1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Cisco Systems. All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * OpenIB.org BSD license below:
12  *
13  *     Redistribution and use in source and binary forms, with or
14  *     without modification, are permitted provided that the following
15  *     conditions are met:
16  *
17  *      - Redistributions of source code must retain the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer.
20  *
21  *      - Redistributions in binary form must reproduce the above
22  *        copyright notice, this list of conditions and the following
23  *        disclaimer in the documentation and/or other materials
24  *        provided with the distribution.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33  * SOFTWARE.
34  */
35 
36 #include <linux/string.h>
37 #include <linux/slab.h>
38 #include <linux/sched.h>
39 
40 #include <asm/io.h>
41 
42 #include <rdma/ib_verbs.h>
43 #include <rdma/ib_cache.h>
44 #include <rdma/ib_pack.h>
45 #include <rdma/uverbs_ioctl.h>
46 
47 #include "mthca_dev.h"
48 #include "mthca_cmd.h"
49 #include "mthca_memfree.h"
50 #include "mthca_wqe.h"
51 
52 enum {
53 	MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
54 	MTHCA_ACK_REQ_FREQ       = 10,
55 	MTHCA_FLIGHT_LIMIT       = 9,
56 	MTHCA_UD_HEADER_SIZE     = 72, /* largest UD header possible */
57 	MTHCA_INLINE_HEADER_SIZE = 4,  /* data segment overhead for inline */
58 	MTHCA_INLINE_CHUNK_SIZE  = 16  /* inline data segment chunk */
59 };
60 
61 enum {
62 	MTHCA_QP_STATE_RST  = 0,
63 	MTHCA_QP_STATE_INIT = 1,
64 	MTHCA_QP_STATE_RTR  = 2,
65 	MTHCA_QP_STATE_RTS  = 3,
66 	MTHCA_QP_STATE_SQE  = 4,
67 	MTHCA_QP_STATE_SQD  = 5,
68 	MTHCA_QP_STATE_ERR  = 6,
69 	MTHCA_QP_STATE_DRAINING = 7
70 };
71 
72 enum {
73 	MTHCA_QP_ST_RC 	= 0x0,
74 	MTHCA_QP_ST_UC 	= 0x1,
75 	MTHCA_QP_ST_RD 	= 0x2,
76 	MTHCA_QP_ST_UD 	= 0x3,
77 	MTHCA_QP_ST_MLX = 0x7
78 };
79 
80 enum {
81 	MTHCA_QP_PM_MIGRATED = 0x3,
82 	MTHCA_QP_PM_ARMED    = 0x0,
83 	MTHCA_QP_PM_REARM    = 0x1
84 };
85 
86 enum {
87 	/* qp_context flags */
88 	MTHCA_QP_BIT_DE  = 1 <<  8,
89 	/* params1 */
90 	MTHCA_QP_BIT_SRE = 1 << 15,
91 	MTHCA_QP_BIT_SWE = 1 << 14,
92 	MTHCA_QP_BIT_SAE = 1 << 13,
93 	MTHCA_QP_BIT_SIC = 1 <<  4,
94 	MTHCA_QP_BIT_SSC = 1 <<  3,
95 	/* params2 */
96 	MTHCA_QP_BIT_RRE = 1 << 15,
97 	MTHCA_QP_BIT_RWE = 1 << 14,
98 	MTHCA_QP_BIT_RAE = 1 << 13,
99 	MTHCA_QP_BIT_RIC = 1 <<  4,
100 	MTHCA_QP_BIT_RSC = 1 <<  3
101 };
102 
103 enum {
104 	MTHCA_SEND_DOORBELL_FENCE = 1 << 5
105 };
106 
107 struct mthca_qp_path {
108 	__be32 port_pkey;
109 	u8     rnr_retry;
110 	u8     g_mylmc;
111 	__be16 rlid;
112 	u8     ackto;
113 	u8     mgid_index;
114 	u8     static_rate;
115 	u8     hop_limit;
116 	__be32 sl_tclass_flowlabel;
117 	u8     rgid[16];
118 } __packed;
119 
120 struct mthca_qp_context {
121 	__be32 flags;
122 	__be32 tavor_sched_queue; /* Reserved on Arbel */
123 	u8     mtu_msgmax;
124 	u8     rq_size_stride;	/* Reserved on Tavor */
125 	u8     sq_size_stride;	/* Reserved on Tavor */
126 	u8     rlkey_arbel_sched_queue;	/* Reserved on Tavor */
127 	__be32 usr_page;
128 	__be32 local_qpn;
129 	__be32 remote_qpn;
130 	u32    reserved1[2];
131 	struct mthca_qp_path pri_path;
132 	struct mthca_qp_path alt_path;
133 	__be32 rdd;
134 	__be32 pd;
135 	__be32 wqe_base;
136 	__be32 wqe_lkey;
137 	__be32 params1;
138 	__be32 reserved2;
139 	__be32 next_send_psn;
140 	__be32 cqn_snd;
141 	__be32 snd_wqe_base_l;	/* Next send WQE on Tavor */
142 	__be32 snd_db_index;	/* (debugging only entries) */
143 	__be32 last_acked_psn;
144 	__be32 ssn;
145 	__be32 params2;
146 	__be32 rnr_nextrecvpsn;
147 	__be32 ra_buff_indx;
148 	__be32 cqn_rcv;
149 	__be32 rcv_wqe_base_l;	/* Next recv WQE on Tavor */
150 	__be32 rcv_db_index;	/* (debugging only entries) */
151 	__be32 qkey;
152 	__be32 srqn;
153 	__be32 rmsn;
154 	__be16 rq_wqe_counter;	/* reserved on Tavor */
155 	__be16 sq_wqe_counter;	/* reserved on Tavor */
156 	u32    reserved3[18];
157 } __packed;
158 
159 struct mthca_qp_param {
160 	__be32 opt_param_mask;
161 	u32    reserved1;
162 	struct mthca_qp_context context;
163 	u32    reserved2[62];
164 } __packed;
165 
166 enum {
167 	MTHCA_QP_OPTPAR_ALT_ADDR_PATH     = 1 << 0,
168 	MTHCA_QP_OPTPAR_RRE               = 1 << 1,
169 	MTHCA_QP_OPTPAR_RAE               = 1 << 2,
170 	MTHCA_QP_OPTPAR_RWE               = 1 << 3,
171 	MTHCA_QP_OPTPAR_PKEY_INDEX        = 1 << 4,
172 	MTHCA_QP_OPTPAR_Q_KEY             = 1 << 5,
173 	MTHCA_QP_OPTPAR_RNR_TIMEOUT       = 1 << 6,
174 	MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
175 	MTHCA_QP_OPTPAR_SRA_MAX           = 1 << 8,
176 	MTHCA_QP_OPTPAR_RRA_MAX           = 1 << 9,
177 	MTHCA_QP_OPTPAR_PM_STATE          = 1 << 10,
178 	MTHCA_QP_OPTPAR_PORT_NUM          = 1 << 11,
179 	MTHCA_QP_OPTPAR_RETRY_COUNT       = 1 << 12,
180 	MTHCA_QP_OPTPAR_ALT_RNR_RETRY     = 1 << 13,
181 	MTHCA_QP_OPTPAR_ACK_TIMEOUT       = 1 << 14,
182 	MTHCA_QP_OPTPAR_RNR_RETRY         = 1 << 15,
183 	MTHCA_QP_OPTPAR_SCHED_QUEUE       = 1 << 16
184 };
185 
186 static const u8 mthca_opcode[] = {
187 	[IB_WR_SEND]                 = MTHCA_OPCODE_SEND,
188 	[IB_WR_SEND_WITH_IMM]        = MTHCA_OPCODE_SEND_IMM,
189 	[IB_WR_RDMA_WRITE]           = MTHCA_OPCODE_RDMA_WRITE,
190 	[IB_WR_RDMA_WRITE_WITH_IMM]  = MTHCA_OPCODE_RDMA_WRITE_IMM,
191 	[IB_WR_RDMA_READ]            = MTHCA_OPCODE_RDMA_READ,
192 	[IB_WR_ATOMIC_CMP_AND_SWP]   = MTHCA_OPCODE_ATOMIC_CS,
193 	[IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
194 };
195 
196 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
197 {
198 	return qp->qpn >= dev->qp_table.sqp_start &&
199 		qp->qpn <= dev->qp_table.sqp_start + 3;
200 }
201 
202 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
203 {
204 	return qp->qpn >= dev->qp_table.sqp_start &&
205 		qp->qpn <= dev->qp_table.sqp_start + 1;
206 }
207 
208 static void *get_recv_wqe(struct mthca_qp *qp, int n)
209 {
210 	if (qp->is_direct)
211 		return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
212 	else
213 		return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
214 			((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
215 }
216 
217 static void *get_send_wqe(struct mthca_qp *qp, int n)
218 {
219 	if (qp->is_direct)
220 		return qp->queue.direct.buf + qp->send_wqe_offset +
221 			(n << qp->sq.wqe_shift);
222 	else
223 		return qp->queue.page_list[(qp->send_wqe_offset +
224 					    (n << qp->sq.wqe_shift)) >>
225 					   PAGE_SHIFT].buf +
226 			((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
227 			 (PAGE_SIZE - 1));
228 }
229 
230 static void mthca_wq_reset(struct mthca_wq *wq)
231 {
232 	wq->next_ind  = 0;
233 	wq->last_comp = wq->max - 1;
234 	wq->head      = 0;
235 	wq->tail      = 0;
236 }
237 
238 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
239 		    enum ib_event_type event_type)
240 {
241 	struct mthca_qp *qp;
242 	struct ib_event event;
243 
244 	spin_lock(&dev->qp_table.lock);
245 	qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
246 	if (qp)
247 		++qp->refcount;
248 	spin_unlock(&dev->qp_table.lock);
249 
250 	if (!qp) {
251 		mthca_warn(dev, "Async event %d for bogus QP %08x\n",
252 			   event_type, 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 	struct mthca_init_ib_param param;
314 
315 	memset(&param, 0, sizeof param);
316 
317 	param.port_width = dev->limits.port_width_cap;
318 	param.vl_cap     = dev->limits.vl_cap;
319 	param.mtu_cap    = dev->limits.mtu_cap;
320 	param.gid_cap    = dev->limits.gid_table_len;
321 	param.pkey_cap   = dev->limits.pkey_table_len;
322 
323 	err = mthca_INIT_IB(dev, &param, port);
324 	if (err)
325 		mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
326 }
327 
328 static __be32 get_hw_access_flags(struct mthca_qp *qp, const struct ib_qp_attr *attr,
329 				  int attr_mask)
330 {
331 	u8 dest_rd_atomic;
332 	u32 access_flags;
333 	u32 hw_access_flags = 0;
334 
335 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
336 		dest_rd_atomic = attr->max_dest_rd_atomic;
337 	else
338 		dest_rd_atomic = qp->resp_depth;
339 
340 	if (attr_mask & IB_QP_ACCESS_FLAGS)
341 		access_flags = attr->qp_access_flags;
342 	else
343 		access_flags = qp->atomic_rd_en;
344 
345 	if (!dest_rd_atomic)
346 		access_flags &= IB_ACCESS_REMOTE_WRITE;
347 
348 	if (access_flags & IB_ACCESS_REMOTE_READ)
349 		hw_access_flags |= MTHCA_QP_BIT_RRE;
350 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
351 		hw_access_flags |= MTHCA_QP_BIT_RAE;
352 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
353 		hw_access_flags |= MTHCA_QP_BIT_RWE;
354 
355 	return cpu_to_be32(hw_access_flags);
356 }
357 
358 static inline enum ib_qp_state to_ib_qp_state(int mthca_state)
359 {
360 	switch (mthca_state) {
361 	case MTHCA_QP_STATE_RST:      return IB_QPS_RESET;
362 	case MTHCA_QP_STATE_INIT:     return IB_QPS_INIT;
363 	case MTHCA_QP_STATE_RTR:      return IB_QPS_RTR;
