   "status %02x, aborting.\n",
2324 				   status);
2325 			err = -EINVAL;
2326 			goto err_out;
2327 		}
2328 	}
2329 	return 0;
2330 
2331  err_out:
2332 	for (i = 0; i < 2; ++i)
2333 		mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2334 
2335 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2336 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2337 
2338 	return err;
2339 }
2340 
2341 void mthca_cleanup_qp_table(struct mthca_dev *dev)
2342 {
2343 	int i;
2344 	u8 status;
2345 
2346 	for (i = 0; i < 2; ++i)
2347 		mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2348 
2349 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2350 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2351 }
2352