364 	case MTHCA_QP_STATE_RTS:      return IB_QPS_RTS;
365 	case MTHCA_QP_STATE_DRAINING:
366 	case MTHCA_QP_STATE_SQD:      return IB_QPS_SQD;
367 	case MTHCA_QP_STATE_SQE:      return IB_QPS_SQE;
368 	case MTHCA_QP_STATE_ERR:      return IB_QPS_ERR;
369 	default:                      return -1;
370 	}
371 }
372 
373 static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state)
374 {
375 	switch (mthca_mig_state) {
376 	case 0:  return IB_MIG_ARMED;
377 	case 1:  return IB_MIG_REARM;
378 	case 3:  return IB_MIG_MIGRATED;
379 	default: return -1;
380 	}
381 }
382 
383 static int to_ib_qp_access_flags(int mthca_flags)
384 {
385 	int ib_flags = 0;
386 
387 	if (mthca_flags & MTHCA_QP_BIT_RRE)
388 		ib_flags |= IB_ACCESS_REMOTE_READ;
389 	if (mthca_flags & MTHCA_QP_BIT_RWE)
390 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
391 	if (mthca_flags & MTHCA_QP_BIT_RAE)
392 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
393 
394 	return ib_flags;
395 }
396 
397 static void to_rdma_ah_attr(struct mthca_dev *dev,
398 			    struct rdma_ah_attr *ah_attr,
399 			    struct mthca_qp_path *path)
400 {
401 	u8 port_num = (be32_to_cpu(path->port_pkey) >> 24) & 0x3;
402 
403 	memset(ah_attr, 0, sizeof(*ah_attr));
404 
405 	if (port_num == 0 || port_num > dev->limits.num_ports)
406 		return;
407 	ah_attr->type = rdma_ah_find_type(&dev->ib_dev, port_num);
408 	rdma_ah_set_port_num(ah_attr, port_num);
409 
410 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
411 	rdma_ah_set_sl(ah_attr, be32_to_cpu(path->sl_tclass_flowlabel) >> 28);
412 	rdma_ah_set_path_bits(ah_attr, path->g_mylmc & 0x7f);
413 	rdma_ah_set_static_rate(ah_attr,
414 				mthca_rate_to_ib(dev,
415 						 path->static_rate & 0xf,
416 						 port_num));
417 	if (path->g_mylmc & (1 << 7)) {
418 		u32 tc_fl = be32_to_cpu(path->sl_tclass_flowlabel);
419 
420 		rdma_ah_set_grh(ah_attr, NULL,
421 				tc_fl & 0xfffff,
422 				path->mgid_index &
423 				(dev->limits.gid_table_len - 1),
424 				path->hop_limit,
425 				(tc_fl >> 20) & 0xff);
426 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
427 	}
428 }
429 
430 int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
431 		   struct ib_qp_init_attr *qp_init_attr)
432 {
433 	struct mthca_dev *dev = to_mdev(ibqp->device);
434 	struct mthca_qp *qp = to_mqp(ibqp);
435 	int err = 0;
436 	struct mthca_mailbox *mailbox = NULL;
437 	struct mthca_qp_param *qp_param;
438 	struct mthca_qp_context *context;
439 	int mthca_state;
440 
441 	mutex_lock(&qp->mutex);
442 
443 	if (qp->state == IB_QPS_RESET) {
444 		qp_attr->qp_state = IB_QPS_RESET;
445 		goto done;
446 	}
447 
448 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
449 	if (IS_ERR(mailbox)) {
450 		err = PTR_ERR(mailbox);
451 		goto out;
452 	}
453 
454 	err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox);
455 	if (err) {
456 		mthca_warn(dev, "QUERY_QP failed (%d)\n", err);
457 		goto out_mailbox;
458 	}
459 
460 	qp_param    = mailbox->buf;
461 	context     = &qp_param->context;
462 	mthca_state = be32_to_cpu(context->flags) >> 28;
463 
464 	qp->state		     = to_ib_qp_state(mthca_state);
465 	qp_attr->qp_state	     = qp->state;
466 	qp_attr->path_mtu 	     = context->mtu_msgmax >> 5;
467 	qp_attr->path_mig_state      =
468 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
469 	qp_attr->qkey 		     = be32_to_cpu(context->qkey);
470 	qp_attr->rq_psn 	     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
471 	qp_attr->sq_psn 	     = be32_to_cpu(context->next_send_psn) & 0xffffff;
472 	qp_attr->dest_qp_num 	     = be32_to_cpu(context->remote_qpn) & 0xffffff;
473 	qp_attr->qp_access_flags     =
474 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
475 
476 	if (qp->transport == RC || qp->transport == UC) {
477 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
478 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
479 		qp_attr->alt_pkey_index =
480 			be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
481 		qp_attr->alt_port_num	=
482 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
483 	}
484 
485 	qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
486 	qp_attr->port_num   =
487 		(be32_to_cpu(context->pri_path.port_pkey) >> 24) & 0x3;
488 
489 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
490 	qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
491 
492 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
493 
494 	qp_attr->max_dest_rd_atomic =
495 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
496 	qp_attr->min_rnr_timer 	    =
497 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
498 	qp_attr->timeout 	    = context->pri_path.ackto >> 3;
499 	qp_attr->retry_cnt 	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
500 	qp_attr->rnr_retry 	    = context->pri_path.rnr_retry >> 5;
501 	qp_attr->alt_timeout 	    = context->alt_path.ackto >> 3;
502 
503 done:
504 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
505 	qp_attr->cap.max_send_wr     = qp->sq.max;
506 	qp_attr->cap.max_recv_wr     = qp->rq.max;
507 	qp_attr->cap.max_send_sge    = qp->sq.max_gs;
508 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
509 	qp_attr->cap.max_inline_data = qp->max_inline_data;
510 
511 	qp_init_attr->cap	     = qp_attr->cap;
512 	qp_init_attr->sq_sig_type    = qp->sq_policy;
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 rdma_ah_attr *ah,
523 			  struct mthca_qp_path *path, u8 port)
524 {
525 	path->g_mylmc     = rdma_ah_get_path_bits(ah) & 0x7f;
526 	path->rlid        = cpu_to_be16(rdma_ah_get_dlid(ah));
527 	path->static_rate = mthca_get_rate(dev, rdma_ah_get_static_rate(ah),
528 					   port);
529 
530 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
531 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
532 
533 		if (grh->sgid_index >= dev->limits.gid_table_len) {
534 			mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
535 				  grh->sgid_index,
536 				  dev->limits.gid_table_len - 1);
537 			return -1;
538 		}
539 
540 		path->g_mylmc   |= 1 << 7;
541 		path->mgid_index = grh->sgid_index;
542 		path->hop_limit  = grh->hop_limit;
543 		path->sl_tclass_flowlabel =
544 			cpu_to_be32((rdma_ah_get_sl(ah) << 28) |
545 				    (grh->traffic_class << 20) |
546 				    (grh->flow_label));
547 		memcpy(path->rgid, grh->dgid.raw, 16);
548 	} else {
549 		path->sl_tclass_flowlabel = cpu_to_be32(rdma_ah_get_sl(ah) <<
550 							28);
551 	}
552 
553 	return 0;
554 }
555 
556 static int __mthca_modify_qp(struct ib_qp *ibqp,
557 			     const struct ib_qp_attr *attr, int attr_mask,
558 			     enum ib_qp_state cur_state,
559 			     enum ib_qp_state new_state,
560 			     struct ib_udata *udata)
561 {
562 	struct mthca_dev *dev = to_mdev(ibqp->device);
563 	struct mthca_qp *qp = to_mqp(ibqp);
564 	struct mthca_ucontext *context = rdma_udata_to_drv_context(
565 		udata, struct mthca_ucontext, ibucontext);
566 	struct mthca_mailbox *mailbox;
567 	struct mthca_qp_param *qp_param;
568 	struct mthca_qp_context *qp_context;
569 	u32 sqd_event = 0;
570 	int err = -EINVAL;
571 
572 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
573 	if (IS_ERR(mailbox)) {
574 		err = PTR_ERR(mailbox);
575 		goto out;
576 	}
577 	qp_param = mailbox->buf;
578 	qp_context = &qp_param->context;
579 	memset(qp_param, 0, sizeof *qp_param);
580 
581 	qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
582 					     (to_mthca_st(qp->transport) << 16));
583 	qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
584 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
585 		qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
586 	else {
587 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
588 		switch (attr->path_mig_state) {
589 		case IB_MIG_MIGRATED:
590 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
591 			break;
592 		case IB_MIG_REARM:
593 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
594 			break;
595 		case IB_MIG_ARMED:
596 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
597 			break;
598 		}
599 	}
600 
601 	/* leave tavor_sched_queue as 0 */
602 
603 	if (qp->transport == MLX || qp->transport == UD)
604 		qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
605 	else if (attr_mask & IB_QP_PATH_MTU) {
606 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
607 			mthca_dbg(dev, "path MTU (%u) is invalid\n",
608 				  attr->path_mtu);
609 			goto out_mailbox;
610 		}
611 		qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
612 	}
613 
614 	if (mthca_is_memfree(dev)) {
615 		if (qp->rq.max)
616 			qp_context->rq_size_stride = ilog2(qp->rq.max) << 3;
617 		qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
618 
619 		if (qp->sq.max)
620 			qp_context->sq_size_stride = ilog2(qp->sq.max) << 3;
621 		qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
622 	}
623 
624 	/* leave arbel_sched_queue as 0 */
625 
626 	if (qp->ibqp.uobject)
627 		qp_context->usr_page = cpu_to_be32(context->uar.index);
628 	else
629 		qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
630 	qp_context->local_qpn  = cpu_to_be32(qp->qpn);
631 	if (attr_mask & IB_QP_DEST_QPN) {
632 		qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
633 	}
634 
635 	if (qp->transport == MLX)
636 		qp_context->pri_path.port_pkey |=
637 			cpu_to_be32(qp->port << 24);
638 	else {
639 		if (attr_mask & IB_QP_PORT) {
640 			qp_context->pri_path.port_pkey |=
641 				cpu_to_be32(attr->port_num << 24);
642 			qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
643 		}
644 	}
645 
646 	if (attr_mask & IB_QP_PKEY_INDEX) {
647 		qp_context->pri_path.port_pkey |=
648 			cpu_to_be32(attr->pkey_index);
649 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
650 	}
651 
652 	if (attr_mask & IB_QP_RNR_RETRY) {
653 		qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
654 			attr->rnr_retry << 5;
655 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
656 							MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
657 	}
658 
659 	if (attr_mask & IB_QP_AV) {
660 		if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
661 				   attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
662 			goto out_mailbox;
663 
664 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
665 	}
666 
667 	if (ibqp->qp_type == IB_QPT_RC &&
668 	    cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
669 		u8 sched_queue = ibqp->uobject ? 0x2 : 0x1;
670 
671 		if (mthca_is_memfree(dev))
672 			qp_context->rlkey_arbel_sched_queue |= sched_queue;
673 		else
674 			qp_context->tavor_sched_queue |= cpu_to_be32(sched_queue);
675 
676 		qp_param->opt_param_mask |=
677 			cpu_to_be32(MTHCA_QP_OPTPAR_SCHED_QUEUE);
678 	}
679 
680 	if (attr_mask & IB_QP_TIMEOUT) {
681 		qp_context->pri_path.ackto = attr->timeout << 3;
682 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
683 	}
684 
685 	if (attr_mask & IB_QP_ALT_PATH) {
686 		if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
687 			mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
688 				  attr->alt_pkey_index, dev->limits.pkey_table_len-1);
689 			goto out_mailbox;
690 		}
691 
692 		if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
693 			mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
694 				attr->alt_port_num);
695 			goto out_mailbox;
696 		}
697 
698 		if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
699 				   rdma_ah_get_port_num(&attr->alt_ah_attr)))
700 			goto out_mailbox;
701 
702 		qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
703 							      attr->alt_port_num << 24);
704 		qp_context->alt_path.ackto = attr->alt_timeout << 3;
705 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
706 	}
707 
708 	/* leave rdd as 0 */
709 	qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
710 	/* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
711 	qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
712 	qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
713 					     (MTHCA_FLIGHT_LIMIT << 24) |
714 					     MTHCA_QP_BIT_SWE);
715 	if (qp->sq_policy == IB_SIGNAL_ALL_WR)
716 		qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
717 	if (attr_mask & IB_QP_RETRY_CNT) {
718 		qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
719 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
720 	}
721 
722 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
723 		if (attr->max_rd_atomic) {
724 			qp_context->params1 |=
725 				cpu_to_be32(MTHCA_QP_BIT_SRE |
726 					    MTHCA_QP_BIT_SAE);
727 			qp_context->params1 |=
728 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
729 		}
730 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
731 	}
732 
733 	if (attr_mask & IB_QP_SQ_PSN)
734 		qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
735 	qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
736 
737 	if (mthca_is_memfree(dev)) {
738 		qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
739 		qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
740 	}
741 
742 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
743 		if (attr->max_dest_rd_atomic)
744 			qp_context->params2 |=
745 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
746 
747 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
748 	}
749 
750 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
751 		qp_context->params2      |= get_hw_access_flags(qp, attr, attr_mask);
752 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
753 							MTHCA_QP_OPTPAR_RRE |
754 							MTHCA_QP_OPTPAR_RAE);
755 	}
756 
757 	qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
758 
759 	if (ibqp->srq)
760 		qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
761 
762 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
763 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
764 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
765 	}
766 	if (attr_mask & IB_QP_RQ_PSN)
767 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
768 
769 	qp_context->ra_buff_indx =
770 		cpu_to_be32(dev->qp_table.rdb_base +
771 			    ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
772 			     dev->qp_table.rdb_shift));
773 
774 	qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
775 
776 	if (mthca_is_memfree(dev))
777 		qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
778 
779 	if (attr_mask & IB_QP_QKEY) {
780 		qp_context->qkey = cpu_to_be32(attr->qkey);
781 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
782 	}
783 
784 	if (ibqp->srq)
785 		qp_context->srqn = cpu_to_be32(1 << 24 |
786 					       to_msrq(ibqp->srq)->srqn);
787 
788 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
789 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY		&&
790 	    attr->en_sqd_async_notify)
791 		sqd_event = 1 << 31;
792 
793 	err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
794 			      mailbox, sqd_event);
795 	if (err) {
796 		mthca_warn(dev, "modify QP %d->%d returned %d.\n",
797 			   cur_state, new_state, err);
798 		goto out_mailbox;
799 	}
800 
801 	qp->state = new_state;
802 	if (attr_mask & IB_QP_ACCESS_FLAGS)
803 		qp->atomic_rd_en = attr->qp_access_flags;
804 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
805 		qp->resp_depth = attr->max_dest_rd_atomic;
806 	if (attr_mask & IB_QP_PORT)
807 		qp->port = attr->port_num;
808 	if (attr_mask & IB_QP_ALT_PATH)
809 		qp->alt_port = attr->alt_port_num;
810 
811 	if (is_sqp(dev, qp))
812 		store_attrs(qp->sqp, attr, attr_mask);
813 
814 	/*
815 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
816 	 * QP0 to RESET or ERROR, bring the link back down.
817 	 */
818 	if (is_qp0(dev, qp)) {
819 		if (cur_state != IB_QPS_RTR &&
820 		    new_state == IB_QPS_RTR)
821 			init_port(dev, qp->port);
822 
823 		if (cur_state != IB_QPS_RESET &&
824 		    cur_state != IB_QPS_ERR &&
825 		    (new_state == IB_QPS_RESET ||
826 		     new_state == IB_QPS_ERR))
827 			mthca_CLOSE_IB(dev, qp->port);
828 	}
829 
830 	/*
831 	 * If we moved a kernel QP to RESET, clean up all old CQ
832 	 * entries and reinitialize the QP.
833 	 */
834 	if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
835 		mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
836 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
837 		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
838 			mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn, NULL);
839 
840 		mthca_wq_reset(&qp->sq);
841 		qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
842 
843 		mthca_wq_reset(&qp->rq);
844 		qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
845 
846 		if (mthca_is_memfree(dev)) {
847 			*qp->sq.db = 0;
848 			*qp->rq.db = 0;
849 		}
850 	}
851 
852 out_mailbox:
853 	mthca_free_mailbox(dev, mailbox);
854 out:
855 	return err;
856 }
857 
858 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
859 		    struct ib_udata *udata)
860 {
861 	struct mthca_dev *dev = to_mdev(ibqp->device);
862 	struct mthca_qp *qp = to_mqp(ibqp);
863 	enum ib_qp_state cur_state, new_state;
864 	int err = -EINVAL;
865 
866 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
867 		return -EOPNOTSUPP;
868 
869 	mutex_lock(&qp->mutex);
870 	if (attr_mask & IB_QP_CUR_STATE) {
871 		cur_state = attr->cur_qp_state;
872 	} else {
873 		spin_lock_irq(&qp->sq.lock);
874 		spin_lock(&qp->rq.lock);
875 		cur_state = qp->state;
876 		spin_unlock(&qp->rq.lock);
877 		spin_unlock_irq(&qp->sq.lock);
878 	}
879 
880 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
881 
882 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
883 				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 	err = __mthca_modify_qp(ibqp, attr, attr_mask, cur_state, new_state,
924 				udata);
925 
926 out:
927 	mutex_unlock(&qp->mutex);
928 	return err;
929 }
930 
931 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
932 {
933 	/*
934 	 * Calculate the maximum size of WQE s/g segments, excluding
935 	 * the next segment and other non-data segments.
936 	 */
937 	int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
938 
939 	switch (qp->transport) {
940 	case MLX:
941 		max_data_size -= 2 * sizeof (struct mthca_data_seg);
942 		break;
943 
944 	case UD:
945 		if (mthca_is_memfree(dev))
946 			max_data_size -= sizeof (struct mthca_arbel_ud_seg);
947 		else
948 			max_data_size -= sizeof (struct mthca_tavor_ud_seg);
949 		break;
950 
951 	default:
952 		max_data_size -= sizeof (struct mthca_raddr_seg);
953 		break;
954 	}
955 
956 	return max_data_size;
957 }
958 
959 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
960 {
961 	/* We don't support inline data for kernel QPs (yet). */
962 	return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
963 }
964 
965 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
966 				 struct mthca_pd *pd,
967 				 struct mthca_qp *qp)
968 {
969 	int max_data_size = mthca_max_data_size(dev, qp,
970 						min(dev->limits.max_desc_sz,
971 						    1 << qp->sq.wqe_shift));
972 
973 	qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
974 
975 	qp->sq.max_gs = min_t(int, dev->limits.max_sg,
976 			      max_data_size / sizeof (struct mthca_data_seg));
977 	qp->rq.max_gs = min_t(int, dev->limits.max_sg,
978 			       (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
979 				sizeof (struct mthca_next_seg)) /
980 			       sizeof (struct mthca_data_seg));
981 }
982 
983 /*
984  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
985  * rq.max_gs and sq.max_gs must all be assigned.
986  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
987  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
988  * queue)
989  */
990 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
991 			       struct mthca_pd *pd,
992 			       struct mthca_qp *qp,
993 			       struct ib_udata *udata)
994 {
995 	int size;
996 	int err = -ENOMEM;
997 
998 	size = sizeof (struct mthca_next_seg) +
999 		qp->rq.max_gs * sizeof (struct mthca_data_seg);
1000 
1001 	if (size > dev->limits.max_desc_sz)
1002 		return -EINVAL;
1003 
1004 	for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
1005 	     qp->rq.wqe_shift++)
1006 		; /* nothing */
1007 
1008 	size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
1009 	switch (qp->transport) {
1010 	case MLX:
1011 		size += 2 * sizeof (struct mthca_data_seg);
1012 		break;
1013 
1014 	case UD:
1015 		size += mthca_is_memfree(dev) ?
1016 			sizeof (struct mthca_arbel_ud_seg) :
1017 			sizeof (struct mthca_tavor_ud_seg);
1018 		break;
1019 
1020 	case UC:
1021 		size += sizeof (struct mthca_raddr_seg);
1022 		break;
1023 
1024 	case RC:
1025 		size += sizeof (struct mthca_raddr_seg);
1026 		/*
1027 		 * An atomic op will require an atomic segment, a
1028 		 * remote address segment and one scatter entry.
1029 		 */
1030 		size = max_t(int, size,
1031 			     sizeof (struct mthca_atomic_seg) +
1032 			     sizeof (struct mthca_raddr_seg) +
1033 			     sizeof (struct mthca_data_seg));
1034 		break;
1035 
1036 	default:
1037 		break;
1038 	}
1039 
1040 	/* Make sure that we have enough space for a bind request */
1041 	size = max_t(int, size, sizeof (struct mthca_bind_seg));
1042 
1043 	size += sizeof (struct mthca_next_seg);
1044 
1045 	if (size > dev->limits.max_desc_sz)
1046 		return -EINVAL;
1047 
1048 	for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
1049 	     qp->sq.wqe_shift++)
1050 		; /* nothing */
1051 
1052 	qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
1053 				    1 << qp->sq.wqe_shift);
1054 
1055 	/*
1056 	 * If this is a userspace QP, we don't actually have to
1057 	 * allocate anything.  All we need is to calculate the WQE
1058 	 * sizes and the send_wqe_offset, so we're done now.
1059 	 */
1060 	if (udata)
1061 		return 0;
1062 
1063 	size = PAGE_ALIGN(qp->send_wqe_offset +
1064 			  (qp->sq.max << qp->sq.wqe_shift));
1065 
1066 	qp->wrid = kmalloc_array(qp->rq.max + qp->sq.max, sizeof(u64),
1067 				 GFP_KERNEL);
1068 	if (!qp->wrid)
1069 		goto err_out;
1070 
1071 	err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1072 			      &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1073 	if (err)
1074 		goto err_out;
1075 
1076 	return 0;
1077 
1078 err_out:
1079 	kfree(qp->wrid);
1080 	return err;
1081 }
1082 
1083 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1084 			       struct mthca_qp *qp)
1085 {
1086 	mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1087 				       (qp->sq.max << qp->sq.wqe_shift)),
1088 		       &qp->queue, qp->is_direct, &qp->mr);
1089 	kfree(qp->wrid);
1090 }
1091 
1092 static int mthca_map_memfree(struct mthca_dev *dev,
1093 			     struct mthca_qp *qp)
1094 {
1095 	int ret;
1096 
1097 	if (mthca_is_memfree(dev)) {
1098 		ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1099 		if (ret)
1100 			return ret;
1101 
1102 		ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1103 		if (ret)
1104 			goto err_qpc;
1105 
1106 		ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1107 				      qp->qpn << dev->qp_table.rdb_shift);
1108 		if (ret)
1109 			goto err_eqpc;
1110 
1111 	}
1112 
1113 	return 0;
1114 
1115 err_eqpc:
1116 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1117 
1118 err_qpc:
1119 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1120 
1121 	return ret;
1122 }
1123 
1124 static void mthca_unmap_memfree(struct mthca_dev *dev,
1125 				struct mthca_qp *qp)
1126 {
1127 	mthca_table_put(dev, dev->qp_table.rdb_table,
1128 			qp->qpn << dev->qp_table.rdb_shift);
1129 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1130 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1131 }
1132 
1133 static int mthca_alloc_memfree(struct mthca_dev *dev,
1134 			       struct mthca_qp *qp)
1135 {
1136 	if (mthca_is_memfree(dev)) {
1137 		qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1138 						 qp->qpn, &qp->rq.db);
1139 		if (qp->rq.db_index < 0)
1140 			return -ENOMEM;
1141 
1142 		qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1143 						 qp->qpn, &qp->sq.db);
1144 		if (qp->sq.db_index < 0) {
1145 			mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1146 			return -ENOMEM;
1147 		}
1148 	}
1149 
1150 	return 0;
1151 }
1152 
1153 static void mthca_free_memfree(struct mthca_dev *dev,
1154 			       struct mthca_qp *qp)
1155 {
1156 	if (mthca_is_memfree(dev)) {
1157 		mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1158 		mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1159 	}
1160 }
1161 
1162 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1163 				 struct mthca_pd *pd,
1164 				 struct mthca_cq *send_cq,
1165 				 struct mthca_cq *recv_cq,
1166 				 enum ib_sig_type send_policy,
1167 				 struct mthca_qp *qp,
1168 				 struct ib_udata *udata)
1169 {
1170 	int ret;
1171 	int i;
1172 	struct mthca_next_seg *next;
1173 
1174 	qp->refcount = 1;
1175 	init_waitqueue_head(&qp->wait);
1176 	mutex_init(&qp->mutex);
1177 	qp->state    	 = IB_QPS_RESET;
1178 	qp->atomic_rd_en = 0;
1179 	qp->resp_depth   = 0;
1180 	qp->sq_policy    = send_policy;
1181 	mthca_wq_reset(&qp->sq);
1182 	mthca_wq_reset(&qp->rq);
1183 
1184 	spin_lock_init(&qp->sq.lock);
1185 	spin_lock_init(&qp->rq.lock);
1186 
1187 	ret = mthca_map_memfree(dev, qp);
1188 	if (ret)
1189 		return ret;
1190 
1191 	ret = mthca_alloc_wqe_buf(dev, pd, qp, udata);
1192 	if (ret) {
1193 		mthca_unmap_memfree(dev, qp);
1194 		return ret;
1195 	}
1196 
1197 	mthca_adjust_qp_caps(dev, pd, qp);
1198 
1199 	/*
1200 	 * If this is a userspace QP, we're done now.  The doorbells
1201 	 * will be allocated and buffers will be initialized in
1202 	 * userspace.
1203 	 */
1204 	if (udata)
1205 		return 0;
1206 
1207 	ret = mthca_alloc_memfree(dev, qp);
1208 	if (ret) {
1209 		mthca_free_wqe_buf(dev, qp);
1210 		mthca_unmap_memfree(dev, qp);
1211 		return ret;
1212 	}
1213 
1214 	if (mthca_is_memfree(dev)) {
1215 		struct mthca_data_seg *scatter;
1216 		int size = (sizeof (struct mthca_next_seg) +
1217 			    qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1218 
1219 		for (i = 0; i < qp->rq.max; ++i) {
1220 			next = get_recv_wqe(qp, i);
1221 			next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1222 						   qp->rq.wqe_shift);
1223 			next->ee_nds = cpu_to_be32(size);
1224 
1225 			for (scatter = (void *) (next + 1);
1226 			     (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1227 			     ++scatter)
1228 				scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1229 		}
1230 
1231 		for (i = 0; i < qp->sq.max; ++i) {
1232 			next = get_send_wqe(qp, i);
1233 			next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1234 						    qp->sq.wqe_shift) +
1235 						   qp->send_wqe_offset);
1236 		}
1237 	} else {
1238 		for (i = 0; i < qp->rq.max; ++i) {
1239 			next = get_recv_wqe(qp, i);
1240 			next->nda_op = htonl((((i + 1) % qp->rq.max) <<
1241 					      qp->rq.wqe_shift) | 1);
1242 		}
1243 
1244 	}
1245 
1246 	qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1247 	qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1248 
1249 	return 0;
1250 }
1251 
1252 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1253 			     struct mthca_pd *pd, struct mthca_qp *qp)
1254 {
1255 	int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1256 
1257 	/* Sanity check QP size before proceeding */
1258 	if (cap->max_send_wr  	 > dev->limits.max_wqes ||
1259 	    cap->max_recv_wr  	 > dev->limits.max_wqes ||
1260 	    cap->max_send_sge 	 > dev->limits.max_sg   ||
1261 	    cap->max_recv_sge 	 > dev->limits.max_sg   ||
1262 	    cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1263 		return -EINVAL;
1264 
1265 	/*
1266 	 * For MLX transport we need 2 extra send gather entries:
1267 	 * one for the header and one for the checksum at the end
1268 	 */
1269 	if (qp->transport == MLX && cap->max_send_sge + 2 > dev->limits.max_sg)
1270 		return -EINVAL;
1271 
1272 	if (mthca_is_memfree(dev)) {
1273 		qp->rq.max = cap->max_recv_wr ?
1274 			roundup_pow_of_two(cap->max_recv_wr) : 0;
1275 		qp->sq.max = cap->max_send_wr ?
1276 			roundup_pow_of_two(cap->max_send_wr) : 0;
1277 	} else {
1278 		qp->rq.max = cap->max_recv_wr;
1279 		qp->sq.max = cap->max_send_wr;
1280 	}
1281 
1282 	qp->rq.max_gs = cap->max_recv_sge;
1283 	qp->sq.max_gs = max_t(int, cap->max_send_sge,
1284 			      ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1285 				    MTHCA_INLINE_CHUNK_SIZE) /
1286 			      sizeof (struct mthca_data_seg));
1287 
1288 	return 0;
1289 }
1290 
1291 int mthca_alloc_qp(struct mthca_dev *dev,
1292 		   struct mthca_pd *pd,
1293 		   struct mthca_cq *send_cq,
1294 		   struct mthca_cq *recv_cq,
1295 		   enum ib_qp_type type,
1296 		   enum ib_sig_type send_policy,
1297 		   struct ib_qp_cap *cap,
1298 		   struct mthca_qp *qp,
1299 		   struct ib_udata *udata)
1300 {
1301 	int err;
1302 
1303 	switch (type) {
1304 	case IB_QPT_RC: qp->transport = RC; break;
1305 	case IB_QPT_UC: qp->transport = UC; break;
1306 	case IB_QPT_UD: qp->transport = UD; break;
1307 	default: return -EINVAL;
1308 	}
1309 
1310 	err = mthca_set_qp_size(dev, cap, pd, qp);
1311 	if (err)
1312 		return err;
1313 
1314 	qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1315 	if (qp->qpn == -1)
1316 		return -ENOMEM;
1317 
1318 	/* initialize port to zero for error-catching. */
1319 	qp->port = 0;
1320 
1321 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1322 				    send_policy, qp, udata);
1323 	if (err) {
1324 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1325 		return err;
1326 	}
1327 
1328 	spin_lock_irq(&dev->qp_table.lock);
1329 	mthca_array_set(&dev->qp_table.qp,
1330 			qp->qpn & (dev->limits.num_qps - 1), qp);
1331 	spin_unlock_irq(&dev->qp_table.lock);
1332 
1333 	return 0;
1334 }
1335 
1336 static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1337 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1338 {
1339 	if (send_cq == recv_cq) {
1340 		spin_lock_irq(&send_cq->lock);
1341 		__acquire(&recv_cq->lock);
1342 	} else if (send_cq->cqn < recv_cq->cqn) {
1343 		spin_lock_irq(&send_cq->lock);
1344 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1345 	} else {
1346 		spin_lock_irq(&recv_cq->lock);
1347 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1348 	}
1349 }
1350 
1351 static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1352 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1353 {
1354 	if (send_cq == recv_cq) {
1355 		__release(&recv_cq->lock);
1356 		spin_unlock_irq(&send_cq->lock);
1357 	} else if (send_cq->cqn < recv_cq->cqn) {
1358 		spin_unlock(&recv_cq->lock);
1359 		spin_unlock_irq(&send_cq->lock);
1360 	} else {
1361 		spin_unlock(&send_cq->lock);
1362 		spin_unlock_irq(&recv_cq->lock);
1363 	}
1364 }
1365 
1366 int mthca_alloc_sqp(struct mthca_dev *dev,
1367 		    struct mthca_pd *pd,
1368 		    struct mthca_cq *send_cq,
1369 		    struct mthca_cq *recv_cq,
1370 		    enum ib_sig_type send_policy,
1371 		    struct ib_qp_cap *cap,
1372 		    int qpn,
1373 		    u32 port,
1374 		    struct mthca_qp *qp,
1375 		    struct ib_udata *udata)
1376 {
1377 	u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1378 	int err;
1379 
1380 	qp->transport = MLX;
1381 	err = mthca_set_qp_size(dev, cap, pd, qp);
1382 	if (err)
1383 		return err;
1384 
1385 	qp->sqp->header_buf_size = qp->sq.max * MTHCA_UD_HEADER_SIZE;
1386 	qp->sqp->header_buf =
1387 		dma_alloc_coherent(&dev->pdev->dev, qp->sqp->header_buf_size,
1388 				   &qp->sqp->header_dma, GFP_KERNEL);
1389 	if (!qp->sqp->header_buf)
1390 		return -ENOMEM;
1391 
1392 	spin_lock_irq(&dev->qp_table.lock);
1393 	if (mthca_array_get(&dev->qp_table.qp, mqpn))
1394 		err = -EBUSY;
1395 	else
1396 		mthca_array_set(&dev->qp_table.qp, mqpn, qp);
1397 	spin_unlock_irq(&dev->qp_table.lock);
1398 
1399 	if (err)
1400 		goto err_out;
1401 
1402 	qp->port      = port;
1403 	qp->qpn       = mqpn;
1404 	qp->transport = MLX;
1405 
1406 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1407 				    send_policy, qp, udata);
1408 	if (err)
1409 		goto err_out_free;
1410 
1411 	atomic_inc(&pd->sqp_count);
1412 
1413 	return 0;
1414 
1415  err_out_free:
1416 	/*
1417 	 * Lock CQs here, so that CQ polling code can do QP lookup
1418 	 * without taking a lock.
1419 	 */
1420 	mthca_lock_cqs(send_cq, recv_cq);
1421 
1422 	spin_lock(&dev->qp_table.lock);
1423 	mthca_array_clear(&dev->qp_table.qp, mqpn);
1424 	spin_unlock(&dev->qp_table.lock);
1425 
1426 	mthca_unlock_cqs(send_cq, recv_cq);
1427 
1428 err_out:
1429 	dma_free_coherent(&dev->pdev->dev, qp->sqp->header_buf_size,
1430 			  qp->sqp->header_buf, qp->sqp->header_dma);
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 	struct mthca_cq *send_cq;
1449 	struct mthca_cq *recv_cq;
1450 
1451 	send_cq = to_mcq(qp->ibqp.send_cq);
1452 	recv_cq = to_mcq(qp->ibqp.recv_cq);
1453 
1454 	/*
1455 	 * Lock CQs here, so that CQ polling code can do QP lookup
1456 	 * without taking a lock.
1457 	 */
1458 	mthca_lock_cqs(send_cq, recv_cq);
1459 
1460 	spin_lock(&dev->qp_table.lock);
1461 	mthca_array_clear(&dev->qp_table.qp,
1462 			  qp->qpn & (dev->limits.num_qps - 1));
1463 	--qp->refcount;
1464 	spin_unlock(&dev->qp_table.lock);
1465 
1466 	mthca_unlock_cqs(send_cq, recv_cq);
1467 
1468 	wait_event(qp->wait, !get_qp_refcount(dev, qp));
1469 
1470 	if (qp->state != IB_QPS_RESET)
1471 		mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1472 				NULL, 0);
1473 
1474 	/*
1475 	 * If this is a userspace QP, the buffers, MR, CQs and so on
1476 	 * will be cleaned up in userspace, so all we have to do is
1477 	 * unref the mem-free tables and free the QPN in our table.
1478 	 */
1479 	if (!qp->ibqp.uobject) {
1480 		mthca_cq_clean(dev, recv_cq, qp->qpn,
1481 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1482 		if (send_cq != recv_cq)
1483 			mthca_cq_clean(dev, send_cq, qp->qpn, NULL);
1484 
1485 		mthca_free_memfree(dev, qp);
1486 		mthca_free_wqe_buf(dev, qp);
1487 	}
1488 
1489 	mthca_unmap_memfree(dev, qp);
1490 
1491 	if (is_sqp(dev, qp)) {
1492 		atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1493 		dma_free_coherent(&dev->pdev->dev, qp->sqp->header_buf_size,
1494 				  qp->sqp->header_buf, qp->sqp->header_dma);
1495 	} else
1496 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1497 }
1498 
1499 /* Create UD header for an MLX send and build a data segment for it */
1500 static int build_mlx_header(struct mthca_dev *dev, struct mthca_qp *qp, int ind,
1501 			    const struct ib_ud_wr *wr,
1502 			    struct mthca_mlx_seg *mlx,
1503 			    struct mthca_data_seg *data)
1504 {
1505 	struct mthca_sqp *sqp = qp->sqp;
1506 	int header_size;
1507 	int err;
1508 	u16 pkey;
1509 
1510 	ib_ud_header_init(256, /* assume a MAD */ 1, 0, 0,
1511 			  mthca_ah_grh_present(to_mah(wr->ah)), 0, 0, 0,
1512 			  &sqp->ud_header);
1513 
1514 	err = mthca_read_ah(dev, to_mah(wr->ah), &sqp->ud_header);
1515 	if (err)
1516 		return err;
1517 	mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1518 	mlx->flags |= cpu_to_be32((!qp->ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1519 				  (sqp->ud_header.lrh.destination_lid ==
1520 				   IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1521 				  (sqp->ud_header.lrh.service_level << 8));
1522 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
1523 	mlx->vcrc = 0;
1524 
1525 	switch (wr->wr.opcode) {
1526 	case IB_WR_SEND:
1527 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1528 		sqp->ud_header.immediate_present = 0;
1529 		break;
1530 	case IB_WR_SEND_WITH_IMM:
1531 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1532 		sqp->ud_header.immediate_present = 1;
1533 		sqp->ud_header.immediate_data = wr->wr.ex.imm_data;
1534 		break;
1535 	default:
1536 		return -EINVAL;
1537 	}
1538 
1539 	sqp->ud_header.lrh.virtual_lane    = !qp->ibqp.qp_num ? 15 : 0;
1540 	if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1541 		sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1542 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
1543 	if (!qp->ibqp.qp_num)
1544 		ib_get_cached_pkey(&dev->ib_dev, qp->port, sqp->pkey_index,
1545 				   &pkey);
1546 	else
1547 		ib_get_cached_pkey(&dev->ib_dev, qp->port, wr->pkey_index,
1548 				   &pkey);
1549 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1550 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
1551 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1552 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
1553 					       sqp->qkey : wr->remote_qkey);
1554 	sqp->ud_header.deth.source_qpn = cpu_to_be32(qp->ibqp.qp_num);
1555 
1556 	header_size = ib_ud_header_pack(&sqp->ud_header,
1557 					sqp->header_buf +
1558 					ind * MTHCA_UD_HEADER_SIZE);
1559 
1560 	data->byte_count = cpu_to_be32(header_size);
1561 	data->lkey       = cpu_to_be32(to_mpd(qp->ibqp.pd)->ntmr.ibmr.lkey);
1562 	data->addr       = cpu_to_be64(sqp->header_dma +
1563 				       ind * MTHCA_UD_HEADER_SIZE);
1564 
1565 	return 0;
1566 }
1567 
1568 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1569 				    struct ib_cq *ib_cq)
1570 {
1571 	unsigned cur;
1572 	struct mthca_cq *cq;
1573 
1574 	cur = wq->head - wq->tail;
1575 	if (likely(cur + nreq < wq->max))
1576 		return 0;
1577 
1578 	cq = to_mcq(ib_cq);
1579 	spin_lock(&cq->lock);
1580 	cur = wq->head - wq->tail;
1581 	spin_unlock(&cq->lock);
1582 
1583 	return cur + nreq >= wq->max;
1584 }
1585 
1586 static __always_inline void set_raddr_seg(struct mthca_raddr_seg *rseg,
1587 					  u64 remote_addr, u32 rkey)
1588 {
1589 	rseg->raddr    = cpu_to_be64(remote_addr);
1590 	rseg->rkey     = cpu_to_be32(rkey);
1591 	rseg->reserved = 0;
1592 }
1593 
1594 static __always_inline void set_atomic_seg(struct mthca_atomic_seg *aseg,
1595 					   const struct ib_atomic_wr *wr)
1596 {
1597 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1598 		aseg->swap_add = cpu_to_be64(wr->swap);
1599 		aseg->compare  = cpu_to_be64(wr->compare_add);
1600 	} else {
1601 		aseg->swap_add = cpu_to_be64(wr->compare_add);
1602 		aseg->compare  = 0;
1603 	}
1604 
1605 }
1606 
1607 static void set_tavor_ud_seg(struct mthca_tavor_ud_seg *useg,
1608 			     const struct ib_ud_wr *wr)
1609 {
1610 	useg->lkey    = cpu_to_be32(to_mah(wr->ah)->key);
1611 	useg->av_addr =	cpu_to_be64(to_mah(wr->ah)->avdma);
1612 	useg->dqpn    =	cpu_to_be32(wr->remote_qpn);
1613 	useg->qkey    =	cpu_to_be32(wr->remote_qkey);
1614 
1615 }
1616 
1617 static void set_arbel_ud_seg(struct mthca_arbel_ud_seg *useg,
1618 			     const struct ib_ud_wr *wr)
1619 {
1620 	memcpy(useg->av, to_mah(wr->ah)->av, MTHCA_AV_SIZE);
1621 	useg->dqpn = cpu_to_be32(wr->remote_qpn);
1622 	useg->qkey = cpu_to_be32(wr->remote_qkey);
1623 }
1624 
1625 int mthca_tavor_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
1626 			  const struct ib_send_wr **bad_wr)
1627 {
1628 	struct mthca_dev *dev = to_mdev(ibqp->device);
1629 	struct mthca_qp *qp = to_mqp(ibqp);
1630 	void *wqe;
1631 	void *prev_wqe;
1632 	unsigned long flags;
1633 	int err = 0;
1634 	int nreq;
1635 	int i;
1636 	int size;
1637 	/*
1638 	 * f0 and size0 are only used if nreq != 0, and they will
1639 	 * always be initialized the first time through the main loop
1640 	 * before nreq is incremented.  So nreq cannot become non-zero
1641 	 * without initializing f0 and size0, and they are in fact
1642 	 * never used uninitialized.
1643 	 */
1644 	int size0;
1645 	u32 f0;
1646 	int ind;
1647 	u8 op0 = 0;
1648 
1649 	spin_lock_irqsave(&qp->sq.lock, flags);
1650 
1651 	/* XXX check that state is OK to post send */
1652 
1653 	ind = qp->sq.next_ind;
1654 
1655 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1656 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1657 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1658 					" %d max, %d nreq)\n", qp->qpn,
1659 					qp->sq.head, qp->sq.tail,
1660 					qp->sq.max, nreq);
1661 			err = -ENOMEM;
1662 			*bad_wr = wr;
1663 			goto out;
1664 		}
1665 
1666 		wqe = get_send_wqe(qp, ind);
1667 		prev_wqe = qp->sq.last;
1668 		qp->sq.last = wqe;
1669 
1670 		((struct mthca_next_seg *) wqe)->nda_op = 0;
1671 		((struct mthca_next_seg *) wqe)->ee_nds = 0;
1672 		((struct mthca_next_seg *) wqe)->flags =
1673 			((wr->send_flags & IB_SEND_SIGNALED) ?
1674 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1675 			((wr->send_flags & IB_SEND_SOLICITED) ?
1676 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1677 			cpu_to_be32(1);
1678 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1679 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1680 			((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
1681 
1682 		wqe += sizeof (struct mthca_next_seg);
1683 		size = sizeof (struct mthca_next_seg) / 16;
1684 
1685 		switch (qp->transport) {
1686 		case RC:
1687 			switch (wr->opcode) {
1688 			case IB_WR_ATOMIC_CMP_AND_SWP:
1689 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1690 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
1691 					      atomic_wr(wr)->rkey);
1692 				wqe += sizeof (struct mthca_raddr_seg);
1693 
1694 				set_atomic_seg(wqe, atomic_wr(wr));
1695 				wqe += sizeof (struct mthca_atomic_seg);
1696 				size += (sizeof (struct mthca_raddr_seg) +
1697 					 sizeof (struct mthca_atomic_seg)) / 16;
1698 				break;
1699 
1700 			case IB_WR_RDMA_WRITE:
1701 			case IB_WR_RDMA_WRITE_WITH_IMM:
1702 			case IB_WR_RDMA_READ:
1703 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
1704 					      rdma_wr(wr)->rkey);
1705 				wqe  += sizeof (struct mthca_raddr_seg);
1706 				size += sizeof (struct mthca_raddr_seg) / 16;
1707 				break;
1708 
1709 			default:
1710 				/* No extra segments required for sends */
1711 				break;
1712 			}
1713 
1714 			break;
1715 
1716 		case UC:
1717 			switch (wr->opcode) {
1718 			case IB_WR_RDMA_WRITE:
1719 			case IB_WR_RDMA_WRITE_WITH_IMM:
1720 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
1721 					      rdma_wr(wr)->rkey);
1722 				wqe  += sizeof (struct mthca_raddr_seg);
1723 				size += sizeof (struct mthca_raddr_seg) / 16;
1724 				break;
1725 
1726 			default:
1727 				/* No extra segments required for sends */
1728 				break;
1729 			}
1730 
1731 			break;
1732 
1733 		case UD:
1734 			set_tavor_ud_seg(wqe, ud_wr(wr));
1735 			wqe  += sizeof (struct mthca_tavor_ud_seg);
1736 			size += sizeof (struct mthca_tavor_ud_seg) / 16;
1737 			break;
1738 
1739 		case MLX:
1740 			err = build_mlx_header(
1741 				dev, qp, ind, ud_wr(wr),
1742 				wqe - sizeof(struct mthca_next_seg), wqe);
1743 			if (err) {
1744 				*bad_wr = wr;
1745 				goto out;
1746 			}
1747 			wqe += sizeof (struct mthca_data_seg);
1748 			size += sizeof (struct mthca_data_seg) / 16;
1749 			break;
1750 		}
1751 
1752 		if (wr->num_sge > qp->sq.max_gs) {
1753 			mthca_err(dev, "too many gathers\n");
1754 			err = -EINVAL;
1755 			*bad_wr = wr;
1756 			goto out;
1757 		}
1758 
1759 		for (i = 0; i < wr->num_sge; ++i) {
1760 			mthca_set_data_seg(wqe, wr->sg_list + i);
1761 			wqe  += sizeof (struct mthca_data_seg);
1762 			size += sizeof (struct mthca_data_seg) / 16;
1763 		}
1764 
1765 		/* Add one more inline data segment for ICRC */
1766 		if (qp->transport == MLX) {
1767 			((struct mthca_data_seg *) wqe)->byte_count =
1768 				cpu_to_be32((1 << 31) | 4);
1769 			((u32 *) wqe)[1] = 0;
1770 			wqe += sizeof (struct mthca_data_seg);
1771 			size += sizeof (struct mthca_data_seg) / 16;
1772 		}
1773 
1774 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
1775 
1776 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1777 			mthca_err(dev, "opcode invalid\n");
1778 			err = -EINVAL;
1779 			*bad_wr = wr;
1780 			goto out;
1781 		}
1782 
1783 		((struct mthca_next_seg *) prev_wqe)->nda_op =
1784 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
1785 				     qp->send_wqe_offset) |
1786 				    mthca_opcode[wr->opcode]);
1787 		wmb();
1788 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1789 			cpu_to_be32((nreq ? 0 : MTHCA_NEXT_DBD) | size |
1790 				    ((wr->send_flags & IB_SEND_FENCE) ?
1791 				    MTHCA_NEXT_FENCE : 0));
1792 
1793 		if (!nreq) {
1794 			size0 = size;
1795 			op0   = mthca_opcode[wr->opcode];
1796 			f0    = wr->send_flags & IB_SEND_FENCE ?
1797 				MTHCA_SEND_DOORBELL_FENCE : 0;
1798 		}
1799 
1800 		++ind;
1801 		if (unlikely(ind >= qp->sq.max))
1802 			ind -= qp->sq.max;
1803 	}
1804 
1805 out:
1806 	if (likely(nreq)) {
1807 		wmb();
1808 
1809 		mthca_write64(((qp->sq.next_ind << qp->sq.wqe_shift) +
1810 			       qp->send_wqe_offset) | f0 | op0,
1811 			      (qp->qpn << 8) | size0,
1812 			      dev->kar + MTHCA_SEND_DOORBELL,
1813 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1814 	}
1815 
1816 	qp->sq.next_ind = ind;
1817 	qp->sq.head    += nreq;
1818 
1819 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1820 	return err;
1821 }
1822 
1823 int mthca_tavor_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
1824 			     const struct ib_recv_wr **bad_wr)
1825 {
1826 	struct mthca_dev *dev = to_mdev(ibqp->device);
1827 	struct mthca_qp *qp = to_mqp(ibqp);
1828 	unsigned long flags;
1829 	int err = 0;
1830 	int nreq;
1831 	int i;
1832 	int size;
1833 	/*
1834 	 * size0 is only used if nreq != 0, and it will always be
1835 	 * initialized the first time through the main loop before
1836 	 * nreq is incremented.  So nreq cannot become non-zero
1837 	 * without initializing size0, and it is in fact never used
1838 	 * uninitialized.
1839 	 */
1840 	int size0;
1841 	int ind;
1842 	void *wqe;
1843 	void *prev_wqe;
1844 
1845 	spin_lock_irqsave(&qp->rq.lock, flags);
1846 
1847 	/* XXX check that state is OK to post receive */
1848 
1849 	ind = qp->rq.next_ind;
1850 
1851 	for (nreq = 0; wr; wr = wr->next) {
1852 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1853 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1854 					" %d max, %d nreq)\n", qp->qpn,
1855 					qp->rq.head, qp->rq.tail,
1856 					qp->rq.max, nreq);
1857 			err = -ENOMEM;
1858 			*bad_wr = wr;
1859 			goto out;
1860 		}
1861 
1862 		wqe = get_recv_wqe(qp, ind);
1863 		prev_wqe = qp->rq.last;
1864 		qp->rq.last = wqe;
1865 
1866 		((struct mthca_next_seg *) wqe)->ee_nds =
1867 			cpu_to_be32(MTHCA_NEXT_DBD);
1868 		((struct mthca_next_seg *) wqe)->flags = 0;
1869 
1870 		wqe += sizeof (struct mthca_next_seg);
1871 		size = sizeof (struct mthca_next_seg) / 16;
1872 
1873 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1874 			err = -EINVAL;
1875 			*bad_wr = wr;
1876 			goto out;
1877 		}
1878 
1879 		for (i = 0; i < wr->num_sge; ++i) {
1880 			mthca_set_data_seg(wqe, wr->sg_list + i);
1881 			wqe  += sizeof (struct mthca_data_seg);
1882 			size += sizeof (struct mthca_data_seg) / 16;
1883 		}
1884 
1885 		qp->wrid[ind] = wr->wr_id;
1886 
1887 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1888 			cpu_to_be32(MTHCA_NEXT_DBD | size);
1889 
1890 		if (!nreq)
1891 			size0 = size;
1892 
1893 		++ind;
1894 		if (unlikely(ind >= qp->rq.max))
1895 			ind -= qp->rq.max;
1896 
1897 		++nreq;
1898 		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1899 			nreq = 0;
1900 
1901 			wmb();
1902 
1903 			mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1904 				      qp->qpn << 8, dev->kar + MTHCA_RECEIVE_DOORBELL,
1905 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1906 
1907 			qp->rq.next_ind = ind;
1908 			qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1909 		}
1910 	}
1911 
1912 out:
1913 	if (likely(nreq)) {
1914 		wmb();
1915 
1916 		mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1917 			      qp->qpn << 8 | nreq, dev->kar + MTHCA_RECEIVE_DOORBELL,
1918 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1919 	}
1920 
1921 	qp->rq.next_ind = ind;
1922 	qp->rq.head    += nreq;
1923 
1924 	spin_unlock_irqrestore(&qp->rq.lock, flags);
1925 	return err;
1926 }
1927 
1928 int mthca_arbel_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
1929 			  const struct ib_send_wr **bad_wr)
1930 {
1931 	struct mthca_dev *dev = to_mdev(ibqp->device);
1932 	struct mthca_qp *qp = to_mqp(ibqp);
1933 	u32 dbhi;
1934 	void *wqe;
1935 	void *prev_wqe;
1936 	unsigned long flags;
1937 	int err = 0;
1938 	int nreq;
1939 	int i;
1940 	int size;
1941 	/*
1942 	 * f0 and size0 are only used if nreq != 0, and they will
1943 	 * always be initialized the first time through the main loop
1944 	 * before nreq is incremented.  So nreq cannot become non-zero
1945 	 * without initializing f0 and size0, and they are in fact
1946 	 * never used uninitialized.
1947 	 */
1948 	int size0;
1949 	u32 f0;
1950 	int ind;
1951 	u8 op0 = 0;
1952 
1953 	spin_lock_irqsave(&qp->sq.lock, flags);
1954 
1955 	/* XXX check that state is OK to post send */
1956 
1957 	ind = qp->sq.head & (qp->sq.max - 1);
1958 
1959 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1960 		if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1961 			nreq = 0;
1962 
1963 			dbhi = (MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1964 				((qp->sq.head & 0xffff) << 8) | f0 | op0;
1965 
1966 			qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1967 
1968 			/*
1969 			 * Make sure that descriptors are written before
1970 			 * doorbell record.
1971 			 */
1972 			wmb();
1973 			*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1974 
1975 			/*
1976 			 * Make sure doorbell record is written before we
1977 			 * write MMIO send doorbell.
1978 			 */
1979 			wmb();
1980 
1981 			mthca_write64(dbhi, (qp->qpn << 8) | size0,
1982 				      dev->kar + MTHCA_SEND_DOORBELL,
1983 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1984 		}
1985 
1986 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1987 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1988 					" %d max, %d nreq)\n", qp->qpn,
1989 					qp->sq.head, qp->sq.tail,
1990 					qp->sq.max, nreq);
1991 			err = -ENOMEM;
1992 			*bad_wr = wr;
1993 			goto out;
1994 		}
1995 
1996 		wqe = get_send_wqe(qp, ind);
1997 		prev_wqe = qp->sq.last;
1998 		qp->sq.last = wqe;
1999 
2000 		((struct mthca_next_seg *) wqe)->flags =
2001 			((wr->send_flags & IB_SEND_SIGNALED) ?
2002 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
2003 			((wr->send_flags & IB_SEND_SOLICITED) ?
2004 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
2005 			((wr->send_flags & IB_SEND_IP_CSUM) ?
2006 			 cpu_to_be32(MTHCA_NEXT_IP_CSUM | MTHCA_NEXT_TCP_UDP_CSUM) : 0) |
2007 			cpu_to_be32(1);
2008 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
2009 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
2010 			((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
2011 
2012 		wqe += sizeof (struct mthca_next_seg);
2013 		size = sizeof (struct mthca_next_seg) / 16;
2014 
2015 		switch (qp->transport) {
2016 		case RC:
2017 			switch (wr->opcode) {
2018 			case IB_WR_ATOMIC_CMP_AND_SWP:
2019 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2020 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
2021 					      atomic_wr(wr)->rkey);
2022 				wqe += sizeof (struct mthca_raddr_seg);
2023 
2024 				set_atomic_seg(wqe, atomic_wr(wr));
2025 				wqe  += sizeof (struct mthca_atomic_seg);
2026 				size += (sizeof (struct mthca_raddr_seg) +
2027 					 sizeof (struct mthca_atomic_seg)) / 16;
2028 				break;
2029 
2030 			case IB_WR_RDMA_READ:
2031 			case IB_WR_RDMA_WRITE:
2032 			case IB_WR_RDMA_WRITE_WITH_IMM:
2033 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
2034 					      rdma_wr(wr)->rkey);
2035 				wqe  += sizeof (struct mthca_raddr_seg);
2036 				size += sizeof (struct mthca_raddr_seg) / 16;
2037 				break;
2038 
2039 			default:
2040 				/* No extra segments required for sends */
2041 				break;
2042 			}
2043 
2044 			break;
2045 
2046 		case UC:
2047 			switch (wr->opcode) {
2048 			case IB_WR_RDMA_WRITE:
2049 			case IB_WR_RDMA_WRITE_WITH_IMM:
2050 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
2051 					      rdma_wr(wr)->rkey);
2052 				wqe  += sizeof (struct mthca_raddr_seg);
2053 				size += sizeof (struct mthca_raddr_seg) / 16;
2054 				break;
2055 
2056 			default:
2057 				/* No extra segments required for sends */
2058 				break;
2059 			}
2060 
2061 			break;
2062 
2063 		case UD:
2064 			set_arbel_ud_seg(wqe, ud_wr(wr));
2065 			wqe  += sizeof (struct mthca_arbel_ud_seg);
2066 			size += sizeof (struct mthca_arbel_ud_seg) / 16;
2067 			break;
2068 
2069 		case MLX:
2070 			err = build_mlx_header(
2071 				dev, qp, ind, ud_wr(wr),
2072 				wqe - sizeof(struct mthca_next_seg), wqe);
2073 			if (err) {
2074 				*bad_wr = wr;
2075 				goto out;
2076 			}
2077 			wqe += sizeof (struct mthca_data_seg);
2078 			size += sizeof (struct mthca_data_seg) / 16;
2079 			break;
2080 		}
2081 
2082 		if (wr->num_sge > qp->sq.max_gs) {
2083 			mthca_err(dev, "too many gathers\n");
2084 			err = -EINVAL;
2085 			*bad_wr = wr;
2086 			goto out;
2087 		}
2088 
2089 		for (i = 0; i < wr->num_sge; ++i) {
2090 			mthca_set_data_seg(wqe, wr->sg_list + i);
2091 			wqe  += sizeof (struct mthca_data_seg);
2092 			size += sizeof (struct mthca_data_seg) / 16;
2093 		}
2094 
2095 		/* Add one more inline data segment for ICRC */
2096 		if (qp->transport == MLX) {
2097 			((struct mthca_data_seg *) wqe)->byte_count =
2098 				cpu_to_be32((1 << 31) | 4);
2099 			((u32 *) wqe)[1] = 0;
2100 			wqe += sizeof (struct mthca_data_seg);
2101 			size += sizeof (struct mthca_data_seg) / 16;
2102 		}
2103 
2104 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
2105 
2106 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2107 			mthca_err(dev, "opcode invalid\n");
2108 			err = -EINVAL;
2109 			*bad_wr = wr;
2110 			goto out;
2111 		}
2112 
2113 		((struct mthca_next_seg *) prev_wqe)->nda_op =
2114 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
2115 				     qp->send_wqe_offset) |
2116 				    mthca_opcode[wr->opcode]);
2117 		wmb();
2118 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
2119 			cpu_to_be32(MTHCA_NEXT_DBD | size |
2120 				    ((wr->send_flags & IB_SEND_FENCE) ?
2121 				     MTHCA_NEXT_FENCE : 0));
2122 
2123 		if (!nreq) {
2124 			size0 = size;
2125 			op0   = mthca_opcode[wr->opcode];
2126 			f0    = wr->send_flags & IB_SEND_FENCE ?
2127 				MTHCA_SEND_DOORBELL_FENCE : 0;
2128 		}
2129 
2130 		++ind;
2131 		if (unlikely(ind >= qp->sq.max))
2132 			ind -= qp->sq.max;
2133 	}
2134 
2135 out:
2136 	if (likely(nreq)) {
2137 		dbhi = (nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0;
2138 
2139 		qp->sq.head += nreq;
2140 
2141 		/*
2142 		 * Make sure that descriptors are written before
2143 		 * doorbell record.
2144 		 */
2145 		wmb();
2146 		*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2147 
2148 		/*
2149 		 * Make sure doorbell record is written before we
2150 		 * write MMIO send doorbell.
2151 		 */
2152 		wmb();
2153 
2154 		mthca_write64(dbhi, (qp->qpn << 8) | size0, dev->kar + MTHCA_SEND_DOORBELL,
2155 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2156 	}
2157 
2158 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2159 	return err;
2160 }
2161 
2162 int mthca_arbel_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
2163 			     const struct ib_recv_wr **bad_wr)
2164 {
2165 	struct mthca_dev *dev = to_mdev(ibqp->device);
2166 	struct mthca_qp *qp = to_mqp(ibqp);
2167 	unsigned long flags;
2168 	int err = 0;
2169 	int nreq;
2170 	int ind;
2171 	int i;
2172 	void *wqe;
2173 
2174 	spin_lock_irqsave(&qp->rq.lock, flags);
2175 
2176 	/* XXX check that state is OK to post receive */
2177 
2178 	ind = qp->rq.head & (qp->rq.max - 1);
2179 
2180 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2181 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2182 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2183 					" %d max, %d nreq)\n", qp->qpn,
2184 					qp->rq.head, qp->rq.tail,
2185 					qp->rq.max, nreq);
2186 			err = -ENOMEM;
2187 			*bad_wr = wr;
2188 			goto out;
2189 		}
2190 
2191 		wqe = get_recv_wqe(qp, ind);
2192 
2193 		((struct mthca_next_seg *) wqe)->flags = 0;
2194 
2195 		wqe += sizeof (struct mthca_next_seg);
2196 
2197 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2198 			err = -EINVAL;
2199 			*bad_wr = wr;
2200 			goto out;
2201 		}
2202 
2203 		for (i = 0; i < wr->num_sge; ++i) {
2204 			mthca_set_data_seg(wqe, wr->sg_list + i);
2205 			wqe += sizeof (struct mthca_data_seg);
2206 		}
2207 
2208 		if (i < qp->rq.max_gs)
2209 			mthca_set_data_seg_inval(wqe);
2210 
2211 		qp->wrid[ind] = wr->wr_id;
2212 
2213 		++ind;
2214 		if (unlikely(ind >= qp->rq.max))
2215 			ind -= qp->rq.max;
2216 	}
2217 out:
2218 	if (likely(nreq)) {
2219 		qp->rq.head += nreq;
2220 
2221 		/*
2222 		 * Make sure that descriptors are written before
2223 		 * doorbell record.
2224 		 */
2225 		wmb();
2226 		*qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2227 	}
2228 
2229 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2230 	return err;
2231 }
2232 
2233 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2234 			int index, int *dbd, __be32 *new_wqe)
2235 {
2236 	struct mthca_next_seg *next;
2237 
2238 	/*
2239 	 * For SRQs, all receive WQEs generate a CQE, so we're always
2240 	 * at the end of the doorbell chain.
2241 	 */
2242 	if (qp->ibqp.srq && !is_send) {
2243 		*new_wqe = 0;
2244 		return;
2245 	}
2246 
2247 	if (is_send)
2248 		next = get_send_wqe(qp, index);
2249 	else
2250 		next = get_recv_wqe(qp, index);
2251 
2252 	*dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2253 	if (next->ee_nds & cpu_to_be32(0x3f))
2254 		*new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2255 			(next->ee_nds & cpu_to_be32(0x3f));
2256 	else
2257 		*new_wqe = 0;
2258 }
2259 
2260 int mthca_init_qp_table(struct mthca_dev *dev)
2261 {
2262 	int err;
2263 	int i;
2264 
2265 	spin_lock_init(&dev->qp_table.lock);
2266 
2267 	/*
2268 	 * We reserve 2 extra QPs per port for the special QPs.  The
2269 	 * special QP for port 1 has to be even, so round up.
2270 	 */
2271 	dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2272 	err = mthca_alloc_init(&dev->qp_table.alloc,
2273 			       dev->limits.num_qps,
2274 			       (1 << 24) - 1,
2275 			       dev->qp_table.sqp_start +
2276 			       MTHCA_MAX_PORTS * 2);
2277 	if (err)
2278 		return err;
2279 
2280 	err = mthca_array_init(&dev->qp_table.qp,
2281 			       dev->limits.num_qps);
2282 	if (err) {
2283 		mthca_alloc_cleanup(&dev->qp_table.alloc);
2284 		return err;
2285 	}
2286 
2287 	for (i = 0; i < 2; ++i) {
2288 		err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2289 				    dev->qp_table.sqp_start + i * 2);
2290 		if (err) {
2291 			mthca_warn(dev, "CONF_SPECIAL_QP returned "
2292 				   "%d, aborting.\n", err);
2293 			goto err_out;
2294 		}
2295 	}
2296 	return 0;
2297 
2298  err_out:
2299 	for (i = 0; i < 2; ++i)
2300 		mthca_CONF_SPECIAL_QP(dev, i, 0);
2301 
2302 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2303 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2304 
2305 	return err;
2306 }
2307 
2308 void mthca_cleanup_qp_table(struct mthca_dev *dev)
2309 {
2310 	int i;
2311 
2312 	for (i = 0; i < 2; ++i)
2313 		mthca_CONF_SPECIAL_QP(dev, i, 0);
2314 
2315 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2316 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2317 }
2318